merge

docs/Makefile

@@ -6,7 +6,7 @@ SPHINXOPTS    =
 SPHINXAPIDOC  = sphinx-apidoc
 SPHINXBUILD   = sphinx-build
 SPHINXPROJ    = fastNLP
-SPHINXEXCLUDE = ../fastNLP/transformers/* ../fastNLP/modules/* ../fastNLP/core/drivers/torch_paddle_driver/* ../fastNLP/core/utils/torch_paddle_utils.py
+SPHINXEXCLUDE = ../fastNLP/transformers/*
 SOURCEDIR     = source
 BUILDDIR      = build
 PORT          = 9000
@@ -16,7 +16,7 @@ help:
 	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS)
 
 apidoc:
-	$(SPHINXAPIDOC) -efM -d 6 -o source ../$(SPHINXPROJ) $(SPHINXEXCLUDE)
+	$(SPHINXAPIDOC) -efM -o source ../$(SPHINXPROJ) $(SPHINXEXCLUDE)
 
 server:
 	cd build/html && python -m http.server $(PORT)
@@ -24,6 +24,9 @@ server:
 delete:
 	rm -f source/$(SPHINXPROJ).* source/modules.rst && rm -rf build
 
+web:
+	make html && make server
+
 dev:
 	make delete && make apidoc  && make html && make server
 

docs/source/conf.py

@@ -42,7 +42,8 @@ extensions = [
     'sphinx.ext.viewcode',
     'sphinx.ext.autosummary',
     'sphinx.ext.mathjax',
-    'sphinx.ext.todo'
+    'sphinx.ext.todo',
+    'sphinx_autodoc_typehints'
 ]
 
 autodoc_default_options = {
@@ -53,8 +54,10 @@ autodoc_default_options = {
 
 add_module_names = False
 autosummary_ignore_module_all = False
-autodoc_typehints = "description"
+# autodoc_typehints = "description"
 autoclass_content = "class"
+typehints_fully_qualified = False
+typehints_defaults = "comma"
 
 # Add any paths that contain templates here, relative to this directory.
 templates_path = ['_templates']
@@ -168,8 +171,8 @@ texinfo_documents = [
 
 # -- Extension configuration -------------------------------------------------
 def maybe_skip_member(app, what, name, obj, skip, options):
-    # if obj.__doc__ is None:
-    #     return True
+    if obj.__doc__ is None:
+        return True
     if name == "__init__":
         return False
     if name.startswith("_"):

docs/source/fastNLP.core.callbacks.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.callbacks.torch_callbacks
 
@@ -18,7 +18,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.callbacks.callback
    fastNLP.core.callbacks.callback_event

docs/source/fastNLP.core.callbacks.torch_callbacks.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.callbacks.torch_callbacks.torch_grad_clip_callback
    fastNLP.core.callbacks.torch_callbacks.torch_lr_sched_callback

docs/source/fastNLP.core.collators.padders.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.collators.padders.exceptions
    fastNLP.core.collators.padders.get_padder

docs/source/fastNLP.core.collators.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.collators.padders
 
@@ -18,7 +18,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.collators.collator
    fastNLP.core.collators.packer_unpacker

docs/source/fastNLP.core.controllers.loops.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.controllers.loops.evaluate_batch_loop
    fastNLP.core.controllers.loops.loop

docs/source/fastNLP.core.controllers.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.controllers.loops
    fastNLP.core.controllers.utils
@@ -19,7 +19,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.controllers.evaluator
    fastNLP.core.controllers.trainer

docs/source/fastNLP.core.controllers.utils.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.controllers.utils.state
    fastNLP.core.controllers.utils.utils

docs/source/fastNLP.core.dataloaders.jittor_dataloader.rst

@@ -10,6 +10,6 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.dataloaders.jittor_dataloader.fdl

docs/source/fastNLP.core.dataloaders.paddle_dataloader.rst

@@ -10,6 +10,6 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.dataloaders.paddle_dataloader.fdl

docs/source/fastNLP.core.dataloaders.prepare_dataloader.rst

@@ -0,0 +1,7 @@
+fastNLP.core.dataloaders.prepare\_dataloader module
+===================================================
+
+.. automodule:: fastNLP.core.dataloaders.prepare_dataloader
+   :members:
+   :undoc-members:
+   :show-inheritance:

docs/source/fastNLP.core.dataloaders.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.dataloaders.jittor_dataloader
    fastNLP.core.dataloaders.paddle_dataloader
@@ -20,7 +20,8 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.dataloaders.mix_dataloader
+   fastNLP.core.dataloaders.prepare_dataloader
    fastNLP.core.dataloaders.utils

docs/source/fastNLP.core.dataloaders.torch_dataloader.rst

@@ -10,6 +10,6 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.dataloaders.torch_dataloader.fdl

docs/source/fastNLP.core.dataset.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.dataset.dataset
    fastNLP.core.dataset.field

docs/source/fastNLP.core.drivers.jittor_driver.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.drivers.jittor_driver.initialize_jittor_driver
    fastNLP.core.drivers.jittor_driver.jittor_driver

docs/source/fastNLP.core.drivers.paddle_driver.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.drivers.paddle_driver.dist_utils
    fastNLP.core.drivers.paddle_driver.fleet

docs/source/fastNLP.core.drivers.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.drivers.jittor_driver
    fastNLP.core.drivers.paddle_driver
@@ -20,7 +20,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.drivers.choose_driver
    fastNLP.core.drivers.driver

docs/source/fastNLP.core.drivers.torch_driver.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.drivers.torch_driver.ddp
    fastNLP.core.drivers.torch_driver.dist_utils

docs/source/fastNLP.core.log.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.log.handler
    fastNLP.core.log.highlighter

docs/source/fastNLP.core.metrics.backend.jittor_backend.rst

@@ -10,6 +10,6 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.metrics.backend.jittor_backend.backend

docs/source/fastNLP.core.metrics.backend.paddle_backend.rst

@@ -10,6 +10,6 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.metrics.backend.paddle_backend.backend

docs/source/fastNLP.core.metrics.backend.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.metrics.backend.jittor_backend
    fastNLP.core.metrics.backend.paddle_backend
@@ -20,7 +20,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.metrics.backend.auto_backend
    fastNLP.core.metrics.backend.backend

docs/source/fastNLP.core.metrics.backend.torch_backend.rst

@@ -10,6 +10,6 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.metrics.backend.torch_backend.backend

docs/source/fastNLP.core.metrics.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.metrics.backend
 
@@ -18,7 +18,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.metrics.accuracy
    fastNLP.core.metrics.classify_f1_pre_rec_metric

docs/source/fastNLP.core.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.callbacks
    fastNLP.core.collators
@@ -27,6 +27,6 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.vocabulary

docs/source/fastNLP.core.samplers.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.samplers.conversion_utils
    fastNLP.core.samplers.mix_sampler

docs/source/fastNLP.core.utils.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core.utils.cache_results
    fastNLP.core.utils.dummy_class

docs/source/fastNLP.envs.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.envs.distributed
    fastNLP.envs.env

docs/source/fastNLP.io.loader.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.io.loader.classification
    fastNLP.io.loader.conll

docs/source/fastNLP.io.pipe.rst

@@ -10,7 +10,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.io.pipe.classification
    fastNLP.io.pipe.conll

docs/source/fastNLP.io.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.io.loader
    fastNLP.io.pipe
@@ -19,7 +19,7 @@ Submodules
 ----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.io.data_bundle
    fastNLP.io.embed_loader

docs/source/fastNLP.rst

@@ -10,7 +10,7 @@ Subpackages
 -----------
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP.core
    fastNLP.envs

docs/source/modules.rst

@@ -2,6 +2,6 @@ fastNLP
 =======
 
 .. toctree::
-   :maxdepth: 6
+   :maxdepth: 4
 
    fastNLP

fastNLP/core/__init__.py

@@ -3,9 +3,7 @@ __all__ = [
     'Callback',
     'Event',
     'Filter',
-    'CallbackManager',
     'CheckpointCallback',
-    'choose_progress_callback',
     'ProgressCallback',
     'RichCallback',
     "LRSchedCallback",
@@ -54,7 +52,6 @@ __all__ = [
     'DataSet',
     'FieldArray',
     'Instance',
-    'ApplyResultException',
 
     # drivers
     "TorchSingleDriver",
@@ -63,7 +60,6 @@ __all__ = [
     "PaddleFleetDriver",
     "JittorSingleDriver",
     "JittorMPIDriver",
-    "TorchPaddleDriver",
 
     # log
     "logger",

fastNLP/core/callbacks/callback_manager.py

@@ -10,8 +10,8 @@ from .callback_event import Event
 from .callback import Callback
 from fastNLP.core.log import logger
 from .progress_callback import ProgressCallback, choose_progress_callback
-from fastNLP.envs import rank_zero_call
-from fastNLP.core.utils.utils import _get_fun_msg
+from ..utils.exceptions import EarlyStopException
+from ..utils.utils import _get_fun_msg
 
 
 def _transfer(func):
@@ -25,6 +25,8 @@ def _transfer(func):
         for callback_fn in manager.callback_fns[func.__name__]:
             try:
                 callback_fn(*arg, **kwargs)
+            except EarlyStopException as e:
+                raise e
             except BaseException as e:
                 logger.error(f"The following callback_fn raise exception:{_get_fun_msg(callback_fn)}.")
                 raise e
@@ -178,14 +180,16 @@ class CallbackManager:
                 states[each_callback.callback_name]["states"] = each_callback.on_save_checkpoint(trainer)
 
         if len(_duplicated_callbacks) > 0:
-            logger.warning(f"Notice these callbacks' `callback_name` are duplicated: {_duplicated_callbacks}, "
-                           f"and we will only save the first callback's state we meet.")
+            logger.warning(f"Notice these callback_name: {_duplicated_callbacks} are duplicated, "
+                           f"fastNLP will only save the first callback's state.")
 
         # 2. 每一个具体的 callback 函数的 filter 的状态；
         _record_duplicated_callback_names = set()
         for each_callback_filters in self._callback_filters:
             if each_callback_filters[0] not in _record_duplicated_callback_names:
                 _record_duplicated_callback_names.add(each_callback_filters[0])
+                if 'filter_states' not in states[each_callback_filters[0]]:
+                    states[each_callback_filters[0]]["filter_states"] = {}
                 states[each_callback_filters[0]]["filter_states"][each_callback_filters[1]] = each_callback_filters[2].state_dict()
 
         # 3. 保存 callback_counter；
@@ -212,13 +216,15 @@ class CallbackManager:
             if each_callback_filters[0] in states:
                 if each_callback_filters[0] not in _already_loaded_callback_names:
                     _already_loaded_callback_names.add(each_callback_filters[0])
-                    each_callback_filters[2].load_state_dict(states[each_callback_filters[0]]["filter_states"][each_callback_filters[1]])
+                    if 'filter_states' in states[each_callback_filters[0]] and \
+                            each_callback_filters[1] in states[each_callback_filters[0]]['filter_states']:
+                        each_callback_filters[2].load_state_dict(states[each_callback_filters[0]]['filter_states'][each_callback_filters[1]])
                 else:
                     _duplicated_callback_names.add(each_callback_filters[0])
 
         if len(_duplicated_callback_names) > 0:
-            logger.warning(f"Notice these callbacks' `callback_name` are duplicated: {_duplicated_callback_names}, "
-                           f"and we will only load the first callback's state we meet.")
+            logger.rank_zero_warning(f"Notice these callback_name: {_duplicated_callback_names} are duplicated, "
+                                     f"fastNLP will only load the first callback's state.")
 
         # 2. 再恢复每一个 callback 的单独的状态；
         # 每一个我们自己提供的类 callback，都需要重写其特定的 `callback_name` 方法，保证如果两个 callback 的 callback_name 一样，
@@ -229,8 +235,6 @@ class CallbackManager:
                 _already_loaded_callback_names.add(each_callback.callback_name)
                 # 这里要注意，我们已经确保每一个 callback 的 `on_load_checkpoint` 函数拿到的就是其自己的状态；
                 each_callback.on_load_checkpoint(trainer, states[each_callback.callback_name]["states"])
-            else:
-                each_callback.on_load_checkpoint(trainer, None)
 
     @property
     def has_trainer_checkpoint(self) -> bool:

fastNLP/core/callbacks/checkpoint_callback.py

@@ -19,7 +19,7 @@ class CheckpointCallback(Callback):
                  only_state_dict: bool = True, model_save_fn: Optional[Callable] = None, save_object: str = 'model',
                  save_evaluate_results=True, **kwargs):
         """
-        保存模型 checkpoint 的  callback ，其保存的文件目录以及文件名命名规则如下::
+        保存 checkpoint 的  callback ，其保存的文件目录以及文件名命名规则如下::
 
             - folder/
                 - YYYY-mm-dd-HH_MM_SS_fffff/  # 自动根据当前脚本的启动时间创建的
@@ -29,8 +29,9 @@ class CheckpointCallback(Callback):
                     - {save_object}-epoch_{epoch_idx}-batch_{global_batch_idx}-exception_{exception_type}/  # exception时保存。
                     - {save_object}-epoch_{epoch_idx}-batch_{global_batch_idx}-{monitor}_{monitor_value}/  # 满足topk条件存储文件名
 
-        model_save_fn 为 None ，则以上每个 folder 中，将生成 fastnlp_model.pkl.tar 文件。
-        若 model_save_fn 不为 None，则 fastNLP 将 folder 绝对路径传递给该函数，fastNLP 在该 folder 下不进行模型保存。
+        model_save_fn 为 None ，则以上每个 folder 中，将生成 fastnlp_model.pkl.tar 文件。若 model_save_fn 不为 None，
+        则 fastNLP 将 folder 绝对路径传递给该函数，fastNLP 在该 folder 下不进行模型保存。默认情况下，本 checkpoint 只保存了 model
+        的状态；如还需保存 Trainer 的状态以断点重训的话，请使用 ``save_object='trainer'`` 。
 
         :param monitor: 监控的 metric 值。如果在 evaluation 结果中没有找到完全一致的名称，将使用 最长公共字符串算法 找到最匹配
                 的那个作为 monitor 。如果为 None，将尝试使用 Trainer 设置的 monitor 。也可以传入一个函数，接受参数为 evaluation 的结
@@ -46,22 +47,14 @@ class CheckpointCallback(Callback):
         :param only_state_dict: 保存模型时是否只保存 state_dict 。当 model_save_fn 不为 None 时，该参数无效。
         :param model_save_fn: 个性化的保存函数，当触发保存操作时，就调用这个函数，这个函数应当接受一个文件夹作为参数，不返回任何东西。
             如果传入了 model_save_fn 函数，fastNLP 将不再进行模型相关的保存。在多卡场景下，我们只在 rank 0 上会运行该函数。
-        :param save_object: 可选 ['trainer', 'model']，表示在保存时的保存对象为 trainer+model 还是 只是model 。
+        :param save_object: 可选 ['trainer', 'model']，表示在保存时的保存对象为 ``trainer+model`` 还是 只是 ``model`` 。如果
+            保存 ``trainer`` 对象的话，将会保存 :class:~fastNLP.Trainer 的相关状态，可以通过 :meth:`Trainer.load` 加载该断
+            点继续训练。如果保存的是 ``Model`` 对象，则可以通过 :meth:`Trainer.load_model` 加载该模型权重。
         :param save_evaluate_results: 是否保存 evaluate 的结果。如果为 True ，在保存 topk 模型的 folder 中还将额外保存一个
             fastnlp_evaluate_results.json 文件，记录当前的 results。仅在设置了 topk 的场景下有用，默认为 True 。
         :param kwargs:
         """
         super().__init__()
-        if folder is None:
-            logger.warning(
-                "Parameter `folder` is None, and we will use the current work directory to find and load your model.")
-            folder = Path.cwd()
-        folder = Path(folder)
-        if not folder.exists():
-            raise NotADirectoryError(f"Path '{folder.absolute()}' is not existed!")
-        elif folder.is_file():
-            raise ValueError("Parameter `folder` should be a directory instead of a file.")
-
         if every_n_epochs is not None:
             if not isinstance(every_n_epochs, int) or every_n_epochs < 1:
                 raise ValueError("Parameter `every_n_epochs` should be an int and greater than or equal to 1.")
@@ -74,12 +67,6 @@ class CheckpointCallback(Callback):
         else:
             every_n_batches = sys.maxsize  # 使得没有数字可以整除
 
-        if topk is not None:
-            if not isinstance(topk, int):
-                raise ValueError("Parameter `topk` should be an int.")
-        else:
-            topk = 0
-
         if on_exceptions is not None:
             if not isinstance(on_exceptions, Sequence):
                 on_exceptions = [on_exceptions]

fastNLP/core/callbacks/load_best_model_callback.py

@@ -19,7 +19,8 @@ class LoadBestModelCallback(HasMonitorCallback):
                  model_load_fn:Optional[Callable] = None,
                  delete_after_train:bool = True):
         """
-        保存最佳的 monitor 值最佳的模型，并在训练结束的时候重新加载模型。仅在训练正常结束的时候才能加载最好的模型。
+        保存最佳的 monitor 值最佳的模型，并在训练结束的时候重新加载模型，默认会在加载之后删除权重文件。仅在训练正常结束的时候才能加载
+        最好的模型。
 
         :param str monitor: 监控的 metric 值。如果在 evaluation 结果中没有找到完全一致的名称，将使用 最长公共字符串算法 找到最匹配
             的那个作为 monitor 。如果为 None，将尝试使用 Trainer 设置的 monitor 。也可以传入一个函数，接受参数为 evaluation 的结

fastNLP/core/callbacks/topk_saver.py

@@ -33,9 +33,8 @@ class Saver:
         :param kwargs: 更多需要传递给 Trainer.save() 或者 Trainer.save_model() 接口的参数。
         """
         if folder is None:
-            logger.rank_zero_warning(
-                "Parameter `folder` is None, and we will use the current work directory to find and load your model.")
-            folder = Path.cwd()
+            folder = Path.cwd().absolute()
+            logger.info(f"Parameter `folder` is None, and we will use {folder} to save and load your model.")
         folder = Path(folder)
         if not folder.exists():
             folder.mkdir(parents=True, exist_ok=True)

fastNLP/core/callbacks/utils.py

@@ -8,8 +8,8 @@ from fastNLP.core.utils.utils import _get_fun_msg
 
 def _get_monitor_value(monitor: Union[callable, str], real_monitor: Optional[str], res: dict) ->Tuple[str, float]:
     """
-    从res中寻找 monitor 并返回。如果 monitor 没找到则尝试用 _real_monitor ,若 _real_monitor 为 None 则尝试使用 monitor 的值进行
-    匹配。
+    从 ``res`` 中寻找 ``monitor`` 并返回。如果 ``monitor`` 没找到则尝试用 ``_real_monitor`` ,若 ``_real_monitor`` 为 ``None``
+    则尝试使用 ``monitor`` 的值进行匹配。
 
     :param monitor:
     :param real_monitor:

fastNLP/core/collators/padders/get_padder.py

@@ -121,7 +121,7 @@ def get_padder(batch_field:Sequence[Any], pad_val, dtype, backend, field_name)->
                 # 这里 ele_dtype 传入为 None 的原因是防止出现 paddle tensor 转换为 torch tensor
                 return TorchTensorPadder(pad_val=pad_val, ele_dtype=None, dtype=dtype)
             elif backend == 'paddle':
-                return PaddleTensorPadder(pad_val=pad_val, ele_dtype=None, dtype=dtype)
+                return PaddleTensorPadder(pad_val=pad_val, ele_dtype=ele_dtype, dtype=dtype)
             elif backend == 'jittor':
                 return JittorTensorPadder(pad_val=pad_val, ele_dtype=ele_dtype, dtype=dtype)
             else:

fastNLP/core/collators/padders/paddle_padder.py

@@ -141,7 +141,10 @@ class PaddleTensorPadder(Padder):
 
         shapes = [field.shape for field in batch_field]
         max_shape = [len(batch_field)] + [max(*_) for _ in zip(*shapes)]
-        array = np.full(max_shape, fill_value=pad_val)
+        if isinstance(batch_field[0], paddle.Tensor):
+            array = paddle.full(max_shape, fill_value=pad_val, dtype=dtype)
+        else:
+            array = np.full(max_shape, fill_value=pad_val, dtype=batch_field[0].dtype)
         for i, field in enumerate(batch_field):
             slices = (i, ) + tuple(slice(0, s) for s in shapes[i])
             array[slices] = field

fastNLP/core/collators/padders/torch_padder.py

@@ -118,8 +118,8 @@ class TorchTensorPadder(Padder):
                 batch_field = [torch.tensor(field.tolist(), dtype=dtype) for field in batch_field]
             else:
                 device = batch_field[0].device
-                if dtype is None:
-                    dtype = batch_field[0].dtype
+            if dtype is None:
+                dtype = batch_field[0].dtype
         except AttributeError:
             raise RuntimeError(f"If the field is not a torch.Tensor (it is {type(batch_field[0])}), "
                                f"it must have tolist() method.")

fastNLP/core/controllers/evaluator.py

@@ -236,8 +236,7 @@ class Evaluator:
         """
         调用所有 metric 的 reset() 方法，清除累积的状态。
 
-        Returns:
-
+        :return:
         """
         self.metrics_wrapper.reset()
 
