迁移transformers ver.4.11.3的bert bart roberta gpt2和cpt模型

fastNLP/core/utils/dummy_class.py

@@ -1,4 +1,5 @@
-
+import functools
 
 class DummyClass:
-    pass
+    def __call__(self, *args, **kwargs):
+        return

fastNLP/transformers/__init__.py

@@ -0,0 +1 @@
+"""基于 transformers-4.11.3 版本迁移"""

fastNLP/transformers/torch/__init__.py

@@ -0,0 +1,9 @@
+"""
+为了防止因 https://github.com/huggingface/transformers 版本变化导致代码不兼容，当前 folder 以及子 folder 
+都复制自 https://github.com/huggingface/transformers 的4.11.3版本。
+In order to avoid the code change of https://github.com/huggingface/transformers to cause version
+mismatch, we copy code from https://github.com/huggingface/transformers(version:4.11.3) in this
+folder and its subfolder.
+"""
+__version__ = "4.11.3"
+from .models import *

fastNLP/transformers/torch/activations.py

@@ -0,0 +1,125 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+
+from packaging import version
+
+from fastNLP.envs.imports import _NEED_IMPORT_TORCH
+from fastNLP.core.log import logger
+
+if _NEED_IMPORT_TORCH:
+    import torch
+    from torch import nn, tanh, sigmoid
+    from torch.nn.functional import relu
+else:
+    from fastNLP.core.utils.dummy_class import (
+        DummyClass as relu, 
+        DummyClass as tanh,
+        DummyClass as sigmoid,
+)
+
+
+def _gelu_python(x):
+    """
+    Original Implementation of the GELU activation function in Google BERT repo when initially created. For
+    information: OpenAI GPT's GELU is slightly different (and gives slightly different results): 0.5 * x * (1 +
+    torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) This is now written in C in nn.functional
+    Also see the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
+    """
+    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
+
+
+def gelu_new(x):
+    """
+    Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT). Also see
+    the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
+    """
+    return 0.5 * x * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
+
+if _NEED_IMPORT_TORCH:
+    if version.parse(torch.__version__) < version.parse("1.4"):
+        gelu = _gelu_python
+    else:
+        gelu = nn.functional.gelu
+else:
+    from fastNLP.core.utils.dummy_class import DummyClass as gelu
+
+def gelu_fast(x):
+    return 0.5 * x * (1.0 + torch.tanh(x * 0.7978845608 * (1.0 + 0.044715 * x * x)))
+
+
+def quick_gelu(x):
+    return x * torch.sigmoid(1.702 * x)
+
+
+def _silu_python(x):
+    """
+    See Gaussian Error Linear Units (Hendrycks et al., https://arxiv.org/abs/1606.08415) where the SiLU (Sigmoid Linear
+    Unit) was originally introduced and coined, and see Sigmoid-Weighted Linear Units for Neural Network Function
+    Approximation in Reinforcement Learning (Elfwing et al., https://arxiv.org/abs/1702.03118) and Swish: a Self-Gated
+    Activation Function (Ramachandran et al., https://arxiv.org/abs/1710.05941v1) where the SiLU was experimented with
+    later.
+    """
+    return x * torch.sigmoid(x)
+
+if _NEED_IMPORT_TORCH:
+    if version.parse(torch.__version__) < version.parse("1.7"):
+        silu = _silu_python
+    else:
+        silu = nn.functional.silu
+else:
+    from fastNLP.core.utils.dummy_class import DummyClass as silu
+
+
+def _mish_python(x):
+    """
+    See Mish: A Self-Regularized Non-Monotonic Activation Function (Misra., https://arxiv.org/abs/1908.08681). Also
+    visit the official repository for the paper: https://github.com/digantamisra98/Mish
+    """
+    return x * torch.tanh(nn.functional.softplus(x))
+
+if _NEED_IMPORT_TORCH:
+    if version.parse(torch.__version__) < version.parse("1.9"):
+        mish = _mish_python
+    else:
+        mish = nn.functional.mish
+else:
+    from fastNLP.core.utils.dummy_class import DummyClass as mish
+
+
+def linear_act(x):
+    return x
+
+
+ACT2FN = {
+    "relu": relu,
+    "silu": silu,
+    "swish": silu,
+    "gelu": gelu,
+    "tanh": tanh,
+    "gelu_new": gelu_new,
+    "gelu_fast": gelu_fast,
+    "quick_gelu": quick_gelu,
+    "mish": mish,
+    "linear": linear_act,
+    "sigmoid": sigmoid,
+}
+
+
+def get_activation(activation_string):
+    if activation_string in ACT2FN:
+        return ACT2FN[activation_string]
+    else:
+        raise KeyError(f"function {activation_string} not found in ACT2FN mapping {list(ACT2FN.keys())}")

fastNLP/transformers/torch/configuration_utils.py

@@ -0,0 +1,777 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Configuration base class and utilities."""
+
+
+import copy
+import json
+import os
+from typing import Any, Dict, Tuple, Union
+
+from . import __version__
+from .file_utils import (
+    CONFIG_NAME,
+    cached_path,
+    hf_bucket_url,
+    is_offline_mode,
+    is_remote_url,
+)
+from fastNLP.envs.imports import _NEED_IMPORT_TORCH
+from fastNLP.core.log import logger
+
+
+class PretrainedConfig:
+    r"""
+    Base class for all configuration classes. Handles a few parameters common to all models' configurations as well as
+    methods for loading/downloading/saving configurations.
+
+    Note:
+        A configuration file can be loaded and saved to disk. Loading the configuration file and using this file to
+        initialize a model does **not** load the model weights. It only affects the model's configuration.
+
+    Class attributes (overridden by derived classes)
+
+        - **model_type** (:obj:`str`) -- An identifier for the model type, serialized into the JSON file, and used to
+          recreate the correct object in :class:`~transformers.AutoConfig`.
+        - **is_composition** (:obj:`bool`) -- Whether the config class is composed of multiple sub-configs. In this
+          case the config has to be initialized from two or more configs of type
+          :class:`~transformers.PretrainedConfig` like: :class:`~transformers.EncoderDecoderConfig` or
+          :class:`~RagConfig`.
+        - **keys_to_ignore_at_inference** (:obj:`List[str]`) -- A list of keys to ignore by default when looking at
+          dictionary outputs of the model during inference.
+        - **attribute_map** (:obj:`Dict[str, str]`) -- A dict that maps model specific attribute names to the
+          standardized naming of attributes.
+
+    Common attributes (present in all subclasses)
+
+        - **vocab_size** (:obj:`int`) -- The number of tokens in the vocabulary, which is also the first dimension of
+          the embeddings matrix (this attribute may be missing for models that don't have a text modality like ViT).
+        - **hidden_size** (:obj:`int`) -- The hidden size of the model.
+        - **num_attention_heads** (:obj:`int`) -- The number of attention heads used in the multi-head attention layers
+          of the model.
+        - **num_hidden_layers** (:obj:`int`) -- The number of blocks in the model.
+
+    Args:
+        name_or_path (:obj:`str`, `optional`, defaults to :obj:`""`):
+            Store the string that was passed to :func:`~transformers.PreTrainedModel.from_pretrained` or
+            :func:`~transformers.TFPreTrainedModel.from_pretrained` as ``pretrained_model_name_or_path`` if the
+            configuration was created with such a method.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not the model should return all hidden-states.
+        output_attentions (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not the model should returns all attentions.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return a :class:`~transformers.file_utils.ModelOutput` instead of a plain
+            tuple.
+        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether the model is used as an encoder/decoder or not.
+        is_decoder (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether the model is used as decoder or not (in which case it's used as an encoder).
+        add_cross_attention (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether cross-attention layers should be added to the model. Note, this option is only relevant for models
+            that can be used as decoder models within the `:class:~transformers.EncoderDecoderModel` class, which
+            consists of all models in ``AUTO_MODELS_FOR_CAUSAL_LM``.
+        tie_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether all encoder weights should be tied to their equivalent decoder weights. This requires the encoder
+            and decoder model to have the exact same parameter names.
+        prune_heads (:obj:`Dict[int, List[int]]`, `optional`, defaults to :obj:`{}`):
+            Pruned heads of the model. The keys are the selected layer indices and the associated values, the list of
+            heads to prune in said layer.
+
+            For instance ``{1: [0, 2], 2: [2, 3]}`` will prune heads 0 and 2 on layer 1 and heads 2 and 3 on layer 2.
+        chunk_size_feed_forward (:obj:`int`, `optional`, defaults to :obj:`0`):
+            The chunk size of all feed forward layers in the residual attention blocks. A chunk size of :obj:`0` means
+            that the feed forward layer is not chunked. A chunk size of n means that the feed forward layer processes
+            :obj:`n` < sequence_length embeddings at a time. For more information on feed forward chunking, see `How
+            does Feed Forward Chunking work? <../glossary.html#feed-forward-chunking>`__ .
+
+    Parameters for sequence generation
+
+        - **max_length** (:obj:`int`, `optional`, defaults to 20) -- Maximum length that will be used by default in the
+          :obj:`generate` method of the model.
+        - **min_length** (:obj:`int`, `optional`, defaults to 10) -- Minimum length that will be used by default in the
+          :obj:`generate` method of the model.
+        - **do_sample** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Flag that will be used by default in the
+          :obj:`generate` method of the model. Whether or not to use sampling ; use greedy decoding otherwise.
+        - **early_stopping** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Flag that will be used by default
+          in the :obj:`generate` method of the model. Whether to stop the beam search when at least ``num_beams``
+          sentences are finished per batch or not.
+        - **num_beams** (:obj:`int`, `optional`, defaults to 1) -- Number of beams for beam search that will be used by
+          default in the :obj:`generate` method of the model. 1 means no beam search.
+        - **num_beam_groups** (:obj:`int`, `optional`, defaults to 1) -- Number of groups to divide :obj:`num_beams`
+          into in order to ensure diversity among different groups of beams that will be used by default in the
+          :obj:`generate` method of the model. 1 means no group beam search.
+        - **diversity_penalty** (:obj:`float`, `optional`, defaults to 0.0) -- Value to control diversity for group
+          beam search. that will be used by default in the :obj:`generate` method of the model. 0 means no diversity
+          penalty. The higher the penalty, the more diverse are the outputs.
+        - **temperature** (:obj:`float`, `optional`, defaults to 1) -- The value used to module the next token
+          probabilities that will be used by default in the :obj:`generate` method of the model. Must be strictly
+          positive.
+        - **top_k** (:obj:`int`, `optional`, defaults to 50) -- Number of highest probability vocabulary tokens to keep
+          for top-k-filtering that will be used by default in the :obj:`generate` method of the model.
+        - **top_p** (:obj:`float`, `optional`, defaults to 1) -- Value that will be used by default in the
+          :obj:`generate` method of the model for ``top_p``. If set to float < 1, only the most probable tokens with
+          probabilities that add up to ``top_p`` or higher are kept for generation.
+        - **repetition_penalty** (:obj:`float`, `optional`, defaults to 1) -- Parameter for repetition penalty that
+          will be used by default in the :obj:`generate` method of the model. 1.0 means no penalty.
+        - **length_penalty** (:obj:`float`, `optional`, defaults to 1) -- Exponential penalty to the length that will
+          be used by default in the :obj:`generate` method of the model.
+        - **no_repeat_ngram_size** (:obj:`int`, `optional`, defaults to 0) -- Value that will be used by default in the
+          :obj:`generate` method of the model for ``no_repeat_ngram_size``. If set to int > 0, all ngrams of that size
+          can only occur once.
+        - **encoder_no_repeat_ngram_size** (:obj:`int`, `optional`, defaults to 0) -- Value that will be used by
+          default in the :obj:`generate` method of the model for ``encoder_no_repeat_ngram_size``. If set to int > 0,
+          all ngrams of that size that occur in the ``encoder_input_ids`` cannot occur in the ``decoder_input_ids``.
+        - **bad_words_ids** (:obj:`List[int]`, `optional`) -- List of token ids that are not allowed to be generated
+          that will be used by default in the :obj:`generate` method of the model. In order to get the tokens of the
+          words that should not appear in the generated text, use :obj:`tokenizer.encode(bad_word,
+          add_prefix_space=True)`.
+        - **num_return_sequences** (:obj:`int`, `optional`, defaults to 1) -- Number of independently computed returned
+          sequences for each element in the batch that will be used by default in the :obj:`generate` method of the
+          model.
+        - **output_scores** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether the model should return the
+          logits when used for generation
+        - **return_dict_in_generate** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether the model should
+          return a :class:`~transformers.file_utils.ModelOutput` instead of a :obj:`torch.LongTensor`
+        - **forced_bos_token_id** (:obj:`int`, `optional`) -- The id of the token to force as the first generated token
+          after the :obj:`decoder_start_token_id`. Useful for multilingual models like :doc:`mBART
+          <../model_doc/mbart>` where the first generated token needs to be the target language token.
+        - **forced_eos_token_id** (:obj:`int`, `optional`) -- The id of the token to force as the last generated token
+          when :obj:`max_length` is reached.
+        - **remove_invalid_values** (:obj:`bool`, `optional`) -- Whether to remove possible `nan` and `inf` outputs of
+          the model to prevent the generation method to crash. Note that using ``remove_invalid_values`` can slow down
+          generation.
+
+
+    Parameters for fine-tuning tasks
+
+        - **architectures** (:obj:`List[str]`, `optional`) -- Model architectures that can be used with the model
+          pretrained weights.
+        - **finetuning_task** (:obj:`str`, `optional`) -- Name of the task used to fine-tune the model. This can be
+          used when converting from an original (TensorFlow or PyTorch) checkpoint.
+        - **id2label** (:obj:`Dict[int, str]`, `optional`) -- A map from index (for instance prediction index, or
+          target index) to label.
+        - **label2id** (:obj:`Dict[str, int]`, `optional`) -- A map from label to index for the model.
+        - **num_labels** (:obj:`int`, `optional`) -- Number of labels to use in the last layer added to the model,
+          typically for a classification task.
+        - **task_specific_params** (:obj:`Dict[str, Any]`, `optional`) -- Additional keyword arguments to store for the
+          current task.
+        - **problem_type** (:obj:`str`, `optional`) -- Problem type for :obj:`XxxForSequenceClassification` models. Can
+          be one of (:obj:`"regression"`, :obj:`"single_label_classification"`, :obj:`"multi_label_classification"`).
+          Please note that this parameter is only available in the following models: `AlbertForSequenceClassification`,
+          `BertForSequenceClassification`, `BigBirdForSequenceClassification`, `ConvBertForSequenceClassification`,
+          `DistilBertForSequenceClassification`, `ElectraForSequenceClassification`, `FunnelForSequenceClassification`,
+          `LongformerForSequenceClassification`, `MobileBertForSequenceClassification`,
+          `ReformerForSequenceClassification`, `RobertaForSequenceClassification`,
+          `SqueezeBertForSequenceClassification`, `XLMForSequenceClassification` and `XLNetForSequenceClassification`.
+
+    Parameters linked to the tokenizer
+
+        - **tokenizer_class** (:obj:`str`, `optional`) -- The name of the associated tokenizer class to use (if none is
+          set, will use the tokenizer associated to the model by default).
+        - **prefix** (:obj:`str`, `optional`) -- A specific prompt that should be added at the beginning of each text
+          before calling the model.
+        - **bos_token_id** (:obj:`int`, `optional`)) -- The id of the `beginning-of-stream` token.
+        - **pad_token_id** (:obj:`int`, `optional`)) -- The id of the `padding` token.
+        - **eos_token_id** (:obj:`int`, `optional`)) -- The id of the `end-of-stream` token.
+        - **decoder_start_token_id** (:obj:`int`, `optional`)) -- If an encoder-decoder model starts decoding with a
+          different token than `bos`, the id of that token.
+        - **sep_token_id** (:obj:`int`, `optional`)) -- The id of the `separation` token.
+
+    PyTorch specific parameters
+
+        - **torchscript** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether or not the model should be
+          used with Torchscript.
+        - **tie_word_embeddings** (:obj:`bool`, `optional`, defaults to :obj:`True`) -- Whether the model's input and
+          output word embeddings should be tied. Note that this is only relevant if the model has a output word
+          embedding layer.
+        - **torch_dtype** (:obj:`str`, `optional`) -- The :obj:`dtype` of the weights. This attribute can be used to
+          initialize the model to a non-default ``dtype`` (which is normally ``float32``) and thus allow for optimal
+          storage allocation. For example, if the saved model is ``float16``, ideally we want to load it back using the
+          minimal amount of memory needed to load ``float16`` weights. Since the config object is stored in plain text,
+          this attribute contains just the floating type string without the ``torch.`` prefix. For example, for
+          ``torch.float16`` ``torch_dtype`` is the ``"float16"`` string.
+
+          This attribute is currently not being used during model loading time, but this may change in the future
+          versions. But we can already start preparing for the future by saving the dtype with save_pretrained.
+
+    TensorFlow specific parameters
+
+        - **use_bfloat16** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether or not the model should use
+          BFloat16 scalars (only used by some TensorFlow models).
+    """
+    model_type: str = ""
+    is_composition: bool = False
+    attribute_map: Dict[str, str] = {}
+
+    def __setattr__(self, key, value):
+        if key in super().__getattribute__("attribute_map"):
+            key = super().__getattribute__("attribute_map")[key]
+        super().__setattr__(key, value)
+
+    def __getattribute__(self, key):
+        if key != "attribute_map" and key in super().__getattribute__("attribute_map"):
+            key = super().__getattribute__("attribute_map")[key]
+        return super().__getattribute__(key)
+
+    def __init__(self, **kwargs):
+        # Attributes with defaults
+        self.return_dict = kwargs.pop("return_dict", True)
+        self.output_hidden_states = kwargs.pop("output_hidden_states", False)
+        self.output_attentions = kwargs.pop("output_attentions", False)
+        self.torchscript = kwargs.pop("torchscript", False)  # Only used by PyTorch models
+        self.torch_dtype = kwargs.pop("torch_dtype", None)  # Only used by PyTorch models
+        self.use_bfloat16 = kwargs.pop("use_bfloat16", False)
+        self.pruned_heads = kwargs.pop("pruned_heads", {})
+        self.tie_word_embeddings = kwargs.pop(
+            "tie_word_embeddings", True
+        )  # Whether input and output word embeddings should be tied for all MLM, LM and Seq2Seq models.
+
+        # Is decoder is used in encoder-decoder models to differentiate encoder from decoder
+        self.is_encoder_decoder = kwargs.pop("is_encoder_decoder", False)
+        self.is_decoder = kwargs.pop("is_decoder", False)
+        self.add_cross_attention = kwargs.pop("add_cross_attention", False)
+        self.tie_encoder_decoder = kwargs.pop("tie_encoder_decoder", False)
+
+        # Parameters for sequence generation
+        self.max_length = kwargs.pop("max_length", 20)
+        self.min_length = kwargs.pop("min_length", 0)
+        self.do_sample = kwargs.pop("do_sample", False)
+        self.early_stopping = kwargs.pop("early_stopping", False)
+        self.num_beams = kwargs.pop("num_beams", 1)
+        self.num_beam_groups = kwargs.pop("num_beam_groups", 1)
+        self.diversity_penalty = kwargs.pop("diversity_penalty", 0.0)
+        self.temperature = kwargs.pop("temperature", 1.0)
+        self.top_k = kwargs.pop("top_k", 50)
+        self.top_p = kwargs.pop("top_p", 1.0)
+        self.repetition_penalty = kwargs.pop("repetition_penalty", 1.0)
+        self.length_penalty = kwargs.pop("length_penalty", 1.0)
+        self.no_repeat_ngram_size = kwargs.pop("no_repeat_ngram_size", 0)
+        self.encoder_no_repeat_ngram_size = kwargs.pop("encoder_no_repeat_ngram_size", 0)
+        self.bad_words_ids = kwargs.pop("bad_words_ids", None)
+        self.num_return_sequences = kwargs.pop("num_return_sequences", 1)
+        self.chunk_size_feed_forward = kwargs.pop("chunk_size_feed_forward", 0)
+        self.output_scores = kwargs.pop("output_scores", False)
+        self.return_dict_in_generate = kwargs.pop("return_dict_in_generate", False)
+        self.forced_bos_token_id = kwargs.pop("forced_bos_token_id", None)
+        self.forced_eos_token_id = kwargs.pop("forced_eos_token_id", None)
+        self.remove_invalid_values = kwargs.pop("remove_invalid_values", False)
+
+        # Fine-tuning task arguments
+        self.architectures = kwargs.pop("architectures", None)
+        self.finetuning_task = kwargs.pop("finetuning_task", None)
+        self.id2label = kwargs.pop("id2label", None)
+        self.label2id = kwargs.pop("label2id", None)
+        if self.id2label is not None:
+            kwargs.pop("num_labels", None)
+            self.id2label = dict((int(key), value) for key, value in self.id2label.items())
+            # Keys are always strings in JSON so convert ids to int here.
+        else:
+            self.num_labels = kwargs.pop("num_labels", 2)
+
+        if self.torch_dtype is not None and isinstance(self.torch_dtype, str):
+            # we will start using self.torch_dtype in v5, but to be consistent with
+            # from_pretrained's torch_dtype arg convert it to an actual torch.dtype object
+            if _NEED_IMPORT_TORCH:
+                import torch
+
+                self.torch_dtype = getattr(torch, self.torch_dtype)
+
+        # Tokenizer arguments TODO: eventually tokenizer and models should share the same config
+        self.tokenizer_class = kwargs.pop("tokenizer_class", None)
+        self.prefix = kwargs.pop("prefix", None)
+        self.bos_token_id = kwargs.pop("bos_token_id", None)
+        self.pad_token_id = kwargs.pop("pad_token_id", None)
+        self.eos_token_id = kwargs.pop("eos_token_id", None)
+        self.sep_token_id = kwargs.pop("sep_token_id", None)
+
+        self.decoder_start_token_id = kwargs.pop("decoder_start_token_id", None)
+
+        # task specific arguments
+        self.task_specific_params = kwargs.pop("task_specific_params", None)
+
+        # regression / multi-label classification
+        self.problem_type = kwargs.pop("problem_type", None)
+        allowed_problem_types = ("regression", "single_label_classification", "multi_label_classification")
+        if self.problem_type is not None and self.problem_type not in allowed_problem_types:
+            raise ValueError(
+                f"The config parameter `problem_type` was not understood: received {self.problem_type}"
+                "but only 'regression', 'single_label_classification' and 'multi_label_classification' are valid."
+            )
+
+        # TPU arguments
+        if kwargs.pop("xla_device", None) is not None:
+            logger.warning(
+                "The `xla_device` argument has been deprecated in v4.4.0 of Transformers. It is ignored and you can "
+                "safely remove it from your `config.json` file."
+            )
+
+        # Name or path to the pretrained checkpoint
+        self._name_or_path = str(kwargs.pop("name_or_path", ""))
+
+        # Drop the transformers version info
+        self.transformers_version = kwargs.pop("transformers_version", None)
+
+        # Deal with gradient checkpointing
+        if kwargs.get("gradient_checkpointing", False):
+            logger.warn(
+                "Passing `gradient_checkpointing` to a config initialization is deprecated and will be removed in v5 "
+                "Transformers. Using `model.gradient_checkpointing_enable()` instead, or if you are using the "
+                "`Trainer` API, pass `gradient_checkpointing=True` in your `TrainingArguments`."
+            )
+
+        # Additional attributes without default values
+        for key, value in kwargs.items():
+            try:
+                setattr(self, key, value)
+            except AttributeError as err:
+                logger.error(f"Can't set {key} with value {value} for {self}")
+                raise err
+
+    @property
+    def name_or_path(self) -> str:
+        return self._name_or_path
+
+    @name_or_path.setter
+    def name_or_path(self, value):
+        self._name_or_path = str(value)  # Make sure that name_or_path is a string (for JSON encoding)
+
+    @property
+    def use_return_dict(self) -> bool:
+        """
+        :obj:`bool`: Whether or not return :class:`~transformers.file_utils.ModelOutput` instead of tuples.
+        """
+        # If torchscript is set, force `return_dict=False` to avoid jit errors
+        return self.return_dict and not self.torchscript
+
+    @property
+    def num_labels(self) -> int:
+        """
+        :obj:`int`: The number of labels for classification models.
+        """
+        return len(self.id2label)
+
+    @num_labels.setter
+    def num_labels(self, num_labels: int):
+        if not hasattr(self, "id2label") or self.id2label is None or len(self.id2label) != num_labels:
+            self.id2label = {i: f"LABEL_{i}" for i in range(num_labels)}
+            self.label2id = dict(zip(self.id2label.values(), self.id2label.keys()))
+
+    def save_pretrained(self, save_directory: Union[str, os.PathLike], **kwargs):
+        """
+        Save a configuration object to the directory ``save_directory``, so that it can be re-loaded using the
+        :func:`~transformers.PretrainedConfig.from_pretrained` class method.
+
+        Args:
+            save_directory (:obj:`str` or :obj:`os.PathLike`):
+                Directory where the configuration JSON file will be saved (will be created if it does not exist).
+            push_to_hub (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to push your model to the Hugging Face model hub after saving it.
+
+                .. warning::
+
+                    Using :obj:`push_to_hub=True` will synchronize the repository you are pushing to with
+                    :obj:`save_directory`, which requires :obj:`save_directory` to be a local clone of the repo you are
+                    pushing to if it's an existing folder. Pass along :obj:`temp_dir=True` to use a temporary directory
+                    instead.
+
+            kwargs:
+                Additional key word arguments passed along to the
+                :meth:`~transformers.file_utils.PushToHubMixin.push_to_hub` method.
+        """
+        if os.path.isfile(save_directory):
+            raise AssertionError(f"Provided path ({save_directory}) should be a directory, not a file")
+
+        os.makedirs(save_directory, exist_ok=True)
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_config_file = os.path.join(save_directory, CONFIG_NAME)
+
+        self.to_json_file(output_config_file, use_diff=True)
+        logger.info(f"Configuration saved in {output_config_file}")
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        r"""
+        Instantiate a :class:`~transformers.PretrainedConfig` (or a derived class) from a pretrained model
+        configuration.
+
+        Args:
+            pretrained_model_name_or_path (:obj:`str` or :obj:`os.PathLike`):
+                This can be either:
+
+                - a string, the `model id` of a pretrained model configuration hosted inside a model repo on
+                  huggingface.co. Valid model ids can be located at the root-level, like ``bert-base-uncased``, or
+                  namespaced under a user or organization name, like ``dbmdz/bert-base-german-cased``.
+                - a path to a `directory` containing a configuration file saved using the
+                  :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g., ``./my_model_directory/``.
+                - a path or url to a saved configuration JSON `file`, e.g.,
+                  ``./my_model_directory/configuration.json``.
+            cache_dir (:obj:`str` or :obj:`os.PathLike`, `optional`):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            force_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to force to (re-)download the configuration files and override the cached versions if
+                they exist.
+            resume_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to delete incompletely received file. Attempts to resume the download if such a file
+                exists.
+            proxies (:obj:`Dict[str, str]`, `optional`):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., :obj:`{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+            use_auth_token (:obj:`str` or `bool`, `optional`):
+                The token to use as HTTP bearer authorization for remote files. If :obj:`True`, will use the token
+                generated when running :obj:`transformers-cli login` (stored in :obj:`~/.huggingface`).
+            revision(:obj:`str`, `optional`, defaults to :obj:`"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so ``revision`` can be any
+                identifier allowed by git.
+            return_unused_kwargs (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                If :obj:`False`, then this function returns just the final configuration object.
+
+                If :obj:`True`, then this functions returns a :obj:`Tuple(config, unused_kwargs)` where `unused_kwargs`
+                is a dictionary consisting of the key/value pairs whose keys are not configuration attributes: i.e.,
+                the part of ``kwargs`` which has not been used to update ``config`` and is otherwise ignored.
+            kwargs (:obj:`Dict[str, Any]`, `optional`):
+                The values in kwargs of any keys which are configuration attributes will be used to override the loaded
+                values. Behavior concerning key/value pairs whose keys are *not* configuration attributes is controlled
+                by the ``return_unused_kwargs`` keyword parameter.
+
+        .. note::
+
+            Passing :obj:`use_auth_token=True` is required when you want to use a private model.
+
+
+        Returns:
+            :class:`PretrainedConfig`: The configuration object instantiated from this pretrained model.
+
+        Examples::
+
+            # We can't instantiate directly the base class `PretrainedConfig` so let's show the examples on a
+            # derived class: BertConfig
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from huggingface.co and cache.
+            config = BertConfig.from_pretrained('./test/saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
+            config = BertConfig.from_pretrained('./test/saved_model/my_configuration.json')
+            config = BertConfig.from_pretrained('bert-base-uncased', output_attentions=True, foo=False)
+            assert config.output_attentions == True
+            config, unused_kwargs = BertConfig.from_pretrained('bert-base-uncased', output_attentions=True,
+                                                               foo=False, return_unused_kwargs=True)
+            assert config.output_attentions == True
+            assert unused_kwargs == {'foo': False}
+
+        """
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warn(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+    @classmethod
+    def get_config_dict(
+        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+    ) -> Tuple[Dict[str, Any], Dict[str, Any]]:
+        """
+        From a ``pretrained_model_name_or_path``, resolve to a dictionary of parameters, to be used for instantiating a
+        :class:`~transformers.PretrainedConfig` using ``from_dict``.
+
+
+
+        Parameters:
+            pretrained_model_name_or_path (:obj:`str` or :obj:`os.PathLike`):
+                The identifier of the pre-trained checkpoint from which we want the dictionary of parameters.
+
+        Returns:
+            :obj:`Tuple[Dict, Dict]`: The dictionary(ies) that will be used to instantiate the configuration object.
+
+        """
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", False)
+        revision = kwargs.pop("revision", None)
+        from_pipeline = kwargs.pop("_from_pipeline", None)
+        from_auto_class = kwargs.pop("_from_auto", False)
+
+        user_agent = {"file_type": "config", "from_auto_class": from_auto_class}
+        if from_pipeline is not None:
+            user_agent["using_pipeline"] = from_pipeline
+
+        if is_offline_mode() and not local_files_only:
+            logger.info("Offline mode: forcing local_files_only=True")
+            local_files_only = True
+
+        pretrained_model_name_or_path = str(pretrained_model_name_or_path)
+        if os.path.isdir(pretrained_model_name_or_path):
+            config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
+        elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
+            config_file = pretrained_model_name_or_path
+        else:
+            config_file = hf_bucket_url(
+                pretrained_model_name_or_path, filename=CONFIG_NAME, revision=revision, mirror=None
+            )
+
+        try:
+            # Load from URL or cache if already cached
+            resolved_config_file = cached_path(
+                config_file,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                user_agent=user_agent,
+            )
+            # Load config dict
+            config_dict = cls._dict_from_json_file(resolved_config_file)
+
+        except EnvironmentError as err:
+            logger.error(err)
+            msg = (
+                f"Can't load config for '{pretrained_model_name_or_path}'. Make sure that:\n\n"
+                f"- '{pretrained_model_name_or_path}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
+                f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a {CONFIG_NAME} file\n\n"
+            )
+
+            if revision is not None:
+                msg += f"- or '{revision}' is a valid git identifier (branch name, a tag name, or a commit id) that exists for this model name as listed on its model page on 'https://huggingface.co/models'\n\n"
+
+            raise EnvironmentError(msg)
+
+        except (json.JSONDecodeError, UnicodeDecodeError):
+            msg = (
+                f"Couldn't reach server at '{config_file}' to download configuration file or "
+                "configuration file is not a valid JSON file. "
+                f"Please check network or file content here: {resolved_config_file}."
+            )
+            raise EnvironmentError(msg)
+
+        if resolved_config_file == config_file:
+            logger.info(f"loading configuration file {config_file}")
+        else:
+            logger.info(f"loading configuration file {config_file} from cache at {resolved_config_file}")
+
+        return config_dict, kwargs
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any], **kwargs) -> "PretrainedConfig":
+        """
+        Instantiates a :class:`~transformers.PretrainedConfig` from a Python dictionary of parameters.
+
+        Args:
+            config_dict (:obj:`Dict[str, Any]`):
+                Dictionary that will be used to instantiate the configuration object. Such a dictionary can be
+                retrieved from a pretrained checkpoint by leveraging the
+                :func:`~transformers.PretrainedConfig.get_config_dict` method.
+            kwargs (:obj:`Dict[str, Any]`):
+                Additional parameters from which to initialize the configuration object.
+
+        Returns:
+            :class:`PretrainedConfig`: The configuration object instantiated from those parameters.
+        """
+        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
+
+        config = cls(**config_dict)
+
+        if hasattr(config, "pruned_heads"):
+            config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
+
+        # Update config with kwargs if needed
+        to_remove = []
+        for key, value in kwargs.items():
+            if hasattr(config, key):
+                setattr(config, key, value)
+                if key != "torch_dtype":
+                    to_remove.append(key)
+        for key in to_remove:
+            kwargs.pop(key, None)
+
+        logger.info(f"Model config {config}")
+        if return_unused_kwargs:
+            return config, kwargs
+        else:
+            return config
+
+    @classmethod
+    def from_json_file(cls, json_file: Union[str, os.PathLike]) -> "PretrainedConfig":
+        """
+        Instantiates a :class:`~transformers.PretrainedConfig` from the path to a JSON file of parameters.
+
+        Args:
+            json_file (:obj:`str` or :obj:`os.PathLike`):
+                Path to the JSON file containing the parameters.
+
+        Returns:
+            :class:`PretrainedConfig`: The configuration object instantiated from that JSON file.
+
+        """
+        config_dict = cls._dict_from_json_file(json_file)
+        return cls(**config_dict)
+
+    @classmethod
+    def _dict_from_json_file(cls, json_file: Union[str, os.PathLike]):
+        with open(json_file, "r", encoding="utf-8") as reader:
+            text = reader.read()
+        return json.loads(text)
+
+    def __eq__(self, other):
+        return self.__dict__ == other.__dict__
+
+    def __repr__(self):
+        return f"{self.__class__.__name__} {self.to_json_string()}"
+
+    def to_diff_dict(self) -> Dict[str, Any]:
+        """
+        Removes all attributes from config which correspond to the default config attributes for better readability and
+        serializes to a Python dictionary.
+
+        Returns:
+            :obj:`Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        config_dict = self.to_dict()
+
+        # get the default config dict
+        default_config_dict = PretrainedConfig().to_dict()
+
+        # get class specific config dict
+        class_config_dict = self.__class__().to_dict() if not self.is_composition else {}
+
+        serializable_config_dict = {}
+
+        # only serialize values that differ from the default config
+        for key, value in config_dict.items():
+            if (
+                key not in default_config_dict
+                or key == "transformers_version"
+                or value != default_config_dict[key]
+                or (key in class_config_dict and value != class_config_dict[key])
+            ):
+                serializable_config_dict[key] = value
+
+        self.dict_torch_dtype_to_str(serializable_config_dict)
+
+        return serializable_config_dict
+
+    def to_dict(self) -> Dict[str, Any]:
+        """
+        Serializes this instance to a Python dictionary.
+
+        Returns:
+            :obj:`Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance.
+        """
+        output = copy.deepcopy(self.__dict__)
+        if hasattr(self.__class__, "model_type"):
+            output["model_type"] = self.__class__.model_type
+
+        # Transformers version when serializing the model
+        output["transformers_version"] = __version__
+
+        self.dict_torch_dtype_to_str(output)
+
+        return output
+
+    def to_json_string(self, use_diff: bool = True) -> str:
+        """
+        Serializes this instance to a JSON string.
+
+        Args:
+            use_diff (:obj:`bool`, `optional`, defaults to :obj:`True`):
+                If set to ``True``, only the difference between the config instance and the default
+                ``PretrainedConfig()`` is serialized to JSON string.
+
+        Returns:
+            :obj:`str`: String containing all the attributes that make up this configuration instance in JSON format.
+        """
+        if use_diff is True:
+            config_dict = self.to_diff_dict()
+        else:
+            config_dict = self.to_dict()
+        return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
+
+    def to_json_file(self, json_file_path: Union[str, os.PathLike], use_diff: bool = True):
+        """
+        Save this instance to a JSON file.
+
+        Args:
+            json_file_path (:obj:`str` or :obj:`os.PathLike`):
+                Path to the JSON file in which this configuration instance's parameters will be saved.
+            use_diff (:obj:`bool`, `optional`, defaults to :obj:`True`):
+                If set to ``True``, only the difference between the config instance and the default
+                ``PretrainedConfig()`` is serialized to JSON file.
+        """
+        with open(json_file_path, "w", encoding="utf-8") as writer:
+            writer.write(self.to_json_string(use_diff=use_diff))
+
+    def update(self, config_dict: Dict[str, Any]):
+        """
+        Updates attributes of this class with attributes from ``config_dict``.
+
+        Args:
+            config_dict (:obj:`Dict[str, Any]`): Dictionary of attributes that should be updated for this class.
+        """
+        for key, value in config_dict.items():
+            setattr(self, key, value)
+
+    def update_from_string(self, update_str: str):
+        """
+        Updates attributes of this class with attributes from ``update_str``.
+
+        The expected format is ints, floats and strings as is, and for booleans use ``true`` or ``false``. For example:
+        "n_embd=10,resid_pdrop=0.2,scale_attn_weights=false,summary_type=cls_index"
+
+        The keys to change have to already exist in the config object.
+
+        Args:
+            update_str (:obj:`str`): String with attributes that should be updated for this class.
+
+        """
+
+        d = dict(x.split("=") for x in update_str.split(","))
+        for k, v in d.items():
+            if not hasattr(self, k):
+                raise ValueError(f"key {k} isn't in the original config dict")
+
+            old_v = getattr(self, k)
+            if isinstance(old_v, bool):
+                if v.lower() in ["true", "1", "y", "yes"]:
+                    v = True
+                elif v.lower() in ["false", "0", "n", "no"]:
+                    v = False
+                else:
+                    raise ValueError(f"can't derive true or false from {v} (key {k})")
+            elif isinstance(old_v, int):
+                v = int(v)
+            elif isinstance(old_v, float):
+                v = float(v)
+            elif not isinstance(old_v, str):
+                raise ValueError(
+                    f"You can only update int, float, bool or string values in the config, got {v} for key {k}"
+                )
+
+            setattr(self, k, v)
+
+    def dict_torch_dtype_to_str(self, d: Dict[str, Any]) -> None:
+        """
+        Checks whether the passed dictionary has a `torch_dtype` key and if it's not None, converts torch.dtype to a
+        string of just the type. For example, :obj:`torch.float32` get converted into `"float32"` string, which can
+        then be stored in the json format.
+        """
+        if d.get("torch_dtype", None) is not None and not isinstance(d["torch_dtype"], str):
+            d["torch_dtype"] = str(d["torch_dtype"]).split(".")[1]