@@ -359,6 +358,11 @@ class _MetricsWrapper:
                 metric.update(res)
 
     def reset(self):
+        """
+        将 Metric 中的状态重新设置。
+
+        :return:
+        """
         for metric in self._metrics:
             if _is_allennlp_metric(metric):
                 metric.get_metric(reset=True)

fastNLP/core/controllers/loops/evaluate_batch_loop.py

@@ -34,7 +34,7 @@ class EvaluateBatchLoop(Loop):
             except BaseException as e:
                 if callable(getattr(dataloader, 'get_batch_indices', None)):
                     indices = dataloader.get_batch_indices()
-                    logger.debug(f"The following exception happens when running on samples: {indices}")
+                    logger.error(f"Exception happens when evaluating on samples: {indices}")
                 raise e
 
             self.batch_step_fn(evaluator, batch)

fastNLP/core/controllers/loops/train_batch_loop.py

@@ -32,7 +32,7 @@ class TrainBatchLoop(Loop):
                 break
             except BaseException as e:
                 if indices and not isinstance(e, EarlyStopException):
-                    logger.debug(f"The following exception happens when running on samples: {indices}")
+                    logger.error(f"Exception happens when running on samples: {indices}")
                 raise e
 
             trainer.on_train_batch_begin(batch, indices)

fastNLP/core/controllers/trainer.py

@@ -282,32 +282,41 @@ class Trainer(TrainerEventTrigger):
 
         :kwargs:
             * *torch_kwargs* -- 用于在指定 ``driver`` 为 'torch' 时设定具体 driver 实例的一些参数：
+                
                 * ddp_kwargs -- 用于在使用 ``TorchDDPDriver`` 时指定 ``DistributedDataParallel`` 初始化时的参数；例如传入
-                 {'find_unused_parameters': True} 来解决有参数不参与前向运算导致的报错等；
+                {'find_unused_parameters': True} 来解决有参数不参与前向运算导致的报错等；
                 * set_grad_to_none -- 是否在训练过程中在每一次 optimizer 更新后将 grad 置为 None；
                 * torch_non_blocking -- 表示用于 pytorch 的 tensor 的 to 方法的参数 non_blocking；
+            * *paddle_kwargs* -- 用于在指定 ``driver`` 为 'paddle' 时设定具体 driver 实例的一些参数：
+                
+                * fleet_kwargs -- 用于在使用 ``PaddleFleetDriver`` 时指定 ``DataParallel`` 和 ``fleet`` 初始化时的参数，包括：
+                    
+                    * is_collective -- 是否使用 paddle 集群式的分布式训练方法，目前仅支持为 True 的情况；
+                    * role_maker -- 初始化 ``fleet`` 分布式训练 API 时使用的 ``RoleMaker``
+                    * 其它用于初始化 ``DataParallel`` 的参数；
             * *data_device* -- 一个具体的 driver 实例中，有 ``model_device`` 和 ``data_device``，前者表示模型所在的设备，后者表示
-             当 ``model_device`` 为 None 时应当将数据迁移到哪个设备；
+            当 ``model_device`` 为 None 时应当将数据迁移到哪个设备；
 
-             .. note::
+                .. note::
 
                 注意您在绝大部分情况下不会用到该参数！
 
                 1. 当 driver 实例的 ``model_device`` 不为 None 时，该参数无效；
                 2. 对于 pytorch，仅当用户自己通过 ``python -m torch.distributed.launch`` 并且自己初始化 ``init_process_group`` 时，
                 driver 实例的 ``model_device`` 才会为 None；
+                3. 对于 paddle，该参数无效；
 
             * *use_dist_sampler* -- 表示是否使用分布式的 ``sampler``。在多卡时，分布式 ``sampler`` 将自动决定每张卡上读取的 sample ，使得一个 epoch
-             内所有卡的 sample 加起来为一整个数据集的 sample。默认会根据 driver 是否为分布式进行设置。
+            内所有卡的 sample 加起来为一整个数据集的 sample。默认会根据 driver 是否为分布式进行设置。
             * *evaluate_use_dist_sampler* -- 表示在 ``Evaluator`` 中在使用分布式的时候是否将 dataloader 的 ``sampler`` 替换为分布式的 ``sampler``；默认为 ``True``；
             * *output_from_new_proc* -- 应当为一个字符串，表示在多进程的 driver 中其它进程的输出流应当被做如何处理；其值应当为以下之一：
-             ["all", "ignore", "only_error"]；当该参数的值不是以上值时，该值应当表示一个文件夹的名字，我们会将其他 rank 的输出流重定向到
-             log 文件中，然后将 log 文件保存在通过该参数值设定的文件夹中；默认为 "only_error"；
+            ["all", "ignore", "only_error"]；当该参数的值不是以上值时，该值应当表示一个文件夹的名字，我们会将其他 rank 的输出流重定向到
+            log 文件中，然后将 log 文件保存在通过该参数值设定的文件夹中；默认为 "only_error"；
 
-             注意该参数仅当使用分布式的 ``driver`` 时才有效，例如 ``TorchDDPDriver``；
+                注意该参数仅当使用分布式的 ``driver`` 时才有效，例如 ``TorchDDPDriver``；
             * *progress_bar* -- 以哪种方式显示 progress ，目前支持[None, 'raw', 'rich', 'auto'] 或者 RichCallback, RawTextCallback对象，
-             默认为 auto , auto 表示如果检测到当前 terminal 为交互型则使用 RichCallback，否则使用 RawTextCallback对象。如果
-             需要定制 progress bar 的参数，例如打印频率等，可以传入 RichCallback, RawTextCallback 对象。
+            默认为 auto , auto 表示如果检测到当前 terminal 为交互型则使用 RichCallback，否则使用 RawTextCallback对象。如果
+            需要定制 progress bar 的参数，例如打印频率等，可以传入 RichCallback, RawTextCallback 对象。
             * *train_input_mapping* -- 与 input_mapping 一致，但是只用于 ``Trainer`` 中。与 input_mapping 互斥。
             * *train_output_mapping* -- 与 output_mapping 一致，但是只用于 ``Trainer`` 中。与 output_mapping 互斥。
             * *evaluate_input_mapping* -- 与 input_mapping 一致，但是只用于 ``Evaluator`` 中。与 input_mapping 互斥。
@@ -558,7 +567,7 @@ class Trainer(TrainerEventTrigger):
             else:
                 raise FileNotFoundError("You are using `resume_from`, but we can not find your specific file.")
 
-        if self.evaluator is not None and num_eval_sanity_batch > 0:
+        if self.evaluator is not None and num_eval_sanity_batch != 0:
             logger.info(f"Running evaluator sanity check for {num_eval_sanity_batch} batches.")
             self.on_sanity_check_begin()
             sanity_check_res = self.evaluator.run(num_eval_batch_per_dl=num_eval_sanity_batch)
@@ -951,7 +960,7 @@ class Trainer(TrainerEventTrigger):
             self.driver.save_model(folder, only_state_dict, **kwargs)
         self.driver.barrier()
 
-    def load_model(self, folder: Union[str, Path, BinaryIO, io.BytesIO], only_state_dict: bool = False,
+    def load_model(self, folder: Union[str, Path, BinaryIO, io.BytesIO], only_state_dict: bool = True,
                    model_load_fn: Optional[Callable] = None, **kwargs):
         """
         加载模型

fastNLP/core/dataloaders/paddle_dataloader/fdl.py

@@ -162,9 +162,9 @@ class PaddleDataLoader(DataLoader):
 
     def get_batch_indices(self) -> List[int]:
         """
-        获取当前 batch 的 idx
+        获取当前 ``batch`` 中每条数据对应的索引。
 
-        :return:
+        :return: 当前 ``batch`` 数据的索引；
         """
         return self.cur_batch_indices
 

fastNLP/core/dataloaders/prepare_dataloader.py

@@ -10,7 +10,7 @@ from ..samplers import RandomBatchSampler, RandomSampler
 from .torch_dataloader import prepare_torch_dataloader
 from .paddle_dataloader import prepare_paddle_dataloader
 from .jittor_dataloader import prepare_jittor_dataloader
-from ...envs import FASTNLP_BACKEND, SUPPORT_BACKENDS, _module_available
+from ...envs import FASTNLP_BACKEND, SUPPORT_BACKENDS
 from ..log import logger
 
 

fastNLP/core/dataloaders/torch_dataloader/fdl.py

@@ -170,9 +170,9 @@ class TorchDataLoader(DataLoader):
 
     def get_batch_indices(self) -> List[int]:
         """
-        获取当前 batch 的 idx
+        获取当前 ``batch`` 中每条数据对应的索引。
 
-        :return:
+        :return: 当前 ``batch`` 数据的索引；
         """
         return self.cur_batch_indices
 

fastNLP/core/dataset/dataset.py

@@ -400,15 +400,22 @@ class DataSet:
                     new_field_name: str = None, num_proc: int = 0,
                     progress_desc: str = None, show_progress_bar: bool = True):
         r"""
-        将 DataSet 中的每个 instance 中的名为 `field_name` 的 field 传给 func，并获取它的返回值。
-
-        :param field_name: 传入 func 的是哪个 field。
-        :param func: input是 instance 中名为 `field_name` 的 field 的内容。
-        :param new_field_name: 将 func 返回的内容放入到 `new_field_name` 这个 field 中，如果名称与已有的 field 相同，则覆
-            盖之前的 field。如果为 None 则不创建新的 field。
-        :param num_proc: 进程的数量。请注意，由于python语言的特性，多少进程就会导致多少倍内存的增长。
-        :param progress_desc: progress_desc 的值，默认为 Main
-        :param show_progress_bar: 是否展示进度条，默认展示进度条
+        将 :class:`~DataSet` 每个 ``instance`` 中为 ``field_name`` 的 ``field`` 传给函数 ``func``，并写入到 ``new_field_name``
+        中。
+
+        :param field_name: 传入 ``func`` 的 ``field`` 名称；
+        :param func: 对指定 ``field`` 进行处理的函数，注意其输入应为 ``instance`` 中名为 ``field_name`` 的 ``field`` 的内容；
+        :param new_field_name: 函数执行结果写入的 ``field`` 名称。该函数会将 ``func`` 返回的内容放入到 ``new_field_name`` 对
+            应的 ``field`` 中，注意如果名称与已有的 ``field`` 相同则会进行覆盖。如果为 ``None`` 则不会覆盖和创建 ``field`` ；
+        :param num_proc: 使用进程的数量。
+        
+            .. note::
+            
+                由于 ``python`` 语言的特性，设置该参数后会导致相应倍数的内存增长，这可能会对您程序的执行带来一定的影响。
+
+        :param progress_desc: 进度条的描述字符，默认为 ``Main``；
+        :param show_progress_bar: 是否在处理过程中展示进度条；
+        :return: 从函数 ``func`` 中得到的返回值；
         """
         assert len(self) != 0, "Null DataSet cannot use apply_field()."
         if not self.has_field(field_name=field_name):
@@ -451,8 +458,8 @@ class DataSet:
         apply_out = self._apply_process(num_proc, func, progress_desc=progress_desc,
                                         show_progress_bar=show_progress_bar, _apply_field=field_name)
         #   只检测第一个数据是否为dict类型，若是则默认所有返回值为dict；否则报错。
-        if not isinstance(apply_out[0], dict):
-            raise Exception("The result of func is not a dict")
+        if not isinstance(apply_out[0], Mapping):
+            raise Exception(f"The result of func is not a Mapping, but a {type(apply_out[0])}")
 
         for key, value in apply_out[0].items():
             results[key] = [value]

fastNLP/core/drivers/__init__.py

@@ -9,7 +9,6 @@ __all__ = [
     "JittorDriver",
     "JittorSingleDriver",
     "JittorMPIDriver",
-    "TorchPaddleDriver",
     'torch_seed_everything',
     'paddle_seed_everything',
     'optimizer_state_to_device'
@@ -18,7 +17,6 @@ __all__ = [
 from .torch_driver import TorchDriver, TorchSingleDriver, TorchDDPDriver, torch_seed_everything, optimizer_state_to_device
 from .jittor_driver import JittorDriver, JittorMPIDriver, JittorSingleDriver
 from .paddle_driver import PaddleDriver, PaddleFleetDriver, PaddleSingleDriver, paddle_seed_everything
-from .torch_paddle_driver import TorchPaddleDriver
 from .driver import Driver
 
 

fastNLP/core/drivers/choose_driver.py

@@ -23,9 +23,9 @@ def choose_driver(model, driver: Union[str, Driver], device: Optional[Union[int,
     elif driver in {"jittor"}:
         from fastNLP.core.drivers.jittor_driver.initialize_jittor_driver import initialize_jittor_driver
         return initialize_jittor_driver(driver, device, model, **kwargs)
-    elif driver in {"paddle", "fleet"}:
+    elif driver in {"paddle"}:
         from fastNLP.core.drivers.paddle_driver.initialize_paddle_driver import initialize_paddle_driver
         return initialize_paddle_driver(driver, device, model, **kwargs)
     else:
         raise ValueError("Parameter `driver` can only be one of these values: ['torch', 'fairscale', "
-                         "'jittor', 'paddle', 'fleet'].")
+                         "'jittor', 'paddle'].")

fastNLP/core/drivers/jittor_driver/initialize_jittor_driver.py

@@ -7,18 +7,22 @@ from fastNLP.envs.imports import _NEED_IMPORT_JITTOR
 if _NEED_IMPORT_JITTOR:
     import jittor
 
+__all__ = []
+
 def initialize_jittor_driver(driver: str, device: Union[str, int, List[int]], model: jittor.Module, **kwargs) -> JittorDriver:
     r"""
-    用来根据参数 `driver` 和 `device` 来确定并且初始化一个具体的 `Driver` 实例然后返回回去；
-    在这个函数中，我们会根据用户设置的device来确定JittorDriver的mode。
+    用来根据参数 ``device`` 来确定并且初始化一个具体的 ``Driver`` 实例然后返回回去。
+
+    .. todo::
+
+        创建多卡的 driver
 
-    :param driver: 该参数的值应为以下之一：["jittor"]；
-    :param device: jittor运行的设备
+    :param driver: 该参数的值应为以下之一：``["jittor"]``；
+    :param device: ``jittor`` 运行的设备；
     :param model: 训练或者评测的具体的模型；
     :param kwargs: 
 
-    :return: 返回一个元组，元组的第一个值是具体的基于 jittor 的 `Driver` 实例，元组的第二个值是该 driver 的名字（用于检测一个脚本中
-     先后 driver 的次序的正确问题）；
+    :return: :class:`~fastNLP.core.JittorSingleDriver` 或 :class:`~fastNLP.core.JittorMPIDriver` 实例；
     """
 
     if driver not in {"jittor"}:

fastNLP/core/drivers/jittor_driver/jittor_driver.py

@@ -24,7 +24,17 @@ if _NEED_IMPORT_JITTOR:
 
 class JittorDriver(Driver):
     r"""
-    Jittor 框架的 Driver
+    ``Jittor`` 框架的 ``Driver``
+
+    .. note::
+
+        这是一个正在开发中的功能，敬请期待。
+
+    .. todo::
+
+        实现 fp16 的设置，且支持 cpu 和 gpu 的切换；
+        实现用于断点重训的 save 和 load 函数；
+
     """
 
     def __init__(self, model, fp16: bool = False, **kwargs):

fastNLP/core/drivers/jittor_driver/mpi.py

@@ -13,6 +13,14 @@ __all__ = [
 ]
 
 class JittorMPIDriver(JittorDriver):
+    """
+    执行 ``Jittor`` 框架下分布式训练的 ``Driver``。
+
+    .. note::
+
+        这是一个正在开发中的功能，敬请期待。
+
+    """
     def __init__(
         self,
         model,

fastNLP/core/drivers/jittor_driver/single_device.py

@@ -16,8 +16,17 @@ __all__ = [
 
 class JittorSingleDriver(JittorDriver):
     r"""
-    用于 cpu 和 单卡 gpu 运算
-    TODO: jittor 的 fp16
+    ``Jittor`` 框架下用于 ``cpu`` 和单卡 ``gpu`` 运算的 ``Driver``。
+
+    .. note::
+
+        这是一个正在开发中的功能，敬请期待。
+
+    .. todo::
+
+        支持 cpu 和 gpu 的切换；
+        实现断点重训中替换 dataloader 的 set_dist_repro_dataloader 函数
+
     """
 
     def __init__(self, model, device=None, fp16: bool = False, **kwargs):
@@ -30,11 +39,6 @@ class JittorSingleDriver(JittorDriver):
         self.world_size = 1
 
     def step(self):
-        """
-        jittor optimizers 的step函数可以传入参数loss
-        此时会同时进行 zero_grad 和 backward
-        为了统一，这里暂不使用这样的方式
-        """
         for optimizer in self.optimizers:
             optimizer.step()
 

fastNLP/core/drivers/jittor_driver/utils.py

@@ -5,10 +5,11 @@ from fastNLP.envs.imports import _NEED_IMPORT_JITTOR
 if _NEED_IMPORT_JITTOR:
     import jittor
 
+__all__ = []
+
 class DummyGradScaler:
     """
     用于仿造的GradScaler对象，防止重复写大量的if判断
-
     """
     def __init__(self, *args, **kwargs):
         pass

fastNLP/core/drivers/paddle_driver/fleet.py

@@ -1,8 +1,6 @@
 import os
 from typing import List, Union, Optional, Dict, Tuple, Callable
 
-from fastNLP.core.utils.paddle_utils import get_device_from_visible
-
 from .paddle_driver import PaddleDriver
 from .fleet_launcher import FleetLauncher
 from .utils import (
@@ -19,7 +17,9 @@ from fastNLP.core.utils import (
     check_user_specific_params,
     is_in_paddle_dist,
     is_in_paddle_dist,
+    get_paddle_device_id,
 )
+from fastNLP.core.utils.paddle_utils import _convert_data_device
 from fastNLP.envs.distributed import rank_zero_rm
 from fastNLP.core.samplers import (
     ReproduceBatchSampler,
@@ -31,7 +31,12 @@ from fastNLP.core.samplers import (
     re_instantiate_sampler,
     conversion_between_reproducible_and_unrepeated_sampler,
 )
-from fastNLP.envs.env import FASTNLP_DISTRIBUTED_CHECK, FASTNLP_GLOBAL_SEED, FASTNLP_NO_SYNC
+from fastNLP.envs.env import (
+    FASTNLP_DISTRIBUTED_CHECK,
+    FASTNLP_GLOBAL_SEED,
+    FASTNLP_NO_SYNC,
+    USER_CUDA_VISIBLE_DEVICES,
+)
 from fastNLP.core.log import logger
 
 if _NEED_IMPORT_PADDLE:
@@ -51,7 +56,7 @@ class PaddleFleetDriver(PaddleDriver):
     def __init__(
             self, 
             model, 
-            parallel_device: Optional[Union[List[int], int]],
+            parallel_device: Optional[Union[List[str], str]],
             is_pull_by_paddle_run: bool = False,
             fp16: bool = False,
             **kwargs
@@ -185,6 +190,8 @@ class PaddleFleetDriver(PaddleDriver):
             不管是什么情况，`PaddleFleetDriver` 在 `setup` 函数的最后，都会将所有进程的 pid 主动记录下来，这样当一个进程出现 exception 后，
              driver 的 on_exception 函数就会被 trainer 调用，其会调用 os.kill 指令将其它进程 kill 掉；
         """
+        if USER_CUDA_VISIBLE_DEVICES not in os.environ:
+            raise RuntimeError("To run paddle distributed training, please set `FASTNLP_BACKEND` to 'paddle' before using FastNLP.")
         super(PaddleFleetDriver, self).__init__(model, fp16=fp16, **kwargs)
 
         # 如果不是通过 launch 启动，要求用户必须传入 parallel_device
@@ -213,25 +220,6 @@ class PaddleFleetDriver(PaddleDriver):
                     "you initialize the paddle distribued process out of our control.")
 
             self.outside_fleet = True
-            # 用户只有将模型上传到对应机器上后才能用 DataParallel 包裹，因此如果用户在外面初始化了 Fleet，那么在 PaddleFleetDriver 中
-            # 我们就直接将 model_device 置为 None；
-            self._model_device = None
-
-        # 当参数 `device` 为 None 时并且该参数不为 None，表示将对应的数据移到指定的机器上；
-        self._data_device = kwargs.get("data_device", None)
-        if self._data_device is not None:
-            if isinstance(self._data_device, int):
-                if self._data_device < 0:
-                    raise ValueError("Parameter `data_device` can not be smaller than 0.")
-                _could_use_device_num = paddle.device.cuda.device_count()
-                if self._data_device >= _could_use_device_num:
-                    raise ValueError("The gpu device that parameter `device` specifies is not existed.")
-                self._data_device = f"gpu:{self._data_device}"
-            elif not isinstance(self._data_device, str):
-                raise ValueError("Parameter `device` is wrong type, please check our documentation for the right use.")
-            if self.outside_fleet and paddle.device.get_device() != self._data_device:
-                logger.warning("`Parameter data_device` is not equal to paddle.deivce.get_device(), "
-                                "please keep them equal to avoid some potential bugs.")
 
         self.world_size = None
         self.global_rank = 0
@@ -304,7 +292,8 @@ class PaddleFleetDriver(PaddleDriver):
             else:
                 # 已经设置过一次，保证参数必须是一样的
                 pre_gpus = os.environ[FASTNLP_DISTRIBUTED_CHECK]
-                pre_gpus = [int (x) for x in pre_gpus.split(",")]
+                pre_gpus = [int(x) for x in pre_gpus.split(",")]
+                cur_gpus = [get_paddle_device_id(g) for g in self.parallel_device]
                 if sorted(pre_gpus) != sorted(self.parallel_device):
                     raise RuntimeError("Notice you are using `PaddleFleetDriver` after one instantiated `PaddleFleetDriver`, it is not"
                                     "allowed that your second `PaddleFleetDriver` has a new setting of parameters `parallel_device`.")
@@ -410,8 +399,6 @@ class PaddleFleetDriver(PaddleDriver):
 
     @property
     def data_device(self):
-        if self.outside_fleet:
-            return self._data_device
         return self.model_device
 
     def model_call(self, batch, fn: Callable, signature_fn: Optional[Callable]) -> Dict:
@@ -565,7 +552,7 @@ class PaddleFleetDriver(PaddleDriver):
 
     def broadcast_object(self, obj, src:int=0, group=None, **kwargs):
         # 因为设置了CUDA_VISIBLE_DEVICES，可能会引起错误
-        device = get_device_from_visible(self.data_device)
+        device = _convert_data_device(self.data_device)
         return fastnlp_paddle_broadcast_object(obj, src, device=device, group=group)
 
     def all_gather(self, obj, group=None) -> List:

fastNLP/core/drivers/paddle_driver/fleet_launcher.py

@@ -11,11 +11,14 @@ from fastNLP.envs.env import (
     FASTNLP_LOG_LEVEL,
     FASTNLP_GLOBAL_SEED,
 )
+from fastNLP.core.utils import get_paddle_device_id
 from .utils import (
     find_free_ports,
     reset_seed,
 )
 
+__all__ = []
+
 # 记录各个进程信息
 class SubTrainer(object):
     """
@@ -34,11 +37,11 @@ class FleetLauncher:
     """
     def __init__(
         self,
-        devices: List[int],
+        devices: List[str],
         output_from_new_proc: str = "only_error"
     ):
 
-        self.devices = devices
+        self.devices = [ get_paddle_device_id(g) for g in devices]
         self.output_from_new_proc = output_from_new_proc
 
         self.setup()

fastNLP/core/drivers/paddle_driver/initialize_paddle_driver.py

@@ -7,50 +7,58 @@ from .single_device import PaddleSingleDriver
 from .fleet import PaddleFleetDriver
 
 from fastNLP.envs.imports import _NEED_IMPORT_PADDLE
-from fastNLP.core.utils import is_in_paddle_launch_dist
+from fastNLP.envs.env import USER_CUDA_VISIBLE_DEVICES
+from fastNLP.core.utils import is_in_paddle_launch_dist, get_paddle_gpu_str
 from fastNLP.core.log import logger
 
 if _NEED_IMPORT_PADDLE:
     import paddle
 
+__all__ = []
+
 def initialize_paddle_driver(driver: str, device: Optional[Union[str, int, List[int]]],
                             model: "paddle.nn.Layer", **kwargs) -> PaddleDriver:
     r"""
-    用来根据参数 `driver` 和 `device` 来确定并且初始化一个具体的 `Driver` 实例然后返回回去；
-    1、如果检测到当前进程为用户通过 `python -m paddle.distributed.launch xxx.py` 方式拉起的，则将
-    设备自动设置为用户指定的设备（由于我们在引入 fastNLP 进行了特殊的设置，因此可以通过 `CUDA_VISIBLE_DEVICES` 获取）
-    2、如果检测到输入的 `driver` 是 `paddle` 但 `device` 包含了多个设备，那么我们会给出警告并且自动返回多卡的 Driver
-    3、如果检测到输入的 `driver` 是 `fleet` 但 `device` 仅有一个设备，那么我们会给出警告但仍旧返回多卡的 Driver
+    用来根据参数 ``device`` 来确定并且初始化一个具体的 ``Driver`` 实例。
+
+    1. 如果检测到当前进程为用户通过 ``python -m paddle.distributed.launch xxx.py`` 方式拉起的，则将
+    设备自动设置为用户指定的设备（由于我们要求分布式训练必须进行 ``backend`` 的设置，因此可以通过 ``CUDA_VISIBLE_DEVICES`` 获取）
+    
+    2. 如果 ``device`` 包含了多个设备，则返回一个 :class:`~fastNLP.core.PaddleFleetDriver` 实例，否则返回
+    单卡的 :class:`~fastNLP.core.PaddleSingleDriver` 实例
 
-    :param driver: 使用的 ``driver`` 类型，在这个函数中仅支持 ``paddle``
-    :param device: 该参数的格式与 `Trainer` 对参数 `device` 的要求一致；
+    :param driver: 使用的 ``driver`` 类型，在这个函数中仅支持 ``paddle``；
+    :param device: 该参数的格式与 ``Trainer`` 对参数 ``device`` 的要求一致；
     :param model: 训练或者评测的具体的模型；
 
-    :return: 返回构造的 `Driver` 实例。
+    :return: 一个 :class:`~fastNLP.core.PaddleSingleDriver` 或 :class:`~fastNLP.core.PaddleFleetDriver` 实例；
     """
     if driver != "paddle":
         raise ValueError("When initialize PaddleDriver, parameter `driver` must be 'paddle'.")
+    user_visible_devices = os.getenv(USER_CUDA_VISIBLE_DEVICES)
     if is_in_paddle_launch_dist():
+        if user_visible_devices is None:
+            raise RuntimeError("To run paddle distributed training, please set `FASTNLP_BACKEND` to 'paddle' before using FastNLP.")
         if device is not None:
             logger.warning_once("Parameter `device` would be ignored when you are using `paddle.distributed.launch` to pull "
-                           "up your script. And we will directly get the local device via "
-                           "and `os.environ['CUDA_VISIBLE_DEVICES']``.")
-        device = [int(g) for g in os.environ["CUDA_VISIBLE_DEVICES"].split(",")]
-        # TODO 目前一个进程仅对应一个卡，所以暂时传入一个 int
+                           "up your script. And we will directly get the local device via environment variables.")
+        _visible_list = user_visible_devices.split(",")
+        device = [ f"gpu:{_visible_list.index(g) }" for g in os.environ["CUDA_VISIBLE_DEVICES"].split(",")]
+        # TODO 目前一个进程仅对应一个卡，所以暂时传入单个
         return PaddleFleetDriver(model, device[0], True, **kwargs)
 
-    user_visible_devices = os.getenv("USER_CUDA_VISIBLE_DEVICES")
     if user_visible_devices is None:
-        raise RuntimeError("`USER_CUDA_VISIBLE_DEVICES` cannot be None, please check if you have set "
-                            "`FASTNLP_BACKEND` to 'paddle' before using FastNLP.")
-    _could_use_device_num = len(user_visible_devices.split(","))
+        _could_use_device_num = paddle.device.cuda.device_count()
+    else:
+        _could_use_device_num = len(user_visible_devices.split(","))
+
     if isinstance(device, int):
         if device < 0 and device != -1:
             raise ValueError("Parameter `device` can only be '-1' when it is smaller than 0.")
         if device >= _could_use_device_num:
             raise ValueError("The gpu device that parameter `device` specifies is not existed.")
         if device == -1:
-            device = list(range(_could_use_device_num))
+            device = [ get_paddle_gpu_str(g) for g in range(_could_use_device_num)]
     elif isinstance(device, Sequence) and not isinstance(device, str):
         device = list(set(device))
         for each in device:
@@ -61,8 +69,10 @@ def initialize_paddle_driver(driver: str, device: Optional[Union[str, int, List[
             elif each >= _could_use_device_num:
                 raise ValueError("When parameter `device` is 'Sequence' type, the value in it should not be bigger than"
                                  " the available gpu number.")
+        device = [get_paddle_gpu_str(g) for g in device]
     elif device is not None and not isinstance(device, str):
         raise ValueError("Parameter `device` is wrong type, please check our documentation for the right use.")
+    
     if isinstance(device, List):
         return PaddleFleetDriver(model, device, **kwargs)
     else:

fastNLP/core/drivers/paddle_driver/paddle_driver.py

@@ -7,10 +7,13 @@ from dataclasses import dataclass
 
 import numpy as np
 
+from fastNLP.envs.env import USER_CUDA_VISIBLE_DEVICES
+
 from .utils import _build_fp16_env, optimizer_state_to_device, DummyGradScaler
 from fastNLP.envs.imports import _NEED_IMPORT_PADDLE
 from fastNLP.core.drivers.driver import Driver
-from fastNLP.core.utils import apply_to_collection, paddle_move_data_to_device, get_device_from_visible
+from fastNLP.core.utils import apply_to_collection, paddle_move_data_to_device
+from fastNLP.core.utils.paddle_utils import _convert_data_device
 from fastNLP.envs import (
     FASTNLP_SEED_WORKERS,
     FASTNLP_MODEL_FILENAME,
@@ -369,7 +372,7 @@ class PaddleDriver(Driver):
 
         :return: 将移动到指定机器上的 batch 对象返回；
         """
-        device = get_device_from_visible(self.data_device)
+        device = _convert_data_device(self.data_device)
         return paddle_move_data_to_device(batch, device)
 
     @staticmethod

fastNLP/core/drivers/paddle_driver/single_device.py

@@ -8,10 +8,10 @@ from fastNLP.envs.imports import _NEED_IMPORT_PADDLE
 from fastNLP.envs.env import USER_CUDA_VISIBLE_DEVICES
 from fastNLP.core.utils import (
     auto_param_call,
-    get_device_from_visible,
     get_paddle_gpu_str,
     get_paddle_device_id,
 )
+from fastNLP.core.utils.paddle_utils import _convert_data_device
 from fastNLP.core.utils.utils import _get_fun_msg
 from fastNLP.core.samplers import (
     ReproducibleBatchSampler,
@@ -40,9 +40,6 @@ class PaddleSingleDriver(PaddleDriver):
             raise ValueError("`paddle.DataParallel` is not supported in `PaddleSingleDriver`")
 
         cuda_visible_devices = os.getenv(USER_CUDA_VISIBLE_DEVICES)
-        if cuda_visible_devices is None:
-            raise RuntimeError("`USER_CUDA_VISIBLE_DEVICES` cannot be None, please check if you have set "
-                            "`FASTNLP_BACKEND` to 'paddle' before using FastNLP.")
         if cuda_visible_devices == "":
             device = "cpu"
             logger.info("You have set `CUDA_VISIBLE_DEVICES` to '' in system environment variable, and we are gonna to"
@@ -54,11 +51,9 @@ class PaddleSingleDriver(PaddleDriver):
             raise ValueError("Parameter `device` can not be None in `PaddleSingleDriver`.")
 
         if device != "cpu":
-            if isinstance(device, int):
-                device_id = device
-            else:
-                device_id = get_paddle_device_id(device)
-            os.environ["CUDA_VISIBLE_DEVICES"] = cuda_visible_devices.split(",")[device_id]
+            device_id = get_paddle_device_id(device)
+            if cuda_visible_devices is not None:
+                os.environ["CUDA_VISIBLE_DEVICES"] = cuda_visible_devices.split(",")[device_id]
         self.model_device = get_paddle_gpu_str(device)
 
         self.local_rank = 0
@@ -69,7 +64,8 @@ class PaddleSingleDriver(PaddleDriver):
         r"""
         该函数用来初始化训练环境，用于设置当前训练的设备，并将模型迁移到对应设备上。
         """
-        device = get_device_from_visible(self.model_device, output_type=str)
+        device = _convert_data_device(self.data_device)
+
         paddle.device.set_device(device)
         with contextlib.redirect_stdout(None):
             self.model.to(device)

fastNLP/core/drivers/torch_driver/initialize_torch_driver.py

@@ -10,19 +10,18 @@ from .ddp import TorchDDPDriver
 from fastNLP.core.log import logger
 from fastNLP.envs import FASTNLP_BACKEND_LAUNCH
 
+__all__ = []
 
 def initialize_torch_driver(driver: str, device: Optional[Union[str, "torch.device", int, List[int]]],
                             model: "torch.nn.Module", **kwargs) -> TorchDriver:
     r"""
-    用来根据参数 `driver` 和 `device` 来确定并且初始化一个具体的 `Driver` 实例然后返回回去；
-    注意如果输入的 `device` 如果和 `driver` 对应不上就直接报错；
+    用来根据参数 ``driver` 和 ``device`` 来确定并且初始化一个具体的 ``Driver`` 实例然后返回回去；
 
-    :param driver: 该参数的值应为以下之一：["torch", "torch_ddp", "fairscale"]；
-    :param device: 该参数的格式与 `Trainer` 对参数 `device` 的要求一致；
+    :param driver: 该参数的值应为以下之一：``["torch", "fairscale"]``；
+    :param device: 该参数的格式与 ``Trainer`` 对参数 ``device`` 的要求一致；
     :param model: 训练或者评测的具体的模型；
 
-    :return: 返回一个元组，元组的第一个值是具体的基于 pytorch 的 `Driver` 实例，元组的第二个值是该 driver 的名字（用于检测一个脚本中
-     先后 driver 的次序的正确问题）；
+    :return: 返回一个 :class:`~fastNLP.core.TorchSingleDriver` 或 :class:`~fastNLP.core.TorchDDPDriver` 实例；
     """
     # world_size 和 rank
     if FASTNLP_BACKEND_LAUNCH in os.environ:
@@ -55,8 +54,8 @@ def initialize_torch_driver(driver: str, device: Optional[Union[str, "torch.devi
             elif each < 0:
                 raise ValueError("When parameter `device` is 'Sequence' type, the value in it should be bigger than 0.")
             elif each >= _could_use_device_num:
-                raise ValueError("When parameter `device` is 'Sequence' type, the value in it should not be bigger than"
-                                 " the available gpu number.")
+                raise ValueError(f"When parameter `device` is 'Sequence' type, the value in it should not be bigger than"
+                                 f" the available gpu number:{_could_use_device_num}.")
         device = [torch.device(f"cuda:{w}") for w in device]
     elif device is not None and not isinstance(device, torch.device):
         raise ValueError("Parameter `device` is wrong type, please check our documentation for the right use.")

fastNLP/core/drivers/torch_driver/torch_driver.py

@@ -167,6 +167,12 @@ class TorchDriver(Driver):
         """
         model = self.unwrap_model()
         res = torch.load(filepath, map_location='cpu')
+        if isinstance(res, dict) and only_state_dict is False:
+            logger.rank_zero_warning(f"It seems like that {filepath} only contains state, you may need to use "
+                                     f"`only_state_dict=True`")
+        elif not isinstance(res, dict) and only_state_dict is True:
+            logger.rank_zero_warning(f"It seems like that {filepath} is not state, you may need to use "
+                                     f"`only_state_dict=False`")
         if only_state_dict:
             model.load_state_dict(res)
         else:

fastNLP/core/drivers/torch_paddle_driver/__init__.py

@@ -1,5 +0,0 @@
-__all__ = [
-    "TorchPaddleDriver",
-]
-
-from .torch_paddle_driver import TorchPaddleDriver

fastNLP/core/drivers/torch_paddle_driver/torch_paddle_driver.py

@@ -1,193 +0,0 @@
-from typing import Optional, Dict, Union, Callable, Tuple
-
-from fastNLP.envs.imports import _NEED_IMPORT_PADDLE, _NEED_IMPORT_TORCH
-from fastNLP.core.utils.utils import _get_fun_msg
-
-
-if _NEED_IMPORT_PADDLE:
-    import paddle
-    from paddle.io import DataLoader as PaddleDataLoader
-    from paddle.optimizer import Optimizer as PaddleOptimizer
-
-if _NEED_IMPORT_TORCH:
-    import torch
-    from torch.utils.data import DataLoader as TorchDataLoader
-    from torch.optim import Optimizer as TorchOptimizer
-
-from fastNLP.core.drivers.driver import Driver
-from fastNLP.envs.distributed import rank_zero_call
-from fastNLP.core.utils.utils import auto_param_call, apply_to_collection
-from fastNLP.core.log.logger import logger
-from fastNLP.modules.mix_modules.mix_module import MixModule
-
-
-__all__ = [
-    "TorchPaddleDriver",
-]
-
-class TorchPaddleDriver(Driver):
-    """
-    针对torch和paddle混合模型的driver
-    由于是两种不同的框架不方便实现多卡，暂时先实现CPU和GPU单卡的功能
-    """
-    def __init__(self, model, device: Optional[str] = None, **kwargs):
-        super(TorchPaddleDriver, self).__init__(model)
-
-        self.model_device = device
-        self.torch_non_blocking = kwargs.get("torch_non_blocking", None)
-        self.paddle_blocking = kwargs.get("paddle_blocking", None)
-
-        self._data_device = kwargs.get("_data_device", None)
-        if isinstance(self._data_device, int):
-            # 将data_device设置为cuda:x的字符串形式
-            if self._data_device < 0:
-                raise ValueError("Parameter `_data_device` can not be smaller than 0.")
-            _could_use_device_num = paddle.device.cuda.device_count()
-            if self._data_device >= _could_use_device_num:
-                raise ValueError("The gpu device that parameter `device` specifies is not existed.")
-            self._data_device = f"cuda:{self._data_device}"
-        elif self._data_device is not None:
-            raise ValueError("Parameter `device` is wrong type, please check our documentation for the right use.")
-
-    def setup(self):
-        if self.model_device is not None:
-            paddle.device.set_device(self.model_device.replace("cuda", "gpu"))
-            self.model.to(self.model_device)
-
-    @staticmethod
-    def check_dataloader_legality(dataloader, dataloader_name, is_train: bool = False):
-        if is_train:
-            if not isinstance(dataloader, (TorchDataLoader, PaddleDataLoader)):
-                raise ValueError(f"Parameter `{dataloader_name}` should be 'torch.util.data.DataLoader' or `paddle.io.dataloader` type, not {type(dataloader)}.")
-        else:
-            if not isinstance(dataloader, Dict):
-                raise ValueError(f"Parameter `{dataloader_name}` should be 'Dict' type, not {type(dataloader)}.")
-            else:
-                for each_dataloader in dataloader.values():
-                    if not isinstance(each_dataloader, (TorchDataLoader, PaddleDataLoader)):
-                        raise ValueError(f"Each dataloader of parameter `{dataloader_name}` should be "
-                                         f"'torch.util.data.DataLoader' or `paddle.io.dataloader` "
-                                         f"type, not {type(each_dataloader)}.")
-
-    @staticmethod
-    def _check_optimizer_legality(optimizers):
-        for each_optimizer in optimizers:
-            if not isinstance(each_optimizer, (TorchOptimizer, PaddleOptimizer)):
-                raise ValueError(f"Each optimizers of parameter `optimizers` should be "
-                                 f"'torch.optim.Optimizer' or 'paddle.optimizers.Optimizer' type, "
-                                 f"not {type(each_optimizer)}.")
-
-    def step(self):
-        for optimizer in self.optimizers:
-            optimizer.step()
-
-    def backward(self, loss):
-        loss.backward()
-
-    def zero_grad(self):
-        for optimizer in self.optimizers:
-            if isinstance(optimizer, TorchOptimizer):
-                optimizer.zero_grad()
-            elif isinstance(optimizer, PaddleOptimizer):
-                optimizer.clear_grad()
-            else:
-                raise ValueError("Unknown optimizers type.")
-
-    def model_call(self, batch, fn: Callable, signature_fn: Optional[Callable]) -> Dict:
-        if isinstance(batch, Dict) and not self.wo_auto_param_call:
-            return auto_param_call(fn, batch, signature_fn=signature_fn)
-        else:
-            return fn(batch)
-
-    def get_model_call_fn(self, fn: str) -> Tuple:
-        if hasattr(self.model, fn):
-            fn = getattr(self.model, fn)
-            if not callable(fn):
-                raise RuntimeError(f"The `{fn}` attribute is not `Callable`.")
-            logger.debug(f'Use {_get_fun_msg(fn, with_fp=False)}...')
-            return fn, None
-        elif fn in {"train_step", "evaluate_step"}:
-            logger.debug(f'Use {_get_fun_msg(self.model.forward, with_fp=False)}...')
-            return self.model, self.model.forward
-        else:
-            raise RuntimeError(f"There is no `{fn}` method in your {type(self.model)}.")
-
-    def predict_step(self, batch):
-        if isinstance(batch, Dict):
-            return auto_param_call(self._predict_step, batch)
-        else:
-            return self._predict_step(batch)
-
-    @rank_zero_call
-    def save_model(self, filepath: str, only_state_dict: bool = True, model_save_fn: Optional[Callable] = None):
-        r"""
-        暂时不提供保存整个模型的方法
-        """
-        if only_state_dict == False:
-            logger.warn("TorchPaddleModule only support saving state dicts now.")
-        if model_save_fn is not None:
-            model_save_fn(filepath)
-        else:
-            model = self.unwrap_model()
-            self.move_model_to_device(model, "cpu")
-            self.model.save(filepath)
-            self.move_model_to_device(model, self.model_device)
-
-    def load_model(self, filepath: str):
-        """
-        加载模型的加载函数；
-
-        :param filepath: 保存文件的文件位置（需要包括文件名）；
-        :return:
-        """
-        return self.model.load(filepath)
-
-    def save(self):
-        ...
-
-    def load(self):
-        ...
-
-    @staticmethod
-    def move_model_to_device(model: MixModule, device: str):
-        if device is not None:
-            model.to(device)
-
-    def unwrap_model(self):
-        return self.model
-
-    @staticmethod
-    def tensor_to_numeric(tensor):
-        if tensor is None:
-            return None
-
-        def _translate(_data):
-            return _data.tolist()
-
-        return apply_to_collection(
-            data=tensor,
-            dtype=(paddle.Tensor, torch.Tensor),
-            function=_translate
-        )
-
-    def set_model_mode(self, mode: str):
-        assert mode in {"train", "eval"}
-        getattr(self.model, mode)()
-
-    def get_model_device(self):
-        return self.model_device
-
-    @property
-    def data_device(self):
-        if self.model_device is not None:
-            return self.model_device
-        else:
-            return self._data_device
-
-    def set_model_mode(self, mode: str):
-        assert mode in {"train", "eval"}
-        getattr(self.model, mode)()
-
-    def set_sampler_epoch(self, dataloader: Union['TorchDataLoader', 'PaddleDataLoader'], cur_epoch_idx):
-        # 保证 ddp 训练时的 shuffle=True 时的正确性，因为需要保证每一个进程上的 sampler 的shuffle 的随机数种子是一样的；
-        return dataloader

fastNLP/core/drivers/torch_paddle_driver/utils.py

@@ -1,4 +0,0 @@
-from fastNLP.envs.imports import _NEED_IMPORT_PADDLE
-
-if _NEED_IMPORT_PADDLE:
-    pass

fastNLP/core/log/print.py

@@ -1,7 +1,7 @@
 __all__ = [
     'print'
 ]
-
+from logging import INFO
 from .logger import logger
 