fastNLP/transformers/torch/deepspeed.py

@@ -0,0 +1,388 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Integration with Deepspeed
+"""
+
+import importlib.util
+import io
+import json
+import weakref
+from copy import deepcopy
+from functools import partialmethod
+
+from .dependency_versions_check import dep_version_check
+from fastNLP.envs.imports import _NEED_IMPORT_TORCH
+from fastNLP.core.log import logger
+
+if _NEED_IMPORT_TORCH:
+    import torch
+
+
+def is_deepspeed_available():
+    return importlib.util.find_spec("deepspeed") is not None
+
+
+class HfDeepSpeedConfig:
+    """
+    This object contains a DeepSpeed configuration dictionary and can be quickly queried for things like zero stage.
+
+    A ``weakref`` of this object is stored in the module's globals to be able to access the config from areas where
+    things like the Trainer object is not available (e.g. ``from_pretrained`` and ``_get_resized_embeddings``).
+    Therefore it's important that this object remains alive while the program is still running.
+
+    :class:`~transformers.Trainer` uses the ``HfTrainerDeepSpeedConfig`` subclass instead. That subclass has logic to
+    sync the configuration with values of :class:`~transformers.TrainingArguments` by replacing special placeholder
+    values: ``"auto"``. Without this special logic the DeepSpeed configuration is not modified in any way.
+
+    Args:
+        config_file_or_dict (:obj:`Union[str, Dict]`): path to DeepSpeed config file or dict.
+
+    """
+
+    def __init__(self, config_file_or_dict):
+        # set global weakref object
+        set_hf_deepspeed_config(self)
+
+        dep_version_check("deepspeed")
+
+        if isinstance(config_file_or_dict, dict):
+            # Don't modify user's data should they want to reuse it (e.g. in tests), because once we
+            # modified it, it will not be accepted here again, since `auto` values would have been overridden
+            config = deepcopy(config_file_or_dict)
+        elif isinstance(config_file_or_dict, str):
+            with io.open(config_file_or_dict, "r", encoding="utf-8") as f:
+                config = json.load(f)
+        else:
+            raise ValueError("expecting either a path to a DeepSpeed config file or a pre-populated dict")
+        self.config = config
+
+        # zero stage - this is done as early as possible, before model is created, to allow
+        # ``is_deepspeed_zero3_enabled`` query and getting to the early deepspeed config object
+        # during ``zero.Init()`` which needs whether fp16 is enabled, dtype, etc.
+        self._stage = self.get_value("zero_optimization.stage", -1)
+
+        # offload
+        self._offload = False
+        if self.is_zero2() or self.is_zero3():
+            offload_devices_valid = set(["cpu", "nvme"])
+            offload_devices = set(
+                [
+                    self.get_value("zero_optimization.offload_optimizer.device"),
+                    self.get_value("zero_optimization.offload_param.device"),
+                ]
+            )
+            if len(offload_devices & offload_devices_valid) > 0:
+                self._offload = True
+
+    def find_config_node(self, ds_key_long):
+        config = self.config
+
+        # find the config node of interest if it exists
+        nodes = ds_key_long.split(".")
+        ds_key = nodes.pop()
+        for node in nodes:
+            config = config.get(node)
+            if config is None:
+                return None, ds_key
+
+        return config, ds_key
+
+    def get_value(self, ds_key_long, default=None):
+        """
+        Returns the set value or ``default`` if no value is set
+        """
+        config, ds_key = self.find_config_node(ds_key_long)
+        if config is None:
+            return default
+        return config.get(ds_key, default)
+
+    def is_true(self, ds_key_long):
+        """
+        Returns :obj:`True`/:obj:`False` only if the value is set, always :obj:`False` otherwise. So use this method to
+        ask the very specific question of whether the value is set to :obj:`True` (and it's not set to :obj:`False` or
+        isn't set).
+
+        """
+        value = self.get_value(ds_key_long)
+        return False if value is None else bool(value)
+
+    def is_false(self, ds_key_long):
+        """
+        Returns :obj:`True`/:obj:`False` only if the value is set, always :obj:`False` otherwise. So use this method to
+        ask the very specific question of whether the value is set to :obj:`False` (and it's not set to :obj:`True` or
+        isn't set).
+        """
+        value = self.get_value(ds_key_long)
+        return False if value is None else not bool(value)
+
+    def is_zero2(self):
+        return self._stage == 2
+
+    def is_zero3(self):
+        return self._stage == 3
+
+    def is_offload(self):
+        return self._offload
+
+
+class HfTrainerDeepSpeedConfig(HfDeepSpeedConfig):
+    """
+    The ``HfTrainerDeepSpeedConfig`` object is meant to be created during ``TrainingArguments`` object creation and has
+    the same lifespan as the latter.
+    """
+
+    def __init__(self, config_file_or_dict):
+        super().__init__(config_file_or_dict)
+        self._dtype = torch.float16
+        self.mismatches = []
+
+    def dtype(self):
+        return self._dtype
+
+    def fill_match(self, ds_key_long, hf_val, hf_key=None, must_match=True):
+        """
+        A utility method that massages the config file and can optionally verify that the values match.
+
+        1. Replace "auto" values with ``TrainingArguments`` value.
+
+        2. If it wasn't "auto" and ``must_match`` is true, then check that DS config matches Trainer
+        config values and if mismatched add the entry to ``self.mismatched`` - will assert during
+        ``trainer_config_finalize`` for one or more mismatches.
+
+        """
+        config, ds_key = self.find_config_node(ds_key_long)
+        if config is None:
+            return
+
+        if config.get(ds_key) == "auto":
+            config[ds_key] = hf_val
+            return
+
+        if not must_match:
+            return
+
+        ds_val = config.get(ds_key)
+        if ds_val is not None and ds_val != hf_val:
+            self.mismatches.append(f"- ds {ds_key_long}={ds_val} vs hf {hf_key}={hf_val}")
+
+    fill_only = partialmethod(fill_match, must_match=False)
+
+    def trainer_config_process(self, args):
+        """
+        Adjust the config with ``TrainingArguments`` values. This stage is run during ``TrainingArguments`` object
+        creation.
+        """
+        # DeepSpeed does:
+        # train_batch_size = world_size * train_micro_batch_size_per_gpu * gradient_accumulation_steps
+        train_batch_size = args.world_size * args.per_device_train_batch_size * args.gradient_accumulation_steps
+        self.fill_match(
+            "train_micro_batch_size_per_gpu", args.per_device_train_batch_size, "per_device_train_batch_size"
+        )
+        self.fill_match("gradient_accumulation_steps", args.gradient_accumulation_steps, "gradient_accumulation_steps")
+        self.fill_match("train_batch_size", train_batch_size, "train_batch_size (calculated)")
+        self.fill_match("gradient_clipping", args.max_grad_norm, "max_grad_norm")
+
+        self.fill_match("optimizer.params.lr", args.learning_rate, "learning_rate")
+        self.fill_match("optimizer.params.betas", [args.adam_beta1, args.adam_beta2], "adam_beta1+adam_beta2")
+        self.fill_match("optimizer.params.eps", args.adam_epsilon, "adam_epsilon")
+        self.fill_match("optimizer.params.weight_decay", args.weight_decay, "weight_decay")
+
+        self.fill_only("scheduler.params.warmup_min_lr", 0)  # not a trainer arg
+        self.fill_match("scheduler.params.warmup_max_lr", args.learning_rate, "learning_rate")
+        # total_num_steps - will get set in trainer_config_finalize
+
+        # fp16
+        if args.fp16:
+            fp16_backend = "apex" if args.fp16_backend == "apex" else "amp"
+        else:
+            fp16_backend = None
+
+        # amp: similar to the pytorch native amp - it has a bunch of optional params but we won't set
+        # any here unless the user did the work
+        self.fill_match("fp16.enabled", fp16_backend == "amp", "fp16+fp16_backend(amp)")
+
+        # apex: delegates amp work to apex (which needs to be available), but it cannot be used with any
+        # ZeRO features
+        self.fill_match("amp.enabled", fp16_backend == "apex", "fp16+fp16_backend(apex)")
+        self.fill_match("amp.opt_level", args.fp16_opt_level, "fp16_opt_level")
+
+        # only if we have an explicit fp16.enabled = False then it's fp32, if it's True or this
+        # whole config section is missing then the fallback is fp16
+        if self.is_false("fp16.enabled"):
+            self._dtype = torch.float32
+        # later there will be other dtypes besides just fp16 and fp32
+        # also not quite sure what dtype should be under apex, defaulting to fp16 for now
+
+    def trainer_config_finalize(self, args, model, num_training_steps):
+        """
+        This stage is run after we have the model and know num_training_steps.
+
+        Now we we can complete the configuration process.
+        """
+        # zero
+        if self.is_zero3():
+            # automatically assign the optimal config values based on model config
+            hidden_size = model.config.hidden_size
+            self.fill_only("zero_optimization.reduce_bucket_size", hidden_size * hidden_size)
+            self.fill_only("zero_optimization.stage3_prefetch_bucket_size", 0.9 * hidden_size * hidden_size)
+            self.fill_only("zero_optimization.stage3_param_persistence_threshold", 10 * hidden_size)
+
+        # scheduler
+        self.fill_match("scheduler.params.total_num_steps", num_training_steps, "num_training_steps (calculated)")
+        self.fill_match("scheduler.params.warmup_num_steps", args.get_warmup_steps(num_training_steps), "warmup_steps")
+
+        if len(self.mismatches) > 0:
+            mismatches = "\n".join(self.mismatches)
+            raise ValueError(
+                f"Please correct the following DeepSpeed config values that mismatch TrainingArguments values:\n{mismatches}\n"
+                "The easiest method is to set these DeepSpeed config values to 'auto'."
+            )
+
+
+# keep the config object global to be able to access it anywhere during TrainingArguments life-cycle
+_hf_deepspeed_config_weak_ref = None
+
+
+def set_hf_deepspeed_config(hf_deepspeed_config_obj):
+    # this is a special weakref global object to allow us to get to Deepspeed config from APIs
+    # that don't have an easy way to get to the Deepspeed config outside of the Trainer domain.
+    global _hf_deepspeed_config_weak_ref
+    # will go away automatically when HfDeepSpeedConfig is destroyed (when TrainingArguments is destroyed)
+    _hf_deepspeed_config_weak_ref = weakref.ref(hf_deepspeed_config_obj)
+
+
+def is_deepspeed_zero3_enabled():
+    if _hf_deepspeed_config_weak_ref is not None and _hf_deepspeed_config_weak_ref() is not None:
+        return _hf_deepspeed_config_weak_ref().is_zero3()
+    else:
+        return False
+
+
+def deepspeed_config():
+    if _hf_deepspeed_config_weak_ref is not None and _hf_deepspeed_config_weak_ref() is not None:
+        return _hf_deepspeed_config_weak_ref().config
+    else:
+        return None
+
+
+def deepspeed_init(trainer, num_training_steps, resume_from_checkpoint=None):
+    """
+    Init DeepSpeed, after updating the DeepSpeed configuration with any relevant Trainer's args.
+
+    If ``resume_from_checkpoint`` was passed then an attempt to resume from a previously saved checkpoint will be made.
+
+    Args:
+        trainer: Trainer object
+        num_training_steps: per single gpu
+        resume_from_checkpoint: path to a checkpoint if to resume from after normal DeepSpeedEngine load
+
+    Returns: model, optimizer, lr_scheduler
+
+    """
+    import deepspeed
+    from deepspeed.utils import logger as ds_logger
+
+    model = trainer.model
+    args = trainer.args
+
+    hf_deepspeed_config = args.hf_deepspeed_config
+    hf_deepspeed_config.trainer_config_finalize(args, model, num_training_steps)
+
+    # resume config update - some bits like `model` and `num_training_steps` only become available during train
+    config = hf_deepspeed_config.config
+
+    # Optimizer + Scheduler
+    # Currently supported combos:
+    # 1. DS scheduler + DS optimizer: Yes
+    # 2. HF scheduler + HF optimizer: Yes
+    # 3. DS scheduler + HF optimizer: Yes
+    # 4. HF scheduler + DS optimizer: Yes
+    #
+    # Unless Offload is enabled in which case it's:
+    # 1. DS scheduler + DS optimizer: Yes
+    # 2. HF scheduler + HF optimizer: Mostly*
+    # 3. DS scheduler + HF optimizer: Mostly*
+    # 4. HF scheduler + DS optimizer: Yes
+    #
+    # Mostly*: All non-native DeepSpeed optimizers that have both CPU and GPU implementation should work (except LAMB)
+
+    optimizer = None
+    if "optimizer" in config:
+        if args.adafactor:
+            raise ValueError(
+                "--adafactor was passed, but also found `optimizer` configured in the DeepSpeed config. "
+                "Only one optimizer can be configured."
+            )
+    else:
+        if hf_deepspeed_config.is_offload():
+            logger.info(
+                "Detected ZeRO Offload and non-DeepSpeed optimizers: This combination should work as long as the custom optimizer has both CPU and GPU implementation (except LAMB)"
+            )
+
+        # ds supports Adam, OneBitAdam, and Lamb optimizers and can import other optimizers from torch.
+        # But trainer uses AdamW by default.
+        optimizer = trainer.create_optimizer()
+        # To use other optimizers requires voiding warranty with: `zero_allow_untested_optimizer`
+        config["zero_allow_untested_optimizer"] = True
+
+    def _lr_scheduler_callable(optimizer):
+        return trainer.create_scheduler(num_training_steps=num_training_steps, optimizer=optimizer)
+
+    lr_scheduler = None
+    if "scheduler" not in config:
+        if optimizer is None:
+            # Optimizer is not available, so use callable to defer lr_scheduler creation to DS init
+            lr_scheduler = _lr_scheduler_callable
+        else:
+            lr_scheduler = trainer.create_scheduler(num_training_steps=num_training_steps, optimizer=optimizer)
+
+    # keep for quick debug:
+    # from pprint import pprint; pprint(config)
+
+    # set the Deepspeed log level consistent with the trainer
+    ds_logger.setLevel(args.get_process_log_level())
+
+    model_parameters = filter(lambda p: p.requires_grad, model.parameters())
+
+    model, optimizer, _, lr_scheduler = deepspeed.initialize(
+        model=model,
+        model_parameters=model_parameters,
+        config_params=config,
+        optimizer=optimizer,
+        lr_scheduler=lr_scheduler,
+    )
+
+    if resume_from_checkpoint is not None:
+
+        # it's possible that the user is trying to resume from model_path, which doesn't necessarily
+        # contain a deepspeed checkpoint. e.g. examples just check if the dir exists and assume it's
+        # a resume from a checkpoint and not just a local pretrained weight. So we check here if the
+        # path contains what looks like a deepspeed checkpoint
+        import glob
+
+        deepspeed_checkpoint_dirs = sorted(glob.glob(f"{resume_from_checkpoint}/global_step*"))
+
+        if len(deepspeed_checkpoint_dirs) > 0:
+            logger.info(f"Attempting to resume from {resume_from_checkpoint}")
+            # this magically updates self.optimizer and self.lr_scheduler
+            load_path, _ = model.load_checkpoint(
+                resume_from_checkpoint, load_optimizer_states=True, load_lr_scheduler_states=True
+            )
+            if load_path is None:
+                raise ValueError(f"[deepspeed] failed to resume from checkpoint {resume_from_checkpoint}")
+        else:
+            logger.info(f"{resume_from_checkpoint} doesn't have deepspeed checkpoints, doing nothing")
+
+    return model, optimizer, lr_scheduler

fastNLP/transformers/torch/dependency_versions_check.py

@@ -0,0 +1,20 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import sys
+
+from .dependency_versions_table import deps
+from .utils.versions import require_version
+
+def dep_version_check(pkg, hint=None):
+    require_version(deps[pkg], hint)

fastNLP/transformers/torch/dependency_versions_table.py

@@ -0,0 +1,76 @@
+# THIS FILE HAS BEEN AUTOGENERATED. To update:
+# 1. modify the `_deps` dict in setup.py
+# 2. run `make deps_table_update``
+deps = {
+    "Pillow": "Pillow",
+    "black": "black==21.4b0",
+    "codecarbon": "codecarbon==1.2.0",
+    "cookiecutter": "cookiecutter==1.7.2",
+    "dataclasses": "dataclasses",
+    "datasets": "datasets",
+    "deepspeed": "deepspeed>=0.5.3",
+    "docutils": "docutils==0.16.0",
+    "fairscale": "fairscale>0.3",
+    "faiss-cpu": "faiss-cpu",
+    "fastapi": "fastapi",
+    "filelock": "filelock",
+    "flake8": "flake8>=3.8.3",
+    "flax": "flax>=0.3.4",
+    "fugashi": "fugashi>=1.0",
+    "GitPython": "GitPython<3.1.19",
+    "huggingface-hub": "huggingface-hub>=0.0.17",
+    "importlib_metadata": "importlib_metadata",
+    "ipadic": "ipadic>=1.0.0,<2.0",
+    "isort": "isort>=5.5.4",
+    "jax": "jax>=0.2.8",
+    "jaxlib": "jaxlib>=0.1.65",
+    "jieba": "jieba",
+    "keras2onnx": "keras2onnx",
+    "nltk": "nltk",
+    "numpy": "numpy>=1.17",
+    "onnxconverter-common": "onnxconverter-common",
+    "onnxruntime-tools": "onnxruntime-tools>=1.4.2",
+    "onnxruntime": "onnxruntime>=1.4.0",
+    "optuna": "optuna",
+    "optax": "optax>=0.0.8",
+    "packaging": "packaging>=20.0",
+    "parameterized": "parameterized",
+    "protobuf": "protobuf",
+    "psutil": "psutil",
+    "pyyaml": "pyyaml>=5.1",
+    "pydantic": "pydantic",
+    "pytest": "pytest",
+    "pytest-timeout": "pytest-timeout",
+    "pytest-xdist": "pytest-xdist",
+    "python": "python>=3.6.0",
+    "ray[tune]": "ray[tune]",
+    "recommonmark": "recommonmark",
+    "regex": "regex!=2019.12.17",
+    "requests": "requests",
+    "rouge-score": "rouge-score",
+    "sacrebleu": "sacrebleu>=1.4.12,<2.0.0",
+    "sacremoses": "sacremoses",
+    "sagemaker": "sagemaker>=2.31.0",
+    "scikit-learn": "scikit-learn",
+    "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",
+    "sigopt": "sigopt",
+    "soundfile": "soundfile",
+    "sphinx-copybutton": "sphinx-copybutton",
+    "sphinx-markdown-tables": "sphinx-markdown-tables",
+    "sphinx-rtd-theme": "sphinx-rtd-theme==0.4.3",
+    "sphinx": "sphinx==3.2.1",
+    "sphinxext-opengraph": "sphinxext-opengraph==0.4.1",
+    "sphinx-intl": "sphinx-intl",
+    "starlette": "starlette",
+    "tensorflow-cpu": "tensorflow-cpu>=2.3",
+    "tensorflow": "tensorflow>=2.3",
+    "timeout-decorator": "timeout-decorator",
+    "timm": "timm",
+    "tokenizers": "tokenizers>=0.10.1,<0.11",
+    "torch": "torch>=1.0",
+    "torchaudio": "torchaudio",
+    "tqdm": "tqdm>=4.27",
+    "unidic": "unidic>=1.0.2",
+    "unidic_lite": "unidic_lite>=1.0.7",
+    "uvicorn": "uvicorn",
+}

fastNLP/transformers/torch/file_utils.py

@@ -0,0 +1,934 @@
+import copy
+import fnmatch
+import importlib.util
+import io
+import json
+import os
+import re
+import shutil
+import sys
+import tarfile
+import tempfile
+import operator
+from collections import OrderedDict, UserDict
+from contextlib import contextmanager
+from dataclasses import fields
+from enum import Enum
+from functools import partial
+from hashlib import sha256
+from pathlib import Path
+from typing import Any, BinaryIO, Dict, Optional, Tuple, Union
+from urllib.parse import urlparse
+from uuid import uuid4
+from zipfile import ZipFile, is_zipfile
+
+import numpy as np
+# from tqdm.auto import tqdm
+
+import requests
+
+from . import __version__
+from .utils.versions import importlib_metadata
+from fastNLP.envs.imports import _NEED_IMPORT_TORCH, _TORCH_GREATER_EQUAL_1_8
+from fastNLP.envs.utils import _compare_version
+from fastNLP.core.log import logger
+
+if _NEED_IMPORT_TORCH:
+    import torch
+    _torch_version = importlib_metadata.version("torch")
+
+hf_cache_home = os.path.expanduser(
+    os.getenv("HF_HOME", os.path.join(os.getenv("XDG_CACHE_HOME", "~/.cache"), "huggingface"))
+)
+default_cache_path = os.path.join(hf_cache_home, "transformers")
+
+PYTORCH_PRETRAINED_BERT_CACHE = os.getenv("PYTORCH_PRETRAINED_BERT_CACHE", default_cache_path)
+PYTORCH_TRANSFORMERS_CACHE = os.getenv("PYTORCH_TRANSFORMERS_CACHE", PYTORCH_PRETRAINED_BERT_CACHE)
+TRANSFORMERS_CACHE = os.getenv("TRANSFORMERS_CACHE", PYTORCH_TRANSFORMERS_CACHE)
+SESSION_ID = uuid4().hex
+
+ENV_VARS_TRUE_VALUES = {"1", "ON", "YES", "TRUE"}
+
+DISABLE_TELEMETRY = os.getenv("DISABLE_TELEMETRY", False) in ENV_VARS_TRUE_VALUES
+
+WEIGHTS_NAME = "pytorch_model.bin"
+DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
+
+_staging_mode = os.environ.get("HUGGINGFACE_CO_STAGING", "NO").upper() in ENV_VARS_TRUE_VALUES
+_default_endpoint = "https://moon-staging.huggingface.co" if _staging_mode else "https://huggingface.co"
+
+HUGGINGFACE_CO_RESOLVE_ENDPOINT = os.environ.get("HUGGINGFACE_CO_RESOLVE_ENDPOINT", _default_endpoint)
+HUGGINGFACE_CO_PREFIX = HUGGINGFACE_CO_RESOLVE_ENDPOINT + "/{model_id}/resolve/{revision}/{filename}"
+
+CONFIG_NAME = "config.json"
+
+_is_offline_mode = True if os.environ.get("TRANSFORMERS_OFFLINE", "0").upper() in ENV_VARS_TRUE_VALUES else False
+
+@contextmanager
+def filelock(path):
+    try:
+        import fcntl
+        open_mode = os.O_RDWR | os.O_CREAT | os.O_TRUNC
+        fd = os.open(path, open_mode)
+        fcntl.flock(fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
+    except:
+        pass
+
+    yield
+
+    try:
+        fcntl.flock(fd, fcntl.LOCK_UN)
+        os.close(fd)
+    except:
+        pass
+
+def is_offline_mode():
+    return _is_offline_mode
+
+def is_training_run_on_sagemaker():
+    return "SAGEMAKER_JOB_NAME" in os.environ
+
+def add_start_docstrings(*docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
+        return fn
+
+    return docstring_decorator
+
+
+def add_start_docstrings_to_model_forward(*docstr):
+    def docstring_decorator(fn):
+        class_name = f":class:`~transformers.{fn.__qualname__.split('.')[0]}`"
+        intro = f"   The {class_name} forward method, overrides the :func:`__call__` special method."
+        note = r"""
+
+    .. note::
+        Although the recipe for forward pass needs to be defined within this function, one should call the
+        :class:`Module` instance afterwards instead of this since the former takes care of running the pre and post
+        processing steps while the latter silently ignores them.
+        """
+        fn.__doc__ = intro + note + "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
+        return fn
+
+    return docstring_decorator
+
+
+def add_end_docstrings(*docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = fn.__doc__ + "".join(docstr)
+        return fn
+
+    return docstring_decorator
+
+PT_RETURN_INTRODUCTION = r"""
+    Returns:
+        :class:`~{full_output_type}` or :obj:`tuple(torch.FloatTensor)`: A :class:`~{full_output_type}` or a tuple of
+        :obj:`torch.FloatTensor` (if ``return_dict=False`` is passed or when ``config.return_dict=False``) comprising
+        various elements depending on the configuration (:class:`~transformers.{config_class}`) and inputs.
+
+"""
+
+def _get_indent(t):
+    """Returns the indentation in the first line of t"""
+    search = re.search(r"^(\s*)\S", t)
+    return "" if search is None else search.groups()[0]
+
+
+def _convert_output_args_doc(output_args_doc):
+    """Convert output_args_doc to display properly."""
+    # Split output_arg_doc in blocks argument/description
+    indent = _get_indent(output_args_doc)
+    blocks = []
+    current_block = ""
+    for line in output_args_doc.split("\n"):
+        # If the indent is the same as the beginning, the line is the name of new arg.
+        if _get_indent(line) == indent:
+            if len(current_block) > 0:
+                blocks.append(current_block[:-1])
+            current_block = f"{line}\n"
+        else:
+            # Otherwise it's part of the description of the current arg.
+            # We need to remove 2 spaces to the indentation.
+            current_block += f"{line[2:]}\n"
+    blocks.append(current_block[:-1])
+
+    # Format each block for proper rendering
+    for i in range(len(blocks)):
+        blocks[i] = re.sub(r"^(\s+)(\S+)(\s+)", r"\1- **\2**\3", blocks[i])
+        blocks[i] = re.sub(r":\s*\n\s*(\S)", r" -- \1", blocks[i])
+
+    return "\n".join(blocks)
+
+def _prepare_output_docstrings(output_type, config_class):
+    """
+    Prepares the return part of the docstring using `output_type`.
+    """
+    docstrings = output_type.__doc__
+
+    # Remove the head of the docstring to keep the list of args only
+    lines = docstrings.split("\n")
+    i = 0
+    while i < len(lines) and re.search(r"^\s*(Args|Parameters):\s*$", lines[i]) is None:
+        i += 1
+    if i < len(lines):
+        docstrings = "\n".join(lines[(i + 1) :])
+        docstrings = _convert_output_args_doc(docstrings)
+
+    # Add the return introduction
+    full_output_type = f"{output_type.__module__}.{output_type.__name__}"
+    intro = PT_RETURN_INTRODUCTION
+    intro = intro.format(full_output_type=full_output_type, config_class=config_class)
+    return intro + docstrings
+
+PT_TOKEN_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> labels = torch.tensor([1] * inputs["input_ids"].size(1)).unsqueeze(0)  # Batch size 1
+
+        >>> outputs = model(**inputs, labels=labels)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_QUESTION_ANSWERING_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        >>> inputs = tokenizer(question, text, return_tensors='pt')
+        >>> start_positions = torch.tensor([1])
+        >>> end_positions = torch.tensor([3])
+
+        >>> outputs = model(**inputs, start_positions=start_positions, end_positions=end_positions)
+        >>> loss = outputs.loss
+        >>> start_scores = outputs.start_logits
+        >>> end_scores = outputs.end_logits
+"""
+
+PT_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        >>> outputs = model(**inputs, labels=labels)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_MASKED_LM_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("The capital of France is {mask}.", return_tensors="pt")
+        >>> labels = tokenizer("The capital of France is Paris.", return_tensors="pt")["input_ids"]
+
+        >>> outputs = model(**inputs, labels=labels)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_BASE_MODEL_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+"""
+
+PT_MULTIPLE_CHOICE_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        >>> choice0 = "It is eaten with a fork and a knife."
+        >>> choice1 = "It is eaten while held in the hand."
+        >>> labels = torch.tensor(0).unsqueeze(0)  # choice0 is correct (according to Wikipedia ;)), batch size 1
+
+        >>> encoding = tokenizer([prompt, prompt], [choice0, choice1], return_tensors='pt', padding=True)
+        >>> outputs = model(**{{k: v.unsqueeze(0) for k,v in encoding.items()}}, labels=labels)  # batch size is 1
+
+        >>> # the linear classifier still needs to be trained
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_CAUSAL_LM_SAMPLE = r"""
+    Example::
+
+        >>> import torch
+        >>> from transformers import {tokenizer_class}, {model_class}
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs, labels=inputs["input_ids"])
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_SAMPLE_DOCSTRINGS = {
+    "SequenceClassification": PT_SEQUENCE_CLASSIFICATION_SAMPLE,
+    "QuestionAnswering": PT_QUESTION_ANSWERING_SAMPLE,
+    "TokenClassification": PT_TOKEN_CLASSIFICATION_SAMPLE,
+    "MultipleChoice": PT_MULTIPLE_CHOICE_SAMPLE,
+    "MaskedLM": PT_MASKED_LM_SAMPLE,
+    "LMHead": PT_CAUSAL_LM_SAMPLE,
+    "BaseModel": PT_BASE_MODEL_SAMPLE,
+}
+
+def add_code_sample_docstrings(
+    *docstr, tokenizer_class=None, checkpoint=None, output_type=None, config_class=None, mask=None, model_cls=None
+):
+    def docstring_decorator(fn):
+        # model_class defaults to function's class if not specified otherwise
+        model_class = fn.__qualname__.split(".")[0] if model_cls is None else model_cls
+
+        sample_docstrings = PT_SAMPLE_DOCSTRINGS
+
+        doc_kwargs = dict(model_class=model_class, tokenizer_class=tokenizer_class, checkpoint=checkpoint)
+
+        if "SequenceClassification" in model_class:
+            code_sample = sample_docstrings["SequenceClassification"]
+        elif "QuestionAnswering" in model_class:
+            code_sample = sample_docstrings["QuestionAnswering"]
+        elif "TokenClassification" in model_class:
+            code_sample = sample_docstrings["TokenClassification"]
+        elif "MultipleChoice" in model_class:
+            code_sample = sample_docstrings["MultipleChoice"]
+        elif "MaskedLM" in model_class or model_class in ["FlaubertWithLMHeadModel", "XLMWithLMHeadModel"]:
+            doc_kwargs["mask"] = "[MASK]" if mask is None else mask
+            code_sample = sample_docstrings["MaskedLM"]
+        elif "LMHead" in model_class or "CausalLM" in model_class:
+            code_sample = sample_docstrings["LMHead"]
+        elif "Model" in model_class or "Encoder" in model_class:
+            code_sample = sample_docstrings["BaseModel"]
+        else:
+            raise ValueError(f"Docstring can't be built for model {model_class}")
+
+        output_doc = _prepare_output_docstrings(output_type, config_class) if output_type is not None else ""
+        built_doc = code_sample.format(**doc_kwargs)
+        fn.__doc__ = (fn.__doc__ or "") + "".join(docstr) + output_doc + built_doc
+        return fn
+
+    return docstring_decorator
+
+def replace_return_docstrings(output_type=None, config_class=None):
+    def docstring_decorator(fn):
+        docstrings = fn.__doc__
+        lines = docstrings.split("\n")
+        i = 0
+        while i < len(lines) and re.search(r"^\s*Returns?:\s*$", lines[i]) is None:
+            i += 1
+        if i < len(lines):
+            lines[i] = _prepare_output_docstrings(output_type, config_class)
+            docstrings = "\n".join(lines)
+        else:
+            raise ValueError(
+                f"The function {fn} should have an empty 'Return:' or 'Returns:' in its docstring as placeholder, current docstring is:\n{docstrings}"
+            )
+        fn.__doc__ = docstrings
+        return fn
+
+    return docstring_decorator
+
+def is_remote_url(url_or_filename):
+    parsed = urlparse(url_or_filename)
+    return parsed.scheme in ("http", "https")
+
+def hf_bucket_url(
+    model_id: str, filename: str, subfolder: Optional[str] = None, revision: Optional[str] = None, mirror=None
+) -> str:
+    """
+    Resolve a model identifier, a file name, and an optional revision id, to a huggingface.co-hosted url, redirecting
+    to Cloudfront (a Content Delivery Network, or CDN) for large files.
+
+    Cloudfront is replicated over the globe so downloads are way faster for the end user (and it also lowers our
+    bandwidth costs).
+
+    Cloudfront aggressively caches files by default (default TTL is 24 hours), however this is not an issue here
+    because we migrated to a git-based versioning system on huggingface.co, so we now store the files on S3/Cloudfront
+    in a content-addressable way (i.e., the file name is its hash). Using content-addressable filenames means cache
+    can't ever be stale.
+
+    In terms of client-side caching from this library, we base our caching on the objects' ETag. An object' ETag is:
+    its sha1 if stored in git, or its sha256 if stored in git-lfs. Files cached locally from transformers before v3.5.0
+    are not shared with those new files, because the cached file's name contains a hash of the url (which changed).
+    """
+    if subfolder is not None:
+        filename = f"{subfolder}/{filename}"
+
+    if mirror:
+        if mirror in ["tuna", "bfsu"]:
+            raise ValueError("The Tuna and BFSU mirrors are no longer available. Try removing the mirror argument.")
+        legacy_format = "/" not in model_id
+        if legacy_format:
+            return f"{mirror}/{model_id}-{filename}"
+        else:
+            return f"{mirror}/{model_id}/{filename}"
+
+    if revision is None:
+        revision = "main"
+    return HUGGINGFACE_CO_PREFIX.format(model_id=model_id, revision=revision, filename=filename)
+
+def url_to_filename(url: str, etag: Optional[str] = None) -> str:
+    """
+    Convert `url` into a hashed filename in a repeatable way. If `etag` is specified, append its hash to the url's,
+    delimited by a period. If the url ends with .h5 (Keras HDF5 weights) adds '.h5' to the name so that TF 2.0 can
+    identify it as a HDF5 file (see
+    https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1380)
+    """
+    url_bytes = url.encode("utf-8")
+    filename = sha256(url_bytes).hexdigest()
+
+    if etag:
+        etag_bytes = etag.encode("utf-8")
+        filename += "." + sha256(etag_bytes).hexdigest()
+
+    if url.endswith(".h5"):
+        filename += ".h5"
+
+    return filename
+
+def cached_path(
+    url_or_filename,
+    cache_dir=None,
+    force_download=False,
+    proxies=None,
+    resume_download=False,
+    user_agent: Union[Dict, str, None] = None,
+    extract_compressed_file=False,
+    force_extract=False,
+    use_auth_token: Union[bool, str, None] = None,
+    local_files_only=False,
+) -> Optional[str]:
+    """
+    Given something that might be a URL (or might be a local path), determine which. If it's a URL, download the file
+    and cache it, and return the path to the cached file. If it's already a local path, make sure the file exists and
+    then return the path
+
+    Args:
+        cache_dir: specify a cache directory to save the file to (overwrite the default cache dir).
+        force_download: if True, re-download the file even if it's already cached in the cache dir.
+        resume_download: if True, resume the download if incompletely received file is found.
+        user_agent: Optional string or dict that will be appended to the user-agent on remote requests.
+        use_auth_token: Optional string or boolean to use as Bearer token for remote files. If True,
+            will get token from ~/.huggingface.
+        extract_compressed_file: if True and the path point to a zip or tar file, extract the compressed
+            file in a folder along the archive.
+        force_extract: if True when extract_compressed_file is True and the archive was already extracted,
+            re-extract the archive and override the folder where it was extracted.
+
+    Return:
+        Local path (string) of file or if networking is off, last version of file cached on disk.
+
+    Raises:
+        In case of non-recoverable file (non-existent or inaccessible url + no cache on disk).
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    if isinstance(url_or_filename, Path):
+        url_or_filename = str(url_or_filename)
+    if isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    if is_offline_mode() and not local_files_only:
+        logger.info("Offline mode: forcing local_files_only=True")
+        local_files_only = True
+
+    if is_remote_url(url_or_filename):
+        # URL, so get it from the cache (downloading if necessary)
+        output_path = get_from_cache(
+            url_or_filename,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            resume_download=resume_download,
+            user_agent=user_agent,
+            use_auth_token=use_auth_token,
+            local_files_only=local_files_only,
+        )
+    elif os.path.exists(url_or_filename):
+        # File, and it exists.
+        output_path = url_or_filename
+    elif urlparse(url_or_filename).scheme == "":
+        # File, but it doesn't exist.
+        raise EnvironmentError(f"file {url_or_filename} not found")
+    else:
+        # Something unknown
+        raise ValueError(f"unable to parse {url_or_filename} as a URL or as a local path")
+
+    if extract_compressed_file:
+        if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
+            return output_path
+
+        # Path where we extract compressed archives
+        # We avoid '.' in dir name and add "-extracted" at the end: "./model.zip" => "./model-zip-extracted/"
+        output_dir, output_file = os.path.split(output_path)
+        output_extract_dir_name = output_file.replace(".", "-") + "-extracted"
+        output_path_extracted = os.path.join(output_dir, output_extract_dir_name)
+
+        if os.path.isdir(output_path_extracted) and os.listdir(output_path_extracted) and not force_extract:
+            return output_path_extracted
+
+        # Prevent parallel extractions
+        lock_path = output_path + ".lock"
+        with filelock(lock_path):
+            shutil.rmtree(output_path_extracted, ignore_errors=True)
+            os.makedirs(output_path_extracted)
+            if is_zipfile(output_path):
+                with ZipFile(output_path, "r") as zip_file:
+                    zip_file.extractall(output_path_extracted)
+                    zip_file.close()
+            elif tarfile.is_tarfile(output_path):
+                tar_file = tarfile.open(output_path)
+                tar_file.extractall(output_path_extracted)
+                tar_file.close()
+            else:
+                raise EnvironmentError(f"Archive format of {output_path} could not be identified")
+
+        return output_path_extracted
+
+    return output_path
+
+def define_sagemaker_information():
+    try:
+        instance_data = requests.get(os.environ["ECS_CONTAINER_METADATA_URI"]).json()
+        dlc_container_used = instance_data["Image"]
+        dlc_tag = instance_data["Image"].split(":")[1]
+    except Exception:
+        dlc_container_used = None
+        dlc_tag = None
+
+    sagemaker_params = json.loads(os.getenv("SM_FRAMEWORK_PARAMS", "{}"))
+    runs_distributed_training = True if "sagemaker_distributed_dataparallel_enabled" in sagemaker_params else False
+    account_id = os.getenv("TRAINING_JOB_ARN").split(":")[4] if "TRAINING_JOB_ARN" in os.environ else None
+
+    sagemaker_object = {
+        "sm_framework": os.getenv("SM_FRAMEWORK_MODULE", None),
+        "sm_region": os.getenv("AWS_REGION", None),
+        "sm_number_gpu": os.getenv("SM_NUM_GPUS", 0),
+        "sm_number_cpu": os.getenv("SM_NUM_CPUS", 0),
+        "sm_distributed_training": runs_distributed_training,
+        "sm_deep_learning_container": dlc_container_used,
+        "sm_deep_learning_container_tag": dlc_tag,
+        "sm_account_id": account_id,
+    }
+    return sagemaker_object
+
+def http_user_agent(user_agent: Union[Dict, str, None] = None) -> str:
+    """
+    Formats a user-agent string with basic info about a request.
+    """
+    ua = f"transformers/{__version__}; python/{sys.version.split()[0]}; session_id/{SESSION_ID}"
+    if _NEED_IMPORT_TORCH:
+        ua += f"; torch/{_torch_version}"
+    if DISABLE_TELEMETRY:
+        return ua + "; telemetry/off"
+    if is_training_run_on_sagemaker():
+        ua += "; " + "; ".join(f"{k}/{v}" for k, v in define_sagemaker_information().items())
+    # CI will set this value to True
+    if os.environ.get("TRANSFORMERS_IS_CI", "").upper() in ENV_VARS_TRUE_VALUES:
+        ua += "; is_ci/true"
+    if isinstance(user_agent, dict):
+        ua += "; " + "; ".join(f"{k}/{v}" for k, v in user_agent.items())
+    elif isinstance(user_agent, str):
+        ua += "; " + user_agent
+    return ua
+
+def http_get(url: str, temp_file: BinaryIO, proxies=None, resume_size=0, headers: Optional[Dict[str, str]] = None):
+    """
+    Download remote file. Do not gobble up errors.
+    """
+    headers = copy.deepcopy(headers)
+    if resume_size > 0:
+        headers["Range"] = f"bytes={resume_size}-"
+    r = requests.get(url, stream=True, proxies=proxies, headers=headers)
+    r.raise_for_status()
+    content_length = r.headers.get("Content-Length")
+    total = resume_size + int(content_length) if content_length is not None else None
+    # progress = tqdm(
+    #     unit="B",
+    #     unit_scale=True,
+    #     unit_divisor=1024,
+    #     total=total,
+    #     initial=resume_size,
+    #     desc="Downloading",
+    #     disable=bool(logging.get_verbosity() == logging.NOTSET),
+    # )
+    for chunk in r.iter_content(chunk_size=1024):
+        if chunk:  # filter out keep-alive new chunks
+            # progress.update(len(chunk))
+            temp_file.write(chunk)
+    # progress.close()
+
+def get_from_cache(
+    url: str,
+    cache_dir=None,
+    force_download=False,
+    proxies=None,
+    etag_timeout=10,
+    resume_download=False,
+    user_agent: Union[Dict, str, None] = None,
+    use_auth_token: Union[bool, str, None] = None,
+    local_files_only=False,
+) -> Optional[str]:
+    """
+    Given a URL, look for the corresponding file in the local cache. If it's not there, download it. Then return the
+    path to the cached file.
+
+    Return:
+        Local path (string) of file or if networking is off, last version of file cached on disk.
+
+    Raises:
+        In case of non-recoverable file (non-existent or inaccessible url + no cache on disk).
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    if isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    os.makedirs(cache_dir, exist_ok=True)
+
+    headers = {"user-agent": http_user_agent(user_agent)}
+    if isinstance(use_auth_token, str):
+        headers["authorization"] = f"Bearer {use_auth_token}"
+    elif use_auth_token:
+        raise RuntimeError("`use_auth_token=True` is not supported in FastNLP now")
+        # token = HfFolder.get_token()
+        # if token is None:
+        #     raise EnvironmentError("You specified use_auth_token=True, but a huggingface token was not found.")
+        # headers["authorization"] = f"Bearer {token}"
+
+    url_to_download = url
+    etag = None
+    if not local_files_only:
+        try:
+            r = requests.head(url, headers=headers, allow_redirects=False, proxies=proxies, timeout=etag_timeout)
+            r.raise_for_status()
+            etag = r.headers.get("X-Linked-Etag") or r.headers.get("ETag")
+            # We favor a custom header indicating the etag of the linked resource, and
+            # we fallback to the regular etag header.
+            # If we don't have any of those, raise an error.
+            if etag is None:
+                raise OSError(
+                    "Distant resource does not have an ETag, we won't be able to reliably ensure reproducibility."
+                )
+            # In case of a redirect,
+            # save an extra redirect on the request.get call,
+            # and ensure we download the exact atomic version even if it changed
+            # between the HEAD and the GET (unlikely, but hey).
+            if 300 <= r.status_code <= 399:
+                url_to_download = r.headers["Location"]
+        except (requests.exceptions.SSLError, requests.exceptions.ProxyError):
+            # Actually raise for those subclasses of ConnectionError
+            raise
+        except (requests.exceptions.ConnectionError, requests.exceptions.Timeout):
+            # Otherwise, our Internet connection is down.
+            # etag is None
+            pass
+
+    filename = url_to_filename(url, etag)
+
+    # get cache path to put the file
+    cache_path = os.path.join(cache_dir, filename)
+
+    # etag is None == we don't have a connection or we passed local_files_only.
+    # try to get the last downloaded one
+    if etag is None:
+        if os.path.exists(cache_path):
+            return cache_path
+        else:
+            matching_files = [
+                file
+                for file in fnmatch.filter(os.listdir(cache_dir), filename.split(".")[0] + ".*")
+                if not file.endswith(".json") and not file.endswith(".lock")
+            ]
+            if len(matching_files) > 0:
+                return os.path.join(cache_dir, matching_files[-1])
+            else:
+                # If files cannot be found and local_files_only=True,
+                # the models might've been found if local_files_only=False
+                # Notify the user about that
+                if local_files_only:
+                    raise FileNotFoundError(
+                        "Cannot find the requested files in the cached path and outgoing traffic has been"
+                        " disabled. To enable model look-ups and downloads online, set 'local_files_only'"
+                        " to False."
+                    )
+                else:
+                    raise ValueError(
+                        "Connection error, and we cannot find the requested files in the cached path."
+                        " Please try again or make sure your Internet connection is on."
+                    )
+
+    # From now on, etag is not None.
+    if os.path.exists(cache_path) and not force_download:
+        return cache_path
+
+    # Prevent parallel downloads of the same file with a lock.
+    lock_path = cache_path + ".lock"
+    with filelock(lock_path):
+
+        # If the download just completed while the lock was activated.
+        if os.path.exists(cache_path) and not force_download:
+            # Even if returning early like here, the lock will be released.
+            return cache_path
+
+        if resume_download:
+            incomplete_path = cache_path + ".incomplete"
+
+            @contextmanager
+            def _resumable_file_manager() -> "io.BufferedWriter":
+                with open(incomplete_path, "ab") as f:
+                    yield f
+
+            temp_file_manager = _resumable_file_manager
+            if os.path.exists(incomplete_path):
+                resume_size = os.stat(incomplete_path).st_size
+            else:
+                resume_size = 0
+        else:
+            temp_file_manager = partial(tempfile.NamedTemporaryFile, mode="wb", dir=cache_dir, delete=False)
+            resume_size = 0
+
+        # Download to temporary file, then copy to cache dir once finished.
+        # Otherwise you get corrupt cache entries if the download gets interrupted.
+        with temp_file_manager() as temp_file:
+            logger.info(f"{url} not found in cache or force_download set to True, downloading to {temp_file.name}")
+
+            http_get(url_to_download, temp_file, proxies=proxies, resume_size=resume_size, headers=headers)
+
+        logger.info(f"storing {url} in cache at {cache_path}")
+        os.replace(temp_file.name, cache_path)
+
+        # NamedTemporaryFile creates a file with hardwired 0600 perms (ignoring umask), so fixing it.
+        umask = os.umask(0o666)
+        os.umask(umask)
+        os.chmod(cache_path, 0o666 & ~umask)
+
+        logger.info(f"creating metadata file for {cache_path}")
+        meta = {"url": url, "etag": etag}
+        meta_path = cache_path + ".json"
+        with open(meta_path, "w") as meta_file:
+            json.dump(meta, meta_file)
+
+    return cache_path
+
+def is_torch_fx_available():
+    return _TORCH_GREATER_EQUAL_1_8 and _compare_version("torch", operator.lt, "1.9.0")
+
+def is_torch_fx_proxy(x):
+    if is_torch_fx_available():
+        import torch.fx
+
+        return isinstance(x, torch.fx.Proxy)
+    return False
+
+def is_sentencepiece_available():
+    return importlib.util.find_spec("sentencepiece") is not None
+
+def is_tokenizers_available():
+    return importlib.util.find_spec("tokenizers") is not None
+
+def is_tensor(x):
+    """
+    Tests if ``x`` is a :obj:`torch.Tensor`, :obj:`tf.Tensor`, obj:`jaxlib.xla_extension.DeviceArray` or
+    :obj:`np.ndarray`.
+    """
+    if is_torch_fx_proxy(x):
+        return True
+
+    if isinstance(x, torch.Tensor):
+        return True
+
+    return isinstance(x, np.ndarray)
+
+def to_py_obj(obj):
+    """
+    Convert a TensorFlow tensor, PyTorch tensor, Numpy array or python list to a python list.
+    """
+    if isinstance(obj, (dict, UserDict)):
+        return {k: to_py_obj(v) for k, v in obj.items()}
+    elif isinstance(obj, (list, tuple)):
+        return [to_py_obj(o) for o in obj]
+    elif _NEED_IMPORT_TORCH and _is_torch(obj):
+        return obj.detach().cpu().tolist()
+    elif isinstance(obj, np.ndarray):
+        return obj.tolist()
+    else:
+        return obj
+
+def _is_numpy(x):
+    return isinstance(x, np.ndarray)
+
+def _is_torch(x):
+    import torch
+
+    return isinstance(x, torch.Tensor)
+
+
+def _is_torch_device(x):
+    import torch
+
+    return isinstance(x, torch.device)
+
+class ModelOutput(OrderedDict):
+    """
+    Base class for all model outputs as dataclass. Has a ``__getitem__`` that allows indexing by integer or slice (like
+    a tuple) or strings (like a dictionary) that will ignore the ``None`` attributes. Otherwise behaves like a regular
+    python dictionary.
+
+    .. warning::
+        You can't unpack a :obj:`ModelOutput` directly. Use the :meth:`~transformers.file_utils.ModelOutput.to_tuple`
+        method to convert it to a tuple before.
+    """
+
+    def __post_init__(self):
+        class_fields = fields(self)
+
+        # Safety and consistency checks
+        assert len(class_fields), f"{self.__class__.__name__} has no fields."
+        assert all(
+            field.default is None for field in class_fields[1:]
+        ), f"{self.__class__.__name__} should not have more than one required field."
+
+        first_field = getattr(self, class_fields[0].name)
+        other_fields_are_none = all(getattr(self, field.name) is None for field in class_fields[1:])
+
+        if other_fields_are_none and not is_tensor(first_field):
+            if isinstance(first_field, dict):
+                iterator = first_field.items()
+                first_field_iterator = True
+            else:
+                try:
+                    iterator = iter(first_field)
+                    first_field_iterator = True
+                except TypeError:
+                    first_field_iterator = False
+
+            # if we provided an iterator as first field and the iterator is a (key, value) iterator
+            # set the associated fields
+            if first_field_iterator:
+                for element in iterator:
+                    if (
+                        not isinstance(element, (list, tuple))
+                        or not len(element) == 2
+                        or not isinstance(element[0], str)
+                    ):
+                        break
+                    setattr(self, element[0], element[1])
+                    if element[1] is not None:
+                        self[element[0]] = element[1]
+            elif first_field is not None:
+                self[class_fields[0].name] = first_field
+        else:
+            for field in class_fields:
+                v = getattr(self, field.name)
+                if v is not None:
+                    self[field.name] = v
+
+    def __delitem__(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.")
+
+    def setdefault(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.")
+
+    def pop(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.")
+
+    def update(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.")
+
+    def __getitem__(self, k):
+        if isinstance(k, str):
+            inner_dict = {k: v for (k, v) in self.items()}
+            return inner_dict[k]
+        else:
+            return self.to_tuple()[k]
+
+    def __setattr__(self, name, value):
+        if name in self.keys() and value is not None:
+            # Don't call self.__setitem__ to avoid recursion errors
+            super().__setitem__(name, value)
+        super().__setattr__(name, value)
+
+    def __setitem__(self, key, value):
+        # Will raise a KeyException if needed
+        super().__setitem__(key, value)
+        # Don't call self.__setattr__ to avoid recursion errors
+        super().__setattr__(key, value)
+
+    def to_tuple(self) -> Tuple[Any]:
+        """
+        Convert self to a tuple containing all the attributes/keys that are not ``None``.
+        """
+        return tuple(self[k] for k in self.keys())
+
+
+class ExplicitEnum(Enum):
+    """
+    Enum with more explicit error message for missing values.
+    """
+
+    @classmethod
+    def _missing_(cls, value):
+        raise ValueError(
+            f"{value} is not a valid {cls.__name__}, please select one of {list(cls._value2member_map_.keys())}"
+        )
+
+
+class PaddingStrategy(ExplicitEnum):
+    """
+    Possible values for the ``padding`` argument in :meth:`PreTrainedTokenizerBase.__call__`. Useful for tab-completion
+    in an IDE.
+    """
+
+    LONGEST = "longest"
+    MAX_LENGTH = "max_length"
+    DO_NOT_PAD = "do_not_pad"
+
+
+class TensorType(ExplicitEnum):
+    """
+    Possible values for the ``return_tensors`` argument in :meth:`PreTrainedTokenizerBase.__call__`. Useful for
+    tab-completion in an IDE.
+    """
+
+    PYTORCH = "pt"
+    NUMPY = "np"