 
@@ -22,4 +22,6 @@ def print(*args, sep=' ', end='\n', file=None, flush=False):
     :return:
     """
     line = sep.join(map(str, args))
-    logger.info(line)
+    if logger.isEnabledFor(INFO):
+        kwargs = logger._add_rank_info({})
+        logger._log(INFO, line, args, **kwargs)

fastNLP/core/metrics/backend/paddle_backend/backend.py

@@ -1,12 +1,14 @@
+import os
 from typing import List, Any
 
 import numpy as np
 
 from fastNLP.core.metrics.backend import Backend
-from fastNLP.core.utils.paddle_utils import paddle_to, get_device_from_visible
+from fastNLP.core.utils.paddle_utils import paddle_to, _convert_data_device
 from fastNLP.core.metrics.utils import AggregateMethodError
 from fastNLP.core.drivers.paddle_driver.dist_utils import fastnlp_paddle_all_gather
 from fastNLP.envs.imports import _NEED_IMPORT_PADDLE
+from fastNLP.envs.env import USER_CUDA_VISIBLE_DEVICES
 
 if _NEED_IMPORT_PADDLE:
     import paddle
@@ -79,7 +81,7 @@ class PaddleBackend(Backend):
             raise ValueError(f"tensor: {tensor} can not convert to ndarray!")
 
     def move_tensor_to_device(self, tensor, device):
-        device = get_device_from_visible(device)
+        device = _convert_data_device(device)
         return paddle_to(tensor, device)
 
     def all_gather_object(self, obj, group=None) -> List:

fastNLP/core/metrics/metric.py

@@ -84,7 +84,7 @@ class Metric:
     def _sync_get_metric(self, get_metric):
         @functools.wraps(get_metric)
         def _wrap_get_metric(*args, **kwargs):
-            assert self._updated, f"You have to call `{self.__class__.__name__}` update() function before calling " \
+            assert self._updated, f"You have to call `{self.__class__.__name__}'s update() function before calling " \
                                   f"get_metric()."
             with self.sync(recover=True, aggregate=self.aggregate_when_get_metric):
                 results = get_metric(*args, **kwargs)

fastNLP/core/samplers/reproducible_batch_sampler.py

@@ -366,17 +366,22 @@ class BucketedBatchSampler(ReproducibleBatchSampler):
     def __init__(self, dataset, length: Union[List[int], str], batch_size:int = 32, num_batch_per_bucket:int = 10,
                  shuffle: bool = True, drop_last: bool = False, seed: int = 0, **kwargs):
         """
-        首先按照 sample 的长度排序，然后按照 batch_size*num_batch_per_bucket 为一个桶的大小，sample 只会在这个桶内进行组合，这样
-        每个 batch 中的 padding 数量会比较少 （因为桶内的数据的长度都接近）。
+        首先按照 ``sample`` 的长度排序，然后按照 batch_size*num_batch_per_bucket 为一个桶的大小，``sample`` 只会在这个桶内进行组
+        合，这样每个 ``batch`` 中的 ``padding`` 数量会比较少 （因为桶内的数据的长度都接近）。
 
         :param dataset: 实现了 __len__ 方法的数据容器。
-        :param length: 如果为 List，应当与 dataset 有一样的长度，表示 dataset 中每个元素的数量；仅当传入的 dataset 为 fastNLP 的
-            DataSet 时支持传入 str，会将该str理解为 dataset 的 field 名称，若 field 中的元素为 int，则认为该值是 sample 的长度。
-            如果否则使用 len() 函数得到每个 sample 中这个 field 的长度。
+        :param length: 每条数据的长度。
+
+            * 为 ``List[int]`` 时
+             应当与 dataset 有一样的长度，表示 dataset 中每个元素的数量；
+            * 为 ``str`` 时
+             仅当传入的 ``dataset`` 是 :class:`fastNLP.DataSet` 时，允许传入 `str` ，该 `str` 将被认为是 ``dataset`` 中的
+              ``field`` 。若 field 中的元素为 ``int``，则认为该值是 sample 的长度；若不为 ``int`` ，则尝试使用 ``len`` 方法
+              获取该 ``field`` 中每个元素的长度。
         :param batch_size: 每个 batch 的大小
-        :param num_batch_per_bucket: 多少个 batch 组成一个桶，数据只会在一个桶内进行 shuffle 。
-        :param shuffle: 如果为 True，将不进行 shuffle，实际上数据会以从长到短的方式输出。
-        :param drop_last: 如果最后一个 batch 的 sample 数量无法凑齐 batch_size 这么多，是否需要丢掉。
+        :param num_batch_per_bucket: 多少个 ``batch`` 组成一个桶，数据只会在一个桶内进行 ``shuffle`` 。
+        :param shuffle: 如果为 True，将不进行 ``shuffle``，实际上数据会以从长到短的方式输出。
+        :param drop_last: 如果最后一个 `batch` 的 ``sample`` 数量无法凑齐 ``batch_size`` 这么多，是否需要丢掉。
         :param seed: 设置的随机数种子
         :param kwargs: fastNLP 保留使用
         """
@@ -386,10 +391,12 @@ class BucketedBatchSampler(ReproducibleBatchSampler):
             if not isinstance(length[0], int):
                 length = list(map(len, length))
         else:
-            assert len(length) == len(dataset), "When the dataset is not fastNLP.DataSet, " \
-                                              "the length parameter can only be List[int]"
+            types = set(map(type, length))
+            assert isinstance(length, list) and len(types)==1 and types.pop()==int, \
+                "When the dataset is not fastNLP.DataSet, the length parameter can only be List[int]"
 
-        assert len(length) == len(dataset), "The length of `data` and `length` should be equal."
+        assert len(length) == len(dataset), f"The length of `dataset`({len(dataset)}) and " \
+                                            f"`length`({len(length)}) should be equal."
 
         self.dataset = dataset
         self.length = np.array(length, dtype=int)  # 按照长到短排列的序号。

fastNLP/core/samplers/reproducible_sampler.py

@@ -55,6 +55,7 @@ class ReproducibleSampler:
 class RandomSampler(ReproducibleSampler):
     def __init__(self, dataset, shuffle: bool = True, seed: int = 0, **kwargs):
         """
+        随机顺序的 Sampler 。
 
         :param dataset: 实现了 __len__ 方法的数据容器
         :param shuffle: 是否在每次 iterate 的时候打乱顺序。
@@ -169,9 +170,8 @@ class RandomSampler(ReproducibleSampler):
     def set_epoch(self, epoch: int) -> None:
         self.epoch = epoch
 
-    def set_distributed(self, num_replicas, rank, pad=True):
+    def set_distributed(self, num_replicas:int, rank:int, pad:bool=True):
         """
-        该方法本质上等同于 ddp 情形下的没有完成的初始化，应当在初始化该 sampler 本身后立即被调用；
 
         :param num_replicas:
         :param rank:
@@ -215,7 +215,7 @@ class RandomSampler(ReproducibleSampler):
 class SequentialSampler(RandomSampler):
     def __init__(self, dataset, **kwargs):
         """
-        按照顺序读取 dataset 。在多卡情况下，间隔读取，例如，在两卡情况下，卡0取 [0,2,4,..], 卡1取 [1,3,5...]。
+        按照顺序读取 ``dataset`` 。在多卡情况下，间隔读取，例如，在两卡情况下，卡 0 取 ``[0,2,4,..]``, 卡1取 ``[1,3,5...]`` 。
 
         :param dataset: 实现了 __len__ 方法的数据容器。
         :param kwargs:
@@ -285,13 +285,20 @@ class SequentialSampler(RandomSampler):
 class SortedSampler(SequentialSampler):
     def __init__(self, dataset, length:Union[str, List], **kwargs):
         """
-        将 dataset 中的数据根据 length 从长到短进行迭代。在多卡情况下，由于padding 最后一个 sample 可能是最长的那个 sample。
+        将 ``dataset`` 中的数据根据 ``length`` 从长到短进行迭代。在多卡情况下，由于 ``padding`` , 最后一个 ``sample`` 可能是最长
+        的那个 ``sample`` 。
 
         :param dataset: 实现了 __len__ 方法的数据容器。
-        :param length: 如果为 List，应当与 dataset 有一样的长度，表示 dataset 中每个元素的数量；仅当传入的 dataset 为 fastNLP 的
-            DataSet 时支持传入 str，会将该str理解为 dataset 的 field 名称，若 field 中的元素为 int，则认为该值是 sample 的长度。
-        :param seed: 设置的随机数种子
-        :param kwargs: fastNLP 保留使用
+        :param length: 每条数据的长度。
+
+            * 为 ``List[int]`` 时
+             应当与 dataset 有一样的长度，表示 dataset 中每个元素的数量；
+            * 为 ``str`` 时
+             仅当传入的 ``dataset`` 是 :class:`fastNLP.DataSet` 时，允许传入 `str` ，该 `str` 将被认为是 ``dataset`` 中的
+              ``field`` 。若 field 中的元素为 ``int``，则认为该值是 sample 的长度；若不为 ``int`` ，则尝试使用 ``len`` 方法
+              获取该 ``field`` 中每个元素的长度。
+        :param seed: 设置的随机数种子。
+        :param kwargs: fastNLP 保留使用。
         """
         super().__init__(dataset=dataset, **kwargs)
         if isinstance(dataset, DataSet) and isinstance(length, str):
@@ -299,8 +306,9 @@ class SortedSampler(SequentialSampler):
             if not isinstance(length[0], int):
                 length = list(map(len, length))
         else:
-            assert len(length) == len(dataset), "When the dataset is not fastNLP.DataSet, " \
-                                                "the length parameter can only be List[int]"
+            types = set(map(type, length))
+            assert isinstance(length, list) and len(types)==1 and types.pop()==int, \
+                "When the dataset is not fastNLP.DataSet, the length parameter can only be List[int]"
 
         assert len(length) == len(dataset), "The length of `data` and `length` should be equal."
 

fastNLP/core/utils/__init__.py

@@ -2,7 +2,6 @@ __all__ = [
     'cache_results',
     'is_jittor_dataset',
     'jittor_collate_wraps',
-    'get_device_from_visible',
     'paddle_to',
     'paddle_move_data_to_device',
     'get_paddle_device_id',
@@ -11,7 +10,6 @@ __all__ = [
     'is_in_fnlp_paddle_dist',
     'is_in_paddle_launch_dist',
     'f_rich_progress',
-    'torch_paddle_move_data_to_device',
     'torch_move_data_to_device',
     'get_fn_arg_names',
     'auto_param_call',
@@ -29,10 +27,9 @@ __all__ = [
 
 from .cache_results import cache_results
 from .jittor_utils import is_jittor_dataset, jittor_collate_wraps
-from .paddle_utils import get_device_from_visible, paddle_to, paddle_move_data_to_device, get_paddle_device_id, get_paddle_gpu_str, is_in_paddle_dist, \
+from .paddle_utils import paddle_to, paddle_move_data_to_device, get_paddle_device_id, get_paddle_gpu_str, is_in_paddle_dist, \
     is_in_fnlp_paddle_dist, is_in_paddle_launch_dist
 from .rich_progress import f_rich_progress
-from .torch_paddle_utils import torch_paddle_move_data_to_device
 from .torch_utils import torch_move_data_to_device
 from .utils import *
 

fastNLP/core/utils/dummy_class.py

@@ -1,4 +1,4 @@
-import functools
+__all__ = []
 
 class DummyClass:
     def __init__(self, *args, **kwargs):

fastNLP/core/utils/jittor_utils.py

@@ -15,6 +15,12 @@ from fastNLP.core.dataset import Instance
 
 
 def is_jittor_dataset(dataset) -> bool:
+    """
+    判断传入的 ``dataset`` 是否是 :class:`jittor.dataset.Dataset` 类型
+
+    :param dataset: 数据集；
+    :return: 当前 ``dataset`` 是否为 ``jittor`` 的数据集类型；
+    """
     try:
         if isinstance(dataset, jt.dataset.Dataset):
             return True
@@ -26,7 +32,8 @@ def is_jittor_dataset(dataset) -> bool:
 
 def jittor_collate_wraps(func, auto_collator: Callable):
     """
-    对jittor的collate_fn进行wrap封装, 如果数据集为mapping类型，那么采用auto_collator，否则还是采用jittor自带的collate_batch
+    对 ``jittor`` 的 ``collate_fn`` 进行 ``wrap`` 封装,。如果数据集为 ``mapping`` 类型，那么采用 ``auto_collator`` ，否则
+    还是采用 ``jittor`` 的 ``collate_batch``。
 
     :param func:
     :param auto_collator:

fastNLP/core/utils/paddle_utils.py

@@ -1,5 +1,4 @@
 __all__ = [
-    "get_device_from_visible",
     "paddle_to",
     "paddle_move_data_to_device",
     "get_paddle_gpu_str",
@@ -21,73 +20,90 @@ if _NEED_IMPORT_PADDLE:
 
 from .utils import apply_to_collection
 
-def get_device_from_visible(device: Union[str, int], output_type=int):
+def _convert_data_device(device: Union[str, int]) -> str:
     """
-    在有 CUDA_VISIBLE_DEVICES 的情况下，获取对应的设备。
-    如 CUDA_VISIBLE_DEVICES=2,3 ，device=3 ，则返回1。
+    用于转换 ``driver`` 的 ``data_device`` 的函数。如果用户设置了 ``FASTNLP_BACKEND=paddle``，那么 ``fastNLP`` 会将
+    可见的设备保存在 ``USER_CUDA_VISIBLE_DEVICES`` 中，并且将 ``CUDA_VISIBLE_DEVICES`` 设置为可见的第一张显卡；这是为
+    了顺利执行 ``paddle`` 的分布式训练而设置的。
+    
+    在这种情况下，单纯使用 ``driver.data_device`` 是无效的。比如在分布式训练中将设备设置为 ``[0,2,3]`` ，且用户设置了
+    ``CUDA_VISIBLE_DEVICES=3,4,5,6`` ，那么在 ``rank1``的进程中有::
 
-    :param device: 未转化的设备名
-    :param output_type: 返回值的类型
-    :return: 转化后的设备id
-    """
-    if output_type not in [int, str]:
-        raise ValueError("Parameter `output_type` should be one of these types: [int, str]")
-    if device == "cpu":
-        return device
-    cuda_visible_devices = os.getenv("CUDA_VISIBLE_DEVICES")
-    user_visible_devices = os.getenv(USER_CUDA_VISIBLE_DEVICES)
-    if user_visible_devices is None:
-        raise RuntimeError("`USER_CUDA_VISIBLE_DEVICES` cannot be None, please check if you have set "
-                            "`FASTNLP_BACKEND` to 'paddle' before using FastNLP.")
-    idx = get_paddle_device_id(device)
-    # 利用 USER_CUDA_VISIBLDE_DEVICES 获取用户期望的设备
-    if user_visible_devices is None:
-        raise RuntimeError("This situation cannot happen, please report a bug to us.")
-    idx = user_visible_devices.split(",")[idx]
-
-    cuda_visible_devices_list = cuda_visible_devices.split(',')
-    if idx not in cuda_visible_devices_list:
-        raise ValueError(f"Can't find your devices {idx} in CUDA_VISIBLE_DEVICES[{cuda_visible_devices}]. ")
-    res = cuda_visible_devices_list.index(idx)
-    if output_type == int:
-        return res
-    else:
-        return f"gpu:{res}"
+        os.environ["CUDA_VISIBLE_DEVICES"] = "5"
+        os.environ["USER_CUDA_VISIBLE_DEVICES"] = "3,4,5,6"
+        driver.data_device = "gpu:2" # 为了向用户正确地反映他们设置的设备减少歧义，因此这里没有设置为 "gpu:5"
+
+    此时我们便需要通过这个函数将 ``data_device`` 转换为 ``gpu:0``。具体过程便是通过索引 **2** 在 ``USER_CUDA_VISIBLE_DEVICES`` 中
+    找到设备 **5**，然后在 ``CUDA_VISIBLE_DEVICES`` 中找到设备 **5** 的索引 **0** 返回。
+
+    .. note::
 
-def paddle_to(data, device: Union[str, int]):
+        在分布式单进程仅支持单卡的情况下中，这个函数实际等同于直接转换为 ``gpu:0`` 返回。
+
+    :param device: 未转化的设备；
+    :return: 转化后的设备，格式为 ``gpu:x``；
+    """
+    try:
+        user_visible_devices = os.getenv(USER_CUDA_VISIBLE_DEVICES)
+        if device == "cpu" or user_visible_devices is None:
+            # 传入的是 CPU，或者没有设置 USER_CUDA_VISIBLE_DEVICES
+            # 此时不需要进行转换
+            return get_paddle_gpu_str(device)
+
+        idx = get_paddle_device_id(device)
+        idx = user_visible_devices.split(",")[idx]
+        # 此时 CUDA_VISIBLE_DEVICES 一定不是 None
+        cuda_visible_devices_list = os.getenv("CUDA_VISIBLE_DEVICES").split(',')
+        return f"gpu:{cuda_visible_devices_list.index(idx)}"
+    except Exception as e:
+        raise ValueError(f"Can't convert device {device} when USER_CUDA_VISIBLE_DEVICES={user_visible_devices} "
+                        "and CUDA_VISIBLE_DEVICES={cuda_visible_devices}. If this situation happens, please report this bug to us.")
+
+def paddle_to(data: "paddle.Tensor", device: Union[str, int]) -> "paddle.Tensor":
     """
-    将 `data` 迁移到指定的 `device` 上
+    将 ``data`` 迁移到指定的 ``device`` 上。``paddle.Tensor`` 没有类似 ``torch.Tensor`` 的 ``to`` 函数，该函数
+    只是集成了 :func:`paddle.Tensor.cpu` 和 :func:`paddle.Tensor.cuda` 两个函数。
 
-    :param data: 要迁移的张量
-    :param device: 目标设备，可以是 `str` 或 `int`
-    :return: 迁移后的张量
+    :param data: 要迁移的张量；
+    :param device: 目标设备，可以是 ``str`` 或 ``int`` 类型；
+    :return: 迁移后的张量；
     """
 
     if device == "cpu":
         return data.cpu()
     else:
-        # device = get_device_from_visible(device, output_type=int)
         return data.cuda(get_paddle_device_id(device))
 
-
-def get_paddle_gpu_str(device: Union[str, int]):
+def get_paddle_gpu_str(device: Union[str, int]) -> str:
     """
-    获得 `gpu:x` 类型的设备名
+    获得 ``gpu:x`` 格式的设备名::
 
-    :param device: 设备编号或设备名
-    :return: 返回对应的 `gpu:x` 格式的设备名
+        >>> get_paddle_gpu_str(1)
+        'gpu:1'
+        >>> get_paddle_gpu_str("cuda:1")
+        'gpu:1'
+
+    :param device: 设备编号或设备名；
+    :return: 返回对应的 ``gpu:x`` 格式的设备名；
     """
     if isinstance(device, str):
         return device.replace("cuda", "gpu")
     return f"gpu:{device}"
 
 
-def get_paddle_device_id(device: Union[str, int]):
+def get_paddle_device_id(device: Union[str, int]) -> int:
     """
-    获得 gpu 的设备id
+    获得 ``device`` 的设备编号::
+
+        >>> get_paddle_device_id("gpu:1")
+        1
+        >>> get_paddle_device_id("gpu")
+        0
+
+    请注意不要向这个函数中传入 ``cpu``。
 
-    :param: device: 设备编号或设备名
-    :return: 设备对应的编号
+    :param: device: 设备编号或设备名；
+    :return: 设备对应的编号；
     """
     if isinstance(device, int):
         return device
@@ -109,21 +125,17 @@ def get_paddle_device_id(device: Union[str, int]):
 
     return device_id
 
-def paddle_move_data_to_device(batch: Any, device: Optional[str] = None,
-                              data_device: Optional[str] = None) -> Any:
+def paddle_move_data_to_device(batch: Any, device: Optional[Union[str, int]]) -> Any:
     r"""
-    将数据集合传输到给定设备。只有paddle.Tensor对象会被传输到设备中，其余保持不变
+    将 ``paddle`` 的数据集合传输到给定设备。只有 :class:`paddle.Tensor` 对象会被传输到设备中，其余保持不变。
 
-    :param batch:
-    :param device: `cpu`, `gpu` or `gpu:x`
-    :param data_device:
-    :return: 相同的集合，但所有包含的张量都驻留在新设备上；
+    :param batch: 需要进行迁移的数据集合；
+    :param device: 目标设备。可以是显卡设备的编号，或是``cpu``, ``gpu`` 或 ``gpu:x`` 格式的字符串；当这个参数
+        为 `None`` 时，不会执行任何操作。
+    :return: 迁移到新设备上的数据集合；
     """
     if device is None:
-        if data_device is not None:
-            device = data_device
-        else:
-            return batch
+        return batch
 
     def batch_to(data: Any) -> Any:
         return paddle_to(data, device)
@@ -131,22 +143,22 @@ def paddle_move_data_to_device(batch: Any, device: Optional[str] = None,
     return apply_to_collection(batch, dtype=paddle.Tensor, function=batch_to)
 
 
-def is_in_paddle_dist():
+def is_in_paddle_dist() -> bool:
     """
-    判断是否处于分布式的进程下，使用 global_rank 和 selected_gpus 判断
+    判断是否处于 ``paddle`` 分布式的进程下，使用 ``PADDLE_RANK_IN_NODE`` 和 ``FLAGS_selected_gpus`` 判断。
     """
     return ('PADDLE_RANK_IN_NODE' in os.environ and 'FLAGS_selected_gpus' in os.environ)
 