fastNLP/transformers/torch/generation_beam_search.py

@@ -0,0 +1,393 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from abc import ABC, abstractmethod
+from collections import UserDict
+from typing import Optional, Tuple
+
+from .file_utils import add_start_docstrings
+from fastNLP.envs.imports import _NEED_IMPORT_TORCH
+from fastNLP.core.log import logger
+
+if _NEED_IMPORT_TORCH:
+    import torch
+
+
+PROCESS_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_beams, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using any class inheriting from :class:`~transformers.PreTrainedTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        next_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            Current scores of the top :obj:`2 * num_beams` non-finished beam hypotheses.
+        next_tokens (:obj:`torch.LongTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            :obj:`input_ids` of the tokens corresponding to the top :obj:`2 * num_beams` non-finished beam hypotheses.
+        next_indices (:obj:`torch.LongTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            Beam indices indicating to which beam hypothesis the :obj:`next_tokens` correspond.
+        pad_token_id (:obj:`int`, `optional`):
+            The id of the `padding` token.
+        eos_token_id (:obj:`int`, `optional`):
+            The id of the `end-of-sequence` token.
+
+    Return:
+        :obj:`UserDict`: A dictionary composed of the fields as defined above:
+
+            - **next_beam_scores** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Updated
+              scores of all non-finished beams.
+            - **next_beam_tokens** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Next tokens
+              to be added to the non-finished beam_hypotheses.
+            - **next_beam_indices** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Beam indices
+              indicating to which beam the next tokens shall be added.
+
+"""
+
+FINALIZE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_beams, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using any class inheriting from :class:`~transformers.PreTrainedTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        final_beam_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The final scores of all non-finished beams.
+        final_beam_tokens (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The last tokens to be added to the non-finished beam_hypotheses.
+        final_beam_indices (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The beam indices indicating to which beam the :obj:`final_beam_tokens` shall be added.
+        pad_token_id (:obj:`int`, `optional`):
+            The id of the `padding` token.
+        eos_token_id (:obj:`int`, `optional`):
+            The id of the `end-of-sequence` token.
+
+    Return:
+        :obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`: The generated
+        sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or shorter if all
+        batches finished early due to the :obj:`eos_token_id`.
+
+"""
+
+
+class BeamScorer(ABC):
+    """
+    Abstract base class for all beam scorers that are used for :meth:`~transformers.PreTrainedModel.beam_search` and
+    :meth:`~transformers.PreTrainedModel.beam_sample`.
+    """
+
+    @abstractmethod
+    @add_start_docstrings(PROCESS_INPUTS_DOCSTRING)
+    def process(
+        self,
+        input_ids: "torch.LongTensor",
+        next_scores: "torch.FloatTensor",
+        next_tokens: "torch.LongTensor",
+        next_indices: "torch.LongTensor",
+        **kwargs
+    ) -> Tuple["torch.Tensor"]:
+        raise NotImplementedError("This is an abstract method.")
+
+    @abstractmethod
+    @add_start_docstrings(FINALIZE_INPUTS_DOCSTRING)
+    def finalize(
+        self,
+        input_ids: "torch.LongTensor",
+        next_scores: "torch.FloatTensor",
+        next_tokens: "torch.LongTensor",
+        next_indices: "torch.LongTensor",
+        max_length: int,
+        **kwargs
+    ) -> "torch.LongTensor":
+        raise NotImplementedError("This is an abstract method.")
+
+
+class BeamSearchScorer(BeamScorer):
+    r"""
+    :class:`transformers.BeamScorer` implementing standard beam search decoding.
+
+    Adapted in part from `Facebook's XLM beam search code
+    <https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529>`__.
+
+    Reference for the diverse beam search algorithm and implementation `Ashwin Kalyan's DBS implementation
+    <https://github.com/ashwinkalyan/dbs/blob/master/dbs/beam_utils.lua>`__
+
+    Args:
+        batch_size (:obj:`int`):
+            Batch Size of :obj:`input_ids` for which standard beam search decoding is run in parallel.
+        max_length (:obj:`int`):
+            The maximum length of the sequence to be generated.
+        num_beams (:obj:`int`):
+            Number of beams for beam search.
+        device (:obj:`torch.device`):
+            Defines the device type (*e.g.*, :obj:`"cpu"` or :obj:`"cuda"`) on which this instance of
+            :obj:`BeamSearchScorer` will be allocated.
+        length_penalty (:obj:`float`, `optional`, defaults to 1.0):
+            Exponential penalty to the length. 1.0 means no penalty. Set to values < 1.0 in order to encourage the
+            model to generate shorter sequences, to a value > 1.0 in order to encourage the model to produce longer
+            sequences.
+        do_early_stopping (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to stop the beam search when at least ``num_beams`` sentences are finished per batch or not.
+        num_beam_hyps_to_keep (:obj:`int`, `optional`, defaults to 1):
+            The number of beam hypotheses that shall be returned upon calling
+            :meth:`~transformer.BeamSearchScorer.finalize`.
+        num_beam_groups (:obj:`int`):
+            Number of groups to divide :obj:`num_beams` into in order to ensure diversity among different groups of
+            beams. See `this paper <https://arxiv.org/pdf/1610.02424.pdf>`__ for more details.
+    """
+
+    def __init__(
+        self,
+        batch_size: int,
+        num_beams: int,
+        device: "torch.device",
+        length_penalty: Optional[float] = 1.0,
+        do_early_stopping: Optional[bool] = False,
+        num_beam_hyps_to_keep: Optional[int] = 1,
+        num_beam_groups: Optional[int] = 1,
+        **kwargs,
+    ):
+        self.num_beams = num_beams
+        self.device = device
+        self.length_penalty = length_penalty
+        self.do_early_stopping = do_early_stopping
+        self.num_beam_hyps_to_keep = num_beam_hyps_to_keep
+        self.num_beam_groups = num_beam_groups
+        self.group_size = self.num_beams // self.num_beam_groups
+
+        self._is_init = False
+        self._beam_hyps = [
+            BeamHypotheses(
+                num_beams=self.num_beams,
+                length_penalty=self.length_penalty,
+                early_stopping=self.do_early_stopping,
+            )
+            for _ in range(batch_size)
+        ]
+        self._done = torch.tensor([False for _ in range(batch_size)], dtype=torch.bool, device=self.device)
+
+        if not isinstance(num_beams, int) or num_beams <= 1:
+            raise ValueError(
+                f"`num_beams` has to be an integer strictly greater than 1, but is {num_beams}. For `num_beams` == 1, one should make use of `greedy_search` instead."
+            )
+
+        if not isinstance(num_beam_groups, int) or (num_beam_groups > num_beams) or (num_beams % num_beam_groups != 0):
+            raise ValueError(
+                f"`num_beam_groups` has to be an integer smaller or equal than `num_beams` and `num_beams` "
+                f"has to be divisible by `num_beam_groups`, but is {num_beam_groups} with `num_beams` being {num_beams}."
+            )
+
+        if "max_length" in kwargs:
+            logger.warn(
+                "Passing `max_length` to BeamSearchScorer is deprecated and has no effect."
+                "`max_length` should be passed directly to `beam_search(...)`, `beam_sample(...)`"
+                ",or `group_beam_search(...)`."
+            )
+
+    @property
+    def is_done(self) -> bool:
+        return self._done.all()
+
+    def process(
+        self,
+        input_ids: "torch.LongTensor",
+        next_scores: "torch.FloatTensor",
+        next_tokens: "torch.LongTensor",
+        next_indices: "torch.LongTensor",
+        pad_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+    ) -> Tuple["torch.Tensor"]:
+        cur_len = input_ids.shape[-1]
+        batch_size = len(self._beam_hyps)
+        assert batch_size == (input_ids.shape[0] // self.group_size)
+
+        device = input_ids.device
+        next_beam_scores = torch.zeros((batch_size, self.group_size), dtype=next_scores.dtype, device=device)
+        next_beam_tokens = torch.zeros((batch_size, self.group_size), dtype=next_tokens.dtype, device=device)
+        next_beam_indices = torch.zeros((batch_size, self.group_size), dtype=next_indices.dtype, device=device)
+
+        for batch_idx, beam_hyp in enumerate(self._beam_hyps):
+            if self._done[batch_idx]:
+                assert (
+                    len(beam_hyp) >= self.num_beams
+                ), f"Batch can only be done if at least {self.num_beams} beams have been generated"
+                assert (
+                    eos_token_id is not None and pad_token_id is not None
+                ), "generated beams >= num_beams -> eos_token_id and pad_token have to be defined"
+                # pad the batch
+                next_beam_scores[batch_idx, :] = 0
+                next_beam_tokens[batch_idx, :] = pad_token_id
+                next_beam_indices[batch_idx, :] = 0
+                continue
+
+            # next tokens for this sentence
+            beam_idx = 0
+            for beam_token_rank, (next_token, next_score, next_index) in enumerate(
+                zip(next_tokens[batch_idx], next_scores[batch_idx], next_indices[batch_idx])
+            ):
+                batch_beam_idx = batch_idx * self.group_size + next_index
+                # add to generated hypotheses if end of sentence
+                if (eos_token_id is not None) and (next_token.item() == eos_token_id):
+                    # if beam_token does not belong to top num_beams tokens, it should not be added
+                    is_beam_token_worse_than_top_num_beams = beam_token_rank >= self.group_size
+                    if is_beam_token_worse_than_top_num_beams:
+                        continue
+                    beam_hyp.add(
+                        input_ids[batch_beam_idx].clone(),
+                        next_score.item(),
+                    )
+                else:
+                    # add next predicted token since it is not eos_token
+                    next_beam_scores[batch_idx, beam_idx] = next_score
+                    next_beam_tokens[batch_idx, beam_idx] = next_token
+                    next_beam_indices[batch_idx, beam_idx] = batch_beam_idx
+                    beam_idx += 1
+
+                # once the beam for next step is full, don't add more tokens to it.
+                if beam_idx == self.group_size:
+                    break
+
+            if beam_idx < self.group_size:
+                raise ValueError(
+                    f"At most {self.group_size} tokens in {next_tokens[batch_idx]} can be equal to `eos_token_id: {eos_token_id}`. Make sure {next_tokens[batch_idx]} are corrected."
+                )
+
+            # Check if we are done so that we can save a pad step if all(done)
+            self._done[batch_idx] = self._done[batch_idx] or beam_hyp.is_done(
+                next_scores[batch_idx].max().item(), cur_len
+            )
+
+        return UserDict(
+            {
+                "next_beam_scores": next_beam_scores.view(-1),
+                "next_beam_tokens": next_beam_tokens.view(-1),
+                "next_beam_indices": next_beam_indices.view(-1),
+            }
+        )
+
+    def finalize(
+        self,
+        input_ids: "torch.LongTensor",
+        final_beam_scores: "torch.FloatTensor",
+        final_beam_tokens: "torch.LongTensor",
+        final_beam_indices: "torch.LongTensor",
+        max_length: int,
+        pad_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+    ) -> Tuple["torch.LongTensor"]:
+        batch_size = len(self._beam_hyps)
+
+        # finalize all open beam hypotheses and add to generated hypotheses
+        for batch_idx, beam_hyp in enumerate(self._beam_hyps):
+            if self._done[batch_idx]:
+                continue
+
+            # all open beam hypotheses are added to the beam hypothesis
+            # beam hypothesis class automatically keeps the best beams
+            for beam_id in range(self.num_beams):
+                batch_beam_idx = batch_idx * self.num_beams + beam_id
+                final_score = final_beam_scores[batch_beam_idx].item()
+                final_tokens = input_ids[batch_beam_idx]
+                beam_hyp.add(final_tokens, final_score)
+
+        # select the best hypotheses
+        sent_lengths = input_ids.new(batch_size * self.num_beam_hyps_to_keep)
+        best = []
+        best_scores = torch.zeros(batch_size * self.num_beam_hyps_to_keep, device=self.device, dtype=torch.float32)
+
+        # retrieve best hypotheses
+        for i, beam_hyp in enumerate(self._beam_hyps):
+            sorted_hyps = sorted(beam_hyp.beams, key=lambda x: x[0])
+            for j in range(self.num_beam_hyps_to_keep):
+                best_hyp_tuple = sorted_hyps.pop()
+                best_score = best_hyp_tuple[0]
+                best_hyp = best_hyp_tuple[1]
+                sent_lengths[self.num_beam_hyps_to_keep * i + j] = len(best_hyp)
+
+                # append to lists
+                best.append(best_hyp)
+                best_scores[i * self.num_beam_hyps_to_keep + j] = best_score
+
+        # prepare for adding eos
+        sent_max_len = min(sent_lengths.max().item() + 1, max_length)
+        decoded: "torch.LongTensor" = input_ids.new(batch_size * self.num_beam_hyps_to_keep, sent_max_len)
+        # shorter batches are padded if needed
+        if sent_lengths.min().item() != sent_lengths.max().item():
+            assert pad_token_id is not None, "`pad_token_id` has to be defined"
+            decoded.fill_(pad_token_id)
+
+        # fill with hypotheses and eos_token_id if the latter fits in
+        for i, hypo in enumerate(best):
+            decoded[i, : sent_lengths[i]] = hypo
+            if sent_lengths[i] < max_length:
+                decoded[i, sent_lengths[i]] = eos_token_id
+        return UserDict(
+            {
+                "sequences": decoded,
+                "sequence_scores": best_scores,
+            }
+        )
+
+
+class BeamHypotheses:
+    def __init__(self, num_beams: int, length_penalty: float, early_stopping: bool):
+        """
+        Initialize n-best list of hypotheses.
+        """
+        self.length_penalty = length_penalty
+        self.early_stopping = early_stopping
+        self.num_beams = num_beams
+        self.beams = []
+        self.worst_score = 1e9
+
+    def __len__(self):
+        """
+        Number of hypotheses in the list.
+        """
+        return len(self.beams)
+
+    def add(self, hyp: "torch.LongTensor", sum_logprobs: float):
+        """
+        Add a new hypothesis to the list.
+        """
+        score = sum_logprobs / (hyp.shape[-1] ** self.length_penalty)
+        if len(self) < self.num_beams or score > self.worst_score:
+            self.beams.append((score, hyp))
+            if len(self) > self.num_beams:
+                sorted_next_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.beams)])
+                del self.beams[sorted_next_scores[0][1]]
+                self.worst_score = sorted_next_scores[1][0]
+            else:
+                self.worst_score = min(score, self.worst_score)
+
+    def is_done(self, best_sum_logprobs: float, cur_len: int) -> bool:
+        """
+        If there are enough hypotheses and that none of the hypotheses being generated can become better than the worst
+        one in the heap, then we are done with this sentence.
+        """
+
+        if len(self) < self.num_beams:
+            return False
+        elif self.early_stopping:
+            return True
+        else:
+            cur_score = best_sum_logprobs / cur_len ** self.length_penalty
+            ret = self.worst_score >= cur_score
+            return ret

fastNLP/transformers/torch/generation_logits_process.py

@@ -0,0 +1,618 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+from abc import ABC
+from typing import Callable, Iterable, List, Optional
+
+import numpy as np
+
+from .file_utils import add_start_docstrings
+from fastNLP.envs.imports import _NEED_IMPORT_TORCH
+from fastNLP.core.log import logger
+
+if _NEED_IMPORT_TORCH:
+    import torch
+
+LOGITS_PROCESSOR_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.BertTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.vocab_size)`):
+            Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam
+            search or log softmax for each vocabulary token when using beam search
+        kwargs:
+            Additional logits processor specific kwargs.
+
+    Return:
+        :obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.vocab_size)`: The processed prediction scores.
+
+"""
+
+
+class LogitsProcessor(ABC):
+    """Abstract base class for all logit processors that can be applied during generation."""
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        """Torch method for processing logits."""
+        raise NotImplementedError(
+            f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
+        )
+
+
+class LogitsWarper(ABC):
+    """Abstract base class for all logit warpers that can be applied during generation with multinomial sampling."""
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        """Torch method for warping logits."""
+        raise NotImplementedError(
+            f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
+        )
+
+
+class LogitsProcessorList(list):
+    """
+    This class can be used to create a list of :class:`~transformers.LogitsProcessor` or
+    :class:`~transformers.LogitsWarper` to subsequently process a :obj:`scores` input tensor. This class inherits from
+    list and adds a specific `__call__` method to apply each :class:`~transformers.LogitsProcessor` or
+    :class:`~transformers.LogitsWarper` to the inputs.
+    """
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor", **kwargs) -> "torch.FloatTensor":
+        for processor in self:
+            function_args = inspect.signature(processor.__call__).parameters
+            if len(function_args) > 2:
+                assert all(
+                    arg in kwargs for arg in list(function_args.keys())[2:]
+                ), f"Make sure that all the required parameters: {list(function_args.keys())} for {processor.__class__} are passed to the logits processor."
+                scores = processor(input_ids, scores, **kwargs)
+            else:
+                scores = processor(input_ids, scores)
+        return scores
+
+
+class MinLengthLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` enforcing a min-length by setting EOS probability to 0.
+
+    Args:
+        min_length (:obj:`int`):
+            The minimum length below which the score of :obj:`eos_token_id` is set to :obj:`-float("Inf")`.
+        eos_token_id (:obj:`int`):
+            The id of the `end-of-sequence` token.
+    """
+
+    def __init__(self, min_length: int, eos_token_id: int):
+        if not isinstance(min_length, int) or min_length < 0:
+            raise ValueError(f"`min_length` has to be a positive integer, but is {min_length}")
+
+        if not isinstance(eos_token_id, int) or eos_token_id < 0:
+            raise ValueError(f"`eos_token_id` has to be a positive integer, but is {eos_token_id}")
+
+        self.min_length = min_length
+        self.eos_token_id = eos_token_id
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        cur_len = input_ids.shape[-1]
+        if cur_len < self.min_length:
+            scores[:, self.eos_token_id] = -float("inf")
+        return scores
+
+
+class TemperatureLogitsWarper(LogitsWarper):
+    r"""
+    :class:`transformers.LogitsWarper` for temperature (exponential scaling output probability distribution).
+
+    Args:
+        temperature (:obj:`float`):
+            The value used to module the logits distribution.
+    """
+
+    def __init__(self, temperature: float):
+        if not isinstance(temperature, float) or not (temperature > 0):
+            raise ValueError(f"`temperature` has to be a strictly positive float, but is {temperature}")
+
+        self.temperature = temperature
+
+    def __call__(self, input_ids: "torch.Tensor", scores: "torch.Tensor") -> "torch.FloatTensor":
+        scores = scores / self.temperature
+        return scores
+
+
+class RepetitionPenaltyLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` enforcing an exponential penalty on repeated sequences.
+
+    Args:
+        repetition_penalty (:obj:`float`):
+            The parameter for repetition penalty. 1.0 means no penalty. See `this paper
+            <https://arxiv.org/pdf/1909.05858.pdf>`__ for more details.
+    """
+
+    def __init__(self, penalty: float):
+        if not isinstance(penalty, float) or not (penalty > 0):
+            raise ValueError(f"`penalty` has to be a strictly positive float, but is {penalty}")
+
+        self.penalty = penalty
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        score = torch.gather(scores, 1, input_ids)
+
+        # if score < 0 then repetition penalty has to be multiplied to reduce the previous token probability
+        score = torch.where(score < 0, score * self.penalty, score / self.penalty)
+
+        scores.scatter_(1, input_ids, score)
+        return scores
+
+
+class TopPLogitsWarper(LogitsWarper):
+    """
+    :class:`transformers.LogitsWarper` that performs top-p, i.e. restricting to top tokens summing to prob_cut_off <=
+    prob_cut_off.
+
+    Args:
+        top_p (:obj:`float`):
+            If set to < 1, only the most probable tokens with probabilities that add up to :obj:`top_p` or higher are
+            kept for generation.
+        filter_value (:obj:`float`, `optional`, defaults to :obj:`-float("Inf")`):
+            All filtered values will be set to this float value.
+        min_tokens_to_keep (:obj:`int`, `optional`, defaults to 1):
+            Minimum number of tokens that cannot be filtered.
+    """
+
+    def __init__(self, top_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        top_p = float(top_p)
+        if top_p < 0 or top_p > 1.0:
+            raise ValueError(f"`top_p` has to be a float > 0 and < 1, but is {top_p}")
+
+        self.top_p = top_p
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        sorted_logits, sorted_indices = torch.sort(scores, descending=True)
+        cumulative_probs = sorted_logits.softmax(dim=-1).cumsum(dim=-1)
+
+        # Remove tokens with cumulative top_p above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > self.top_p
+        if self.min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove[..., : self.min_tokens_to_keep - 1] = 0
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+        scores = scores.masked_fill(indices_to_remove, self.filter_value)
+        return scores
+
+
+class TopKLogitsWarper(LogitsWarper):
+    r"""
+    :class:`transformers.LogitsWarper` that performs top-k, i.e. restricting to the k highest probability elements.
+
+    Args:
+        top_k (:obj:`int`):
+            The number of highest probability vocabulary tokens to keep for top-k-filtering.
+        filter_value (:obj:`float`, `optional`, defaults to :obj:`-float("Inf")`):
+            All filtered values will be set to this float value.
+        min_tokens_to_keep (:obj:`int`, `optional`, defaults to 1):
+            Minimum number of tokens that cannot be filtered.
+    """
+
+    def __init__(self, top_k: int, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        if not isinstance(top_k, int) or top_k <= 0:
+            raise ValueError(f"`top_k` has to be a strictly positive integer, but is {top_k}")
+
+        self.top_k = top_k
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        top_k = min(max(self.top_k, self.min_tokens_to_keep), scores.size(-1))  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = scores < torch.topk(scores, top_k)[0][..., -1, None]
+        scores = scores.masked_fill(indices_to_remove, self.filter_value)
+        return scores
+
+
+def _get_ngrams(ngram_size: int, prev_input_ids: "torch.Tensor", num_hypos: int):
+    generated_ngrams = [{} for _ in range(num_hypos)]
+    for idx in range(num_hypos):
+        gen_tokens = prev_input_ids[idx].tolist()
+        generated_ngram = generated_ngrams[idx]
+        for ngram in zip(*[gen_tokens[i:] for i in range(ngram_size)]):
+            prev_ngram_tuple = tuple(ngram[:-1])
+            generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
+    return generated_ngrams
+
+
+def _get_generated_ngrams(banned_ngrams, prev_input_ids, ngram_size, cur_len):
+    # Before decoding the next token, prevent decoding of ngrams that have already appeared
+    start_idx = cur_len + 1 - ngram_size
+    ngram_idx = tuple(prev_input_ids[start_idx:cur_len].tolist())
+    return banned_ngrams.get(ngram_idx, [])
+
+
+def _calc_banned_ngram_tokens(
+    ngram_size: int, prev_input_ids: "torch.Tensor", num_hypos: int, cur_len: int
+) -> List[Iterable[int]]:
+    """Copied from fairseq for no_repeat_ngram in beam_search"""
+    if cur_len + 1 < ngram_size:
+        # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
+        return [[] for _ in range(num_hypos)]
+
+    generated_ngrams = _get_ngrams(ngram_size, prev_input_ids, num_hypos)
+
+    banned_tokens = [
+        _get_generated_ngrams(generated_ngrams[hypo_idx], prev_input_ids[hypo_idx], ngram_size, cur_len)
+        for hypo_idx in range(num_hypos)
+    ]
+    return banned_tokens
+
+
+class NoRepeatNGramLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces no repetition of n-grams. See `Fairseq
+    <https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345>`__.
+
+    Args:
+        ngram_size (:obj:`int`):
+            All ngrams of size :obj:`ngram_size` can only occur once.
+    """
+
+    def __init__(self, ngram_size: int):
+        if not isinstance(ngram_size, int) or ngram_size <= 0:
+            raise ValueError(f"`ngram_size` has to be a strictly positive integer, but is {ngram_size}")
+        self.ngram_size = ngram_size
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        num_batch_hypotheses = scores.shape[0]
+        cur_len = input_ids.shape[-1]
+        banned_batch_tokens = _calc_banned_ngram_tokens(self.ngram_size, input_ids, num_batch_hypotheses, cur_len)
+
+        for i, banned_tokens in enumerate(banned_batch_tokens):
+            scores[i, banned_tokens] = -float("inf")
+
+        return scores
+
+
+class EncoderNoRepeatNGramLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces no repetition of encoder input ids n-grams for the decoder ids.
+    See `ParlAI <https://github.com/facebookresearch/ParlAI/blob/master/parlai/core/torch_generator_agent.py#L1350>`__.
+
+    Args:
+        encoder_ngram_size (:obj:`int`):
+            All ngrams of size :obj:`ngram_size` can only occur within the encoder input ids.
+        encoder_input_ids (:obj:`int`):
+            The encoder_input_ids that should not be repeated within the decoder ids.
+    """
+
+    def __init__(self, encoder_ngram_size: int, encoder_input_ids: "torch.LongTensor"):
+        if not isinstance(encoder_ngram_size, int) or encoder_ngram_size <= 0:
+            raise ValueError(
+                f"`encoder_ngram_size` has to be a strictly positive integer, but is {encoder_ngram_size}"
+            )
+        self.ngram_size = encoder_ngram_size
+        if len(encoder_input_ids.shape) == 1:
+            encoder_input_ids = encoder_input_ids.unsqueeze(0)
+        self.batch_size = encoder_input_ids.shape[0]
+        self.generated_ngrams = _get_ngrams(encoder_ngram_size, encoder_input_ids, self.batch_size)
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        # B x num_beams
+        num_hypos = scores.shape[0]
+        num_beams = num_hypos // self.batch_size
+        cur_len = input_ids.shape[-1]
+        banned_batch_tokens = [
+            _get_generated_ngrams(
+                self.generated_ngrams[hypo_idx // num_beams], input_ids[hypo_idx], self.ngram_size, cur_len
+            )
+            for hypo_idx in range(num_hypos)
+        ]
+
+        for i, banned_tokens in enumerate(banned_batch_tokens):
+            scores[i, banned_tokens] = -float("inf")
+
+        return scores
+
+
+class NoBadWordsLogitsProcessor(LogitsProcessor):
+    """
+    :class:`transformers.LogitsProcessor` that enforces that specified sequences will never be sampled.
+
+    Args:
+        bad_words_ids (:obj:`List[List[int]]`):
+            List of list of token ids that are not allowed to be generated. In order to get the tokens of the words
+            that should not appear in the generated text, use :obj:`tokenizer(bad_word,
+            add_prefix_space=True).input_ids`.
+        eos_token_id (:obj:`int`):
+            The id of the `end-of-sequence` token.
+    """
+
+    def __init__(self, bad_words_ids: List[List[int]], eos_token_id: int):
+
+        if not isinstance(bad_words_ids, List) or len(bad_words_ids) == 0:
+            raise ValueError(f"`bad_words_ids` has to be a non-emtpy list, but is {bad_words_ids}.")
+        if any(not isinstance(bad_word_ids, list) for bad_word_ids in bad_words_ids):
+            raise ValueError(f"`bad_words_ids` has to be a list of lists, but is {bad_words_ids}.")
+        if any(
+            any((not isinstance(token_id, (int, np.integer)) or token_id < 0) for token_id in bad_word_ids)
+            for bad_word_ids in bad_words_ids
+        ):
+            raise ValueError(
+                f"Each list in `bad_words_ids` has to be a list of positive integers, but is {bad_words_ids}."
+            )
+
+        bad_words_ids = list(filter(lambda bad_token_seq: bad_token_seq != [eos_token_id], bad_words_ids))
+        self.bad_words_id_length_1 = []
+        self.bad_words_id_length_greater_than_1 = []
+        for word in bad_words_ids:
+            if len(word) == 1:
+                self.bad_words_id_length_1.append(word[0])
+            else:
+                self.bad_words_id_length_greater_than_1.append(word)
+
+        self.static_bad_words_mask: Optional[torch.LongTensor] = None
+
+        for banned_token_seq in self.bad_words_id_length_greater_than_1:
+            assert len(banned_token_seq) > 0, f"Banned words token sequences {bad_words_ids} cannot have an empty list"
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        if self.static_bad_words_mask is None and len(self.bad_words_id_length_1) > 0:
+            self.static_bad_words_mask = self._calc_static_bad_word_mask(scores)
+
+        dynamic_banned_tokens = self._calc_banned_bad_words_ids(input_ids.tolist())
+        scores = self._set_scores_to_inf_for_banned_tokens(scores, dynamic_banned_tokens)
+
+        return scores
+
+    def _calc_static_bad_word_mask(self, scores: "torch.FloatTensor") -> "torch.BoolTensor":
+        static_bad_words_mask = torch.zeros(scores.shape[1])
+        static_bad_words_mask[self.bad_words_id_length_1] = 1
+        return static_bad_words_mask.unsqueeze(0).to(scores.device).bool()
+
+    def _tokens_match(self, prev_tokens: List[int], tokens: List[int]) -> bool:
+        if len(tokens) == 0:
+            # if bad word tokens is just one token always ban it
+            return True
+        elif len(tokens) > len(prev_tokens):
+            # if bad word tokens are longer then prev input_ids they can't be equal
+            return False
+        else:
+            return prev_tokens[-len(tokens) :] == tokens
+
+    def _calc_banned_bad_words_ids(self, prev_input_ids: List[List[int]]) -> Iterable[int]:
+        banned_tokens = []
+        for prev_input_ids_slice in prev_input_ids:
+            banned_tokens_slice = []
+            for banned_token_seq in self.bad_words_id_length_greater_than_1:
+                if self._tokens_match(prev_input_ids_slice, banned_token_seq[:-1]):
+                    banned_tokens_slice.append(banned_token_seq[-1])
+
+            banned_tokens.append(banned_tokens_slice)
+
+        return banned_tokens
+
+    def _set_scores_to_inf_for_banned_tokens(
+        self, scores: "torch.Tensor", banned_tokens: List[List[int]]
+    ) -> "torch.Tensor":
+        """
+        Modifies the scores in place by setting the banned token positions to `-inf`. Banned token is expected to be a
+        list of list of banned tokens to ban in the format [[batch index, vocabulary position],...
+
+        Args:
+            scores: logits distribution of shape (batch size, vocabulary size)
+            banned_tokens: list of list of tokens to ban of length (batch_size)
+        """
+        banned_mask_list = []
+        for idx, batch_banned_tokens in enumerate(banned_tokens):
+            for token in batch_banned_tokens:
+                # Eliminates invalid bad word IDs that are over the vocabulary size.
+                if token <= scores.shape[1]:
+                    banned_mask_list.append([idx, token])
+                else:
+                    logger.error(
+                        f"An invalid bad word ID is defined: {token}. This ID is not contained in the"
+                        f"vocabulary, and is therefore ignored."
+                    )
+        if not banned_mask_list and self.static_bad_words_mask is None:
+            return scores
+
+        else:
+            if banned_mask_list:
+                banned_mask = torch.LongTensor(banned_mask_list)
+                indices = torch.ones(len(banned_mask))
+                # A sparse tensor is generated from a list of coordinates: [[0, 1], [0, 2], [2, 0]]. A conversion to dense tensor generates:
+                # [ 0  1  1 ]
+                # [ 0  0  0 ]
+                # [ 1  0  0 ]
+
+                banned_mask = (
+                    torch.sparse.LongTensor(banned_mask.t(), indices, scores.size())
+                    .to(scores.device)
+                    .to_dense()
+                    .bool()
+                )
+
+                if self.static_bad_words_mask is not None:
+                    banned_mask = torch.bitwise_or(banned_mask, self.static_bad_words_mask)
+            else:
+                banned_mask = self.static_bad_words_mask
+
+            scores = scores.masked_fill(banned_mask, -float("inf"))
+            return scores
+
+
+class PrefixConstrainedLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces constrained generation and is useful for prefix-conditioned
+    constrained generation. See `Autoregressive Entity Retrieval <https://arxiv.org/abs/2010.00904>`__ for more
+    information.
+
+    Args:
+        prefix_allowed_tokens_fn: (:obj:`Callable[[int, torch.Tensor], List[int]]`):
+            This function constraints the beam search to allowed tokens only at each step. This function takes 2
+            arguments :obj:`inputs_ids` and the batch ID :obj:`batch_id`. It has to return a list with the allowed
+            tokens for the next generation step conditioned on the previously generated tokens :obj:`inputs_ids` and
+            the batch ID :obj:`batch_id`.
+    """
+
+    def __init__(self, prefix_allowed_tokens_fn: Callable[[int, "torch.Tensor"], List[int]], num_beams: int):
+        self._prefix_allowed_tokens_fn = prefix_allowed_tokens_fn
+        self._num_beams = num_beams
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        mask = torch.full_like(scores, -math.inf)
+        for batch_id, beam_sent in enumerate(input_ids.view(-1, self._num_beams, input_ids.shape[-1])):
+            for beam_id, sent in enumerate(beam_sent):
+                mask[batch_id * self._num_beams + beam_id, self._prefix_allowed_tokens_fn(batch_id, sent)] = 0
+
+        return scores + mask
+
+
+class HammingDiversityLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces diverse beam search. Note that this logits processor is only
+    effective for :meth:`transformers.PreTrainedModel.group_beam_search`. See `Diverse Beam Search: Decoding Diverse
+    Solutions from Neural Sequence Models <https://arxiv.org/pdf/1610.02424.pdf>`__ for more details.
+
+    Args:
+        diversity_penalty (:obj:`float`):
+            This value is subtracted from a beam's score if it generates a token same as any beam from other group at a
+            particular time. Note that :obj:`diversity_penalty` is only effective if ``group beam search`` is enabled.
+        num_beams (:obj:`int`):
+            Number of beams used for group beam search. See `this paper <https://arxiv.org/pdf/1610.02424.pdf>`__ for
+            more details.
+        num_beam_groups (:obj:`int`):
+            Number of groups to divide :obj:`num_beams` into in order to ensure diversity among different groups of
+            beams. See `this paper <https://arxiv.org/pdf/1610.02424.pdf>`__ for more details.
+    """
+
+    def __init__(self, diversity_penalty: float, num_beams: int, num_beam_groups: int):
+        if not isinstance(diversity_penalty, float) or (not diversity_penalty > 0.0):
+            raise ValueError("`diversity_penalty` should be a float strictly larger than 0.")
+        self._diversity_penalty = diversity_penalty
+        if not isinstance(num_beams, int) or num_beams < 2:
+            raise ValueError("`num_beams` should be an integer strictly larger than 1.")
+        self._num_beams = num_beams
+        if not isinstance(num_beam_groups, int) or num_beam_groups < 2:
+            raise ValueError("`num_beam_groups` should be an integer strictly larger than 1.")
+        if num_beam_groups > num_beams:
+            raise ValueError("`beam_groups` has to be smaller or equal to `num_beams`.")
+        self._num_sub_beams = num_beams // num_beam_groups
+
+    def __call__(
+        self,
+        input_ids: "torch.LongTensor",
+        scores: "torch.FloatTensor",
+        current_tokens: "torch.LongTensor",
+        beam_group_idx: int,
+    ) -> "torch.FloatTensor":
+        # hamming diversity: penalise using same token in current group which was used in previous groups at
+        # the same time step
+        batch_size = current_tokens.shape[0] // self._num_beams
+        group_start_idx = beam_group_idx * self._num_sub_beams
+        group_end_idx = min(group_start_idx + self._num_sub_beams, self._num_beams)
+        group_size = group_end_idx - group_start_idx
+        vocab_size = scores.shape[-1]
+
+        if group_start_idx == 0:
+            return scores
+
+        for batch_idx in range(batch_size):
+            # predicted tokens of last time step of previous groups
+            previous_group_tokens = current_tokens[
+                batch_idx * self._num_beams : batch_idx * self._num_beams + group_start_idx
+            ]
+            token_frequency = torch.bincount(previous_group_tokens, minlength=vocab_size).to(scores.device)
+            scores[batch_idx * group_size : (batch_idx + 1) * group_size] -= self._diversity_penalty * token_frequency
+
+        return scores
+
+
+class ForcedBOSTokenLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`~transformers.LogitsProcessor` that enforces the specified token as the first generated token.
+
+    Args:
+        bos_token_id (:obj:`int`):
+            The id of the token to force as the first generated token.
+    """
+
+    def __init__(self, bos_token_id: int):
+        self.bos_token_id = bos_token_id
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        cur_len = input_ids.shape[-1]
+        if cur_len == 1:
+            num_tokens = scores.shape[1]
+            scores[:, [i for i in range(num_tokens) if i != self.bos_token_id]] = -float("inf")
+            scores[:, self.bos_token_id] = 0
+        return scores
+
+
+class ForcedEOSTokenLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`~transformers.LogitsProcessor` that enforces the specified token as the last generated token when
+    :obj:`max_length` is reached.
+
+    Args:
+        max_length (:obj:`int`):
+            The maximum length of the sequence to be generated.
+        eos_token_id (:obj:`int`):
+            The id of the token to force as the last generated token when :obj:`max_length` is reached.
+    """
+
+    def __init__(self, max_length: int, eos_token_id: int):
+        self.max_length = max_length
+        self.eos_token_id = eos_token_id
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        cur_len = input_ids.shape[-1]
+        if cur_len == self.max_length - 1:
+            num_tokens = scores.shape[1]
+            scores[:, [i for i in range(num_tokens) if i != self.eos_token_id]] = -float("inf")
+            scores[:, self.eos_token_id] = 0
+        return scores
+
+
+class InfNanRemoveLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`~transformers.LogitsProcessor` that removes all :obj:`nan` and :obj:`inf` values to avoid the generation
+    method to fail. Note that using the logits processor should only be used if necessary since it can slow down the
+    generation method. :obj:`max_length` is reached.
+    """
+
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor") -> "torch.FloatTensor":
+        # set all nan values to 0.0
+        scores[scores != scores] = 0.0
+
+        # set all inf values to max possible value
+        scores[scores == float("inf")] = torch.finfo(scores.dtype).max
+
+        return scores

fastNLP/transformers/torch/generation_stopping_criteria.py

@@ -0,0 +1,128 @@
+import time
+from abc import ABC
+from copy import deepcopy
+from typing import Optional
+
+from .file_utils import add_start_docstrings
+from fastNLP.envs.imports import _NEED_IMPORT_TORCH
+from fastNLP.core.log import logger
+
+if _NEED_IMPORT_TORCH:
+    import torch
+
+STOPPING_CRITERIA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.BertTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.vocab_size)`):
+            Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax
+            or scores for each vocabulary token after SoftMax.
+        kwargs:
+            Additional stopping criteria specific kwargs.
+
+    Return:
+        :obj:`bool`. :obj:`False` indicates we should continue, :obj:`True` indicates we should stop.
+
+"""
+
+
+class StoppingCriteria(ABC):
+    """Abstract base class for all stopping criteria that can be applied during generation."""
+
+    @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor", **kwargs) -> bool:
+        raise NotImplementedError("StoppingCriteria needs to be subclassed")
+
+
+class MaxLengthCriteria(StoppingCriteria):
+    """
+    This class can be used to stop generation whenever the full generated number of tokens exceeds :obj:`max_length`.
+    Keep in mind for decoder-only type of transformers, this will include the initial prompted tokens.
+
+    Args:
+        max_length (:obj:`int`):
+            The maximum length that the output sequence can have in number of tokens.
+    """
+
+    def __init__(self, max_length: int):
+        self.max_length = max_length
+
+    @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor", **kwargs) -> bool:
+        return input_ids.shape[-1] >= self.max_length
+
+
+class MaxNewTokensCriteria(StoppingCriteria):
+    """
+    This class can be used to stop generation whenever the generated number of tokens exceeds :obj:`max_new_tokens`.
+    Keep in mind for decoder-only type of transformers, this will **not** include the initial prompted tokens. This is
+    very close to :obj:`MaxLengthCriteria` but ignores the number of initial tokens.
+
+    Args:
+        start_length (:obj:`int`):
+            The number of initial tokens.
+        max_new_tokens (:obj:`int`):
+            The maximum number of tokens to generate.
+    """
+
+    def __init__(self, start_length: int, max_new_tokens: int):
+        self.start_length = start_length
+        self.max_new_tokens = max_new_tokens
+        self.max_length = start_length + max_new_tokens
+
+    @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor", **kwargs) -> bool:
+        return input_ids.shape[-1] >= self.max_length
+
+
+class MaxTimeCriteria(StoppingCriteria):
+    """
+    This class can be used to stop generation whenever the full generation exceeds some amount of time. By default, the
+    time will start being counted when you initialize this function. You can override this by passing an
+    :obj:`initial_time`.
+
+    Args:
+        max_time (:obj:`float`):
+            The maximum allowed time in seconds for the generation.
+        initial_time (:obj:`float`, `optional`, defaults to :obj:`time.time()`):
+            The start of the generation allowed time.
+    """
+
+    def __init__(self, max_time: float, initial_timestamp: Optional[float] = None):
+        self.max_time = max_time
+        self.initial_timestamp = time.time() if initial_timestamp is None else initial_timestamp
+
+    @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor", **kwargs) -> bool:
+        return time.time() - self.initial_timestamp > self.max_time
+
+
+class StoppingCriteriaList(list):
+    @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: "torch.LongTensor", scores: "torch.FloatTensor", **kwargs) -> bool:
+        return any(criteria(input_ids, scores) for criteria in self)
+
+    @property
+    def max_length(self) -> Optional[int]:
+        for stopping_criterium in self:
+            if isinstance(stopping_criterium, MaxLengthCriteria):
+                return stopping_criterium.max_length
+            elif isinstance(stopping_criterium, MaxNewTokensCriteria):
+                return stopping_criterium.max_length
+        return None
+
+
+def validate_stopping_criteria(stopping_criteria: StoppingCriteriaList, max_length: int) -> StoppingCriteriaList:
+    stopping_max_length = stopping_criteria.max_length
+    new_stopping_criteria = deepcopy(stopping_criteria)
+    if stopping_max_length is not None and stopping_max_length != max_length:
+        logger.warn("You set different `max_length` for stopping criteria and `max_length` parameter", UserWarning)
+    elif stopping_max_length is None:
+        new_stopping_criteria.append(MaxLengthCriteria(max_length=max_length))
+    return new_stopping_criteria

fastNLP/transformers/torch/modeling_outputs.py

@@ -0,0 +1,816 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+from .file_utils import ModelOutput
+from fastNLP.envs.imports import _NEED_IMPORT_TORCH
+
+if _NEED_IMPORT_TORCH:
+    import torch
+
+
+@dataclass
+class BaseModelOutput(ModelOutput):
+    """
+    Base class for model's outputs, with potential hidden states and attentions.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class BaseModelOutputWithPooling(ModelOutput):
+    """
+    Base class for model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, hidden_size)`):
+            Last layer hidden-state of the first token of the sequence (classification token) after further processing
+            through the layers used for the auxiliary pretraining task. E.g. for BERT-family of models, this returns
+            the classification token after processing through a linear layer and a tanh activation function. The linear
+            layer weights are trained from the next sentence prediction (classification) objective during pretraining.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: "torch.FloatTensor" = None
+    pooler_output: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class BaseModelOutputWithPast(ModelOutput):
+    """
+    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+
+            If :obj:`past_key_values` is used only the last hidden-state of the sequences of shape :obj:`(batch_size,
+            1, hidden_size)` is output.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            ``config.is_encoder_decoder=True`` 2 additional tensors of shape :obj:`(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            ``config.is_encoder_decoder=True`` in the cross-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class BaseModelOutputWithCrossAttentions(ModelOutput):
+    """
+    Base class for model's outputs, with potential hidden states and attentions.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` and ``config.add_cross_attention=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+    """
+
+    last_hidden_state: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    cross_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class BaseModelOutputWithPoolingAndCrossAttentions(ModelOutput):
+    """
+    Base class for model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, hidden_size)`):
+            Last layer hidden-state of the first token of the sequence (classification token) after further processing
+            through the layers used for the auxiliary pretraining task. E.g. for BERT-family of models, this returns
+            the classification token after processing through a linear layer and a tanh activation function. The linear
+            layer weights are trained from the next sentence prediction (classification) objective during pretraining.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` and ``config.add_cross_attention=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            ``config.is_encoder_decoder=True`` 2 additional tensors of shape :obj:`(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            ``config.is_encoder_decoder=True`` in the cross-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+    """
+
+    last_hidden_state: "torch.FloatTensor" = None
+    pooler_output: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    cross_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class BaseModelOutputWithPastAndCrossAttentions(ModelOutput):
+    """
+    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+
+            If :obj:`past_key_values` is used only the last hidden-state of the sequences of shape :obj:`(batch_size,
+            1, hidden_size)` is output.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            ``config.is_encoder_decoder=True`` 2 additional tensors of shape :obj:`(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            ``config.is_encoder_decoder=True`` in the cross-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` and ``config.add_cross_attention=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+    """
+
+    last_hidden_state: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    cross_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class Seq2SeqModelOutput(ModelOutput):
+    """
+    Base class for model encoder's outputs that also contains : pre-computed hidden states that can speed up sequential
+    decoding.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the decoder of the model.
+
+            If :obj:`past_key_values` is used only the last hidden-state of the sequences of shape :obj:`(batch_size,
+            1, hidden_size)` is output.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+            shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    last_hidden_state: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    decoder_hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    decoder_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    cross_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    encoder_last_hidden_state: Optional["torch.FloatTensor"] = None
+    encoder_hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    encoder_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class CausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class CausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class CausalLMOutputWithCrossAttentions(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Cross attentions weights after the attention softmax, used to compute the weighted average in the
+            cross-attention heads.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`torch.FloatTensor` tuples of length :obj:`config.n_layers`, with each tuple containing the
+            cached key, value states of the self-attention and the cross-attention layers if model is used in
+            encoder-decoder setting. Only relevant if ``config.is_decoder = True``.
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    cross_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class SequenceClassifierOutputWithPast(ModelOutput):
+    """
+    Base class for outputs of sentence classification models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class MaskedLMOutput(ModelOutput):
+    """
+    Base class for masked language models outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Masked language modeling (MLM) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class Seq2SeqLMOutput(ModelOutput):
+    """
+    Base class for sequence-to-sequence language models outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+            shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    decoder_hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    decoder_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    cross_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    encoder_last_hidden_state: Optional["torch.FloatTensor"] = None
+    encoder_hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    encoder_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class NextSentencePredictorOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`next_sentence_label` is provided):
+            Next sequence prediction (classification) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class SequenceClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of sentence classification models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class Seq2SeqSequenceClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of sequence-to-sequence sentence classification models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+            shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    decoder_hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    decoder_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    cross_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    encoder_last_hidden_state: Optional["torch.FloatTensor"] = None
+    encoder_hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    encoder_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class MultipleChoiceModelOutput(ModelOutput):
+    """
+    Base class for outputs of multiple choice models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
+            `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
+
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class TokenClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of token classification models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided) :
+            Classification loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    logits: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class QuestionAnsweringModelOutput(ModelOutput):
+    """
+    Base class for outputs of question answering models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Span-start scores (before SoftMax).
+        end_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Span-end scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    start_logits: "torch.FloatTensor" = None
+    end_logits: "torch.FloatTensor" = None
+    hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    attentions: Optional[Tuple["torch.FloatTensor"]] = None
+
+
+@dataclass
+class Seq2SeqQuestionAnsweringModelOutput(ModelOutput):
+    """
+    Base class for outputs of sequence-to-sequence question answering models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Span-start scores (before SoftMax).
+        end_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Span-end scores (before SoftMax).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+            shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional["torch.FloatTensor"] = None
+    start_logits: "torch.FloatTensor" = None
+    end_logits: "torch.FloatTensor" = None
+    past_key_values: Optional[Tuple[Tuple["torch.FloatTensor"]]] = None
+    decoder_hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    decoder_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    cross_attentions: Optional[Tuple["torch.FloatTensor"]] = None
+    encoder_last_hidden_state: Optional["torch.FloatTensor"] = None
+    encoder_hidden_states: Optional[Tuple["torch.FloatTensor"]] = None
+    encoder_attentions: Optional[Tuple["torch.FloatTensor"]] = None

fastNLP/transformers/torch/models/__init__.py

@@ -0,0 +1,5 @@
+from .bart import *
+from .bert import *
+from .cpt import *
+from .gpt2 import *
+from .roberta import *

fastNLP/transformers/torch/models/auto/configuration_auto.py

@@ -0,0 +1,541 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Auto Config class. """
+import importlib
+import re
+from collections import OrderedDict
+from typing import List, Union
+
+from fastNLP.transformers.torch.configuration_utils import PretrainedConfig
+from fastNLP.transformers.torch.file_utils import CONFIG_NAME
+from fastNLP.core.log import logger
+
+
+CONFIG_MAPPING_NAMES = OrderedDict(
+    [
+        # Add configs here
+        ("fnet", "FNetConfig"),
+        ("gptj", "GPTJConfig"),
+        ("layoutlmv2", "LayoutLMv2Config"),
+        ("beit", "BeitConfig"),
+        ("rembert", "RemBertConfig"),
+        ("visual_bert", "VisualBertConfig"),
+        ("canine", "CanineConfig"),
+        ("roformer", "RoFormerConfig"),
+        ("clip", "CLIPConfig"),
+        ("bigbird_pegasus", "BigBirdPegasusConfig"),
+        ("deit", "DeiTConfig"),
+        ("luke", "LukeConfig"),
+        ("detr", "DetrConfig"),
+        ("gpt_neo", "GPTNeoConfig"),
+        ("big_bird", "BigBirdConfig"),
+        ("speech_to_text_2", "Speech2Text2Config"),
+        ("speech_to_text", "Speech2TextConfig"),
+        ("vit", "ViTConfig"),
+        ("wav2vec2", "Wav2Vec2Config"),
+        ("m2m_100", "M2M100Config"),
+        ("convbert", "ConvBertConfig"),
+        ("led", "LEDConfig"),
+        ("blenderbot-small", "BlenderbotSmallConfig"),
+        ("retribert", "RetriBertConfig"),
+        ("ibert", "IBertConfig"),
+        ("mt5", "MT5Config"),
+        ("t5", "T5Config"),
+        ("mobilebert", "MobileBertConfig"),
+        ("distilbert", "DistilBertConfig"),
+        ("albert", "AlbertConfig"),
+        ("bert-generation", "BertGenerationConfig"),
+        ("camembert", "CamembertConfig"),
+        ("xlm-roberta", "XLMRobertaConfig"),
+        ("pegasus", "PegasusConfig"),
+        ("marian", "MarianConfig"),
+        ("mbart", "MBartConfig"),
+        ("megatron-bert", "MegatronBertConfig"),
+        ("mpnet", "MPNetConfig"),
+        ("bart", "BartConfig"),
+        ("blenderbot", "BlenderbotConfig"),
+        ("reformer", "ReformerConfig"),
+        ("longformer", "LongformerConfig"),
+        ("roberta", "RobertaConfig"),
+        ("deberta-v2", "DebertaV2Config"),
+        ("deberta", "DebertaConfig"),
+        ("flaubert", "FlaubertConfig"),
+        ("fsmt", "FSMTConfig"),
+        ("squeezebert", "SqueezeBertConfig"),
+        ("hubert", "HubertConfig"),
+        ("bert", "BertConfig"),
+        ("openai-gpt", "OpenAIGPTConfig"),
+        ("gpt2", "GPT2Config"),
+        ("transfo-xl", "TransfoXLConfig"),
+        ("xlnet", "XLNetConfig"),
+        ("xlm-prophetnet", "XLMProphetNetConfig"),
+        ("prophetnet", "ProphetNetConfig"),
+        ("xlm", "XLMConfig"),
+        ("ctrl", "CTRLConfig"),
+        ("electra", "ElectraConfig"),
+        ("speech-encoder-decoder", "SpeechEncoderDecoderConfig"),
+        ("encoder-decoder", "EncoderDecoderConfig"),
+        ("funnel", "FunnelConfig"),
+        ("lxmert", "LxmertConfig"),
+        ("dpr", "DPRConfig"),
+        ("layoutlm", "LayoutLMConfig"),
+        ("rag", "RagConfig"),
+        ("tapas", "TapasConfig"),
+        ("splinter", "SplinterConfig"),
+    ]
+)
+
+CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
+    [
+        # Add archive maps here
+        ("fnet", "FNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("pegasus", "PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gptj", "GPTJ_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("layoutlmv2", "LAYOUTLMV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("beit", "BEIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("rembert", "REMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("visual_bert", "VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("canine", "CANINE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("roformer", "ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("clip", "CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bigbird_pegasus", "BIGBIRD_PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deit", "DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("luke", "LUKE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("detr", "DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gpt_neo", "GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("big_bird", "BIG_BIRD_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("megatron-bert", "MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("speech_to_text", "SPEECH_TO_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("speech_to_text_2", "SPEECH_TO_TEXT_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("vit", "VIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("wav2vec2", "WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("m2m_100", "M2M_100_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("convbert", "CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("led", "LED_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("blenderbot-small", "BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bert", "BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bart", "BART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("blenderbot", "BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mbart", "MBART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("openai-gpt", "OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("transfo-xl", "TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gpt2", "GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("ctrl", "CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xlnet", "XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xlm", "XLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("roberta", "ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("distilbert", "DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("albert", "ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("camembert", "CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("t5", "T5_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xlm-roberta", "XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("flaubert", "FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("fsmt", "FSMT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("electra", "ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("longformer", "LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("retribert", "RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("funnel", "FUNNEL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("lxmert", "LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("layoutlm", "LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("dpr", "DPR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deberta", "DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deberta-v2", "DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("squeezebert", "SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xlm-prophetnet", "XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("prophetnet", "PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mpnet", "MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("tapas", "TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("ibert", "IBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("hubert", "HUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("splinter", "SPLINTER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+    ]
+)
+
+MODEL_NAMES_MAPPING = OrderedDict(
+    [
+        # Add full (and cased) model names here
+        ("fnet", "FNet"),
+        ("gptj", "GPT-J"),
+        ("beit", "BeiT"),
+        ("rembert", "RemBERT"),
+        ("layoutlmv2", "LayoutLMv2"),
+        ("visual_bert", "VisualBert"),
+        ("canine", "Canine"),
+        ("roformer", "RoFormer"),
+        ("clip", "CLIP"),
+        ("bigbird_pegasus", "BigBirdPegasus"),
+        ("deit", "DeiT"),
+        ("luke", "LUKE"),
+        ("detr", "DETR"),
+        ("gpt_neo", "GPT Neo"),
+        ("big_bird", "BigBird"),
+        ("speech_to_text_2", "Speech2Text2"),
+        ("speech_to_text", "Speech2Text"),
+        ("vit", "ViT"),
+        ("wav2vec2", "Wav2Vec2"),
+        ("m2m_100", "M2M100"),
+        ("convbert", "ConvBERT"),
+        ("led", "LED"),
+        ("blenderbot-small", "BlenderbotSmall"),
+        ("retribert", "RetriBERT"),
+        ("ibert", "I-BERT"),
+        ("t5", "T5"),
+        ("mobilebert", "MobileBERT"),
+        ("distilbert", "DistilBERT"),
+        ("albert", "ALBERT"),
+        ("bert-generation", "Bert Generation"),
+        ("camembert", "CamemBERT"),
+        ("xlm-roberta", "XLM-RoBERTa"),
+        ("pegasus", "Pegasus"),
+        ("blenderbot", "Blenderbot"),
+        ("marian", "Marian"),
+        ("mbart", "mBART"),
+        ("megatron-bert", "MegatronBert"),
+        ("bart", "BART"),
+        ("reformer", "Reformer"),
+        ("longformer", "Longformer"),
+        ("roberta", "RoBERTa"),
+        ("flaubert", "FlauBERT"),
+        ("fsmt", "FairSeq Machine-Translation"),
+        ("squeezebert", "SqueezeBERT"),
+        ("bert", "BERT"),
+        ("openai-gpt", "OpenAI GPT"),
+        ("gpt2", "OpenAI GPT-2"),
+        ("transfo-xl", "Transformer-XL"),
+        ("xlnet", "XLNet"),
+        ("xlm", "XLM"),
+        ("ctrl", "CTRL"),
+        ("electra", "ELECTRA"),
+        ("encoder-decoder", "Encoder decoder"),
+        ("speech-encoder-decoder", "Speech Encoder decoder"),
+        ("funnel", "Funnel Transformer"),
+        ("lxmert", "LXMERT"),
+        ("deberta-v2", "DeBERTa-v2"),
+        ("deberta", "DeBERTa"),
+        ("layoutlm", "LayoutLM"),
+        ("dpr", "DPR"),
+        ("rag", "RAG"),
+        ("xlm-prophetnet", "XLMProphetNet"),
+        ("prophetnet", "ProphetNet"),
+        ("mt5", "mT5"),
+        ("mpnet", "MPNet"),
+        ("tapas", "TAPAS"),
+        ("hubert", "Hubert"),
+        ("barthez", "BARThez"),
+        ("phobert", "PhoBERT"),
+        ("cpm", "CPM"),
+        ("bertweet", "Bertweet"),
+        ("bert-japanese", "BertJapanese"),
+        ("byt5", "ByT5"),
+        ("mbart50", "mBART-50"),
+        ("splinter", "Splinter"),
+    ]
+)
+
+SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict([("openai-gpt", "openai")])
+
+
+def model_type_to_module_name(key):
+    """Converts a config key to the corresponding module."""
+    # Special treatment
+    if key in SPECIAL_MODEL_TYPE_TO_MODULE_NAME:
+        return SPECIAL_MODEL_TYPE_TO_MODULE_NAME[key]
+
+    return key.replace("-", "_")
+
+
+def config_class_to_model_type(config):
+    """Converts a config class name to the corresponding model type"""
+    for key, cls in CONFIG_MAPPING_NAMES.items():
+        if cls == config:
+            return key
+    return None
+
+
+class _LazyConfigMapping(OrderedDict):
+    """
+    A dictionary that lazily load its values when they are requested.
+    """
+
+    def __init__(self, mapping):
+        self._mapping = mapping
+        self._modules = {}
+
+    def __getitem__(self, key):
+        if key not in self._mapping:
+            raise KeyError(key)
+        value = self._mapping[key]
+        module_name = model_type_to_module_name(key)
+        if module_name not in self._modules:
+            self._modules[module_name] = importlib.import_module(f".{module_name}", "transformers.models")
+        return getattr(self._modules[module_name], value)
+
+    def keys(self):
+        return self._mapping.keys()
+
+    def values(self):
+        return [self[k] for k in self._mapping.keys()]
+
+    def items(self):
+        return [(k, self[k]) for k in self._mapping.keys()]
+
+    def __iter__(self):
+        return iter(self._mapping.keys())
+
+    def __contains__(self, item):
+        return item in self._mapping
+
+
+CONFIG_MAPPING = _LazyConfigMapping(CONFIG_MAPPING_NAMES)
+
+
+class _LazyLoadAllMappings(OrderedDict):
+    """
+    A mapping that will load all pairs of key values at the first access (either by indexing, requestions keys, values,
+    etc.)
+
+    Args:
+        mapping: The mapping to load.
+    """
+
+    def __init__(self, mapping):
+        self._mapping = mapping
+        self._initialized = False
+        self._data = {}
+
+    def _initialize(self):
+        if self._initialized:
+            return
+        logger.warn(
+            "ALL_PRETRAINED_CONFIG_ARCHIVE_MAP is deprecated and will be removed in v5 of Transformers. "
+            "It does not contain all available model checkpoints, far from it. Checkout hf.co/models for that.",
+            FutureWarning,
+        )
+
+        for model_type, map_name in self._mapping.items():
+            module_name = model_type_to_module_name(model_type)
+            module = importlib.import_module(f".{module_name}", "transformers.models")
+            mapping = getattr(module, map_name)
+            self._data.update(mapping)
+
+        self._initialized = True
+
+    def __getitem__(self, key):
+        self._initialize()
+        return self._data[key]
+
+    def keys(self):
+        self._initialize()
+        return self._data.keys()
+
+    def values(self):
+        self._initialize()
+        return self._data.values()
+
+    def items(self):
+        self._initialize()
+        return self._data.keys()
+
+    def __iter__(self):
+        self._initialize()
+        return iter(self._data)
+
+    def __contains__(self, item):
+        self._initialize()
+        return item in self._data
+
+
+ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = _LazyLoadAllMappings(CONFIG_ARCHIVE_MAP_MAPPING_NAMES)
+
+
+def _get_class_name(model_class: Union[str, List[str]]):
+    if isinstance(model_class, (list, tuple)):
+        return " or ".join([f":class:`~transformers.{c}`" for c in model_class if c is not None])
+    return f":class:`~transformers.{model_class}`"
+
+
+def _list_model_options(indent, config_to_class=None, use_model_types=True):
+    if config_to_class is None and not use_model_types:
+        raise ValueError("Using `use_model_types=False` requires a `config_to_class` dictionary.")
+    if use_model_types:
+        if config_to_class is None:
+            model_type_to_name = {
+                model_type: f":class:`~transformers.{config}`" for model_type, config in CONFIG_MAPPING_NAMES.items()
+            }
+        else:
+            model_type_to_name = {
+                model_type: _get_class_name(model_class)
+                for model_type, model_class in config_to_class.items()
+                if model_type in MODEL_NAMES_MAPPING
+            }
+        lines = [
+            f"{indent}- **{model_type}** -- {model_type_to_name[model_type]} ({MODEL_NAMES_MAPPING[model_type]} model)"
+            for model_type in sorted(model_type_to_name.keys())
+        ]
+    else:
+        config_to_name = {
+            CONFIG_MAPPING_NAMES[config]: _get_class_name(clas)
+            for config, clas in config_to_class.items()
+            if config in CONFIG_MAPPING_NAMES
+        }
+        config_to_model_name = {
+            config: MODEL_NAMES_MAPPING[model_type] for model_type, config in CONFIG_MAPPING_NAMES.items()
+        }
+        lines = [
+            f"{indent}- :class:`~transformers.{config_name}` configuration class: {config_to_name[config_name]} ({config_to_model_name[config_name]} model)"
+            for config_name in sorted(config_to_name.keys())
+        ]
+    return "\n".join(lines)
+
+
+def replace_list_option_in_docstrings(config_to_class=None, use_model_types=True):
+    def docstring_decorator(fn):
+        docstrings = fn.__doc__
+        lines = docstrings.split("\n")
+        i = 0
+        while i < len(lines) and re.search(r"^(\s*)List options\s*$", lines[i]) is None:
+            i += 1
+        if i < len(lines):
+            indent = re.search(r"^(\s*)List options\s*$", lines[i]).groups()[0]
+            if use_model_types:
+                indent = f"{indent}    "
+            lines[i] = _list_model_options(indent, config_to_class=config_to_class, use_model_types=use_model_types)
+            docstrings = "\n".join(lines)
+        else:
+            raise ValueError(
+                f"The function {fn} should have an empty 'List options' in its docstring as placeholder, current docstring is:\n{docstrings}"
+            )
+        fn.__doc__ = docstrings
+        return fn
+
+    return docstring_decorator
+
+
+class AutoConfig:
+    r"""
+    This is a generic configuration class that will be instantiated as one of the configuration classes of the library
+    when created with the :meth:`~transformers.AutoConfig.from_pretrained` class method.
+
+    This class cannot be instantiated directly using ``__init__()`` (throws an error).
+    """
+
+    def __init__(self):
+        raise EnvironmentError(
+            "AutoConfig is designed to be instantiated "
+            "using the `AutoConfig.from_pretrained(pretrained_model_name_or_path)` method."
+        )
+
+    @classmethod
+    def for_model(cls, model_type: str, *args, **kwargs):
+        if model_type in CONFIG_MAPPING:
+            config_class = CONFIG_MAPPING[model_type]
+            return config_class(*args, **kwargs)
+        raise ValueError(
+            f"Unrecognized model identifier: {model_type}. Should contain one of {', '.join(CONFIG_MAPPING.keys())}"
+        )
+
+    @classmethod
+    @replace_list_option_in_docstrings()
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r"""
+        Instantiate one of the configuration classes of the library from a pretrained model configuration.
+
+        The configuration class to instantiate is selected based on the :obj:`model_type` property of the config object
+        that is loaded, or when it's missing, by falling back to using pattern matching on
+        :obj:`pretrained_model_name_or_path`:
+
+        List options
+
+        Args:
+            pretrained_model_name_or_path (:obj:`str` or :obj:`os.PathLike`):
+                Can be either:
+
+                    - A string, the `model id` of a pretrained model configuration hosted inside a model repo on
+                      huggingface.co. Valid model ids can be located at the root-level, like ``bert-base-uncased``, or
+                      namespaced under a user or organization name, like ``dbmdz/bert-base-german-cased``.
+                    - A path to a `directory` containing a configuration file saved using the
+                      :meth:`~transformers.PretrainedConfig.save_pretrained` method, or the
+                      :meth:`~transformers.PreTrainedModel.save_pretrained` method, e.g., ``./my_model_directory/``.
+                    - A path or url to a saved configuration JSON `file`, e.g.,
+                      ``./my_model_directory/configuration.json``.
+            cache_dir (:obj:`str` or :obj:`os.PathLike`, `optional`):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            force_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to force the (re-)download the model weights and configuration files and override the
+                cached versions if they exist.
+            resume_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+                file exists.
+            proxies (:obj:`Dict[str, str]`, `optional`):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., :obj:`{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            revision(:obj:`str`, `optional`, defaults to :obj:`"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so ``revision`` can be any
+                identifier allowed by git.
+            return_unused_kwargs (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                If :obj:`False`, then this function returns just the final configuration object.
+
+                If :obj:`True`, then this functions returns a :obj:`Tuple(config, unused_kwargs)` where `unused_kwargs`
+                is a dictionary consisting of the key/value pairs whose keys are not configuration attributes: i.e.,
+                the part of ``kwargs`` which has not been used to update ``config`` and is otherwise ignored.
+            kwargs(additional keyword arguments, `optional`):
+                The values in kwargs of any keys which are configuration attributes will be used to override the loaded
+                values. Behavior concerning key/value pairs whose keys are *not* configuration attributes is controlled
+                by the ``return_unused_kwargs`` keyword parameter.
+
+        Examples::
+
+            >>> from transformers import AutoConfig
+
+            >>> # Download configuration from huggingface.co and cache.
+            >>> config = AutoConfig.from_pretrained('bert-base-uncased')
+
+            >>> # Download configuration from huggingface.co (user-uploaded) and cache.
+            >>> config = AutoConfig.from_pretrained('dbmdz/bert-base-german-cased')
+
+            >>> # If configuration file is in a directory (e.g., was saved using `save_pretrained('./test/saved_model/')`).
+            >>> config = AutoConfig.from_pretrained('./test/bert_saved_model/')
+
+            >>> # Load a specific configuration file.
+            >>> config = AutoConfig.from_pretrained('./test/bert_saved_model/my_configuration.json')
+
+            >>> # Change some config attributes when loading a pretrained config.
+            >>> config = AutoConfig.from_pretrained('bert-base-uncased', output_attentions=True, foo=False)
+            >>> config.output_attentions
+            True
+            >>> config, unused_kwargs = AutoConfig.from_pretrained('bert-base-uncased', output_attentions=True, foo=False, return_unused_kwargs=True)
+            >>> config.output_attentions
+            True
+            >>> config.unused_kwargs
+            {'foo': False}
+        """
+        kwargs["_from_auto"] = True
+        config_dict, _ = PretrainedConfig.get_config_dict(pretrained_model_name_or_path, **kwargs)
+        if "model_type" in config_dict:
+            config_class = CONFIG_MAPPING[config_dict["model_type"]]
+            return config_class.from_dict(config_dict, **kwargs)
+        else:
+            # Fallback: use pattern matching on the string.
+            for pattern, config_class in CONFIG_MAPPING.items():
+                if pattern in str(pretrained_model_name_or_path):
+                    return config_class.from_dict(config_dict, **kwargs)
+
+        raise ValueError(
+            f"Unrecognized model in {pretrained_model_name_or_path}. "
+            f"Should have a `model_type` key in its {CONFIG_NAME}, or contain one of the following strings "
+            f"in its name: {', '.join(CONFIG_MAPPING.keys())}"
+        )