 
-def is_in_fnlp_paddle_dist():
+def is_in_fnlp_paddle_dist() -> bool:
     """
-    判断是否处于 FastNLP 拉起的分布式进程中
+    判断是否处于 ``fastNLP`` 拉起的 ``paddle`` 分布式进程中
     """
     return FASTNLP_DISTRIBUTED_CHECK in os.environ
 
 
-def is_in_paddle_launch_dist():
+def is_in_paddle_launch_dist() -> bool:
     """
-    判断是否处于 launch 启动的分布式进程中
+    判断是否处于 ``python -m paddle.distributed.launch`` 方法启动的 ``paddle`` 分布式进程中
     """
     return FASTNLP_BACKEND_LAUNCH in os.environ

fastNLP/core/utils/rich_progress.py

@@ -1,7 +1,6 @@
 """
-该文件用于为fastNLP提供一个统一的progress bar管理，通过共用一个Task对象，trainer中的progress bar和evaluation中的progress bar才能
-    不冲突
-
+该文件用于为 ``fastNLP`` 提供一个统一的 ``progress bar`` 管理，通过共用一个``Task`` 对象， :class:`~fastNLP.core.Trainer` 中
+的 ``progress bar`` 和 :class:`~fastNLP.core.Evaluator` 中的 ``progress bar`` 才能不冲突
 """
 import sys
 from typing import Any, Union, Optional

fastNLP/core/utils/torch_paddle_utils.py

@@ -1,49 +0,0 @@
-from typing import Any, Optional
-
-from fastNLP.envs.imports import _NEED_IMPORT_PADDLE, _NEED_IMPORT_TORCH
-
-if _NEED_IMPORT_PADDLE:
-    import paddle
-
-if _NEED_IMPORT_TORCH:
-    import torch
-
-__all__ = [
-    "torch_paddle_move_data_to_device",
-]
-
-from .utils import apply_to_collection
-from .paddle_utils import paddle_to
-
-
-def torch_paddle_move_data_to_device(batch: Any, device: Optional[str] = None, non_blocking: Optional[bool] = True, 
-                                    data_device: Optional[str] = None) -> Any:
-    
-    r"""
-    将数据集合传输到给定设备。只有paddle.Tensor和torch.Tensor对象会被传输到设备中，其余保持不变
-
-    :param batch:
-    :param device:
-    :param non_blocking:
-    :param data_device:
-    :return: 相同的集合，但所有包含的张量都驻留在新设备上；
-    """
-
-    if device is None:
-        if data_device is not None:
-            device = data_device
-        else:
-            return batch
-
-    torch_device = device.replace("gpu", "cuda")
-    paddle_device = device.replace("cuda", "gpu")
-
-    def batch_to(data: Any) -> Any:
-        if isinstance(data, torch.Tensor):
-            data = data.to(torch_device, non_blocking=non_blocking)
-        elif isinstance(data, paddle.Tensor):
-            data = paddle_to(data, paddle_device)
-        
-        return data
-
-    return apply_to_collection(batch, dtype=(paddle.Tensor, torch.Tensor), function=batch_to)

fastNLP/core/utils/torch_utils.py

@@ -44,12 +44,12 @@ class TorchTransferableDataType(ABC):
 def torch_move_data_to_device(batch: Any, device: Optional[Union[str, "torch.device"]] = None,
                               non_blocking: Optional[bool] = True) -> Any:
     r"""
-    将数据集合传输到给定设备。任何定义方法 “to(device)” 的对象都将被移动并且集合中的所有其他对象将保持不变；
+    在 ``pytorch`` 中将数据集合 ``batch`` 传输到给定设备。任何定义方法 ``to(device)`` 的对象都将被移动并且集合中的所有其他对象将保持不变；
 
-    :param batch: 应当迁移的数据；
-    :param device: 数据应当迁移到的设备；当该参数的值为 None 时，表示迁移数据的操作由用户自己完成，我们不需要经管；
-    :param non_blocking: pytorch 的迁移数据方法 `to` 的参数；
-    :return: 相同的集合，但所有包含的张量都驻留在新设备上；
+    :param batch: 需要迁移的数据；
+    :param device: 数据应当迁移到的设备；当该参数的值为 ``None`` 时则不执行任何操作；
+    :param non_blocking: ``pytorch`` 的数据迁移方法 ``to`` 的参数；
+    :return: 迁移到新设备上的数据集合；
     """
     if device is None:
         return batch

fastNLP/core/utils/utils.py

@@ -10,10 +10,6 @@ from typing import Callable, List, Any, Dict, AnyStr, Union, Mapping, Sequence
 from typing import Tuple, Optional
 from time import sleep
 
-try:
-    from typing import Literal, Final
-except ImportError:
-    from typing_extensions import Literal, Final
 import os
 from contextlib import contextmanager
 from functools import wraps
@@ -22,7 +18,6 @@ import numpy as np
 from pathlib import Path
 
 from fastNLP.core.log import logger
-from ...envs import SUPPORT_BACKENDS
 
 
 __all__ = [
@@ -43,10 +38,16 @@ __all__ = [
 
 def get_fn_arg_names(fn: Callable) -> List[str]:
     r"""
-    返回一个函数的所有参数的名字；
+    该函数可以返回一个函数所有参数的名字::
+
+        >>> def function(a, b=1):
+        ...     return a
+        ...
+        >>> get_fn_arg_names(function)
+        ['a', 'b']
 
     :param fn: 需要查询的函数；
-    :return: 一个列表，其中的元素则是查询函数的参数的字符串名字；
+    :return: 包含函数 ``fn`` 参数名的列表；
     """
     return list(inspect.signature(fn).parameters)
 
@@ -54,24 +55,18 @@ def get_fn_arg_names(fn: Callable) -> List[str]:
 def auto_param_call(fn: Callable, *args, signature_fn: Optional[Callable] = None,
                     mapping: Optional[Dict[AnyStr, AnyStr]] = None) -> Any:
     r"""
-    该函数会根据输入函数的形参名从*args（因此都需要是dict类型）中找到匹配的值进行调用，如果传入的数据与fn的形参不匹配，可以通过mapping
-    参数进行转换。mapping参数中的一对（key，value）表示以这个key在*args中找到值，并将这个值传递给形参名为value的参数。
-
-    1.该函数用来提供给用户根据字符串匹配从而实现自动调用；
-    2.注意 mapping 默认为 None，如果你希望指定输入和运行函数的参数的对应方式，那么你应当让 mapping 为一个这样的字典传入进来；
-    如果 mapping 不为 None，那么我们一定会先使用 mapping 将输入的字典的 keys 修改过来，因此请务必亲自检查 mapping 的正确性；
-    3.如果输入的函数的参数有默认值，那么如果之后的输入中没有该参数对应的值，我们就会使用该参数对应的默认值，否则也会使用之后的输入的值；
-    4.如果输入的函数是一个 `partial` 函数，情况同 '3.'，即和默认参数的情况相同；
-
-    :param fn: 用来进行实际计算的函数，其参数可以包含有默认值；
-    :param args: 一系列的位置参数，应当为一系列的字典，我们需要从这些输入中提取 `fn` 计算所需要的实际参数；
-    :param signature_fn: 函数，用来替换 `fn` 的函数签名，如果该参数不为 None，那么我们首先会从该函数中提取函数签名，然后通过该函数签名提取
-     参数值后，再传给 `fn` 进行实际的运算；
-    :param mapping: 一个字典，用来更改其前面的字典的键值；
+    该函数会根据输入函数的形参名从 ``*args`` （均为 ``dict`` 类型）中找到匹配的值进行调用，如果传入的数据与 ``fn`` 的形参不匹配，可以通过
+    ``mapping`` 参数进行转换。``mapping`` 参数中的一对 ``(key, value)`` 表示在 ``*args`` 中找到 ``key`` 对应的值，并将这个值传递给形参中名为
+    ``value`` 的参数。
 
-    :return: 返回 `fn` 运行的结果；
+    1. 该函数用来提供给用户根据字符串匹配从而实现自动调用；
+    2. 注意 ``mapping`` 默认为 ``None``，如果你希望指定输入和运行函数的参数的对应方式，那么你应当让 ``mapping`` 为一个字典传入进来；
+       如果 ``mapping`` 不为 ``None``，那么我们一定会先使用 ``mapping`` 将输入的字典的 ``keys`` 修改过来，因此请务必亲自检查 ``mapping`` 的正确性；
+    3. 如果输入的函数的参数有默认值，那么如果之后的输入中没有该参数对应的值，我们就会使用该参数对应的默认值，否则也会使用之后的输入的值；
+    4. 如果输入的函数是一个 ``partial`` 函数，情况同第三点，即和默认参数的情况相同；
 
     Examples::
+
         >>> # 1
         >>> loss_fn = CrossEntropyLoss()  # 如果其需要的参数为 def CrossEntropyLoss(y, pred)；
         >>> batch = {"x": 20, "y": 1}
@@ -84,6 +79,14 @@ def auto_param_call(fn: Callable, *args, signature_fn: Optional[Callable] = None
         >>> print(auto_param_call(test_fn, {"x": 10}, {"y": 20, "a": 30}))  # res: 70
         >>> print(auto_param_call(partial(test_fn, a=100), {"x": 10}, {"y": 20}))  # res: 140
         >>> print(auto_param_call(partial(test_fn, a=100), {"x": 10}, {"y": 20, "a": 200}))  # res: 240
+
+    :param fn: 用来进行实际计算的函数，其参数可以包含有默认值；
+    :param args: 一系列的位置参数，应当为一系列的字典，我们需要从这些输入中提取 ``fn`` 计算所需要的实际参数；
+    :param signature_fn: 函数，用来替换 ``fn`` 的函数签名，如果该参数不为 ``None``，那么我们首先会从该函数中提取函数签名，然后通过该函数签名提取
+        参数值后，再传给 ``fn`` 进行实际的运算；
+    :param mapping: 一个字典，用来更改其前面的字典的键值；
+
+    :return: 返回 ``fn`` 运行的结果；
     """
 
     if signature_fn is not None:
@@ -164,13 +167,13 @@ def _get_keys(args:List[Dict]) -> List[List[str]]:
 
 def _get_fun_msg(fn, with_fp=True)->str:
     """
-    获取函数的基本信息，帮助报错。
-    ex:
-        print(_get_fun_msg(_get_fun_msg))
-        # `_get_fun_msg(fn) -> str`(In file:/Users/hnyan/Desktop/projects/fastNLP/fastNLP/fastNLP/core/utils/utils.py)
+    获取函数的基本信息，帮助报错::
+
+        >>>> print(_get_fun_msg(_get_fun_msg))
+        `_get_fun_msg(fn) -> str`(In file:/Users/hnyan/Desktop/projects/fastNLP/fastNLP/fastNLP/core/utils/utils.py)
 
     :param callable fn:
-    :param with_fp: 是否包含函数所在的文件信息。
+    :param with_fp: 是否包含函数所在的文件信息；
     :return:
     """
     if isinstance(fn, functools.partial):
@@ -226,13 +229,13 @@ def _check_valid_parameters_number(fn, expected_params:List[str], fn_name=None):
 
 def check_user_specific_params(user_params: Dict, fn: Callable):
     """
-    该函数使用用户的输入来对指定函数的参数进行赋值；
-    主要用于一些用户无法直接调用函数的情况；
-    该函数主要的作用在于帮助检查用户对使用函数 fn 的参数输入是否有误；
+    该函数使用用户的输入来对指定函数的参数进行赋值，主要用于一些用户无法直接调用函数的情况；
+    主要作用在于帮助检查用户对使用函数 ``fn`` 的参数输入是否有误；
 
-    :param user_params: 用户指定的参数的值，应当是一个字典，其中 key 表示每一个参数的名字，value 为每一个参数应当的值；
-    :param fn: 会被调用的函数；
-    :return: 返回一个字典，其中为在之后调用函数 fn 时真正会被传进去的参数的值；
+    :param user_params: 用户指定的参数的值，应当是一个字典，其中 ``key`` 表示每一个参数的名字，
+        ``value`` 为每一个参数的值；
+    :param fn: 将要被调用的函数；
+    :return: 返回一个字典，其中为在之后调用函数 ``fn`` 时真正会被传进去的参数的值；
     """
 
     fn_arg_names = get_fn_arg_names(fn)
@@ -243,6 +246,9 @@ def check_user_specific_params(user_params: Dict, fn: Callable):
 
 
 def dataclass_to_dict(data: "dataclasses.dataclass") -> Dict:
+    """
+    将传入的 ``dataclass`` 实例转换为字典。
+    """
     if not is_dataclass(data):
         raise TypeError(f"Parameter `data` can only be `dataclass` type instead of {type(data)}.")
     _dict = dict()
@@ -253,23 +259,33 @@ def dataclass_to_dict(data: "dataclasses.dataclass") -> Dict:
 
 def match_and_substitute_params(mapping: Optional[Union[Callable, Dict]] = None, data: Optional[Any] = None) -> Any:
     r"""
-    用来实现将输入：batch，或者输出：outputs，通过 `mapping` 将键值进行更换的功能；
-    该函数应用于 `input_mapping` 和 `output_mapping`；
-    对于 `input_mapping`，该函数会在 `TrainBatchLoop` 中取完数据后立刻被调用；
-    对于 `output_mapping`，该函数会在 `Trainer.train_step` 以及 `Evaluator.train_step` 中得到结果后立刻被调用；
+    用来实现将输入的 ``batch`` 或者输出的 ``outputs`` 通过 ``mapping`` 将键值进行更换的功能；
+    该函数应用于 ``input_mapping`` 和 ``output_mapping``；
 
-    转换的逻辑按优先级依次为：
+    * 对于 ``input_mapping``，该函数会在 :class:`~fastNLP.core.controllers.TrainBatchLoop` 中取完数据后立刻被调用；
+    * 对于 ``output_mapping``，该函数会在 :class:`~fastNLP.core.Trainer` 的 :meth:`~fastNLP.core.Trainer.train_step`
+     以及 :class:`~fastNLP.core.Evaluator` 的 :meth:`~fastNLP.core.Evaluator.train_step` 中得到结果后立刻被调用；
 
-    1. 如果 `mapping` 是一个函数，那么会直接返回 `mapping(data)`；
-    2. 如果 `mapping` 是一个 `Dict`，那么 `data` 的类型只能为以下三种： [`Dict`, `dataclass`, `Sequence`]；
-    如果 `data` 是 `Dict`，那么该函数会将 `data` 的 key 替换为 mapping[key]；
-    如果 `data` 是 `dataclass`，那么该函数会先使用 `dataclasses.asdict` 函数将其转换为 `Dict`，然后进行转换；
-    如果 `data` 是 `Sequence`，那么该函数会先将其转换成一个对应的 `Dict`：{"_0": list[0], "_1": list[1], ...}，然后使用
-    mapping对这个 `Dict` 进行转换，如果没有匹配上mapping中的key则保持"_number"这个形式。
+    转换的逻辑按优先级依次为：
 
-    :param mapping: 用于转换的字典或者函数；mapping是函数时，返回值必须为字典类型。
+    1. 如果 ``mapping`` 是一个函数，那么会直接返回 ``mapping(data)``；
+    2. 如果 ``mapping`` 是一个 ``Dict``，那么 ``data`` 的类型只能为以下三种： ``[Dict, dataclass, Sequence]``；
+        
+        * 如果 ``data`` 是 ``Dict``，那么该函数会将 ``data`` 的 ``key`` 替换为 ``mapping[key]``；
+        * 如果 ``data`` 是 ``dataclass``，那么该函数会先使用 :func:`dataclasses.asdict` 函数将其转换为 ``Dict``，然后进行转换；
+        * 如果 ``data`` 是 ``Sequence``，那么该函数会先将其转换成一个对应的字典::
+        
+            {
+                "_0": list[0],
+                "_1": list[1],
+                ...
+            }
+
+          然后使用 ``mapping`` 对这个 ``Dict`` 进行转换，如果没有匹配上 ``mapping`` 中的 ``key`` 则保持 ``\'\_number\'`` 这个形式。
+
+    :param mapping: 用于转换的字典或者函数；当 ``mapping`` 是函数时，返回值必须为字典类型；
     :param data: 需要被转换的对象；
-    :return: 返回转换好的结果；
+    :return: 返回转换后的结果；
     """
     if mapping is None:
         return data
@@ -320,21 +336,20 @@ def apply_to_collection(
         include_none: bool = True,
         **kwargs: Any,
 ) -> Any:
-    """将函数 function 递归地在 data 中的元素执行，但是仅在满足元素为 dtype 时执行。
-
-    this function credit to: https://github.com/PyTorchLightning/pytorch-lightning
-    Args:
-        data: the collection to apply the function to
-        dtype: the given function will be applied to all elements of this dtype
-        function: the function to apply
-        *args: positional arguments (will be forwarded to calls of ``function``)
-        wrong_dtype: the given function won't be applied if this type is specified and the given collections
-            is of the ``wrong_dtype`` even if it is of type ``dtype``
-        include_none: Whether to include an element if the output of ``function`` is ``None``.
-        **kwargs: keyword arguments (will be forwarded to calls of ``function``)
-
-    Returns:
-        The resulting collection
+    """
+    递归地对 ``data`` 中的元素执行函数 ``function``，且仅在满足元素为 ``dtype`` 时执行。
+
+    该函数参考了 `pytorch-lightning <https://github.com/PyTorchLightning/pytorch-lightning>`_ 的实现
+
+    :param data: 需要进行处理的数据集合或数据；
+    :param dtype: 数据的类型，函数 ``function`` 只会被应用于 ``data`` 中类型为 ``dtype`` 的数据；
+    :param function: 对数据进行处理的函数；
+    :param args: ``function`` 所需要的其它参数；
+    :param wrong_dtype: ``function`` 一定不会生效的数据类型。如果数据既是 ``wrong_dtype`` 类型又是 ``dtype`` 类型
+        那么也不会生效；
+    :param include_none: 是否包含执行结果为 ``None`` 的数据，默认为 ``True``；
+    :param kwargs: ``function`` 所需要的其它参数；
+    :return: 经过 ``function`` 处理后的数据集合；
     """
     # Breaking condition
     if isinstance(data, dtype) and (wrong_dtype is None or not isinstance(data, wrong_dtype)):
@@ -402,18 +417,20 @@ def apply_to_collection(
 @contextmanager
 def nullcontext():
     r"""
-    用来实现一个什么 dummy 的 context 上下文环境；
+    实现一个什么都不做的上下文环境。
     """
     yield
 
 
 def sub_column(string: str, c: int, c_size: int, title: str) -> str:
     r"""
-    :param string: 要被截断的字符串
-    :param c: 命令行列数
-    :param c_size: instance或dataset field数
-    :param title: 列名
-    :return: 对一个过长的列进行截断的结果
+    对传入的字符串进行截断，方便在命令行中显示。
+
+    :param string: 要被截断的字符串；
+    :param c: 命令行列数；
+    :param c_size: :class:`~fastNLP.core.Instance` 或 :class:`fastNLP.core.DataSet` 的 ``field`` 数目；
+    :param title: 列名；
+    :return: 对一个过长的列进行截断的结果；
     """
     avg = max(int(c / c_size / 2), len(title))
     string = str(string)
@@ -442,18 +459,17 @@ def _is_iterable(value):
 
 def pretty_table_printer(dataset_or_ins) -> PrettyTable:
     r"""
-    :param dataset_or_ins: 传入一个dataSet或者instance
+    用于在 ``fastNLP`` 中展示数据的函数::
 
-    .. code-block::
-
-        ins = Instance(field_1=[1, 1, 1], field_2=[2, 2, 2], field_3=["a", "b", "c"])
+        >>> ins = Instance(field_1=[1, 1, 1], field_2=[2, 2, 2], field_3=["a", "b", "c"])
         +-----------+-----------+-----------------+
         |  field_1  |  field_2  |     field_3     |
         +-----------+-----------+-----------------+
         | [1, 1, 1] | [2, 2, 2] | ['a', 'b', 'c'] |
         +-----------+-----------+-----------------+
 
-    :return: 以 pretty table的形式返回根据terminal大小进行自动截断
+    :param dataset_or_ins: 要展示的 :class:`~fastNLP.core.DataSet` 或者 :class:`~fastNLP.core.Instance` 实例；
+    :return: 根据命令行大小进行自动截断的数据表格；
     """
     x = PrettyTable()
     try:
@@ -486,7 +502,7 @@ def pretty_table_printer(dataset_or_ins) -> PrettyTable:
 
 
 class Option(dict):
-    r"""a dict can treat keys as attributes"""
+    r"""将键转化为属性的字典类型"""
 
     def __getattr__(self, item):
         try:
@@ -516,11 +532,10 @@ _emitted_deprecation_warnings = set()
 
 
 def deprecated(help_message: Optional[str] = None):
-    """Decorator to mark a function as deprecated.
+    """
+    标记当前功能已经过时的装饰器。
 
-    Args:
-        help_message (`Optional[str]`): An optional message to guide the user on how to
-            switch to non-deprecated usage of the library.
+    :param help_message: 一段指引信息，告知用户如何将代码切换为当前版本提倡的用法；
     """
 
     def decorator(deprecated_function: Callable):
@@ -549,11 +564,10 @@ def deprecated(help_message: Optional[str] = None):
     return decorator
 
 
-def seq_len_to_mask(seq_len, max_len=None):
+def seq_len_to_mask(seq_len, max_len: Optional[int]):
     r"""
 
-    将一个表示sequence length的一维数组转换为二维的mask，不包含的位置为0。
-    转变 1-d seq_len到2-d mask.
+    将一个表示 ``sequence length`` 的一维数组转换为二维的 ``mask`` ，不包含的位置为 **0**。
 
     .. code-block::
 
@@ -570,10 +584,11 @@ def seq_len_to_mask(seq_len, max_len=None):
         >>>print(mask.size())
         torch.Size([14, 100])
 
-    :param np.ndarray,torch.LongTensor seq_len: shape将是(B,)
-    :param int max_len: 将长度pad到这个长度。默认(None)使用的是seq_len中最长的长度。但在nn.DataParallel的场景下可能不同卡的seq_len会有
-        区别，所以需要传入一个max_len使得mask的长度是pad到该长度。
-    :return: np.ndarray, torch.Tensor 。shape将是(B, max_length)， 元素类似为bool或torch.uint8
+    :param seq_len: 大小为 ``(B,)`` 的长度序列；
+    :param int max_len: 将长度补齐或截断到 ``max_len``。默认情况（为 ``None``）使用的是 ``seq_len`` 中最长的长度；
+        但在 :class:`torch.nn.DataParallel` 等分布式的场景下可能不同卡的 ``seq_len`` 会有区别，所以需要传入
+        一个 ``max_len`` 使得 ``mask`` 的补齐或截断到该长度。
+    :return: 大小为 ``(B, max_len)`` 的 ``mask``， 元素类型为 ``bool`` 或 ``uint8``
     """
     if isinstance(seq_len, np.ndarray):
         assert len(np.shape(seq_len)) == 1, f"seq_len can only have one dimension, got {len(np.shape(seq_len))}."

fastNLP/envs/set_backend.py

@@ -51,23 +51,33 @@ def _set_backend():
         assert _module_available(backend), f"You must have {backend} available to use {backend} backend."
         assert 'paddle' not in sys.modules, "You have to use `set_backend()` before `import paddle`."
         user_visible_devices = os.getenv(USER_CUDA_VISIBLE_DEVICES)
+        cuda_visible_devices = os.getenv("CUDA_VISIBLE_DEVICES")
         if 'PADDLE_RANK_IN_NODE' in os.environ and 'FLAGS_selected_gpus' in os.environ:
             # 在分布式子进程下，根据 USER_VISIBLE_DEVICES 得到进程真正占有的设备
             selected_gpus = os.environ['FLAGS_selected_gpus'].split(',')
             if user_visible_devices is not None:
-                # 用户通过 CUDA_VISIBLE_DEVICES 启动了分布式训练
+                # 用户使用 fastNLP 启动了分布式训练
                 # 此时经过 set_backend，用户的设置会保存在 USER_CUDA_VISIBLE_DEVICES 中
-                # 我们需要从中找到真正使用的设备编号
+                # 我们需要从中转换为用户找到真正使用的设备编号
                 user_visible_devices = user_visible_devices.split(",")
-                selected_gpus = ",".join([user_visible_devices[int(i)] for i in selected_gpus])
+                selected_gpus = [user_visible_devices[int(i)] for i in selected_gpus]
+            # 没有找到 USER_CUDA_VISIBLE_DEVICES，说明用户是直接用 launch 启动的
+            elif cuda_visible_devices:
+                # 用户设置了可见设备，需要进行转换
+                # 如 CUDA_VISIBLE_DEVICES = 0,2,3 --gpus=0,2,3
+                # 在 rank1 中此时 selected_gpus = ['1']，需要转换为设备 2
+                os.environ[USER_CUDA_VISIBLE_DEVICES] = cuda_visible_devices
+                cuda_visible_devices = cuda_visible_devices.split(",")
+                selected_gpus = [cuda_visible_devices[int(i)] for i in selected_gpus]
             else:
-                # 没有找到 USER_CUDA_VISIBLE_DEVICES，则将之设置为所有的设备
+                # 用户没有设置可见设备，则赋值成所有的设备
                 os.environ[USER_CUDA_VISIBLE_DEVICES] = ",".join(map(str, list(
                     range(get_gpu_count())
                 )))
             os.environ['CUDA_VISIBLE_DEVICES'] = ",".join(selected_gpus)
             os.environ['FLAGS_selected_gpus'] = ",".join([str(g) for g in range(len(selected_gpus))])
             os.environ['FLAGS_selected_accelerators'] = ",".join([str(g) for g in range(len(selected_gpus))])
+        
         elif 'CUDA_VISIBLE_DEVICES' in os.environ:
             # 主进程中，用户设置了 CUDA_VISIBLE_DEVICES
             # 将用户设置的 CUDA_VISIBLE_DEVICES hack 掉
@@ -91,6 +101,11 @@ def _set_backend():
     elif backend == 'torch':
         assert _module_available(backend), f"You must have {backend} available to use {backend} backend."
 
+    if 'PADDLE_RANK_IN_NODE' in os.environ and 'FLAGS_selected_gpus' in os.environ \
+        and "USER_CUDA_VISIBLE_DEVICES" not in os.environ:
+        # 当用户没有设置 backend 并且使用 launch 启动了多卡，应该提醒用户进行设置
+        raise RuntimeError("To run paddle distributed training, please set `FASTNLP_BACKEND` to 'paddle' before using FastNLP.")
+
 
 def set_env(global_seed=None):
     """

fastNLP/envs/utils.py

@@ -6,6 +6,7 @@ from packaging.version import Version
 import subprocess
 import pkg_resources
 
+__all__ = []
 
 def _module_available(module_path: str) -> bool:
     """Check if a path is available in your environment.
@@ -48,10 +49,11 @@ def _compare_version(package: str, op: Callable, version: str, use_base_version:
         pkg_version = Version(pkg_version.base_version)
     return op(pkg_version, Version(version))
 
-def get_gpu_count():
+def get_gpu_count() -> int:
     """
-    利用命令行获取gpu数目的函数
-    :return: gpu数目，如果没有显卡设备则为-1
+    利用命令行获取 ``gpu`` 数目的函数
+
+    :return: 显卡数目，如果没有显卡设备则为-1
     """
     try:
         lines = subprocess.check_output(['nvidia-smi', '--query-gpu=memory.used', '--format=csv'])

fastNLP/modules/__init__.py

@@ -1,9 +1,9 @@
 __all__ = [
-    "MixModule",
+    # "MixModule",
     "torch2paddle",
     "paddle2torch",
     "torch2jittor",
     "jittor2torch",
 ]
 
-from .mix_modules import MixModule, torch2paddle, paddle2torch, torch2jittor, jittor2torch
+from .mix_modules import torch2paddle, paddle2torch, torch2jittor, jittor2torch

fastNLP/modules/mix_modules/__init__.py

@@ -1,10 +1,10 @@
 __all__ = [
-    "MixModule",
+    # "MixModule",
     "torch2paddle",
     "paddle2torch",
     "torch2jittor",
     "jittor2torch",
 ]
 
-from .mix_module import MixModule
+# from .mix_module import MixModule
 from .utils import *

fastNLP/modules/mix_modules/mix_module.py

@@ -1,310 +0,0 @@
-import os
-import io
-import pickle
-from typing import Dict
-from collections import OrderedDict
-
-import numpy as np
-
-from fastNLP.envs.imports import _NEED_IMPORT_JITTOR, _NEED_IMPORT_PADDLE, _NEED_IMPORT_TORCH
-from fastNLP.core.utils.paddle_utils import paddle_to
-
-if _NEED_IMPORT_PADDLE:
-    import paddle
-    from paddle.nn import Layer as PaddleLayer
-
-if _NEED_IMPORT_TORCH:
-    import torch
-    from torch.nn import Module as TorchModule, Parameter as TorchParameter
-
-if _NEED_IMPORT_JITTOR:
-    import jittor
-
-
-__all__ = [
-    "MixModule",
-]
-
-class MixModule:
-    """
-    TODO: 支持不同的混合方式；添加state_dict的支持；如果参数里有List of Tensors该怎么处理；
-        是否需要仿照Module那样在初始化的时候给各种模型分类
-    可以同时使用Torch和Paddle框架的混合模型
-    """
-    def __init__(self, *args, **kwargs):
-        pass
-
-    def __call__(self, *args, **kwargs):
-        return self.forward(*args, **kwargs)
-
-    def named_parameters(self, prefix='', recurse: bool=True, backend=None):
-        """
-        返回模型的名字和参数
-
-        :param prefix: 输出时在参数名前加上的前缀
-        :param recurse: 是否递归地输出参数
-        :param backend: `backend`=`None`时，将所有模型和张量的参数返回；
-                        `backend`=`torch`时，返回`torch`的参数；
-                        `backend`=`paddle`时，返回`paddle`的参数。
-        """
-        if backend is None:
-            generator = self.attributes(TorchModule, TorchParameter, PaddleLayer)
-        elif backend == "torch":
-            generator = self.attributes(TorchModule, TorchParameter)
-        elif backend == "paddle":
-            generator = self.attributes(PaddleLayer)
-        else:
-            raise ValueError("Unknown backend parameter.")
-
-        for name, value in generator:
-            name = prefix + ('.' if prefix else '') + name
-            if isinstance(value, TorchParameter):
-                # 非Module/Layer类型，直接输出名字和值
-                yield name, value
-            elif recurse:
-                # 递归地调用named_parameters
-                for name_r, value_r in value.named_parameters(name, recurse):
-                    yield name_r, value_r
-
-    def parameters(self, recurse: bool = True, backend: str = None):
-        """
-        返回模型的参数
-
-        :param recurse:
-        :param backend: `backend`=`None`时，将所有模型和张量的参数返回；
-                        `backend`=`torch`时，返回`torch`的参数；
-                        `backend`=`paddle`时，返回`paddle`的参数。
-        """
-        for name, value in self.named_parameters(recurse=recurse, backend=backend):
-            yield value
-    
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError
-
-    def train_step(self, batch):
-        raise NotImplementedError
-
-    def test_step(self, batch):
-        raise NotImplementedError
-
-    def evaluate_step(self, batch):
-        raise NotImplementedError
-
-    def train(self):
-        for name, value in self.attributes(TorchModule, PaddleLayer):
-            value.train()
-
-    def eval(self):
-        for name, value in self.attributes(TorchModule, PaddleLayer):
-            value.eval()
-
-    def to(self, device):
-        """
-        :param device: 设备名
-        """
-        # 有jittor的话 warning
-        if device == "cpu":
-            paddle_device = device
-        elif device.startswith("cuda"):
-            paddle_device = device.replace("cuda", "gpu")
-        elif device.startswith("gpu"):
-            paddle_device = device
-            device = device.replace("gpu", "cuda")
-        else:
-            raise ValueError("Device value error")
-
-        for name, value in self.attributes(TorchModule):
-            # torch的to函数不影响Tensor
-            vars(self)[name] = value.to(device)
-        for name, value in self.attributes(TorchParameter):
-            # Parameter在经过to函数后会变成Tensor类型
-            vars(self)[name] = TorchParameter(value.to(device), requires_grad=value.requires_grad)
-
-        for name, value in self.attributes(PaddleLayer):
-            vars(self)[name] = value.to(paddle_device)
-        for name, value in self.attributes(paddle.Tensor):
-        # paddle的to函数会影响到Tensor
-            vars(self)[name] = paddle_to(value, paddle_device)
-
-        return self
-
-    def state_dict(self, backend: str = None) -> Dict:
-        """
-        返回模型的state_dict。
-
-        .. note:: torch的destination参数会在将来删除，因此不提供destination参数
-
-        :param backend: `backend`=`None`时，将所有模型和张量的state dict返回；
-                        `backend`=`torch`时，返回`torch`的state dict；
-                        `backend`=`paddle`时，返回`paddle`的state dict。
-        """
-        if backend is None:
-            generator = self.attributes(TorchModule, TorchParameter, PaddleLayer)
-        elif backend == "torch":
-            generator = self.attributes(TorchModule, TorchParameter)
-        elif backend == "paddle":
-            generator = self.attributes(PaddleLayer)
-        else:
-            raise ValueError(f"Unknown backend {backend}.")
-
-        destination = OrderedDict()
-
-        for name, value in generator:
-            if value is None:
-                continue
-            if isinstance(value, TorchParameter):
-                destination[name] = value
-            else:
-                # 不同框架state_dict函数的参数名和顺序不同
-                if isinstance(value, PaddleLayer):
-                    kwargs = {
-                        "structured_name_prefix": name + ".",
-                    }
-                elif isinstance(value, TorchModule):
-                    kwargs = {
-                        "prefix": name + ".",
-                    }
-                else:
-                    raise ValueError(f"Unknown item type {type(value)}")
-                destination.update(value.state_dict(**kwargs))
-
-        return destination
-
-    def save_state_dict_to_file(self, path: str):
-        """
-        保存模型的state dict到path
-        """
-        # TODO 设备限制
-        filename = os.path.basename(path)
-        if filename == "":
-            raise ValueError("Received empty filename.")
-        dirname = os.path.dirname(path)
-        if dirname and not os.path.exists(dirname):
-            os.makedirs(dirname)
-        protocol = 4
-
-        saved = {}
-        paddle_dict = self.state_dict(backend="paddle")
-        torch_dict = self.state_dict(backend="torch")
-        # 保存paddle部分
-        # 调用paddle保存时的处理函数
-        paddle_saved_obj = paddle.framework.io._build_saved_state_dict(paddle_dict)
-        paddle_saved_obj = paddle.fluid.io._unpack_saved_dict(paddle_saved_obj, protocol)
-        # 将返回的dict保存
-        saved["paddle"] = paddle_saved_obj
-
-        # 保存torch部分
-        buffer = io.BytesIO()
-        torch.save(torch_dict, buffer)
-        saved["torch"] = buffer.getvalue()
-
-        # 保存
-        with open(path, "wb") as f:
-            pickle.dump(saved, f, protocol)
-
-    def load_state_dict_from_file(self, path: str):
-        """
-        从 `path` 中加载保存的state dict
-        """
-        state_dict = {}
-        with open(path, "rb") as f:
-            loaded = pickle.load(f)
-        # 加载paddle的数据
-        paddle_loaded_obj = loaded["paddle"]
-        paddle_load_result = paddle.fluid.io._pack_loaded_dict(paddle_loaded_obj)
-        if "StructuredToParameterName@@" in paddle_load_result:
-            for key in paddle_load_result["StructuredToParameterName@@"]:
-                if isinstance(paddle_load_result[key], np.ndarray):
-                    paddle_load_result[key] = paddle.to_tensor(paddle_load_result[key])
-        state_dict.update(paddle_load_result)
-        # 加载torch的数据
-        torch_loaded_obj = loaded["torch"]
-        torch_bytes = io.BytesIO(torch_loaded_obj)
-        torch_load_result = torch.load(torch_bytes)
-        state_dict.update(torch_load_result)
-
-        self.load_state_dict(state_dict)
-
-    def load_state_dict(self, state_dict):
-        """
-        从state dict中加载数据
-        """
-        missing_keys = []
-        unexpected_keys = []
-        error_msgs = []
-        new_state = {}
-
-        local_state = self.state_dict()
-
-        # 对字典内容按前缀进行归类
-        for key, value in state_dict.items():
-            splited = key.split(".", 1)
-            if len(splited) == 1:
-                # 没有前缀，实际上只有torch.nn.Parameter会进入这种情况
-                new_state[key] = value
-            else:
-                prefix, name = splited
-                if prefix not in new_state:
-                    new_state[prefix] = {}
-                new_state[prefix][name] = value
-
-        for key, param in self.attributes(TorchModule, TorchParameter, PaddleLayer):
-            if key in new_state:
-                # 在传入的字典中找到了对应的值
-                input_param = new_state[key]
-                if not isinstance(input_param, dict):
-                    # 且不是字典，即上述没有前缀的情况
-                    # 按照torch.nn.Module._load_from_state_dict进行赋值
-                    if not torch.overrides.is_tensor_like(input_param):
-                        error_msgs.append('While copying the parameter named "{}", '
-                                        'expected torch.Tensor or Tensor-like object from checkpoint but '
-                                        'received {}'
-                                        .format(key, type(input_param)))
-                        continue
-
-                    # This is used to avoid copying uninitialized parameters into
-                    # non-lazy modules, since they dont have the hook to do the checks
-                    # in such case, it will error when accessing the .shape attribute.
-                    is_param_lazy = torch.nn.parameter.is_lazy(param)
-                    # Backward compatibility: loading 1-dim tensor from 0.3.* to version 0.4+
-                    if not is_param_lazy and len(param.shape) == 0 and len(input_param.shape) == 1:
-                        input_param = input_param[0]
-
-                    if not is_param_lazy and input_param.shape != param.shape:
-                        # local shape should match the one in checkpoint
-                        error_msgs.append('size mismatch for {}: copying a param with shape {} from checkpoint, '
-                                        'the shape in current model is {}.'
-                                        .format(key, input_param.shape, param.shape))
-                        continue
-                    try:
-                        with torch.no_grad():
-                            param.copy_(input_param)
-                    except Exception as ex:
-                        error_msgs.append('While copying the parameter named "{}", '
-                                        'whose dimensions in the model are {} and '
-                                        'whose dimensions in the checkpoint are {}, '
-                                        'an exception occurred : {}.'
-                                        .format(key, param.size(), input_param.size(), ex.args))
-                else:
-                    # 否则在子模块中
-                    if isinstance(param, TorchModule):
-                        # torch模块
-                        # 由于paddle没有提供类似strict的参数，因此也不对torch作要求
-                        param.load_state_dict(input_param, strict=False)
-                    elif isinstance(param, PaddleLayer):
-                        # paddle模块
-                        param.load_dict(input_param)
-            else:
-                missing_keys.append(key)
-
-        if len(error_msgs) > 0:
-            raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
-                               self.__class__.__name__, "\n\t".join(error_msgs)))
-
-    def attributes(self, *types):
-        """
-        查找对应类型的成员
-        """
-        for name, value in vars(self).items():
-            if isinstance(value, types):
-                yield name, value