fastNLP/transformers/torch/models/auto/tokenization_auto.py

@@ -0,0 +1,199 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Auto Tokenizer class. """
+
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Dict, Optional, Tuple, Union
+
+from ...file_utils import (
+    is_sentencepiece_available,
+    is_tokenizers_available,
+)
+
+if TYPE_CHECKING:
+    # This significantly improves completion suggestion performance when
+    # the transformers package is used with Microsoft's Pylance language server.
+    TOKENIZER_MAPPING_NAMES: OrderedDict[str, Tuple[Optional[str], Optional[str]]] = OrderedDict()
+else:
+    TOKENIZER_MAPPING_NAMES = OrderedDict(
+        [
+            ("fnet", ("FNetTokenizer", "FNetTokenizerFast" if is_tokenizers_available() else None)),
+            ("retribert", ("RetriBertTokenizer", "RetriBertTokenizerFast" if is_tokenizers_available() else None)),
+            ("roformer", ("RoFormerTokenizer", "RoFormerTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "t5",
+                (
+                    "T5Tokenizer" if is_sentencepiece_available() else None,
+                    "T5TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "mt5",
+                (
+                    "MT5Tokenizer" if is_sentencepiece_available() else None,
+                    "MT5TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("mobilebert", ("MobileBertTokenizer", "MobileBertTokenizerFast" if is_tokenizers_available() else None)),
+            ("distilbert", ("DistilBertTokenizer", "DistilBertTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "albert",
+                (
+                    "AlbertTokenizer" if is_sentencepiece_available() else None,
+                    "AlbertTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "camembert",
+                (
+                    "CamembertTokenizer" if is_sentencepiece_available() else None,
+                    "CamembertTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "pegasus",
+                (
+                    "PegasusTokenizer" if is_sentencepiece_available() else None,
+                    "PegasusTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "mbart",
+                (
+                    "MBartTokenizer" if is_sentencepiece_available() else None,
+                    "MBartTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "xlm-roberta",
+                (
+                    "XLMRobertaTokenizer" if is_sentencepiece_available() else None,
+                    "XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("marian", ("MarianTokenizer" if is_sentencepiece_available() else None, None)),
+            ("blenderbot-small", ("BlenderbotSmallTokenizer", None)),
+            ("blenderbot", ("BlenderbotTokenizer", None)),
+            ("bart", ("BartTokenizer", "BartTokenizerFast")),
+            ("longformer", ("LongformerTokenizer", "LongformerTokenizerFast" if is_tokenizers_available() else None)),
+            ("roberta", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "reformer",
+                (
+                    "ReformerTokenizer" if is_sentencepiece_available() else None,
+                    "ReformerTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("electra", ("ElectraTokenizer", "ElectraTokenizerFast" if is_tokenizers_available() else None)),
+            ("funnel", ("FunnelTokenizer", "FunnelTokenizerFast" if is_tokenizers_available() else None)),
+            ("lxmert", ("LxmertTokenizer", "LxmertTokenizerFast" if is_tokenizers_available() else None)),
+            ("layoutlm", ("LayoutLMTokenizer", "LayoutLMTokenizerFast" if is_tokenizers_available() else None)),
+            ("layoutlmv2", ("LayoutLMv2Tokenizer", "LayoutLMv2TokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "dpr",
+                (
+                    "DPRQuestionEncoderTokenizer",
+                    "DPRQuestionEncoderTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "squeezebert",
+                ("SqueezeBertTokenizer", "SqueezeBertTokenizerFast" if is_tokenizers_available() else None),
+            ),
+            ("bert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("openai-gpt", ("OpenAIGPTTokenizer", "OpenAIGPTTokenizerFast" if is_tokenizers_available() else None)),
+            ("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("transfo-xl", ("TransfoXLTokenizer", None)),
+            (
+                "xlnet",
+                (
+                    "XLNetTokenizer" if is_sentencepiece_available() else None,
+                    "XLNetTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("flaubert", ("FlaubertTokenizer", None)),
+            ("xlm", ("XLMTokenizer", None)),
+            ("ctrl", ("CTRLTokenizer", None)),
+            ("fsmt", ("FSMTTokenizer", None)),
+            ("bert-generation", ("BertGenerationTokenizer" if is_sentencepiece_available() else None, None)),
+            ("deberta", ("DebertaTokenizer", "DebertaTokenizerFast" if is_tokenizers_available() else None)),
+            ("deberta-v2", ("DebertaV2Tokenizer" if is_sentencepiece_available() else None, None)),
+            ("rag", ("RagTokenizer", None)),
+            ("xlm-prophetnet", ("XLMProphetNetTokenizer" if is_sentencepiece_available() else None, None)),
+            ("speech_to_text", ("Speech2TextTokenizer" if is_sentencepiece_available() else None, None)),
+            ("speech_to_text_2", ("Speech2Text2Tokenizer", None)),
+            ("m2m_100", ("M2M100Tokenizer" if is_sentencepiece_available() else None, None)),
+            ("prophetnet", ("ProphetNetTokenizer", None)),
+            ("mpnet", ("MPNetTokenizer", "MPNetTokenizerFast" if is_tokenizers_available() else None)),
+            ("tapas", ("TapasTokenizer", None)),
+            ("led", ("LEDTokenizer", "LEDTokenizerFast" if is_tokenizers_available() else None)),
+            ("convbert", ("ConvBertTokenizer", "ConvBertTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "big_bird",
+                (
+                    "BigBirdTokenizer" if is_sentencepiece_available() else None,
+                    "BigBirdTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("ibert", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
+            ("wav2vec2", ("Wav2Vec2CTCTokenizer", None)),
+            ("hubert", ("Wav2Vec2CTCTokenizer", None)),
+            ("gpt_neo", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("luke", ("LukeTokenizer", None)),
+            ("bigbird_pegasus", ("PegasusTokenizer", "PegasusTokenizerFast" if is_tokenizers_available() else None)),
+            ("canine", ("CanineTokenizer", None)),
+            ("bertweet", ("BertweetTokenizer", None)),
+            ("bert-japanese", ("BertJapaneseTokenizer", None)),
+            ("splinter", ("SplinterTokenizer", "SplinterTokenizerFast")),
+            ("byt5", ("ByT5Tokenizer", None)),
+            (
+                "cpm",
+                (
+                    "CpmTokenizer" if is_sentencepiece_available() else None,
+                    "CpmTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("herbert", ("HerbertTokenizer", "HerbertTokenizerFast" if is_tokenizers_available() else None)),
+            ("phobert", ("PhobertTokenizer", None)),
+            (
+                "barthez",
+                (
+                    "BarthezTokenizer" if is_sentencepiece_available() else None,
+                    "BarthezTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "mbart50",
+                (
+                    "MBart50Tokenizer" if is_sentencepiece_available() else None,
+                    "MBart50TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "rembert",
+                (
+                    "RemBertTokenizer" if is_sentencepiece_available() else None,
+                    "RemBertTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "clip",
+                (
+                    "CLIPTokenizer",
+                    "CLIPTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+        ]
+    )

fastNLP/transformers/torch/models/bart/__init__.py

@@ -0,0 +1,20 @@
+__all__ = [
+    "BartConfig",
+    "BART_PRETRAINED_CONFIG_ARCHIVE_MAP",
+
+    "BART_PRETRAINED_MODEL_ARCHIVE_LIST",
+    "BartForCausalLM",
+    "BartForConditionalGeneration",
+    "BartForQuestionAnswering",
+    "BartForSequenceClassification",
+    "BartModel",
+    "BartPretrainedModel",
+    "PretrainedBartModel",
+
+    "BartTokenizer",
+]
+
+from .configuration_bart import BartConfig, BART_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .tokenization_bart import BartTokenizer
+from .modeling_bart import BartForCausalLM, BartForConditionalGeneration, BartModel, BartForQuestionAnswering, \
+    BartForSequenceClassification, BartPretrainedModel, PretrainedBartModel, BART_PRETRAINED_MODEL_ARCHIVE_LIST

fastNLP/transformers/torch/models/bart/configuration_bart.py

@@ -0,0 +1,177 @@
+# coding=utf-8
+# Copyright 2021 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" BART model configuration """
+from fastNLP.transformers.torch.configuration_utils import PretrainedConfig
+from fastNLP.core.log import logger
+
+__all__ = [
+    "BartConfig",
+    "BART_PRETRAINED_CONFIG_ARCHIVE_MAP",
+]
+
+BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/bart-large": "https://huggingface.co/facebook/bart-large/resolve/main/config.json",
+    # See all BART models at https://huggingface.co/models?filter=bart
+}
+
+
+class BartConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.BartModel`. It is used to
+    instantiate a BART model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the BART `facebook/bart-large
+    <https://huggingface.co/facebook/bart-large>`__ architecture.
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 50265):
+            Vocabulary size of the BART model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.BartModel` or
+            :class:`~transformers.TFBartModel`.
+        d_model (:obj:`int`, `optional`, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of encoder layers.
+        decoder_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of decoder layers.
+        encoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (:obj:`int`, `optional`, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (:obj:`int`, `optional`, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string,
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
+        dropout (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        init_std (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the `LayerDrop paper <see
+            https://arxiv.org/abs/1909.11556>`__ for more details.
+        decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the `LayerDrop paper <see
+            https://arxiv.org/abs/1909.11556>`__ for more details.
+        scale_embedding (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        num_labels: (:obj:`int`, `optional`, defaults to 3):
+            The number of labels to use in :class:`~transformers.BartForSequenceClassification`.
+        forced_eos_token_id (:obj:`int`, `optional`, defaults to 2):
+            The id of the token to force as the last generated token when :obj:`max_length` is reached. Usually set to
+            :obj:`eos_token_id`.
+
+    Example::
+
+        >>> from transformers import BartModel, BartConfig
+
+        >>> # Initializing a BART facebook/bart-large style configuration
+        >>> configuration = BartConfig()
+
+        >>> # Initializing a model from the facebook/bart-large style configuration
+        >>> model = BartModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+    model_type = "bart"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        max_position_embeddings=1024,
+        encoder_layers=12,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        decoder_layers=12,
+        decoder_ffn_dim=4096,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        activation_function="gelu",
+        d_model=1024,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        classifier_dropout=0.0,
+        scale_embedding=False,
+        use_cache=True,
+        num_labels=3,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        is_encoder_decoder=True,
+        decoder_start_token_id=2,
+        forced_eos_token_id=2,
+        **kwargs
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+        super().__init__(
+            num_labels=num_labels,
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            forced_eos_token_id=forced_eos_token_id,
+            **kwargs,
+        )
+
+        # ensure backward compatibility for BART CNN models
+        if self.forced_bos_token_id is None and kwargs.get("force_bos_token_to_be_generated", False):
+            self.forced_bos_token_id = self.bos_token_id
+            logger.warn(
+                f"Please make sure the config includes `forced_bos_token_id={self.bos_token_id}` in future versions."
+                "The config can simply be saved and uploaded again to be fixed."
+            )

fastNLP/transformers/torch/models/bart/tokenization_bart.py

@@ -0,0 +1,65 @@
+# coding=utf-8
+# Copyright 2020 The Facebook AI Research Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..roberta.tokenization_roberta import RobertaTokenizer
+from fastNLP.core.log import logger
+
+__all__ = [
+    "BartTokenizer",
+]
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt"}
+
+# See all BART models at https://huggingface.co/models?filter=bart
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "facebook/bart-base": "https://huggingface.co/facebook/bart-base/resolve/main/vocab.json",
+        "facebook/bart-large": "https://huggingface.co/facebook/bart-large/resolve/main/vocab.json",
+        "facebook/bart-large-mnli": "https://huggingface.co/facebook/bart-large-mnli/resolve/main/vocab.json",
+        "facebook/bart-large-cnn": "https://huggingface.co/facebook/bart-large-cnn/resolve/main/vocab.json",
+        "facebook/bart-large-xsum": "https://huggingface.co/facebook/bart-large-xsum/resolve/main/vocab.json",
+        "yjernite/bart_eli5": "https://huggingface.co/yjernite/bart_eli5/resolve/main/vocab.json",
+    },
+    "merges_file": {
+        "facebook/bart-base": "https://huggingface.co/facebook/bart-base/resolve/main/merges.txt",
+        "facebook/bart-large": "https://huggingface.co/facebook/bart-large/resolve/main/merges.txt",
+        "facebook/bart-large-mnli": "https://huggingface.co/facebook/bart-large-mnli/resolve/main/merges.txt",
+        "facebook/bart-large-cnn": "https://huggingface.co/facebook/bart-large-cnn/resolve/main/merges.txt",
+        "facebook/bart-large-xsum": "https://huggingface.co/facebook/bart-large-xsum/resolve/main/merges.txt",
+        "yjernite/bart_eli5": "https://huggingface.co/yjernite/bart_eli5/resolve/main/merges.txt",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "facebook/bart-base": 1024,
+    "facebook/bart-large": 1024,
+    "facebook/bart-large-mnli": 1024,
+    "facebook/bart-large-cnn": 1024,
+    "facebook/bart-large-xsum": 1024,
+    "yjernite/bart_eli5": 1024,
+}
+
+
+class BartTokenizer(RobertaTokenizer):
+    r"""
+    Construct a BART tokenizer.
+
+    :class:`~transformers.BartTokenizer` is identical to :class:`~transformers.RobertaTokenizer`. Refer to superclass
+    :class:`~transformers.RobertaTokenizer` for usage examples and documentation concerning the initialization
+    parameters and other methods.
+    """
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES

fastNLP/transformers/torch/models/bert/__init__.py

@@ -0,0 +1,27 @@
+__all__ = [
+    "BERT_PRETRAINED_CONFIG_ARCHIVE_MAP",
+    "BertConfig",
+
+    "BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+    "BertForMaskedLM",
+    "BertForMultipleChoice",
+    "BertForNextSentencePrediction",
+    "BertForPreTraining",
+    "BertForQuestionAnswering",
+    "BertForSequenceClassification",
+    "BertForTokenClassification",
+    "BertLayer",
+    "BertLMHeadModel",
+    "BertModel",
+    "BertPreTrainedModel",
+
+    "BasicTokenizer",
+    "BertTokenizer",
+    "WordpieceTokenizer",
+]
+
+from .configuration_bert import BertConfig, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .tokenization_bert import BasicTokenizer, BertTokenizer, WordpieceTokenizer
+from .modeling_bert import BERT_PRETRAINED_MODEL_ARCHIVE_LIST, BertForMaskedLM, BertForMultipleChoice, BertForPreTraining, \
+    BertForNextSentencePrediction, BertForQuestionAnswering, BertForSequenceClassification, BertForTokenClassification, \
+        BertLayer, BertLMHeadModel, BertModel, BertPreTrainedModel

fastNLP/transformers/torch/models/bert/configuration_bert.py

@@ -0,0 +1,158 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" BERT model configuration """
+
+from fastNLP.transformers.torch.configuration_utils import PretrainedConfig
+from fastNLP.core.log import logger
+
+__all__ = [
+    "BERT_PRETRAINED_CONFIG_ARCHIVE_MAP",
+    "BertConfig",
+]
+
+BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "bert-base-uncased": "https://huggingface.co/bert-base-uncased/resolve/main/config.json",
+    "bert-large-uncased": "https://huggingface.co/bert-large-uncased/resolve/main/config.json",
+    "bert-base-cased": "https://huggingface.co/bert-base-cased/resolve/main/config.json",
+    "bert-large-cased": "https://huggingface.co/bert-large-cased/resolve/main/config.json",
+    "bert-base-multilingual-uncased": "https://huggingface.co/bert-base-multilingual-uncased/resolve/main/config.json",
+    "bert-base-multilingual-cased": "https://huggingface.co/bert-base-multilingual-cased/resolve/main/config.json",
+    "bert-base-chinese": "https://huggingface.co/bert-base-chinese/resolve/main/config.json",
+    "bert-base-german-cased": "https://huggingface.co/bert-base-german-cased/resolve/main/config.json",
+    "bert-large-uncased-whole-word-masking": "https://huggingface.co/bert-large-uncased-whole-word-masking/resolve/main/config.json",
+    "bert-large-cased-whole-word-masking": "https://huggingface.co/bert-large-cased-whole-word-masking/resolve/main/config.json",
+    "bert-large-uncased-whole-word-masking-finetuned-squad": "https://huggingface.co/bert-large-uncased-whole-word-masking-finetuned-squad/resolve/main/config.json",
+    "bert-large-cased-whole-word-masking-finetuned-squad": "https://huggingface.co/bert-large-cased-whole-word-masking-finetuned-squad/resolve/main/config.json",
+    "bert-base-cased-finetuned-mrpc": "https://huggingface.co/bert-base-cased-finetuned-mrpc/resolve/main/config.json",
+    "bert-base-german-dbmdz-cased": "https://huggingface.co/bert-base-german-dbmdz-cased/resolve/main/config.json",
+    "bert-base-german-dbmdz-uncased": "https://huggingface.co/bert-base-german-dbmdz-uncased/resolve/main/config.json",
+    "cl-tohoku/bert-base-japanese": "https://huggingface.co/cl-tohoku/bert-base-japanese/resolve/main/config.json",
+    "cl-tohoku/bert-base-japanese-whole-word-masking": "https://huggingface.co/cl-tohoku/bert-base-japanese-whole-word-masking/resolve/main/config.json",
+    "cl-tohoku/bert-base-japanese-char": "https://huggingface.co/cl-tohoku/bert-base-japanese-char/resolve/main/config.json",
+    "cl-tohoku/bert-base-japanese-char-whole-word-masking": "https://huggingface.co/cl-tohoku/bert-base-japanese-char-whole-word-masking/resolve/main/config.json",
+    "TurkuNLP/bert-base-finnish-cased-v1": "https://huggingface.co/TurkuNLP/bert-base-finnish-cased-v1/resolve/main/config.json",
+    "TurkuNLP/bert-base-finnish-uncased-v1": "https://huggingface.co/TurkuNLP/bert-base-finnish-uncased-v1/resolve/main/config.json",
+    "wietsedv/bert-base-dutch-cased": "https://huggingface.co/wietsedv/bert-base-dutch-cased/resolve/main/config.json",
+    # See all BERT models at https://huggingface.co/models?filter=bert
+}
+
+
+class BertConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.BertModel` or a
+    :class:`~transformers.TFBertModel`. It is used to instantiate a BERT model according to the specified arguments,
+    defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the BERT `bert-base-uncased <https://huggingface.co/bert-base-uncased>`__ architecture.
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 30522):
+            Vocabulary size of the BERT model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.BertModel` or
+            :class:`~transformers.TFBertModel`.
+        hidden_size (:obj:`int`, `optional`, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (:obj:`int`, `optional`, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, `optional`, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string,
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
+        hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (:obj:`int`, `optional`, defaults to 2):
+            The vocabulary size of the :obj:`token_type_ids` passed when calling :class:`~transformers.BertModel` or
+            :class:`~transformers.TFBertModel`.
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (:obj:`str`, `optional`, defaults to :obj:`"absolute"`):
+            Type of position embedding. Choose one of :obj:`"absolute"`, :obj:`"relative_key"`,
+            :obj:`"relative_key_query"`. For positional embeddings use :obj:`"absolute"`. For more information on
+            :obj:`"relative_key"`, please refer to `Self-Attention with Relative Position Representations (Shaw et al.)
+            <https://arxiv.org/abs/1803.02155>`__. For more information on :obj:`"relative_key_query"`, please refer to
+            `Method 4` in `Improve Transformer Models with Better Relative Position Embeddings (Huang et al.)
+            <https://arxiv.org/abs/2009.13658>`__.
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if ``config.is_decoder=True``.
+        classifier_dropout (:obj:`float`, `optional`):
+            The dropout ratio for the classification head.
+
+    Examples::
+
+        >>> from transformers import BertModel, BertConfig
+
+        >>> # Initializing a BERT bert-base-uncased style configuration
+        >>> configuration = BertConfig()
+
+        >>> # Initializing a model from the bert-base-uncased style configuration
+        >>> model = BertModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+    model_type = "bert"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        use_cache=True,
+        classifier_dropout=None,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.classifier_dropout = classifier_dropout