fastNLP/modules/mix_modules/utils.py

@@ -1,233 +0,0 @@
-import warnings
-import os
-from typing import Any, Optional, Union
-
-import numpy as np
-
-from fastNLP.core.utils.utils import apply_to_collection
-from fastNLP.core.utils.paddle_utils import paddle_to
-from fastNLP.envs.imports import _NEED_IMPORT_JITTOR, _NEED_IMPORT_TORCH, _NEED_IMPORT_PADDLE
-
-if _NEED_IMPORT_PADDLE:
-    import paddle
-
-if _NEED_IMPORT_JITTOR:
-    import jittor
-
-if _NEED_IMPORT_TORCH:
-    import torch
-
-__all__ = [
-    "paddle2torch",
-    "torch2paddle",
-    "jittor2torch",
-    "torch2jittor",
-]
-
-def _paddle2torch(paddle_tensor: 'paddle.Tensor', target_device: Optional[Union[str, int]] = None, no_gradient: bool = None) -> 'torch.Tensor':
-    """
-    将paddle tensor转换为torch tensor，并且能够保留梯度进行反向传播
-    :param paddle_tensor: 要转换的paddle张量
-    :param target_device: 是否将转换后的张量迁移到特定设备上，输入为`None`时，和输入的张量相同。
-    :param no_gradient: 是否保留原张量的梯度。为`None`时，新的张量与输入张量保持一致；
-                        为`True`时，全部不保留梯度；为`False`时，全部保留梯度。
-    :return: 转换后的torch张量
-    """
-    no_gradient = paddle_tensor.stop_gradient if no_gradient is None else no_gradient
-    paddle_numpy = paddle_tensor.numpy()
-    if not np.issubdtype(paddle_numpy.dtype, np.inexact):
-        no_gradient = True
-
-    if target_device is None:
-        if paddle_tensor.place.is_gpu_place():
-            # paddlepaddle有两种Place，对应不同的device id获取方式
-            if hasattr(paddle_tensor.place, "gpu_device_id"):
-                # paddle.fluid.core_avx.Place
-                # 在gpu环境下创建张量的话，张量的place是这一类型
-                target_device = f"cuda:{paddle_tensor.place.gpu_device_id()}"
-            else:
-                # paddle.CUDAPlace
-                target_device = f"cuda:{paddle_tensor.place.get_device_id()}"
-        else:
-            # TODO: 可能需要支持xpu等设备
-            target_device = "cpu"
-
-    if not no_gradient:
-        # 保持梯度，并保持反向传播
-        # torch.tensor会保留numpy数组的类型
-        torch_tensor = torch.tensor(paddle_numpy, requires_grad=True, device=target_device)
-        hook = torch_tensor.register_hook(
-            lambda grad: paddle.autograd.backward(paddle_tensor, paddle.to_tensor(grad.cpu().numpy()))
-        )
-    else:
-        # 不保留梯度
-        torch_tensor = torch.tensor(paddle_numpy, requires_grad=False, device=target_device)
-
-    return torch_tensor
-
-
-def _torch2paddle(torch_tensor: 'torch.Tensor', target_device: str = None, no_gradient: bool = None) -> 'paddle.Tensor':
-    """
-    将torch tensor转换为paddle tensor，并且能够保留梯度进行反向传播。
-    :param torch_tensor: 要转换的torch张量
-    :param target_device: 是否将转换后的张量迁移到特定设备上，输入为`None`时，和输入的张量相同。
-    :param no_gradient: 是否保留原张量的梯度。为`None`时，新的张量与输入张量保持一致；
-                        为`True`时，全部不保留梯度；为`False`时，全部保留梯度。
-    :return: 转换后的paddle张量
-    """
-    no_gradient = not torch_tensor.requires_grad if no_gradient is None else no_gradient
-    if target_device is None:
-        if torch_tensor.is_cuda:
-            target_device = f"gpu:{torch_tensor.device.index}"
-        else:
-            target_device = "cpu"
-
-    if not no_gradient:
-        # 保持梯度并保持反向传播
-        # paddle的stop_gradient和torch的requires_grad表现是相反的
-        paddle_tensor = paddle.to_tensor(torch_tensor.detach().numpy(), stop_gradient=False)
-        hook = paddle_tensor.register_hook(
-            lambda grad: torch.autograd.backward(torch_tensor, torch.tensor(grad.numpy()))
-        )
-    else:
-        paddle_tensor = paddle.to_tensor(torch_tensor.detach().numpy(), stop_gradient=True)
-
-    paddle_tensor = paddle_to(paddle_tensor, target_device)
-
-    return paddle_tensor
-
-
-def _jittor2torch(jittor_var: 'jittor.Var', target_device: Optional[Union[str, int]] = None, no_gradient: bool = None) -> 'torch.Tensor':
-    """
-    将jittor Var转换为torch tensor，并且能够保留梯度进行反向传播
-    :param jittor_var: 要转换的jittor变量
-    :param target_device: 是否将转换后的张量迁移到特定设备上，输入为`None`时，根据jittor.flags.use_cuda决定。
-    :param no_gradient: 是否保留原张量的梯度。为`None`时，新的张量与输入张量保持一致；
-                        为`True`时，全部不保留梯度；为`False`时，全部保留梯度。
-    :return: 转换后的torch张量
-    """
-    # TODO: warning：无法保留梯度
-    # jittor的grad可以通过callback进行传递
-    # 如果outputs有_grad键，可以实现求导
-    no_gradient = not jittor_var.requires_grad if no_gradient is None else no_gradient
-    if no_gradient == False:
-        warnings.warn("The result tensor will not keep gradients due to differences between jittor and pytorch.")
-    jittor_numpy = jittor_var.numpy()
-    if not np.issubdtype(jittor_numpy.dtype, np.inexact):
-        no_gradient = True
-
-    if target_device is None:
-        # jittor的设备分配是自动的
-        # 根据use_cuda判断
-        if jittor.flags.use_cuda:
-            target_device = "cuda:0"
-        else:
-            target_device = "cpu"
-
-    torch_tensor = torch.tensor(jittor_numpy, requires_grad=not no_gradient, device=target_device)
-
-    return torch_tensor
-
-
-def _torch2jittor(torch_tensor: 'torch.Tensor', no_gradient: bool = None) -> 'jittor.Var':
-    """
-    将torch tensor转换为jittor Var，并且能够保留梯度进行反向传播
-    :param torch_tensor: 要转换的torch张量
-    :param no_gradient: 是否保留原张量的梯度。为`None`时，新的张量与输入张量保持一致；
-                        为`True`时，全部不保留梯度；为`False`时，全部保留梯度。
-    :return: 转换后的jittor变量
-    """
-    no_gradient = not torch_tensor.requires_grad if no_gradient is None else no_gradient
-
-    if not no_gradient:
-        # 保持梯度并保持反向传播
-        jittor_var = jittor.Var(torch_tensor.detach().numpy())
-        jittor_var.requires_grad = True
-        hook = jittor_var.register_hook(
-            lambda grad: torch.autograd.backward(torch_tensor, torch.tensor(grad.numpy()))
-        )
-    else:
-        jittor_var = jittor.Var(torch_tensor.detach().numpy())
-        jittor_var.requires_grad = False
-
-    return jittor_var
-
-
-def torch2paddle(torch_in: Any, target_device: str = None, no_gradient: bool = None) -> Any:
-    """
-    递归地将输入中包含的torch张量转换为paddle张量
-
-    :param torch_in: 要转换的包含torch.Tensor类型的变量
-    :param target_device: 是否将转换后的张量迁移到特定设备上，
-                          输入为`None`时，和输入的张量相同，
-    :param no_gradient: 是否保留原张量的梯度。为`None`时，新的张量与输入张量保持一致；
-                        为`True`时，全部不保留梯度；为`False`时，全部保留梯度。
-    :return: 将所有torch.Tensor转换为paddle.Tensor的张量           
-    """
-
-    return apply_to_collection(
-        torch_in,
-        dtype=torch.Tensor,
-        function=_torch2paddle,
-        target_device=target_device,
-        no_gradient=no_gradient,
-    )
-
-
-def paddle2torch(paddle_in: Any, target_device: str = None, no_gradient: bool = None) -> Any:
-    """
-    递归地将输入中包含的paddle张量转换为torch张量
-
-    :param torch_in: 要转换的包含paddle.Tensor类型的变量
-    :param target_device: 是否将转换后的张量迁移到特定设备上，
-                          输入为`None`时，和输入的张量相同，
-    :param no_gradient: 是否保留原张量的梯度。为`None`时，新的张量与输入张量保持一致；
-                        为`True`时，全部不保留梯度；为`False`时，全部保留梯度。
-    :return: 将所有paddle.Tensor转换为torch.Tensor后的变量          
-    """
-
-    return apply_to_collection(
-        paddle_in,
-        dtype=paddle.Tensor,
-        function=_paddle2torch,
-        target_device=target_device,
-        no_gradient=no_gradient,
-    )
-
-
-def jittor2torch(jittor_in: Any, target_device: str = None, no_gradient: bool = None) -> Any:
-    """
-    递归地将输入中包含的jittor变量转换为torch张量
-
-    :param jittor_in: 要转换的jittor变量
-    :param target_device: 是否将转换后的张量迁移到特定设备上，输入为`None`时，默认为cuda:0。
-    :param no_gradient: 是否保留原张量的梯度。为`None`时，新的张量与输入张量保持一致；
-                        为`True`时，全部不保留梯度；为`False`时，全部保留梯度。
-    :return: 转换后的torch张量
-    """
-
-    return apply_to_collection(
-        jittor_in,
-        dtype=jittor.Var,
-        function=_jittor2torch,
-        target_device=target_device,
-        no_gradient=no_gradient,
-    )
-
-
-def torch2jittor(torch_in: Any, no_gradient: bool = None) -> Any:
-    """
-    递归地将输入中包含的torch张量转换为jittor变量
-
-    :param torch_tensor: 要转换的torch张量
-    :param no_gradient: 是否保留原张量的梯度。为`None`时，新的张量与输入张量保持一致；
-                        为`True`时，全部不保留梯度；为`False`时，全部保留梯度。
-    :return: 转换后的jittor变量
-    """
-    
-    return apply_to_collection(
-        torch_in,
-        dtype=torch.Tensor,
-        function=_torch2jittor,
-        no_gradient=no_gradient,
-    )

tests/core/callbacks/test_checkpoint_callback_torch.py

@@ -14,6 +14,7 @@ from tests.helpers.utils import magic_argv_env_context
 from fastNLP.envs.distributed import rank_zero_rm
 from tests.helpers.models.torch_model import TorchNormalModel_Classification_1
 from tests.helpers.datasets.torch_data import TorchArgMaxDataset
+from tests.helpers.utils import Capturing
 from torchmetrics import Accuracy
 from fastNLP.core.log import logger
 
@@ -428,6 +429,78 @@ def test_trainer_checkpoint_callback_1(
         dist.destroy_process_group()
 
 
+@pytest.mark.torch
+def test_load_state(model_and_optimizers):
+    try:
+        path = Path.cwd().joinpath(f"test_model_checkpoint")
+        path.mkdir(exist_ok=True, parents=True)
+        from fastNLP import Event, Callback
+        @Trainer.on(Event.on_before_backward(every=3), marker='all')
+        def print_outputs(*args):
+            print("????")
+
+        class StateCallback(Callback):
+            def __init__(self, name):
+                self.name = name
+
+            def on_save_checkpoint(self, trainer):
+                return {'name': self.name}
+
+            def on_load_checkpoint(self, trainer, states):
+                self.name = states['name']
+
+            def on_train_end(self, trainer):
+                print(self.name)
+
+        callbacks = [StateCallback('old_callback1'), StateCallback('old_callback2'),
+                     CheckpointCallback(folder=path, every_n_epochs=1, save_object='trainer')]
+
+        trainer = Trainer(
+            model=model_and_optimizers.model,
+            driver='torch',
+            device='cpu',
+            optimizers=model_and_optimizers.optimizers,
+            train_dataloader=model_and_optimizers.train_dataloader,
+            evaluate_dataloaders=model_and_optimizers.evaluate_dataloaders,
+            input_mapping=model_and_optimizers.input_mapping,
+            output_mapping=model_and_optimizers.output_mapping,
+            metrics=model_and_optimizers.metrics,
+            n_epochs=3,
+            callbacks=callbacks,
+            output_from_new_proc="all"
+        )
+        trainer.run(num_eval_sanity_batch=0, num_train_batch_per_epoch=2)
+
+        all_saved_model_paths = {w.name: w for w in path.joinpath(os.environ[FASTNLP_LAUNCH_TIME]).iterdir()}
+        epoch_2_path = all_saved_model_paths['trainer-epoch_2']
+
+        callbacks = [StateCallback('new_callback1'), StateCallback('new_callback2')]
+        trainer = Trainer(
+            model=model_and_optimizers.model,
+            driver='torch',
+            device='cpu',
+            optimizers=model_and_optimizers.optimizers,
+            train_dataloader=model_and_optimizers.train_dataloader,
+            evaluate_dataloaders=model_and_optimizers.evaluate_dataloaders,
+            input_mapping=model_and_optimizers.input_mapping,
+            output_mapping=model_and_optimizers.output_mapping,
+            metrics=model_and_optimizers.metrics,
+            n_epochs=3,
+            callbacks=callbacks,
+            output_from_new_proc="all"
+        )
+        trainer.load(folder=epoch_2_path)
+        with Capturing() as output:
+            trainer.run(num_eval_sanity_batch=0, num_train_batch_per_epoch=2)
+
+        assert 'old_callback1' in output[0]
+        assert 'new_callback2' in output[0]
+        assert output[0].count('???')==1
+
+    finally:
+        rank_zero_rm(path)
+
+
 @pytest.mark.torch
 # 通过自己编写 model_save_fn 和 model_load_fn 来测试 huggingface 的 transformers 的模型的保存和加载；
 @pytest.mark.parametrize("driver,device", [("torch_ddp", [6, 7]), ("torch", 7)])  # ("torch", "cpu"), ("torch_ddp", [0, 1]), ("torch", 1)

tests/core/collators/test_collator.py

@@ -334,9 +334,9 @@ def test_torch_dl():
     dl = TorchDataLoader(ds, batch_size=2)
     batch = next(iter(dl))
     assert 'x' in batch and 'y' in batch and 'z' in batch and 'i' in batch and 'j' in batch
-    assert isinstance(batch['z'], torch.Tensor)
+    assert isinstance(batch['z'], torch.FloatTensor)
     assert isinstance(batch['j'], list)
-    assert isinstance(batch['i']['j'], torch.Tensor)
+    assert isinstance(batch['i']['j'], torch.LongTensor)
 
     dl.set_ignore('x')
     batch = next(iter(dl))

tests/core/controllers/_test_trainer_fleet.py

@@ -1,7 +1,15 @@
 """
-这个文件测试用户以python -m paddle.distributed.launch 启动的情况
-看看有没有用pytest执行的机会
-FASTNLP_BACKEND=paddle python -m paddle.distributed.launch --gpus=0,2,3 _test_trainer_fleet.py
+这个文件测试多卡情况下使用 paddle 的情况::
+
+    >>> # 测试用 python -m paddle.distributed.launch 启动
+    >>> FASTNLP_BACKEND=paddle python -m paddle.distributed.launch --gpus=0,2,3 _test_trainer_fleet.py
+    >>> # 测试在限制 GPU 的情况下用 python -m paddle.distributed.launch 启动
+    >>> CUDA_VISIBLE_DEVICES=0,2,3 FASTNLP_BACKEND=paddle python -m paddle.distributed.launch --gpus=0,2,3 _test_trainer_fleet.py
+    >>> # 测试直接使用多卡
+    >>> FASTNLP_BACKEND=paddle python _test_trainer_fleet.py
+    >>> # 测试在限制 GPU 的情况下直接使用多卡
+    >>> CUDA_VISIBLE_DEVICES=3,4,5,6 FASTNLP_BACKEND=paddle python _test_trainer_fleet.py
+
 """
 import os
 import sys
@@ -71,14 +79,13 @@ def test_trainer_fleet(
 
         n_epochs=n_epochs,
         callbacks=callbacks,
-        output_from_new_proc="logs",
+        # output_from_new_proc="logs",
     )
     trainer.run()
 
 if __name__ == "__main__":
     driver = "paddle"
-    device = [0,2,3]
-    # driver = "paddle"
+    device = [0,1,3]
     # device = 2
     callbacks = [
         # RecordMetricCallback(monitor="acc#acc", metric_threshold=0.0, larger_better=True), 

tests/core/controllers/_test_trainer_fleet_outside.py

@@ -1,7 +1,11 @@
 """
-这个文件测试用户以python -m paddle.distributed.launch 启动的情况
-并且自己初始化了 fleet
-FASTNLP_BACKEND=paddle python -m paddle.distributed.launch --gpus=0,2,3 _test_trainer_fleet_outside.py
+这个文件测试用户自己初始化分布式环境后使用 paddle 的情况:
+
+    >>> # 测试用 python -m paddle.distributed.launch 启动
+    >>> FASTNLP_BACKEND=paddle python -m paddle.distributed.launch --gpus=0,2,3 _test_trainer_fleet_outside.py
+    >>> # 测试在限制 GPU 的情况下用 python -m paddle.distributed.launch 启动
+    >>> CUDA_VISIBLE_DEVICES=0,2,3 FASTNLP_BACKEND=paddle python -m paddle.distributed.launch --gpus=0,2,3 _test_trainer_fleet_outside.py
+
 """
 import os
 import sys
@@ -77,14 +81,13 @@ def test_trainer_fleet(
 
         n_epochs=n_epochs,
         callbacks=callbacks,
-        output_from_new_proc="logs",
-        data_device=f"gpu:{os.environ['CUDA_VISIBLE_DEVICES']}"
+        # output_from_new_proc="logs",
     )
     trainer.run()
 
 if __name__ == "__main__":
     driver = "paddle"
-    device = [0,2,3]
+    device = [0,1,3]
     callbacks = [
         # RecordMetricCallback(monitor="acc#acc", metric_threshold=0.0, larger_better=True), 
         RichCallback(5),

tests/core/controllers/_test_trainer_jittor.py

@@ -0,0 +1,237 @@
+import os
+import sys
+import time
+# os.environ["cuda_archs"] = "61"
+# os.environ["FAS"]
+os.environ["log_silent"] = "1"
+sys.path.append("../../../")
+
+from datasets import load_dataset
+from datasets import DatasetDict
+import jittor as jt
+from jittor import nn, Module
+from jittor.dataset import Dataset
+jt.flags.use_cuda = True
+
+from fastNLP.core.controllers.trainer import Trainer
+from fastNLP.core.metrics.accuracy import Accuracy
+from fastNLP.core.vocabulary import Vocabulary
+from fastNLP.core.callbacks.progress_callback import RichCallback
+from fastNLP.core.callbacks.callback import Callback
+from fastNLP.core.dataloaders.jittor_dataloader.fdl import JittorDataLoader
+
+class TextClassificationDataset(Dataset):
+    def __init__(self, dataset):
+        super(TextClassificationDataset, self).__init__()
+        self.dataset = dataset
+        self.set_attrs(total_len=len(dataset))
+
+    def __getitem__(self, idx):
+        return {"x": self.dataset["input_ids"][idx], "y": self.dataset["label"][idx]}
+
+
+class LSTM(Module):
+    
+    def __init__(self, num_of_words, hidden_size, features):
+
+        self.embedding = nn.Embedding(num_of_words, features)
+        self.lstm = nn.LSTM(features, hidden_size, batch_first=True)
+        self.layer = nn.Linear(hidden_size, 2)
+        self.softmax = nn.Softmax(dim=1)
+        self.loss_fn = nn.CrossEntropyLoss()
+        
+        self.hidden_size = hidden_size
+        self.features = features
+
+    def init_hidden(self, x):
+        # batch_first
+        batch_size = x.shape[0]
+        h0 = jt.randn(1, batch_size, hidden_size)
+        c0 = jt.randn(1, batch_size, hidden_size)
+
+        return h0, c0
+
+    def execute(self, input_ids):
+
+        output = self.embedding(input_ids)
+        # TODO 去除padding
+        output, (h, c) = self.lstm(output, self.init_hidden(output))
+        # len, batch, hidden_size
+        output = self.layer(output[-1])
+
+        return output
+
+    def train_step(self, x, y):
+        x = self(x)
+        outputs = self.loss_fn(x, y)
+        return {"loss": outputs}
+
+    def evaluate_step(self, x, y):
+        x = self(x)
+        return {"pred": x, "target": y.reshape((-1,))}
+
+
+class PrintWhileTrainingCallBack(Callback):
+    """
+    通过该Callback实现训练过程中loss的输出
+    """
+
+    def __init__(self, print_every_epoch, print_every_batch):
+        self.print_every_epoch = print_every_epoch
+        self.print_every_batch = print_every_batch
+
+        self.loss = 0
+        self.start = 0
+        self.epoch_start = 0
+
+    def on_train_begin(self, trainer):
+        """
+        在训练开始前输出信息
+        """
+        print("Start training. Total {} epochs and {} batches in each epoch.".format(
+            trainer.n_epochs, trainer.num_batches_per_epoch
+        ))
+        self.start = time.time()
+
+    def on_before_backward(self, trainer, outputs):
+        """
+        每次反向传播前统计loss，用于计算平均值
+        """
+        loss = trainer.extract_loss_from_outputs(outputs)
+        loss = trainer.driver.tensor_to_numeric(loss)
+        self.loss += loss
+
+    def on_train_epoch_begin(self, trainer):
+        self.epoch_start = time.time()
+
+    def on_train_epoch_end(self, trainer):
+        """
+        在每经过一定epoch或最后一个epoch时输出当前epoch的平均loss和使用时间
+        """
+        if trainer.cur_epoch_idx % self.print_every_epoch == 0 \
+            or trainer.cur_epoch_idx == trainer.n_epochs:
+            print("Epoch: {} Loss: {} Current epoch training time: {}s".format(
+                trainer.cur_epoch_idx, self.loss / trainer.num_batches_per_epoch, time.time() - self.epoch_start
+            ))
+        # 将loss清零
+        self.loss = 0
+    
+    def on_train_batch_end(self, trainer):
+        """
+        在每经过一定batch或最后一个batch时输出当前epoch截止目前的平均loss
+        """
+        if trainer.batch_idx_in_epoch % self.print_every_batch == 0 \
+            or trainer.batch_idx_in_epoch == trainer.num_batches_per_epoch:
+            print("\tBatch: {} Loss: {}".format(
+                trainer.batch_idx_in_epoch, self.loss / trainer.batch_idx_in_epoch
+            ))
+
+    def on_train_end(self, trainer):
+        print("Total training time: {}s".format(time.time() - self.start))
+
+    
+def process_data(ds: DatasetDict, vocabulary: Vocabulary, max_len=256) -> DatasetDict:
+    # 分词
+    ds = ds.map(lambda x: {"input_ids": text_to_id(vocabulary, x["text"], max_len)})
+    ds.set_format(type="numpy", columns=ds.column_names)
+    return ds
+
+def set_vocabulary(vocab, dataset):
+
+    for data in dataset:
+        vocab.update(data["text"].split())
+    return vocab
+
+def text_to_id(vocab, text: str, max_len):
+    text = text.split()
+    # to index
+    ids = [vocab.to_index(word) for word in text]
+    # padding
+    ids += [vocab.padding_idx] * (max_len - len(text))
+    return ids[:max_len]
+
+def get_dataset(name, max_len, train_format="", test_format=""):
+
+    # datasets
+    train_dataset = load_dataset(name, split="train" + train_format).shuffle(seed=123)
+    test_dataset = load_dataset(name, split="test" + test_format).shuffle(seed=321)
+    split = train_dataset.train_test_split(test_size=0.2, seed=123)
+    train_dataset = split["train"]
+    val_dataset = split["test"]
+
+    vocab = Vocabulary()
+    vocab = set_vocabulary(vocab, train_dataset)
+    vocab = set_vocabulary(vocab, val_dataset)
+
+    train_dataset = process_data(train_dataset, vocab, max_len)
+    val_dataset = process_data(val_dataset, vocab, max_len)
+    test_dataset = process_data(test_dataset, vocab, max_len)
+
+    return TextClassificationDataset(train_dataset), TextClassificationDataset(val_dataset), \
+            TextClassificationDataset(test_dataset), vocab
+
+if __name__ == "__main__":
+
+    # 训练参数
+    max_len = 20
+    epochs = 40
+    lr = 1
+    batch_size = 64
+
+    features = 100
+    hidden_size = 128
+
+    # 获取数据集
+    # imdb.py SetFit/sst2
+    train_data, val_data, test_data, vocab = get_dataset("SetFit/sst2", max_len, "", "")
+    # 使用dataloader
+    train_dataloader = JittorDataLoader(
+        dataset=train_data,
+        batch_size=batch_size,
+        shuffle=True,
+        num_workers=4,
+    )
+    val_dataloader = JittorDataLoader(
+        dataset=val_data,
+        batch_size=batch_size,
+        shuffle=True,
+        num_workers=4,
+    )
+    test_dataloader = JittorDataLoader(
+        dataset=test_data,
+        batch_size=1,
+        shuffle=False,
+    )
+
+    # 初始化模型
+    model = LSTM(len(vocab), hidden_size, features)
+
+    # 优化器
+    # 也可以是多个优化器的list
+    optimizer = nn.SGD(model.parameters(), lr)
+
+    # Metrics
+    metrics = {"acc": Accuracy()}
+
+    # callbacks
+    callbacks = [
+        PrintWhileTrainingCallBack(print_every_epoch=1, print_every_batch=10),
+        # RichCallback(), # print_every参数默认为1，即每一个batch更新一次进度条
+    ]
+
+    trainer = Trainer(
+        model=model,
+        driver="jittor",
+        device=[0,1,2,3,4],
+        optimizers=optimizer,
+        train_dataloader=train_dataloader,
+        validate_dataloaders=val_dataloader,
+        validate_every=-1,
+        input_mapping=None,
+        output_mapping=None,
+        metrics=metrics,
+        n_epochs=epochs,
+        callbacks=callbacks,
+        # progress_bar="raw"
+    )
+    trainer.run()