fastNLP/transformers/torch/models/bert/tokenization_bert.py

@@ -0,0 +1,558 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for Bert."""
+
+
+import collections
+import os
+import unicodedata
+from typing import List, Optional, Tuple
+
+from fastNLP.transformers.torch.tokenization_utils import PreTrainedTokenizer, _is_control, _is_punctuation, _is_whitespace
+from fastNLP.core.log import logger
+
+__all__ = [
+    "BasicTokenizer",
+    "BertTokenizer",
+    "WordpieceTokenizer",
+]
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "bert-base-uncased": "https://huggingface.co/bert-base-uncased/resolve/main/vocab.txt",
+        "bert-large-uncased": "https://huggingface.co/bert-large-uncased/resolve/main/vocab.txt",
+        "bert-base-cased": "https://huggingface.co/bert-base-cased/resolve/main/vocab.txt",
+        "bert-large-cased": "https://huggingface.co/bert-large-cased/resolve/main/vocab.txt",
+        "bert-base-multilingual-uncased": "https://huggingface.co/bert-base-multilingual-uncased/resolve/main/vocab.txt",
+        "bert-base-multilingual-cased": "https://huggingface.co/bert-base-multilingual-cased/resolve/main/vocab.txt",
+        "bert-base-chinese": "https://huggingface.co/bert-base-chinese/resolve/main/vocab.txt",
+        "bert-base-german-cased": "https://huggingface.co/bert-base-german-cased/resolve/main/vocab.txt",
+        "bert-large-uncased-whole-word-masking": "https://huggingface.co/bert-large-uncased-whole-word-masking/resolve/main/vocab.txt",
+        "bert-large-cased-whole-word-masking": "https://huggingface.co/bert-large-cased-whole-word-masking/resolve/main/vocab.txt",
+        "bert-large-uncased-whole-word-masking-finetuned-squad": "https://huggingface.co/bert-large-uncased-whole-word-masking-finetuned-squad/resolve/main/vocab.txt",
+        "bert-large-cased-whole-word-masking-finetuned-squad": "https://huggingface.co/bert-large-cased-whole-word-masking-finetuned-squad/resolve/main/vocab.txt",
+        "bert-base-cased-finetuned-mrpc": "https://huggingface.co/bert-base-cased-finetuned-mrpc/resolve/main/vocab.txt",
+        "bert-base-german-dbmdz-cased": "https://huggingface.co/bert-base-german-dbmdz-cased/resolve/main/vocab.txt",
+        "bert-base-german-dbmdz-uncased": "https://huggingface.co/bert-base-german-dbmdz-uncased/resolve/main/vocab.txt",
+        "TurkuNLP/bert-base-finnish-cased-v1": "https://huggingface.co/TurkuNLP/bert-base-finnish-cased-v1/resolve/main/vocab.txt",
+        "TurkuNLP/bert-base-finnish-uncased-v1": "https://huggingface.co/TurkuNLP/bert-base-finnish-uncased-v1/resolve/main/vocab.txt",
+        "wietsedv/bert-base-dutch-cased": "https://huggingface.co/wietsedv/bert-base-dutch-cased/resolve/main/vocab.txt",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "bert-base-uncased": 512,
+    "bert-large-uncased": 512,
+    "bert-base-cased": 512,
+    "bert-large-cased": 512,
+    "bert-base-multilingual-uncased": 512,
+    "bert-base-multilingual-cased": 512,
+    "bert-base-chinese": 512,
+    "bert-base-german-cased": 512,
+    "bert-large-uncased-whole-word-masking": 512,
+    "bert-large-cased-whole-word-masking": 512,
+    "bert-large-uncased-whole-word-masking-finetuned-squad": 512,
+    "bert-large-cased-whole-word-masking-finetuned-squad": 512,
+    "bert-base-cased-finetuned-mrpc": 512,
+    "bert-base-german-dbmdz-cased": 512,
+    "bert-base-german-dbmdz-uncased": 512,
+    "TurkuNLP/bert-base-finnish-cased-v1": 512,
+    "TurkuNLP/bert-base-finnish-uncased-v1": 512,
+    "wietsedv/bert-base-dutch-cased": 512,
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    "bert-base-uncased": {"do_lower_case": True},
+    "bert-large-uncased": {"do_lower_case": True},
+    "bert-base-cased": {"do_lower_case": False},
+    "bert-large-cased": {"do_lower_case": False},
+    "bert-base-multilingual-uncased": {"do_lower_case": True},
+    "bert-base-multilingual-cased": {"do_lower_case": False},
+    "bert-base-chinese": {"do_lower_case": False},
+    "bert-base-german-cased": {"do_lower_case": False},
+    "bert-large-uncased-whole-word-masking": {"do_lower_case": True},
+    "bert-large-cased-whole-word-masking": {"do_lower_case": False},
+    "bert-large-uncased-whole-word-masking-finetuned-squad": {"do_lower_case": True},
+    "bert-large-cased-whole-word-masking-finetuned-squad": {"do_lower_case": False},
+    "bert-base-cased-finetuned-mrpc": {"do_lower_case": False},
+    "bert-base-german-dbmdz-cased": {"do_lower_case": False},
+    "bert-base-german-dbmdz-uncased": {"do_lower_case": True},
+    "TurkuNLP/bert-base-finnish-cased-v1": {"do_lower_case": False},
+    "TurkuNLP/bert-base-finnish-uncased-v1": {"do_lower_case": True},
+    "wietsedv/bert-base-dutch-cased": {"do_lower_case": False},
+}
+
+
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    with open(vocab_file, "r", encoding="utf-8") as reader:
+        tokens = reader.readlines()
+    for index, token in enumerate(tokens):
+        token = token.rstrip("\n")
+        vocab[token] = index
+    return vocab
+
+
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a piece of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+class BertTokenizer(PreTrainedTokenizer):
+    r"""
+    Construct a BERT tokenizer. Based on WordPiece.
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
+    Users should refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (:obj:`str`):
+            File containing the vocabulary.
+        do_lower_case (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to lowercase the input when tokenizing.
+        do_basic_tokenize (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to do basic tokenization before WordPiece.
+        never_split (:obj:`Iterable`, `optional`):
+            Collection of tokens which will never be split during tokenization. Only has an effect when
+            :obj:`do_basic_tokenize=True`
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`"[UNK]"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        sep_token (:obj:`str`, `optional`, defaults to :obj:`"[SEP]"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        pad_token (:obj:`str`, `optional`, defaults to :obj:`"[PAD]"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (:obj:`str`, `optional`, defaults to :obj:`"[CLS]"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        mask_token (:obj:`str`, `optional`, defaults to :obj:`"[MASK]"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        tokenize_chinese_chars (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to tokenize Chinese characters.
+
+            This should likely be deactivated for Japanese (see this `issue
+            <https://github.com/huggingface/transformers/issues/328>`__).
+        strip_accents: (:obj:`bool`, `optional`):
+            Whether or not to strip all accents. If this option is not specified, then it will be determined by the
+            value for :obj:`lowercase` (as in the original BERT).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(
+        self,
+        vocab_file,
+        do_lower_case=True,
+        do_basic_tokenize=True,
+        never_split=None,
+        unk_token="[UNK]",
+        sep_token="[SEP]",
+        pad_token="[PAD]",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        tokenize_chinese_chars=True,
+        strip_accents=None,
+        **kwargs
+    ):
+        super().__init__(
+            do_lower_case=do_lower_case,
+            do_basic_tokenize=do_basic_tokenize,
+            never_split=never_split,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            tokenize_chinese_chars=tokenize_chinese_chars,
+            strip_accents=strip_accents,
+            **kwargs,
+        )
+
+        if not os.path.isfile(vocab_file):
+            raise ValueError(
+                f"Can't find a vocabulary file at path '{vocab_file}'. To load the vocabulary from a Google pretrained "
+                "model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"
+            )
+        self.vocab = load_vocab(vocab_file)
+        self.ids_to_tokens = collections.OrderedDict([(ids, tok) for tok, ids in self.vocab.items()])
+        self.do_basic_tokenize = do_basic_tokenize
+        if do_basic_tokenize:
+            self.basic_tokenizer = BasicTokenizer(
+                do_lower_case=do_lower_case,
+                never_split=never_split,
+                tokenize_chinese_chars=tokenize_chinese_chars,
+                strip_accents=strip_accents,
+            )
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab, unk_token=self.unk_token)
+
+    @property
+    def do_lower_case(self):
+        return self.basic_tokenizer.do_lower_case
+
+    @property
+    def vocab_size(self):
+        return len(self.vocab)
+
+    def get_vocab(self):
+        return dict(self.vocab, **self.added_tokens_encoder)
+
+    def _tokenize(self, text):
+        split_tokens = []
+        if self.do_basic_tokenize:
+            for token in self.basic_tokenizer.tokenize(text, never_split=self.all_special_tokens):
+
+                # If the token is part of the never_split set
+                if token in self.basic_tokenizer.never_split:
+                    split_tokens.append(token)
+                else:
+                    split_tokens += self.wordpiece_tokenizer.tokenize(token)
+        else:
+            split_tokens = self.wordpiece_tokenizer.tokenize(text)
+        return split_tokens
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.vocab.get(token, self.vocab.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.ids_to_tokens.get(index, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        out_string = " ".join(tokens).replace(" ##", "").strip()
+        return out_string
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A BERT sequence has the following format:
+
+        - single sequence: ``[CLS] X [SEP]``
+        - pair of sequences: ``[CLS] A [SEP] B [SEP]``
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` method.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            :obj:`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. A BERT sequence
+        pair mask has the following format:
+
+        ::
+
+            0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+            | first sequence    | second sequence |
+
+        If :obj:`token_ids_1` is :obj:`None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `token type IDs <../glossary.html#token-type-ids>`_ according to the given
+            sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        index = 0
+        if os.path.isdir(save_directory):
+            vocab_file = os.path.join(
+                save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+            )
+        else:
+            vocab_file = (filename_prefix + "-" if filename_prefix else "") + save_directory
+        with open(vocab_file, "w", encoding="utf-8") as writer:
+            for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        f"Saving vocabulary to {vocab_file}: vocabulary indices are not consecutive."
+                        " Please check that the vocabulary is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(token + "\n")
+                index += 1
+        return (vocab_file,)
+
+
+class BasicTokenizer(object):
+    """
+    Constructs a BasicTokenizer that will run basic tokenization (punctuation splitting, lower casing, etc.).
+
+    Args:
+        do_lower_case (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to lowercase the input when tokenizing.
+        never_split (:obj:`Iterable`, `optional`):
+            Collection of tokens which will never be split during tokenization. Only has an effect when
+            :obj:`do_basic_tokenize=True`
+        tokenize_chinese_chars (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to tokenize Chinese characters.
+
+            This should likely be deactivated for Japanese (see this `issue
+            <https://github.com/huggingface/transformers/issues/328>`__).
+        strip_accents: (:obj:`bool`, `optional`):
+            Whether or not to strip all accents. If this option is not specified, then it will be determined by the
+            value for :obj:`lowercase` (as in the original BERT).
+    """
+
+    def __init__(self, do_lower_case=True, never_split=None, tokenize_chinese_chars=True, strip_accents=None):
+        if never_split is None:
+            never_split = []
+        self.do_lower_case = do_lower_case
+        self.never_split = set(never_split)
+        self.tokenize_chinese_chars = tokenize_chinese_chars
+        self.strip_accents = strip_accents
+
+    def tokenize(self, text, never_split=None):
+        """
+        Basic Tokenization of a piece of text. Split on "white spaces" only, for sub-word tokenization, see
+        WordPieceTokenizer.
+
+        Args:
+            **never_split**: (`optional`) list of str
+                Kept for backward compatibility purposes. Now implemented directly at the base class level (see
+                :func:`PreTrainedTokenizer.tokenize`) List of token not to split.
+        """
+        # union() returns a new set by concatenating the two sets.
+        never_split = self.never_split.union(set(never_split)) if never_split else self.never_split
+        text = self._clean_text(text)
+
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        if self.tokenize_chinese_chars:
+            text = self._tokenize_chinese_chars(text)
+        orig_tokens = whitespace_tokenize(text)
+        split_tokens = []
+        for token in orig_tokens:
+            if token not in never_split:
+                if self.do_lower_case:
+                    token = token.lower()
+                    if self.strip_accents is not False:
+                        token = self._run_strip_accents(token)
+                elif self.strip_accents:
+                    token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token, never_split))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text, never_split=None):
+        """Splits punctuation on a piece of text."""
+        if never_split is not None and text in never_split:
+            return [text]
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if (
+            (cp >= 0x4E00 and cp <= 0x9FFF)
+            or (cp >= 0x3400 and cp <= 0x4DBF)  #
+            or (cp >= 0x20000 and cp <= 0x2A6DF)  #
+            or (cp >= 0x2A700 and cp <= 0x2B73F)  #
+            or (cp >= 0x2B740 and cp <= 0x2B81F)  #
+            or (cp >= 0x2B820 and cp <= 0x2CEAF)  #
+            or (cp >= 0xF900 and cp <= 0xFAFF)
+            or (cp >= 0x2F800 and cp <= 0x2FA1F)  #
+        ):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xFFFD or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+class WordpieceTokenizer(object):
+    """Runs WordPiece tokenization."""
+
+    def __init__(self, vocab, unk_token, max_input_chars_per_word=100):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.max_input_chars_per_word = max_input_chars_per_word
+
+    def tokenize(self, text):
+        """
+        Tokenizes a piece of text into its word pieces. This uses a greedy longest-match-first algorithm to perform
+        tokenization using the given vocabulary.
+
+        For example, :obj:`input = "unaffable"` wil return as output :obj:`["un", "##aff", "##able"]`.
+
+        Args:
+          text: A single token or whitespace separated tokens. This should have
+            already been passed through `BasicTokenizer`.
+
+        Returns:
+          A list of wordpiece tokens.
+        """
+
+        output_tokens = []
+        for token in whitespace_tokenize(text):
+            chars = list(token)
+            if len(chars) > self.max_input_chars_per_word:
+                output_tokens.append(self.unk_token)
+                continue
+
+            is_bad = False
+            start = 0
+            sub_tokens = []
+            while start < len(chars):
+                end = len(chars)
+                cur_substr = None
+                while start < end:
+                    substr = "".join(chars[start:end])
+                    if start > 0:
+                        substr = "##" + substr
+                    if substr in self.vocab:
+                        cur_substr = substr
+                        break
+                    end -= 1
+                if cur_substr is None:
+                    is_bad = True
+                    break
+                sub_tokens.append(cur_substr)
+                start = end
+
+            if is_bad:
+                output_tokens.append(self.unk_token)
+            else:
+                output_tokens.extend(sub_tokens)
+        return output_tokens

fastNLP/transformers/torch/models/cpt/__init__.py

@@ -0,0 +1,12 @@
+__all__ = [
+    "CPT_PRETRAINED_MODEL_ARCHIVE_LIST",
+    "CPTForConditionalGeneration",
+    "CPTForSequenceClassification",
+    "CPTForMaskedLM",
+    "CPTForQuestionAnswering",
+    "CPTModel",
+    "CPTPretrainedModel",
+]
+
+from .modeling_cpt import CPT_PRETRAINED_MODEL_ARCHIVE_LIST, CPTForConditionalGeneration, CPTForSequenceClassification, \
+    CPTForMaskedLM, CPTForQuestionAnswering, CPTModel, CPTPretrainedModel

fastNLP/transformers/torch/models/gpt2/__init__.py

@@ -0,0 +1,19 @@
+__all__ = [
+    "GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP",
+    "GPT2Config",
+
+    "GPT2_PRETRAINED_MODEL_ARCHIVE_LIST",
+    "GPT2DoubleHeadsModel",
+    "GPT2ForSequenceClassification",
+    "GPT2ForTokenClassification",
+    "GPT2LMHeadModel",
+    "GPT2Model",
+    "GPT2PreTrainedModel",
+
+    "GPT2Tokenizer",
+]
+
+from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
+from .tokenization_gpt2 import GPT2Tokenizer
+from .modeling_gpt2 import GPT2_PRETRAINED_MODEL_ARCHIVE_LIST, GPT2DoubleHeadsModel, GPT2ForSequenceClassification, \
+    GPT2ForTokenClassification, GPT2LMHeadModel, GPT2Model, GPT2PreTrainedModel

fastNLP/transformers/torch/models/gpt2/configuration_gpt2.py

@@ -0,0 +1,184 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" OpenAI GPT-2 configuration """
+
+from fastNLP.transformers.torch.configuration_utils import PretrainedConfig
+
+__all__ = [
+    "GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP",
+    "GPT2Config",
+]
+
+GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "gpt2": "https://huggingface.co/gpt2/resolve/main/config.json",
+    "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/config.json",
+    "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/config.json",
+    "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/config.json",
+    "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/config.json",
+}
+
+
+class GPT2Config(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a :class:`~transformers.GPT2Model` or a
+    :class:`~transformers.TFGPT2Model`. It is used to instantiate a GPT-2 model according to the specified arguments,
+    defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the GPT-2 `small <https://huggingface.co/gpt2>`__ architecture.
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 50257):
+            Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.GPT2Model` or
+            :class:`~transformers.TFGPT2Model`.
+        n_positions (:obj:`int`, `optional`, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        n_ctx (:obj:`int`, `optional`, defaults to 1024):
+            Dimensionality of the causal mask (usually same as n_positions).
+        n_embd (:obj:`int`, `optional`, defaults to 768):
+            Dimensionality of the embeddings and hidden states.
+        n_layer (:obj:`int`, `optional`, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        n_head (:obj:`int`, `optional`, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        n_inner (:obj:`int`, `optional`, defaults to None):
+            Dimensionality of the inner feed-forward layers. :obj:`None` will set it to 4 times n_embd
+        activation_function (:obj:`str`, `optional`, defaults to :obj:`"gelu"`):
+            Activation function, to be selected in the list :obj:`["relu", "silu", "gelu", "tanh", "gelu_new"]`.
+        resid_pdrop (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        embd_pdrop (:obj:`int`, `optional`, defaults to 0.1):
+            The dropout ratio for the embeddings.
+        attn_pdrop (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout ratio for the attention.
+        layer_norm_epsilon (:obj:`float`, `optional`, defaults to 1e-5):
+            The epsilon to use in the layer normalization layers
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        summary_type (:obj:`string`, `optional`, defaults to :obj:`"cls_index"`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+
+            Has to be one of the following options:
+
+                - :obj:`"last"`: Take the last token hidden state (like XLNet).
+                - :obj:`"first"`: Take the first token hidden state (like BERT).
+                - :obj:`"mean"`: Take the mean of all tokens hidden states.
+                - :obj:`"cls_index"`: Supply a Tensor of classification token position (like GPT/GPT-2).
+                - :obj:`"attn"`: Not implemented now, use multi-head attention.
+        summary_use_proj (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+
+            Whether or not to add a projection after the vector extraction.
+        summary_activation (:obj:`str`, `optional`):
+            Argument used when doing sequence summary. Used in for the multiple choice head in
+            :class:`~transformers.GPT2DoubleHeadsModel`.
+
+            Pass :obj:`"tanh"` for a tanh activation to the output, any other value will result in no activation.
+        summary_proj_to_labels (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+
+            Whether the projection outputs should have :obj:`config.num_labels` or :obj:`config.hidden_size` classes.
+        summary_first_dropout (:obj:`float`, `optional`, defaults to 0.1):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+
+            The dropout ratio to be used after the projection and activation.
+        scale_attn_weights (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Scale attention weights by dividing by sqrt(hidden_size)..
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+
+    Example::
+
+        >>> from transformers import GPT2Model, GPT2Config
+
+        >>> # Initializing a GPT2 configuration
+        >>> configuration = GPT2Config()
+
+        >>> # Initializing a model from the configuration
+        >>> model = GPT2Model(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+
+    model_type = "gpt2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "hidden_size": "n_embd",
+        "max_position_embeddings": "n_positions",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size=50257,
+        n_positions=1024,
+        n_ctx=1024,
+        n_embd=768,
+        n_layer=12,
+        n_head=12,
+        n_inner=None,
+        activation_function="gelu_new",
+        resid_pdrop=0.1,
+        embd_pdrop=0.1,
+        attn_pdrop=0.1,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        summary_type="cls_index",
+        summary_use_proj=True,
+        summary_activation=None,
+        summary_proj_to_labels=True,
+        summary_first_dropout=0.1,
+        scale_attn_weights=True,
+        use_cache=True,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        **kwargs
+    ):
+        self.vocab_size = vocab_size
+        self.n_ctx = n_ctx
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_inner = n_inner
+        self.activation_function = activation_function
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_first_dropout = summary_first_dropout
+        self.summary_proj_to_labels = summary_proj_to_labels
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)

fastNLP/transformers/torch/models/gpt2/tokenization_gpt2.py

@@ -0,0 +1,308 @@
+# coding=utf-8
+# Copyright 2018 The Open AI Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for OpenAI GPT."""
+
+import json
+import os
+from functools import lru_cache
+from typing import TYPE_CHECKING, List, Optional, Tuple
+
+import regex as re
+
+from fastNLP.transformers.torch.tokenization_utils import AddedToken, PreTrainedTokenizer
+# if TYPE_CHECKING:
+#     from transformers.pipelines.conversational import Conversation
+from fastNLP.core.log import logger
+
+__all__ = [
+    "GPT2Tokenizer",
+]
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "gpt2": "https://huggingface.co/gpt2/resolve/main/vocab.json",
+        "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/vocab.json",
+        "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/vocab.json",
+        "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/vocab.json",
+        "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/vocab.json",
+    },
+    "merges_file": {
+        "gpt2": "https://huggingface.co/gpt2/resolve/main/merges.txt",
+        "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/merges.txt",
+        "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/merges.txt",
+        "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/merges.txt",
+        "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/merges.txt",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "gpt2": 1024,
+    "gpt2-medium": 1024,
+    "gpt2-large": 1024,
+    "gpt2-xl": 1024,
+    "distilgpt2": 1024,
+}
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
+    characters the bpe code barfs on.
+
+    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
+    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
+    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
+    tables between utf-8 bytes and unicode strings.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2 ** 8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2 ** 8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """
+    Return set of symbol pairs in a word.
+
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+class GPT2Tokenizer(PreTrainedTokenizer):
+    """
+    Construct a GPT-2 tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ::
+
+        >>> from transformers import GPT2Tokenizer
+        >>> tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+        >>> tokenizer("Hello world")['input_ids']
+        [15496, 995]
+        >>> tokenizer(" Hello world")['input_ids']
+        [18435, 995]
+
+    You can get around that behavior by passing ``add_prefix_space=True`` when instantiating this tokenizer or when you
+    call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.
+
+    .. note::
+
+        When used with ``is_split_into_words=True``, this tokenizer will add a space before each word (even the first
+        one).
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
+    Users should refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (:obj:`str`):
+            Path to the vocabulary file.
+        merges_file (:obj:`str`):
+            Path to the merges file.
+        errors (:obj:`str`, `optional`, defaults to :obj:`"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See `bytes.decode
+            <https://docs.python.org/3/library/stdtypes.html#bytes.decode>`__ for more information.
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The beginning of sequence token.
+        eos_token (:obj:`str`, `optional`, defaults to :obj:`<|endoftext|>`):
+            The end of sequence token.
+        add_prefix_space (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
+            other word. (GPT2 tokenizer detect beginning of words by the preceding space).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        unk_token="<|endoftext|>",
+        bos_token="<|endoftext|>",
+        eos_token="<|endoftext|>",
+        add_prefix_space=False,
+        **kwargs
+    ):
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        super().__init__(
+            errors=errors,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            add_prefix_space=add_prefix_space,
+            **kwargs,
+        )
+
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.errors = errors  # how to handle errors in decoding
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            bpe_merges = merges_handle.read().split("\n")[1:-1]
+        bpe_merges = [tuple(merge.split()) for merge in bpe_merges]
+        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
+        self.cache = {}
+        self.add_prefix_space = add_prefix_space
+
+        # Should have added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions
+        self.pat = re.compile(r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""")
+
+    @property
+    def vocab_size(self):
+        return len(self.encoder)
+
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                except ValueError:
+                    new_word.extend(word[i:])
+                    break
+                else:
+                    new_word.extend(word[i:j])
+                    i = j
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        bpe_tokens = []
+        for token in re.findall(self.pat, text):
+            token = "".join(
+                self.byte_encoder[b] for b in token.encode("utf-8")
+            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
+            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
+        return bpe_tokens
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        text = "".join(tokens)
+        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
+        return text
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            f.write(json.dumps(self.encoder, ensure_ascii=False))
+
+        index = 0
+        with open(merge_file, "w", encoding="utf-8") as writer:
+            writer.write("#version: 0.2\n")
+            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+    def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs):
+        add_prefix_space = kwargs.pop("add_prefix_space", self.add_prefix_space)
+        if is_split_into_words or add_prefix_space:
+            text = " " + text
+        return (text, kwargs)
+
+    # def _build_conversation_input_ids(self, conversation: "Conversation") -> List[int]:
+    #     input_ids = []
+    #     for is_user, text in conversation.iter_texts():
+    #         input_ids.extend(self.encode(text, add_special_tokens=False) + [self.eos_token_id])
+    #     if len(input_ids) > self.model_max_length:
+    #         input_ids = input_ids[-self.model_max_length :]
+    #     return input_ids

fastNLP/transformers/torch/models/roberta/__init__.py

@@ -0,0 +1,21 @@
+__all__ = [
+    "ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP",
+    "RobertaConfig",
+
+    "ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
+    "RobertaForCausalLM",
+    "RobertaForMaskedLM",
+    "RobertaForMultipleChoice",
+    "RobertaForQuestionAnswering",
+    "RobertaForSequenceClassification",
+    "RobertaForTokenClassification",
+    "RobertaModel",
+    "RobertaPreTrainedModel",
+
+    "RobertaTokenizer",
+]
+
+from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
+from .tokenization_roberta import RobertaTokenizer
+from .modeling_roberta import ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST, RobertaForCausalLM, RobertaForMaskedLM, RobertaForMultipleChoice, \
+    RobertaForQuestionAnswering, RobertaForSequenceClassification, RobertaForTokenClassification, RobertaModel, RobertaPreTrainedModel

fastNLP/transformers/torch/models/roberta/configuration_roberta.py

@@ -0,0 +1,65 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" RoBERTa configuration """
+from ..bert.configuration_bert import BertConfig
+from fastNLP.core.log import logger
+
+__all__ = [
+    "ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP",
+    "RobertaConfig",
+]
+
+ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "roberta-base": "https://huggingface.co/roberta-base/resolve/main/config.json",
+    "roberta-large": "https://huggingface.co/roberta-large/resolve/main/config.json",
+    "roberta-large-mnli": "https://huggingface.co/roberta-large-mnli/resolve/main/config.json",
+    "distilroberta-base": "https://huggingface.co/distilroberta-base/resolve/main/config.json",
+    "roberta-base-openai-detector": "https://huggingface.co/roberta-base-openai-detector/resolve/main/config.json",
+    "roberta-large-openai-detector": "https://huggingface.co/roberta-large-openai-detector/resolve/main/config.json",
+}
+
+
+class RobertaConfig(BertConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.RobertaModel` or a
+    :class:`~transformers.TFRobertaModel`. It is used to instantiate a RoBERTa model according to the specified
+    arguments, defining the model architecture.
+
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+    The :class:`~transformers.RobertaConfig` class directly inherits :class:`~transformers.BertConfig`. It reuses the
+    same defaults. Please check the parent class for more information.
+
+    Examples::
+
+        >>> from transformers import RobertaConfig, RobertaModel
+
+        >>> # Initializing a RoBERTa configuration
+        >>> configuration = RobertaConfig()
+
+        >>> # Initializing a model from the configuration
+        >>> model = RobertaModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+    model_type = "roberta"
+
+    def __init__(self, pad_token_id=1, bos_token_id=0, eos_token_id=2, **kwargs):
+        """Constructs RobertaConfig."""
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)

fastNLP/transformers/torch/models/roberta/tokenization_roberta.py

@@ -0,0 +1,254 @@
+# coding=utf-8
+# Copyright 2018 The Open AI Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for RoBERTa."""
+
+from typing import List, Optional
+
+from fastNLP.transformers.torch.tokenization_utils import AddedToken
+from ..gpt2.tokenization_gpt2 import GPT2Tokenizer
+from fastNLP.core.log import logger
+
+__all__ = [
+    "RobertaTokenizer",
+]
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "roberta-base": "https://huggingface.co/roberta-base/resolve/main/vocab.json",
+        "roberta-large": "https://huggingface.co/roberta-large/resolve/main/vocab.json",
+        "roberta-large-mnli": "https://huggingface.co/roberta-large-mnli/resolve/main/vocab.json",
+        "distilroberta-base": "https://huggingface.co/distilroberta-base/resolve/main/vocab.json",
+        "roberta-base-openai-detector": "https://huggingface.co/roberta-base-openai-detector/resolve/main/vocab.json",
+        "roberta-large-openai-detector": "https://huggingface.co/roberta-large-openai-detector/resolve/main/vocab.json",
+    },
+    "merges_file": {
+        "roberta-base": "https://huggingface.co/roberta-base/resolve/main/merges.txt",
+        "roberta-large": "https://huggingface.co/roberta-large/resolve/main/merges.txt",
+        "roberta-large-mnli": "https://huggingface.co/roberta-large-mnli/resolve/main/merges.txt",
+        "distilroberta-base": "https://huggingface.co/distilroberta-base/resolve/main/merges.txt",
+        "roberta-base-openai-detector": "https://huggingface.co/roberta-base-openai-detector/resolve/main/merges.txt",
+        "roberta-large-openai-detector": "https://huggingface.co/roberta-large-openai-detector/resolve/main/merges.txt",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "roberta-base": 512,
+    "roberta-large": 512,
+    "roberta-large-mnli": 512,
+    "distilroberta-base": 512,
+    "roberta-base-openai-detector": 512,
+    "roberta-large-openai-detector": 512,
+}
+
+
+class RobertaTokenizer(GPT2Tokenizer):
+    """
+    Constructs a RoBERTa tokenizer, derived from the GPT-2 tokenizer, using byte-level Byte-Pair-Encoding.
+
+    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ::
+
+        >>> from transformers import RobertaTokenizer
+        >>> tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
+        >>> tokenizer("Hello world")['input_ids']
+        [0, 31414, 232, 328, 2]
+        >>> tokenizer(" Hello world")['input_ids']
+        [0, 20920, 232, 2]
+
+    You can get around that behavior by passing ``add_prefix_space=True`` when instantiating this tokenizer or when you
+    call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.
+
+    .. note::
+
+        When used with ``is_split_into_words=True``, this tokenizer will add a space before each word (even the first
+        one).
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
+    Users should refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (:obj:`str`):
+            Path to the vocabulary file.
+        merges_file (:obj:`str`):
+            Path to the merges file.
+        errors (:obj:`str`, `optional`, defaults to :obj:`"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See `bytes.decode
+            <https://docs.python.org/3/library/stdtypes.html#bytes.decode>`__ for more information.
+        bos_token (:obj:`str`, `optional`, defaults to :obj:`"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            .. note::
+
+                When building a sequence using special tokens, this is not the token that is used for the beginning of
+                sequence. The token used is the :obj:`cls_token`.
+        eos_token (:obj:`str`, `optional`, defaults to :obj:`"</s>"`):
+            The end of sequence token.
+
+            .. note::
+
+                When building a sequence using special tokens, this is not the token that is used for the end of
+                sequence. The token used is the :obj:`sep_token`.
+        sep_token (:obj:`str`, `optional`, defaults to :obj:`"</s>"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        cls_token (:obj:`str`, `optional`, defaults to :obj:`"<s>"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        pad_token (:obj:`str`, `optional`, defaults to :obj:`"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        mask_token (:obj:`str`, `optional`, defaults to :obj:`"<mask>"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        add_prefix_space (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
+            other word. (RoBERTa tokenizer detect beginning of words by the preceding space).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        bos_token="<s>",
+        eos_token="</s>",
+        sep_token="</s>",
+        cls_token="<s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        mask_token="<mask>",
+        add_prefix_space=False,
+        **kwargs
+    ):
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token
+        cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+
+        # Mask token behave like a normal word, i.e. include the space before it
+        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+
+        super().__init__(
+            vocab_file=vocab_file,
+            merges_file=merges_file,
+            errors=errors,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            cls_token=cls_token,
+            pad_token=pad_token,
+            mask_token=mask_token,
+            add_prefix_space=add_prefix_space,
+            **kwargs,
+        )
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A RoBERTa sequence has the following format:
+
+        - single sequence: ``<s> X </s>``
+        - pair of sequences: ``<s> A </s></s> B </s>``
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` method.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            :obj:`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. RoBERTa does not
+        make use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of zeros.
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
+
+    def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs):
+        add_prefix_space = kwargs.pop("add_prefix_space", self.add_prefix_space)
+        if (is_split_into_words or add_prefix_space) and (len(text) > 0 and not text[0].isspace()):
+            text = " " + text
+        return (text, kwargs)

fastNLP/transformers/torch/tokenization_utils.py

@@ -0,0 +1,915 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+ Tokenization classes for python tokenizers. For fast tokenizers (provided by HuggingFace's tokenizers library) see
+ tokenization_utils_fast.py
+"""
+import bisect
+import itertools
+import re
+import unicodedata
+from collections import OrderedDict
+from typing import Any, Dict, List, Optional, Tuple, Union, overload
+
+from .file_utils import PaddingStrategy, TensorType, add_end_docstrings
+from .tokenization_utils_base import (
+    ENCODE_KWARGS_DOCSTRING,
+    ENCODE_PLUS_ADDITIONAL_KWARGS_DOCSTRING,
+    INIT_TOKENIZER_DOCSTRING,
+    AddedToken,
+    BatchEncoding,
+    EncodedInput,
+    EncodedInputPair,
+    PreTokenizedInput,
+    PreTokenizedInputPair,
+    PreTrainedTokenizerBase,
+    TextInput,
+    TextInputPair,
+    TruncationStrategy,
+)
+
+from fastNLP.core.log import logger
+
+# Slow tokenizers are saved in a vocabulary plus three separated files
+SPECIAL_TOKENS_MAP_FILE = "special_tokens_map.json"
+ADDED_TOKENS_FILE = "added_tokens.json"
+TOKENIZER_CONFIG_FILE = "tokenizer_config.json"
+
+
+class Trie:
+    """
+    Trie in Python. Creates a Trie out of a list of words. The trie is used to split on `added_tokens` in one pass
+    Loose reference https://en.wikipedia.org/wiki/Trie
+    """
+
+    def __init__(self):
+        self.data = {}
+
+    def add(self, word: str):
+        """
+        Passes over every char (utf-8 char) on word and recursively adds it to the internal `data` trie representation.
+        The special key `""` is used to represent termination.
+
+        This function is idempotent, adding twice the same word will leave the trie unchanged
+
+        Example::
+
+            >>> trie = Trie()
+            >>> trie.add("Hello 友達")
+            >>> trie.data
+            {"H": {"e": {"l": {"l": {"o": {" ": {"友": {"達": {"": 1}}}}}}}}}
+            >>> trie.add("Hello")
+            >>> trie.data
+            {"H": {"e": {"l": {"l": {"o": {"": 1, " ": {"友": {"達": {"": 1}}}}}}}}}
+        """
+        if not word:
+            # Prevent empty string
+            return
+        ref = self.data
+        for char in word:
+            ref[char] = char in ref and ref[char] or {}
+            ref = ref[char]
+        ref[""] = 1
+
+    def split(self, text: str) -> List[str]:
+        """
+        Will look for the words added to the trie within `text`. Output is the original string splitted along the
+        boundaries of the words found.
+
+        This trie will match the longest possible word first !
+
+        Example::
+
+            >>> trie = Trie()
+            >>> trie.split("[CLS] This is a extra_id_100")
+            ["[CLS] This is a extra_id_100"]
+            >>> trie.add("[CLS]")
+            >>> trie.add("extra_id_1")
+            >>> trie.add("extra_id_100")
+            >>> trie.split("[CLS] This is a extra_id_100")
+            ["[CLS]", " This is a ", "extra_id_100"]
+        """
+        # indexes are counted left of the chars index.
+        # "hello", index 0, is left of h, index 1 is between h and e.
+        # index 5 is right of the "o".
+
+        # States are going to capture every possible start (indexes as above)
+        # as keys, and have as values, a pointer to the position in the trie
+        # where we're at. This is a partial match for now.
+        # This enables to keep track of multiple matches while we're iterating
+        # the string
+        # If the trie contains, "blowing", and "lower" and we encounter the
+        # string "blower", we need to split into ["b", "lower"].
+        # This is where we need to keep track of multiple possible starts.
+        states = OrderedDict()
+
+        # This will contain every indices where we need
+        # to cut.
+        # We force to cut at offset 0 and len(text) (added later)
+        offsets = [0]
+
+        # This is used by the lookahead which needs to skip over
+        # some text where the full match exceeded the place in the initial
+        # for loop
+        skip = None
+        # Main loop, Giving this algorithm O(n) complexity
+        for current, current_char in enumerate(text):
+            if skip and current < skip:
+                # Prevents the lookahead for matching twice
+                # like extra_id_100 and id_100
+                continue
+
+            # This will track every state
+            # that stop matching, we need to stop tracking them.
+            # If we look at "lowball", we're going to match "l" (add it to states), "o", "w", then
+            # fail on "b", we need to remove 0 from the valid states.
+            to_remove = set()
+            # Whenever we found a match, we need to drop everything
+            # this is a greedy algorithm, it will match on the first found token
+            reset = False
+
+            # In this case, we already have partial matches (But unfinished)
+            for start, trie_pointer in states.items():
+                if "" in trie_pointer:
+                    # This is a final match, we need to reset and
+                    # store the results in `offsets`.
+
+                    # Lookahead to match longest first
+                    # Important in case of extra_id_1 vs extra_id_100
+                    lookahead_index = current
+                    end = current
+                    next_char = text[lookahead_index] if lookahead_index < len(text) else None
+                    while next_char in trie_pointer:
+                        trie_pointer = trie_pointer[next_char]
+                        lookahead_index += 1
+                        if "" in trie_pointer:
+                            end = lookahead_index
+                            skip = lookahead_index
+
+                        if lookahead_index == len(text):
+                            # End of string
+                            break
+                        next_char = text[lookahead_index]
+                    # End lookahead
+
+                    # Storing and resetting
+                    offsets.append(start)
+                    offsets.append(end)
+                    reset = True
+                elif current_char in trie_pointer:
+                    # The current character being looked at has a match within the trie
+                    # update the pointer (it will be stored back into states later).
+                    trie_pointer = trie_pointer[current_char]
+
+                    # Storing back the new pointer into the states.
+                    # Partial matches got longer by one.
+                    states[start] = trie_pointer
+                else:
+                    # The new character has not match in the trie, we need
+                    # to stop keeping track of this partial match.
+                    # We can't do it directly within the loop because of how
+                    # python iteration works
+                    to_remove.add(start)
+
+            # Either clearing the full start (we found a real match)
+            # Or clearing only the partial matches that didn't work.
+            if reset:
+                states = {}
+            else:
+                for start in to_remove:
+                    del states[start]
+
+            # If this character is a starting character within the trie
+            # start keeping track of this partial match.
+            if current_char in self.data:
+                states[current] = self.data[current_char]
+
+        # We have a cut at the end with states.
+        for start, trie_pointer in states.items():
+            if "" in trie_pointer:
+                # This is a final match, we need to reset and
+                # store the results in `offsets`.
+                end = len(text)
+                offsets.append(start)
+                offsets.append(end)
+                # Longest cut is always the one with lower start so the first
+                # item so we need to break.
+                break
+
+        # We have all the offsets now, we just need to do the actual splitting.
+        # We need to eventually add the first part of the string and the eventual
+        # last part.
+        offsets.append(len(text))
+        tokens = []
+        start = 0
+        for end in offsets:
+            if start == end:
+                # This might happen if there's a match at index 0
+                # we're also preventing zero-width cuts in case of two
+                # consecutive matches
+                continue
+            tokens.append(text[start:end])
+            start = end
+
+        return tokens
+
+
+def _is_whitespace(char):
+    """Checks whether `char` is a whitespace character."""
+    # \t, \n, and \r are technically control characters but we treat them
+    # as whitespace since they are generally considered as such.
+    if char == " " or char == "\t" or char == "\n" or char == "\r":
+        return True
+    cat = unicodedata.category(char)
+    if cat == "Zs":
+        return True
+    return False
+
+
+def _is_control(char):
+    """Checks whether `char` is a control character."""
+    # These are technically control characters but we count them as whitespace
+    # characters.
+    if char == "\t" or char == "\n" or char == "\r":
+        return False
+    cat = unicodedata.category(char)
+    if cat.startswith("C"):
+        return True
+    return False
+
+
+def _is_punctuation(char):
+    """Checks whether `char` is a punctuation character."""
+    cp = ord(char)
+    # We treat all non-letter/number ASCII as punctuation.
+    # Characters such as "^", "$", and "`" are not in the Unicode
+    # Punctuation class but we treat them as punctuation anyways, for
+    # consistency.
+    if (cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126):
+        return True
+    cat = unicodedata.category(char)
+    if cat.startswith("P"):
+        return True
+    return False
+
+
+def _is_end_of_word(text):
+    """Checks whether the last character in text is one of a punctuation, control or whitespace character."""
+    last_char = text[-1]
+    return bool(_is_control(last_char) | _is_punctuation(last_char) | _is_whitespace(last_char))
+
+
+def _is_start_of_word(text):
+    """Checks whether the first character in text is one of a punctuation, control or whitespace character."""
+    first_char = text[0]
+    return bool(_is_control(first_char) | _is_punctuation(first_char) | _is_whitespace(first_char))
+
+
+def _insert_one_token_to_ordered_list(token_list: List[str], new_token: str):
+    """
+    Inserts one token to an ordered list if it does not already exist. Note: token_list must be sorted.
+    """
+    insertion_idx = bisect.bisect_left(token_list, new_token)
+    # Checks if new_token is already in the ordered token_list
+    if insertion_idx < len(token_list) and token_list[insertion_idx] == new_token:
+        # new_token is in token_list, don't add
+        return
+    else:
+        token_list.insert(insertion_idx, new_token)
+
+
+@add_end_docstrings(INIT_TOKENIZER_DOCSTRING)
+class PreTrainedTokenizer(PreTrainedTokenizerBase):
+    """
+    Base class for all slow tokenizers.
+
+    Inherits from :class:`~transformers.tokenization_utils_base.PreTrainedTokenizerBase`.
+
+    Handle all the shared methods for tokenization and special tokens as well as methods downloading/caching/loading
+    pretrained tokenizers as well as adding tokens to the vocabulary.
+
+    This class also contain the added tokens in a unified way on top of all tokenizers so we don't have to handle the
+    specific vocabulary augmentation methods of the various underlying dictionary structures (BPE, sentencepiece...).
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        # Added tokens - We store this for both slow and fast tokenizers
+        # until the serialization of Fast tokenizers is updated
+        self.added_tokens_encoder: Dict[str, int] = {}
+        self.added_tokens_decoder: Dict[int, str] = {}
+        self.unique_no_split_tokens: List[str] = []
+        self.tokens_trie = Trie()
+
+        self._decode_use_source_tokenizer = False
+
+    @property
+    def is_fast(self) -> bool:
+        return False
+
+    @property
+    def vocab_size(self) -> int:
+        """
+        :obj:`int`: Size of the base vocabulary (without the added tokens).
+        """
+        raise NotImplementedError
+
+    def get_added_vocab(self) -> Dict[str, int]:
+        """
+        Returns the added tokens in the vocabulary as a dictionary of token to index.
+
+        Returns:
+            :obj:`Dict[str, int]`: The added tokens.
+        """
+        return self.added_tokens_encoder
+
+    def __len__(self):
+        """
+        Size of the full vocabulary with the added tokens.
+        """
+        return self.vocab_size + len(self.added_tokens_encoder)
+
+    def _add_tokens(self, new_tokens: Union[List[str], List[AddedToken]], special_tokens: bool = False) -> int:
+        """
+        Add a list of new tokens to the tokenizer class. If the new tokens are not in the vocabulary, they are added to
+        it with indices starting from length of the current vocabulary.
+
+        Args:
+            new_tokens (:obj:`List[str]`or :obj:`List[tokenizers.AddedToken]`):
+                Token(s) to add in vocabulary. A token is only added if it's not already in the vocabulary (tested by
+                checking if the tokenizer assign the index of the ``unk_token`` to them).
+            special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the tokens should be added as special tokens.
+
+        Returns:
+            :obj:`int`: The number of tokens actually added to the vocabulary.
+
+        Examples::
+
+            # Let's see how to increase the vocabulary of Bert model and tokenizer
+            tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+            model = BertModel.from_pretrained('bert-base-uncased')
+
+            num_added_toks = tokenizer.add_tokens(['new_tok1', 'my_new-tok2'])
+            print('We have added', num_added_toks, 'tokens')
+            # Note: resize_token_embeddings expects to receive the full size of the new vocabulary, i.e. the length of the tokenizer.
+            model.resize_token_embeddings(len(tokenizer))
+        """
+        new_tokens = [str(tok) for tok in new_tokens]
+
+        tokens_to_add = []
+        for token in new_tokens:
+            if not isinstance(token, str):
+                raise TypeError(f"Token {token} is not a string but a {type(token)}.")
+            if not special_tokens and hasattr(self, "do_lower_case") and self.do_lower_case:
+                token = token.lower()
+            if (
+                token != self.unk_token
+                and self.convert_tokens_to_ids(token) == self.convert_tokens_to_ids(self.unk_token)
+                and token not in tokens_to_add
+            ):
+                tokens_to_add.append(token)
+                if self.verbose:
+                    logger.info(f"Adding {token} to the vocabulary")
+
+        added_tok_encoder = dict((tok, len(self) + i) for i, tok in enumerate(tokens_to_add))
+        added_tok_decoder = {v: k for k, v in added_tok_encoder.items()}
+        self.added_tokens_encoder.update(added_tok_encoder)
+        self.added_tokens_decoder.update(added_tok_decoder)
+
+        # Make sure we don't split on any special tokens (even they were already in the vocab before e.g. for Albert)
+        if special_tokens:
+            if len(new_tokens) == 1:
+                _insert_one_token_to_ordered_list(self.unique_no_split_tokens, new_tokens[0])
+            else:
+                self.unique_no_split_tokens = sorted(set(self.unique_no_split_tokens).union(set(new_tokens)))
+        else:
+            # Or on the newly added tokens
+            if len(tokens_to_add) == 1:
+                _insert_one_token_to_ordered_list(self.unique_no_split_tokens, tokens_to_add[0])
+            else:
+                self.unique_no_split_tokens = sorted(set(self.unique_no_split_tokens).union(set(tokens_to_add)))
+        self._create_trie(self.unique_no_split_tokens)
+
+        return len(tokens_to_add)
+
+    def _create_trie(self, unique_no_split_tokens):
+        trie = Trie()
+        for token in unique_no_split_tokens:
+            if hasattr(self, "do_lower_case") and self.do_lower_case and token not in self.all_special_tokens:
+                trie.add(token.lower())
+            else:
+                trie.add(token)
+        self.tokens_trie = trie
+
+    def num_special_tokens_to_add(self, pair: bool = False) -> int:
+        """
+        Returns the number of added tokens when encoding a sequence with special tokens.
+
+        .. note::
+            This encodes a dummy input and checks the number of added tokens, and is therefore not efficient. Do not
+            put this inside your training loop.
+
+        Args:
+            pair (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether the number of added tokens should be computed in the case of a sequence pair or a single
+                sequence.
+
+        Returns:
+            :obj:`int`: Number of special tokens added to sequences.
+        """
+        token_ids_0 = []
+        token_ids_1 = []
+        return len(self.build_inputs_with_special_tokens(token_ids_0, token_ids_1 if pair else None))
+
+    def tokenize(self, text: TextInput, **kwargs) -> List[str]:
+        """
+        Converts a string in a sequence of tokens, using the tokenizer.
+
+        Split in words for word-based vocabulary or sub-words for sub-word-based vocabularies
+        (BPE/SentencePieces/WordPieces). Takes care of added tokens.
+
+        Args:
+            text (:obj:`str`):
+                The sequence to be encoded.
+            **kwargs (additional keyword arguments):
+                Passed along to the model-specific ``prepare_for_tokenization`` preprocessing method.
+
+        Returns:
+            :obj:`List[str]`: The list of tokens.
+        """
+        # Simple mapping string => AddedToken for special tokens with specific tokenization behaviors
+        all_special_tokens_extended = dict(
+            (str(t), t) for t in self.all_special_tokens_extended if isinstance(t, AddedToken)
+        )
+
+        text, kwargs = self.prepare_for_tokenization(text, **kwargs)
+
+        if kwargs:
+            logger.warning(f"Keyword arguments {kwargs} not recognized.")
+
+        # TODO: should this be in the base class?
+        if hasattr(self, "do_lower_case") and self.do_lower_case:
+            # convert non-special tokens to lowercase
+            escaped_special_toks = [
+                re.escape(s_tok) for s_tok in (self.unique_no_split_tokens + self.all_special_tokens)
+            ]
+            pattern = r"(" + r"|".join(escaped_special_toks) + r")|" + r"(.+?)"
+            text = re.sub(pattern, lambda m: m.groups()[0] or m.groups()[1].lower(), text)
+
+        no_split_token = set(self.unique_no_split_tokens)
+        tokens = self.tokens_trie.split(text)
+        # ["This is something", "<special_token_1>", "  else"]
+        for i, token in enumerate(tokens):
+            if token in no_split_token:
+                tok_extended = all_special_tokens_extended.get(token, None)
+                left = tokens[i - 1] if i > 0 else None
+                right = tokens[i + 1] if i < len(tokens) - 1 else None
+                if isinstance(tok_extended, AddedToken):
+                    if tok_extended.rstrip and right:
+                        # A bit counter-intuitive but we strip the left of the string
+                        # since tok_extended.rstrip means the special token is eating all white spaces on its right
+                        tokens[i + 1] = right.lstrip()
+                    # Strip white spaces on the left
+                    if tok_extended.lstrip and left:
+                        tokens[i - 1] = left.rstrip()  # Opposite here
+                else:
+                    # We strip left and right by default
+                    if right:
+                        tokens[i + 1] = right.lstrip()
+                    if left:
+                        tokens[i - 1] = left.rstrip()
+        # ["This is something", "<special_token_1>", "else"]
+        tokenized_text = []
+        for token in tokens:
+            # Need to skip eventual empty (fully stripped) tokens
+            if not token:
+                continue
+            if token in no_split_token:
+                tokenized_text.append(token)
+            else:
+                tokenized_text.extend(self._tokenize(token))
+        # ["This", " is", " something", "<special_token_1>", "else"]
+        return tokenized_text
+
+    def _tokenize(self, text, **kwargs):
+        """
+        Converts a string in a sequence of tokens (string), using the tokenizer. Split in words for word-based
+        vocabulary or sub-words for sub-word-based vocabularies (BPE/SentencePieces/WordPieces).
+
+        Do NOT take care of added tokens.
+        """
+        raise NotImplementedError
+
+    def convert_tokens_to_ids(self, tokens: Union[str, List[str]]) -> Union[int, List[int]]:
+        """
+        Converts a token string (or a sequence of tokens) in a single integer id (or a sequence of ids), using the
+        vocabulary.
+
+        Args:
+            tokens (:obj:`str` or :obj:`List[str]`): One or several token(s) to convert to token id(s).
+
+        Returns:
+            :obj:`int` or :obj:`List[int]`: The token id or list of token ids.
+        """
+        if tokens is None:
+            return None
+
+        if isinstance(tokens, str):
+            return self._convert_token_to_id_with_added_voc(tokens)
+
+        ids = []
+        for token in tokens:
+            ids.append(self._convert_token_to_id_with_added_voc(token))
+        return ids
+
+    def _convert_token_to_id_with_added_voc(self, token):
+        if token is None:
+            return None
+
+        if token in self.added_tokens_encoder:
+            return self.added_tokens_encoder[token]
+        return self._convert_token_to_id(token)
+
+    def _convert_token_to_id(self, token):
+        raise NotImplementedError
+
+    def _encode_plus(
+        self,
+        text: Union[TextInput, PreTokenizedInput, EncodedInput],
+        text_pair: Optional[Union[TextInput, PreTokenizedInput, EncodedInput]] = None,
+        add_special_tokens: bool = True,
+        padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+        truncation_strategy: TruncationStrategy = TruncationStrategy.DO_NOT_TRUNCATE,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        is_split_into_words: bool = False,
+        pad_to_multiple_of: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        return_token_type_ids: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        **kwargs
+    ) -> BatchEncoding:
+        def get_input_ids(text):
+            if isinstance(text, str):
+                tokens = self.tokenize(text, **kwargs)
+                return self.convert_tokens_to_ids(tokens)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], str):
+                if is_split_into_words:
+                    tokens = list(
+                        itertools.chain(*(self.tokenize(t, is_split_into_words=True, **kwargs) for t in text))
+                    )
+                    return self.convert_tokens_to_ids(tokens)
+                else:
+                    return self.convert_tokens_to_ids(text)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], int):
+                return text
+            else:
+                if is_split_into_words:
+                    raise ValueError(
+                        f"Input {text} is not valid. Should be a string or a list/tuple of strings when `is_split_into_words=True`."
+                    )
+                else:
+                    raise ValueError(
+                        f"Input {text} is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
+                    )
+
+        if return_offsets_mapping:
+            raise NotImplementedError(
+                "return_offset_mapping is not available when using Python tokenizers."
+                "To use this feature, change your tokenizer to one deriving from "
+                "transformers.PreTrainedTokenizerFast."
+                "More information on available tokenizers at "
+                "https://github.com/huggingface/transformers/pull/2674"
+            )
+
+        first_ids = get_input_ids(text)
+        second_ids = get_input_ids(text_pair) if text_pair is not None else None
+
+        return self.prepare_for_model(
+            first_ids,
+            pair_ids=second_ids,
+            add_special_tokens=add_special_tokens,
+            padding=padding_strategy.value,
+            truncation=truncation_strategy.value,
+            max_length=max_length,
+            stride=stride,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_tensors=return_tensors,
+            prepend_batch_axis=True,
+            return_attention_mask=return_attention_mask,
+            return_token_type_ids=return_token_type_ids,
+            return_overflowing_tokens=return_overflowing_tokens,
+            return_special_tokens_mask=return_special_tokens_mask,
+            return_length=return_length,
+            verbose=verbose,
+        )
+
+    def _batch_encode_plus(
+        self,
+        batch_text_or_text_pairs: Union[
+            List[TextInput],
+            List[TextInputPair],
+            List[PreTokenizedInput],
+            List[PreTokenizedInputPair],
+            List[EncodedInput],
+            List[EncodedInputPair],
+        ],
+        add_special_tokens: bool = True,
+        padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+        truncation_strategy: TruncationStrategy = TruncationStrategy.DO_NOT_TRUNCATE,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        is_split_into_words: bool = False,
+        pad_to_multiple_of: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        return_token_type_ids: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        **kwargs
+    ) -> BatchEncoding:
+        def get_input_ids(text):
+            if isinstance(text, str):
+                tokens = self.tokenize(text, **kwargs)
+                return self.convert_tokens_to_ids(tokens)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], str):
+                if is_split_into_words:
+                    tokens = list(
+                        itertools.chain(*(self.tokenize(t, is_split_into_words=True, **kwargs) for t in text))
+                    )
+                    return self.convert_tokens_to_ids(tokens)
+                else:
+                    return self.convert_tokens_to_ids(text)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], int):
+                return text
+            else:
+                raise ValueError(
+                    "Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
+                )
+
+        if return_offsets_mapping:
+            raise NotImplementedError(
+                "return_offset_mapping is not available when using Python tokenizers."
+                "To use this feature, change your tokenizer to one deriving from "
+                "transformers.PreTrainedTokenizerFast."
+            )
+
+        input_ids = []
+        for ids_or_pair_ids in batch_text_or_text_pairs:
+            if not isinstance(ids_or_pair_ids, (list, tuple)):
+                ids, pair_ids = ids_or_pair_ids, None
+            elif is_split_into_words and not isinstance(ids_or_pair_ids[0], (list, tuple)):
+                ids, pair_ids = ids_or_pair_ids, None
+            else:
+                ids, pair_ids = ids_or_pair_ids
+
+            first_ids = get_input_ids(ids)
+            second_ids = get_input_ids(pair_ids) if pair_ids is not None else None
+            input_ids.append((first_ids, second_ids))
+
+        batch_outputs = self._batch_prepare_for_model(
+            input_ids,
+            add_special_tokens=add_special_tokens,
+            padding_strategy=padding_strategy,
+            truncation_strategy=truncation_strategy,
+            max_length=max_length,
+            stride=stride,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_attention_mask=return_attention_mask,
+            return_token_type_ids=return_token_type_ids,
+            return_overflowing_tokens=return_overflowing_tokens,
+            return_special_tokens_mask=return_special_tokens_mask,
+            return_length=return_length,
+            return_tensors=return_tensors,
+            verbose=verbose,
+        )
+
+        return BatchEncoding(batch_outputs)
+
+    @add_end_docstrings(ENCODE_KWARGS_DOCSTRING, ENCODE_PLUS_ADDITIONAL_KWARGS_DOCSTRING)
+    def _batch_prepare_for_model(
+        self,
+        batch_ids_pairs: List[Union[PreTokenizedInputPair, Tuple[List[int], None]]],
+        add_special_tokens: bool = True,
+        padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+        truncation_strategy: TruncationStrategy = TruncationStrategy.DO_NOT_TRUNCATE,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        pad_to_multiple_of: Optional[int] = None,
+        return_tensors: Optional[str] = None,
+        return_token_type_ids: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+    ) -> BatchEncoding:
+        """
+        Prepares a sequence of input id, or a pair of sequences of inputs ids so that it can be used by the model. It
+        adds special tokens, truncates sequences if overflowing while taking into account the special tokens and
+        manages a moving window (with user defined stride) for overflowing tokens
+
+        Args:
+            batch_ids_pairs: list of tokenized input ids or input ids pairs
+        """
+
+        batch_outputs = {}
+        for first_ids, second_ids in batch_ids_pairs:
+            outputs = self.prepare_for_model(
+                first_ids,
+                second_ids,
+                add_special_tokens=add_special_tokens,
+                padding=PaddingStrategy.DO_NOT_PAD.value,  # we pad in batch afterward
+                truncation=truncation_strategy.value,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=None,  # we pad in batch afterward
+                return_attention_mask=False,  # we pad in batch afterward
+                return_token_type_ids=return_token_type_ids,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_length=return_length,
+                return_tensors=None,  # We convert the whole batch to tensors at the end
+                prepend_batch_axis=False,
+                verbose=verbose,
+            )
+
+            for key, value in outputs.items():
+                if key not in batch_outputs:
+                    batch_outputs[key] = []
+                batch_outputs[key].append(value)
+
+        batch_outputs = self.pad(
+            batch_outputs,
+            padding=padding_strategy.value,
+            max_length=max_length,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_attention_mask=return_attention_mask,
+        )
+
+        batch_outputs = BatchEncoding(batch_outputs, tensor_type=return_tensors)
+
+        return batch_outputs
+
+    def prepare_for_tokenization(
+        self, text: str, is_split_into_words: bool = False, **kwargs
+    ) -> Tuple[str, Dict[str, Any]]:
+        """
+        Performs any necessary transformations before tokenization.
+
+        This method should pop the arguments from kwargs and return the remaining :obj:`kwargs` as well. We test the
+        :obj:`kwargs` at the end of the encoding process to be sure all the arguments have been used.
+
+        Args:
+            text (:obj:`str`):
+                The text to prepare.
+            is_split_into_words (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the input is already pre-tokenized (e.g., split into words). If set to :obj:`True`, the
+                tokenizer assumes the input is already split into words (for instance, by splitting it on whitespace)
+                which it will tokenize. This is useful for NER or token classification.
+            kwargs:
+                Keyword arguments to use for the tokenization.
+
+        Returns:
+            :obj:`Tuple[str, Dict[str, Any]]`: The prepared text and the unused kwargs.
+        """
+        return (text, kwargs)
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List, token_ids_1: Optional[List] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` or ``encode_plus`` methods.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of ids of the first sequence.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                List of ids of the second sequence.
+            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            if token_ids_1 is not None:
+                raise ValueError(
+                    "You should not supply a second sequence if the provided sequence of "
+                    "ids is already formatted with special tokens for the model."
+                )
+
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+        return [0] * ((len(token_ids_1) if token_ids_1 else 0) + len(token_ids_0))
+
+    @overload
+    def convert_ids_to_tokens(self, ids: int, skip_special_tokens: bool = False) -> str:
+        ...
+
+    @overload
+    def convert_ids_to_tokens(self, ids: List[int], skip_special_tokens: bool = False) -> List[str]:
+        ...
+
+    def convert_ids_to_tokens(
+        self, ids: Union[int, List[int]], skip_special_tokens: bool = False
+    ) -> Union[str, List[str]]:
+        """
+        Converts a single index or a sequence of indices in a token or a sequence of tokens, using the vocabulary and
+        added tokens.
+
+        Args:
+            ids (:obj:`int` or :obj:`List[int]`):
+                The token id (or token ids) to convert to tokens.
+            skip_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to remove special tokens in the decoding.
+
+        Returns:
+            :obj:`str` or :obj:`List[str]`: The decoded token(s).
+        """
+        if isinstance(ids, int):
+            if ids in self.added_tokens_decoder:
+                return self.added_tokens_decoder[ids]
+            else:
+                return self._convert_id_to_token(ids)
+        tokens = []
+        for index in ids:
+            index = int(index)
+            if skip_special_tokens and index in self.all_special_ids:
+                continue
+            if index in self.added_tokens_decoder:
+                tokens.append(self.added_tokens_decoder[index])
+            else:
+                tokens.append(self._convert_id_to_token(index))
+        return tokens
+
+    def _convert_id_to_token(self, index: int) -> str:
+        raise NotImplementedError
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        return " ".join(tokens)
+
+    def _decode(
+        self,
+        token_ids: List[int],
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: bool = True,
+        spaces_between_special_tokens: bool = True,
+        **kwargs
+    ) -> str:
+        self._decode_use_source_tokenizer = kwargs.pop("use_source_tokenizer", False)
+
+        filtered_tokens = self.convert_ids_to_tokens(token_ids, skip_special_tokens=skip_special_tokens)
+
+        # To avoid mixing byte-level and unicode for byte-level BPT
+        # we need to build string separately for added tokens and byte-level tokens
+        # cf. https://github.com/huggingface/transformers/issues/1133
+        sub_texts = []
+        current_sub_text = []
+        for token in filtered_tokens:
+            if skip_special_tokens and token in self.all_special_ids:
+                continue
+            if token in self.added_tokens_encoder:
+                if current_sub_text:
+                    sub_texts.append(self.convert_tokens_to_string(current_sub_text))
+                    current_sub_text = []
+                sub_texts.append(token)
+            else:
+                current_sub_text.append(token)
+        if current_sub_text:
+            sub_texts.append(self.convert_tokens_to_string(current_sub_text))
+
+        if spaces_between_special_tokens:
+            text = " ".join(sub_texts)
+        else:
+            text = "".join(sub_texts)
+
+        if clean_up_tokenization_spaces:
+            clean_text = self.clean_up_tokenization(text)
+            return clean_text
+        else:
+            return text

fastNLP/transformers/torch/utils/model_parallel_utils.py

@@ -0,0 +1,54 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from math import ceil
+
+
+def assert_device_map(device_map, num_blocks):
+    blocks = list(range(0, num_blocks))
+
+    device_map_blocks = [item for sublist in list(device_map.values()) for item in sublist]
+
+    # Duplicate check
+    duplicate_blocks = []
+    for i in device_map_blocks:
+        if device_map_blocks.count(i) > 1 and i not in duplicate_blocks:
+            duplicate_blocks.append(i)
+    # Missing blocks
+    missing_blocks = [i for i in blocks if i not in device_map_blocks]
+    extra_blocks = [i for i in device_map_blocks if i not in blocks]
+
+    assert len(duplicate_blocks) == 0, (
+        "Duplicate attention blocks specified in device_map. Attention blocks must be specified to one device. These "
+        "attention blocks were specified more than once: " + str(duplicate_blocks)
+    )
+    assert len(missing_blocks) == 0, (
+        "There are attention blocks for this model that are not specified in the device_map. Add these attention "
+        "blocks to a device on the device_map: " + str(missing_blocks)
+    )
+    assert (
+        len(extra_blocks) == 0
+    ), "The device_map contains more attention blocks than this model has. Remove these from the device_map:" + str(
+        extra_blocks
+    )
+
+
+def get_device_map(n_layers, devices):
+    """Returns a dictionary of layers distributed evenly across all devices."""
+    layers = list(range(n_layers))
+    n_blocks = int(ceil(n_layers / len(devices)))
+    layers_list = list(layers[i : i + n_blocks] for i in range(0, n_layers, n_blocks))
+
+    return dict(zip(devices, layers_list))

fastNLP/transformers/torch/utils/versions.py

@@ -0,0 +1,120 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Utilities for working with package versions
+"""
+
+import operator
+import re
+import sys
+from typing import Optional
+
+from packaging import version
+
+
+# The package importlib_metadata is in a different place, depending on the python version.
+if sys.version_info < (3, 8):
+    import importlib_metadata
+else:
+    import importlib.metadata as importlib_metadata
+
+
+ops = {
+    "<": operator.lt,
+    "<=": operator.le,
+    "==": operator.eq,
+    "!=": operator.ne,
+    ">=": operator.ge,
+    ">": operator.gt,
+}
+
+
+def _compare_versions(op, got_ver, want_ver, requirement, pkg, hint):
+    if got_ver is None:
+        raise ValueError("got_ver is None")
+    if want_ver is None:
+        raise ValueError("want_ver is None")
+    if not ops[op](version.parse(got_ver), version.parse(want_ver)):
+        raise ImportError(
+            f"{requirement} is required for a normal functioning of this module, but found {pkg}=={got_ver}.{hint}"
+        )
+
+
+def require_version(requirement: str, hint: Optional[str] = None) -> None:
+    """
+    Perform a runtime check of the dependency versions, using the exact same syntax used by pip.
+
+    The installed module version comes from the `site-packages` dir via `importlib_metadata`.
+
+    Args:
+        requirement (:obj:`str`): pip style definition, e.g.,  "tokenizers==0.9.4", "tqdm>=4.27", "numpy"
+        hint (:obj:`str`, `optional`): what suggestion to print in case of requirements not being met
+
+    Example::
+
+       require_version("pandas>1.1.2")
+       require_version("numpy>1.18.5", "this is important to have for whatever reason")
+
+    """
+
+    hint = f"\n{hint}" if hint is not None else ""
+
+    # non-versioned check
+    if re.match(r"^[\w_\-\d]+$", requirement):
+        pkg, op, want_ver = requirement, None, None
+    else:
+        match = re.findall(r"^([^!=<>\s]+)([\s!=<>]{1,2}.+)", requirement)
+        if not match:
+            raise ValueError(
+                f"requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23, but got {requirement}"
+            )
+        pkg, want_full = match[0]
+        want_range = want_full.split(",")  # there could be multiple requirements
+        wanted = {}
+        for w in want_range:
+            match = re.findall(r"^([\s!=<>]{1,2})(.+)", w)
+            if not match:
+                raise ValueError(
+                    f"requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23, but got {requirement}"
+                )
+            op, want_ver = match[0]
+            wanted[op] = want_ver
+            if op not in ops:
+                raise ValueError(f"{requirement}: need one of {list(ops.keys())}, but got {op}")
+
+    # special case
+    if pkg == "python":
+        got_ver = ".".join([str(x) for x in sys.version_info[:3]])
+        for op, want_ver in wanted.items():
+            _compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
+        return
+
+    # check if any version is installed
+    try:
+        got_ver = importlib_metadata.version(pkg)
+    except importlib_metadata.PackageNotFoundError:
+        raise importlib_metadata.PackageNotFoundError(
+            f"The '{requirement}' distribution was not found and is required by this application. {hint}"
+        )
+
+    # check that the right version is installed if version number or a range was provided
+    if want_ver is not None:
+        for op, want_ver in wanted.items():
+            _compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
+
+
+def require_version_core(requirement):
+    """require_version wrapper which emits a core-specific hint on failure"""
+    hint = "Try: pip install transformers -U or pip install -e '.[dev]' if you're working with git master"
+    return require_version(requirement, hint)