tests/core/controllers/imdb.py

@@ -0,0 +1,110 @@
+# coding=utf-8
+# Copyright 2020 The TensorFlow Datasets Authors and the HuggingFace Datasets Authors.
+#
+# 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.
+
+# Lint as: python3
+"""IMDB movie reviews dataset."""
+
+import datasets
+from datasets.tasks import TextClassification
+
+
+_DESCRIPTION = """\
+Large Movie Review Dataset.
+This is a dataset for binary sentiment classification containing substantially \
+more data than previous benchmark datasets. We provide a set of 25,000 highly \
+polar movie reviews for training, and 25,000 for testing. There is additional \
+unlabeled data for use as well.\
+"""
+
+_CITATION = """\
+@InProceedings{maas-EtAl:2011:ACL-HLT2011,
+  author    = {Maas, Andrew L.  and  Daly, Raymond E.  and  Pham, Peter T.  and  Huang, Dan  and  Ng, Andrew Y.  and  Potts, Christopher},
+  title     = {Learning Word Vectors for Sentiment Analysis},
+  booktitle = {Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies},
+  month     = {June},
+  year      = {2011},
+  address   = {Portland, Oregon, USA},
+  publisher = {Association for Computational Linguistics},
+  pages     = {142--150},
+  url       = {http://www.aclweb.org/anthology/P11-1015}
+}
+"""
+
+_DOWNLOAD_URL = "http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz"
+
+
+class IMDBReviewsConfig(datasets.BuilderConfig):
+    """BuilderConfig for IMDBReviews."""
+
+    def __init__(self, **kwargs):
+        """BuilderConfig for IMDBReviews.
+        Args:
+          **kwargs: keyword arguments forwarded to super.
+        """
+        super(IMDBReviewsConfig, self).__init__(version=datasets.Version("1.0.0", ""), **kwargs)
+
+
+class Imdb(datasets.GeneratorBasedBuilder):
+    """IMDB movie reviews dataset."""
+
+    BUILDER_CONFIGS = [
+        IMDBReviewsConfig(
+            name="plain_text",
+            description="Plain text",
+        )
+    ]
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features(
+                {"text": datasets.Value("string"), "label": datasets.features.ClassLabel(names=["neg", "pos"])}
+            ),
+            supervised_keys=None,
+            homepage="http://ai.stanford.edu/~amaas/data/sentiment/",
+            citation=_CITATION,
+            task_templates=[TextClassification(text_column="text", label_column="label")],
+        )
+
+    def _split_generators(self, dl_manager):
+        archive = dl_manager.download(_DOWNLOAD_URL)
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN, gen_kwargs={"files": dl_manager.iter_archive(archive), "split": "train"}
+            ),
+            datasets.SplitGenerator(
+                name=datasets.Split.TEST, gen_kwargs={"files": dl_manager.iter_archive(archive), "split": "test"}
+            ),
+            datasets.SplitGenerator(
+                name=datasets.Split("unsupervised"),
+                gen_kwargs={"files": dl_manager.iter_archive(archive), "split": "train", "labeled": False},
+            ),
+        ]
+
+    def _generate_examples(self, files, split, labeled=True):
+        """Generate aclImdb examples."""
+        # For labeled examples, extract the label from the path.
+        if labeled:
+            label_mapping = {"pos": 1, "neg": 0}
+            for path, f in files:
+                if path.startswith(f"aclImdb/{split}"):
+                    label = label_mapping.get(path.split("/")[2])
+                    if label is not None:
+                        yield path, {"text": f.read().decode("utf-8"), "label": label}
+        else:
+            for path, f in files:
+                if path.startswith(f"aclImdb/{split}"):
+                    if path.split("/")[2] == "unsup":
+                        yield path, {"text": f.read().decode("utf-8"), "label": -1}

tests/core/controllers/test_trainer_jittor.py

@@ -11,6 +11,9 @@ if _NEED_IMPORT_JITTOR:
     import jittor as jt
     from jittor import nn, Module
     from jittor.dataset import Dataset
+else:
+    from fastNLP.core.utils.dummy_class import DummyClass as Module
+    from fastNLP.core.utils.dummy_class import DummyClass as Dataset
 
 
 class JittorNormalModel_Classification(Module):
@@ -68,6 +71,7 @@ class TrainJittorConfig:
 
 @pytest.mark.parametrize("driver,device", [("jittor", None)])
 @pytest.mark.parametrize("callbacks", [[RichCallback(100)]])
+@pytest.mark.jittor
 def test_trainer_jittor(
         driver,
         device,

tests/core/controllers/test_trainer_paddle.py

@@ -1,3 +1,5 @@
+import os
+from typing import List
 import pytest
 from dataclasses import dataclass
 
@@ -5,6 +7,7 @@ from fastNLP.core.controllers.trainer import Trainer
 from fastNLP.core.metrics.accuracy import Accuracy
 from fastNLP.core.callbacks.progress_callback import RichCallback
 from fastNLP.envs.imports import _NEED_IMPORT_PADDLE
+from fastNLP.envs.env import USER_CUDA_VISIBLE_DEVICES
 
 if _NEED_IMPORT_PADDLE:
     from paddle.optimizer import Adam
@@ -34,6 +37,8 @@ def test_trainer_paddle(
         callbacks,
         n_epochs=2,
 ):
+    if isinstance(device, List) and USER_CUDA_VISIBLE_DEVICES not in os.environ:
+        pytest.skip("Skip test fleet if FASTNLP_BACKEND is not set to paddle.")
     model = PaddleNormalModel_Classification_1(
         num_labels=TrainPaddleConfig.num_labels,
         feature_dimension=TrainPaddleConfig.feature_dimension

tests/core/drivers/torch_paddle_driver/_test_torch_paddle_driver.py

@@ -1,122 +0,0 @@
-import pytest
-
-from fastNLP.modules.mix_modules.mix_module import MixModule
-from fastNLP.core.drivers.torch_paddle_driver.torch_paddle_driver import TorchPaddleDriver
-from fastNLP.modules.mix_modules.utils import paddle2torch, torch2paddle
-
-import torch
-import paddle
-from paddle.io import Dataset, DataLoader
-import numpy as np
-
-############################################################################
-#
-# 测试在MNIST数据集上的表现
-#
-############################################################################
-
-class MNISTDataset(Dataset):
-    def __init__(self, dataset):
-
-        self.dataset = [
-            (
-                np.array(img).astype('float32').reshape(-1),
-                label
-            ) for img, label in dataset
-        ]
-
-    def __getitem__(self, idx):
-        return self.dataset[idx]
-
-    def __len__(self):
-        return len(self.dataset)
-
-class MixMNISTModel(MixModule):
-    def __init__(self):
-        super(MixMNISTModel, self).__init__()
-
-        self.fc1 = paddle.nn.Linear(784, 64)
-        self.fc2 = paddle.nn.Linear(64, 32)
-        self.fc3 = torch.nn.Linear(32, 10)
-        self.fc4 = torch.nn.Linear(10, 10)
-
-    def forward(self, x):
-
-        paddle_out = self.fc1(x)
-        paddle_out = self.fc2(paddle_out)
-        torch_in = paddle2torch(paddle_out)
-        torch_out = self.fc3(torch_in)
-        torch_out = self.fc4(torch_out)
-
-        return torch_out
-
-    def train_step(self, x):
-        return self.forward(x)
-
-    def test_step(self, x):
-        return self.forward(x)
-
-@pytest.mark.torchpaddle
-class TestMNIST:
-
-    @classmethod
-    def setup_class(self):
-
-        self.train_dataset = paddle.vision.datasets.MNIST(mode='train')
-        self.test_dataset = paddle.vision.datasets.MNIST(mode='test')
-        self.train_dataset = MNISTDataset(self.train_dataset)
-
-        self.lr = 0.0003
-        self.epochs = 20
-
-        self.dataloader = DataLoader(self.train_dataset, batch_size=100, shuffle=True)
-
-    def setup_method(self):
-        
-        model = MixMNISTModel()
-        self.torch_loss_func = torch.nn.CrossEntropyLoss()
-
-        torch_opt = torch.optim.Adam(model.parameters(backend="torch"), self.lr)
-        paddle_opt = paddle.optimizer.Adam(parameters=model.parameters(backend="paddle"), learning_rate=self.lr)
-
-        self.driver = TorchPaddleDriver(model=model, device="cuda:0")
-        self.driver.set_optimizers([torch_opt, paddle_opt])
-
-    def test_case1(self):
-
-        epochs = 20
-
-        self.driver.setup()
-        self.driver.zero_grad()
-        # 开始训练
-        current_epoch_idx = 0
-        while current_epoch_idx < epochs:
-            epoch_loss, batch = 0, 0
-            self.driver.set_model_mode("train")
-            self.driver.set_sampler_epoch(self.dataloader, current_epoch_idx)
-            for batch, (img, label) in enumerate(self.dataloader):
-                img = paddle.to_tensor(img).cuda()
-                torch_out = self.driver.train_step(img)
-                label = torch.from_numpy(label.numpy()).reshape(-1)
-                loss = self.torch_loss_func(torch_out.cpu(), label)
-                epoch_loss += loss.item()
-
-                self.driver.backward(loss)
-                self.driver.step()
-                self.driver.zero_grad()
-
-            current_epoch_idx += 1
-
-        # 开始测试
-        correct = 0
-        for img, label in self.test_dataset:
-
-            img = paddle.to_tensor(np.array(img).astype('float32').reshape(1, -1))
-            torch_out = self.driver.test_step(img)
-            res = torch_out.softmax(-1).argmax().item()
-            label = label.item()
-            if res == label:
-                correct += 1
-
-        acc = correct / len(self.test_dataset)
-        assert acc > 0.85

tests/core/utils/_test_torch_paddle_utils.py

@@ -1,204 +0,0 @@
-import paddle
-import pytest
-import torch
-
-from fastNLP.core.utils.torch_paddle_utils import torch_paddle_move_data_to_device
-
-############################################################################
-#
-# 测试将参数中包含的所有torch和paddle张量迁移到指定设备
-#
-############################################################################
-
-@pytest.mark.torchpaddle
-class TestTorchPaddleMoveDataToDevice:
-
-    def check_gpu(self, tensor, idx):
-        """
-        检查张量是否在指定显卡上的工具函数
-        """
-
-        if isinstance(tensor, paddle.Tensor):
-            assert tensor.place.is_gpu_place()
-            assert tensor.place.gpu_device_id() == idx
-        elif isinstance(tensor, torch.Tensor):
-            assert tensor.is_cuda
-            assert tensor.device.index == idx
-
-    def check_cpu(self, tensor):
-        if isinstance(tensor, paddle.Tensor):
-            assert tensor.place.is_cpu_place()
-        elif isinstance(tensor, torch.Tensor):
-            assert not tensor.is_cuda
-
-    def test_tensor_transfer(self):
-        """
-        测试迁移单个张量
-        """
-
-        paddle_tensor = paddle.rand((3, 4, 5)).cpu()
-        res = torch_paddle_move_data_to_device(paddle_tensor, device=None, data_device=None)
-        self.check_cpu(res)
-
-        res = torch_paddle_move_data_to_device(paddle_tensor, device="gpu:0", data_device=None)
-        self.check_gpu(res, 0)
-
-        res = torch_paddle_move_data_to_device(paddle_tensor, device="gpu:1", data_device=None)
-        self.check_gpu(res, 1)
-
-        res = torch_paddle_move_data_to_device(paddle_tensor, device="cuda:0", data_device="cpu")
-        self.check_gpu(res, 0)
-
-        res = torch_paddle_move_data_to_device(paddle_tensor, device=None, data_device="gpu:0")
-        self.check_gpu(res, 0)
-
-        res = torch_paddle_move_data_to_device(paddle_tensor, device=None, data_device="cuda:1")
-        self.check_gpu(res, 1)
-
-        torch_tensor = torch.rand(3, 4, 5)
-        res = torch_paddle_move_data_to_device(torch_tensor, device=None, data_device=None)
-        self.check_cpu(res)
-
-        res = torch_paddle_move_data_to_device(torch_tensor, device="gpu:0", data_device=None)
-        self.check_gpu(res, 0)
-
-        res = torch_paddle_move_data_to_device(torch_tensor, device="gpu:1", data_device=None)
-        self.check_gpu(res, 1)
-
-        res = torch_paddle_move_data_to_device(torch_tensor, device="gpu:0", data_device="cpu")
-        self.check_gpu(res, 0)
-
-        res = torch_paddle_move_data_to_device(torch_tensor, device=None, data_device="gpu:0")
-        self.check_gpu(res, 0)
-
-        res = torch_paddle_move_data_to_device(torch_tensor, device=None, data_device="gpu:1")
-        self.check_gpu(res, 1)
-
-    def test_list_transfer(self):
-        """
-        测试迁移张量的列表
-        """
-
-        paddle_list = [paddle.rand((6, 4, 2)) for i in range(5)] + [torch.rand((6, 4, 2)) for i in range(5)]
-        res = torch_paddle_move_data_to_device(paddle_list, device=None, data_device="gpu:1")
-        assert isinstance(res, list)
-        for r in res:
-            self.check_gpu(r, 1)
-
-        res = torch_paddle_move_data_to_device(paddle_list, device="cpu", data_device="gpu:1")
-        assert isinstance(res, list)
-        for r in res:
-            self.check_cpu(r)
-
-        res = torch_paddle_move_data_to_device(paddle_list, device="gpu:0", data_device=None)
-        assert isinstance(res, list)
-        for r in res:
-            self.check_gpu(r, 0)
-
-        res = torch_paddle_move_data_to_device(paddle_list, device="gpu:1", data_device="cpu")
-        assert isinstance(res, list)
-        for r in res:
-            self.check_gpu(r, 1)
-
-    def test_tensor_tuple_transfer(self):
-        """
-        测试迁移张量的元组
-        """
-
-        paddle_list = [paddle.rand((6, 4, 2)) for i in range(10)] + [torch.rand((6, 4, 2)) for i in range(5)]
-        paddle_tuple = tuple(paddle_list)
-        res = torch_paddle_move_data_to_device(paddle_tuple, device=None, data_device="gpu:1")
-        assert isinstance(res, tuple)
-        for r in res:
-            self.check_gpu(r, 1)
-
-        res = torch_paddle_move_data_to_device(paddle_tuple, device="cpu", data_device="gpu:1")
-        assert isinstance(res, tuple)
-        for r in res:
-            self.check_cpu(r)
-
-        res = torch_paddle_move_data_to_device(paddle_tuple, device="gpu:0", data_device=None)
-        assert isinstance(res, tuple)
-        for r in res:
-            self.check_gpu(r, 0)
-
-        res = torch_paddle_move_data_to_device(paddle_tuple, device="gpu:1", data_device="cpu")
-        assert isinstance(res, tuple)
-        for r in res:
-            self.check_gpu(r, 1)
-
-    def test_dict_transfer(self):
-        """
-        测试迁移复杂的字典结构
-        """
-
-        paddle_dict = {
-            "torch_tensor": torch.rand((3, 4)),
-            "torch_list": [torch.rand((6, 4, 2)) for i in range(10)],
-            "dict":{
-                "list": [paddle.rand((6, 4, 2)) for i in range(5)] + [torch.rand((6, 4, 2)) for i in range(5)],
-                "torch_tensor": torch.rand((3, 4)),
-                "paddle_tensor": paddle.rand((3, 4))
-            },
-            "paddle_tensor": paddle.rand((3, 4)),
-            "list": [paddle.rand((6, 4, 2)) for i in range(10)] ,
-            "int": 2,
-            "string": "test string"
-        }
-
-        res = torch_paddle_move_data_to_device(paddle_dict, device="gpu:0", data_device=None)
-        assert isinstance(res, dict)
-        self.check_gpu(res["torch_tensor"], 0)
-        self.check_gpu(res["paddle_tensor"], 0)
-        assert isinstance(res["torch_list"], list)
-        for t in res["torch_list"]:
-            self.check_gpu(t, 0)
-        assert isinstance(res["list"], list)
-        for t in res["list"]:
-            self.check_gpu(t, 0)
-        assert isinstance(res["int"], int)
-        assert isinstance(res["string"], str)
-        assert isinstance(res["dict"], dict)
-        assert isinstance(res["dict"]["list"], list)
-        for t in res["dict"]["list"]:
-            self.check_gpu(t, 0)
-        self.check_gpu(res["dict"]["torch_tensor"], 0)
-        self.check_gpu(res["dict"]["paddle_tensor"], 0)
-
-        res = torch_paddle_move_data_to_device(paddle_dict, device=None, data_device="gpu:1")
-        assert isinstance(res, dict)
-        self.check_gpu(res["torch_tensor"], 1)
-        self.check_gpu(res["paddle_tensor"], 1)
-        assert isinstance(res["torch_list"], list)
-        for t in res["torch_list"]:
-            self.check_gpu(t, 1)
-        assert isinstance(res["list"], list)
-        for t in res["list"]:
-            self.check_gpu(t, 1)
-        assert isinstance(res["int"], int)
-        assert isinstance(res["string"], str)
-        assert isinstance(res["dict"], dict)
-        assert isinstance(res["dict"]["list"], list)
-        for t in res["dict"]["list"]:
-            self.check_gpu(t, 1)
-        self.check_gpu(res["dict"]["torch_tensor"], 1)
-        self.check_gpu(res["dict"]["paddle_tensor"], 1)
-
-        res = torch_paddle_move_data_to_device(paddle_dict, device="cpu", data_device="gpu:0")
-        assert isinstance(res, dict)
-        self.check_cpu(res["torch_tensor"])
-        self.check_cpu(res["paddle_tensor"])
-        assert isinstance(res["torch_list"], list)
-        for t in res["torch_list"]:
-            self.check_cpu(t)
-        assert isinstance(res["list"], list)
-        for t in res["list"]:
-            self.check_cpu(t)
-        assert isinstance(res["int"], int)
-        assert isinstance(res["string"], str)
-        assert isinstance(res["dict"], dict)
-        assert isinstance(res["dict"]["list"], list)
-        for t in res["dict"]["list"]:
-            self.check_cpu(t)
-        self.check_cpu(res["dict"]["torch_tensor"])
-        self.check_cpu(res["dict"]["paddle_tensor"])

tests/core/utils/test_paddle_utils.py

@@ -2,37 +2,42 @@ import os
 
 import pytest
 
-from fastNLP.core.utils.paddle_utils import get_device_from_visible, paddle_to, paddle_move_data_to_device
+from fastNLP.core.utils.paddle_utils import _convert_data_device, paddle_to, paddle_move_data_to_device
 from fastNLP.envs.imports import _NEED_IMPORT_PADDLE
 if _NEED_IMPORT_PADDLE:
     import paddle
+
 @pytest.mark.parametrize(
-    ("user_visible_devices, cuda_visible_devices, device, output_type, correct"),
+    ("user_visible_devices, cuda_visible_devices, device, correct"),
     (
-        ("0,1,2,3,4,5,6,7", "0", "cpu", str, "cpu"),
-        ("0,1,2,3,4,5,6,7", "0", "cpu", int, "cpu"),
-        ("0,1,2,3,4,5,6,7", "3,4,5", "gpu:4", int, 1),
-        ("0,1,2,3,4,5,6,7", "3,4,5", "gpu:5", str, "gpu:2"),
-        ("3,4,5,6", "3,5", 0, int, 0),
-        ("3,6,7,8", "6,7,8", "gpu:2", str, "gpu:1"),
+        (None, None, 1, "gpu:1"),
+        (None, "2,4,5,6", 2, "gpu:2"),
+        (None, "3,4,5", 1, "gpu:1"),
+        ("0,1,2,3,4,5,6,7", "0", "cpu", "cpu"),
+        ("3,4,5,6,7", "0", "cpu", "cpu"),
+        ("0,1,2,3,4,5,6,7", "3,4,5", "gpu:4", "gpu:1"),
+        ("0,1,2,3,4,5,6,7", "3,4,5", "gpu:5", "gpu:2"),
+        ("3,4,5,6", "3,5", 0, "gpu:0"),
+        ("3,6,7,8", "6,7,8", "gpu:2", "gpu:1"),
     )
 )
-@pytest.mark.paddle
-def test_get_device_from_visible(user_visible_devices, cuda_visible_devices, device, output_type, correct):
+def test_convert_data_device(user_visible_devices, cuda_visible_devices, device, correct):
     _cuda_visible_devices = os.getenv("CUDA_VISIBLE_DEVICES")
     _user_visible_devices = os.getenv("USER_CUDA_VISIBLE_DEVICES")
-    os.environ["CUDA_VISIBLE_DEVICES"] = cuda_visible_devices
-    os.environ["USER_CUDA_VISIBLE_DEVICES"] = user_visible_devices
-    res = get_device_from_visible(device, output_type)
+    if cuda_visible_devices is not None:
+        os.environ["CUDA_VISIBLE_DEVICES"] = cuda_visible_devices
+    if user_visible_devices is not None:
+        os.environ["USER_CUDA_VISIBLE_DEVICES"] = user_visible_devices
+    res = _convert_data_device(device)
     assert res == correct
 
     # 还原环境变量
     if _cuda_visible_devices is None:
-        del os.environ["CUDA_VISIBLE_DEVICES"]
+        os.environ.pop("CUDA_VISIBLE_DEVICES", None)
     else:
         os.environ["CUDA_VISIBLE_DEVICES"] = _cuda_visible_devices
     if _user_visible_devices is None:
-        del os.environ["USER_CUDA_VISIBLE_DEVICES"]
+        os.environ.pop("USER_CUDA_VISIBLE_DEVICES", None)
     else:
         os.environ["USER_CUDA_VISIBLE_DEVICES"] = _user_visible_devices