diff --git a/fastNLP/core/vocabulary.py b/fastNLP/core/vocabulary.py index 5e8ae989..8c3050ba 100644 --- a/fastNLP/core/vocabulary.py +++ b/fastNLP/core/vocabulary.py @@ -117,6 +117,8 @@ class Vocabulary(object): :param str word: 新词 """ + if word in self._no_create_word: + self._no_create_word.pop(word) self.add(word) @_check_build_status @@ -126,6 +128,9 @@ class Vocabulary(object): :param list[str] word_lst: 词的序列 """ + for word in word_lst: + if word in self._no_create_word: + self._no_create_word.pop(word) self.update(word_lst) def build_vocab(self): diff --git a/fastNLP/io/data_loader/sst.py b/fastNLP/io/data_loader/sst.py index 1e1b8bef..8d0d005f 100644 --- a/fastNLP/io/data_loader/sst.py +++ b/fastNLP/io/data_loader/sst.py @@ -1,10 +1,11 @@ from typing import Iterable from nltk import Tree +import spacy from ..base_loader import DataInfo, DataSetLoader from ...core.vocabulary import VocabularyOption, Vocabulary from ...core.dataset import DataSet from ...core.instance import Instance -from ..embed_loader import EmbeddingOption, EmbedLoader +from ..utils import check_dataloader_paths, get_tokenizer class SSTLoader(DataSetLoader): @@ -34,6 +35,7 @@ class SSTLoader(DataSetLoader): tag_v['0'] = tag_v['1'] tag_v['4'] = tag_v['3'] self.tag_v = tag_v + self.tokenizer = get_tokenizer() def _load(self, path): """ @@ -52,29 +54,37 @@ class SSTLoader(DataSetLoader): ds.append(Instance(words=words, target=tag)) return ds - @staticmethod - def _get_one(data, subtree): + def _get_one(self, data, subtree): tree = Tree.fromstring(data) if subtree: - return [(t.leaves(), t.label()) for t in tree.subtrees()] - return [(tree.leaves(), tree.label())] + return [([x.text for x in self.tokenizer(' '.join(t.leaves()))], t.label()) for t in tree.subtrees() ] + return [([x.text for x in self.tokenizer(' '.join(tree.leaves()))], tree.label())] def process(self, - paths, - train_ds: Iterable[str] = None, + paths, train_subtree=True, src_vocab_op: VocabularyOption = None, - tgt_vocab_op: VocabularyOption = None, - src_embed_op: EmbeddingOption = None): + tgt_vocab_op: VocabularyOption = None,): + paths = check_dataloader_paths(paths) input_name, target_name = 'words', 'target' src_vocab = Vocabulary() if src_vocab_op is None else Vocabulary(**src_vocab_op) tgt_vocab = Vocabulary(unknown=None, padding=None) \ if tgt_vocab_op is None else Vocabulary(**tgt_vocab_op) - info = DataInfo(datasets=self.load(paths)) - _train_ds = [info.datasets[name] - for name in train_ds] if train_ds else info.datasets.values() - src_vocab.from_dataset(*_train_ds, field_name=input_name) - tgt_vocab.from_dataset(*_train_ds, field_name=target_name) + info = DataInfo() + origin_subtree = self.subtree + self.subtree = train_subtree + info.datasets['train'] = self._load(paths['train']) + self.subtree = origin_subtree + for n, p in paths.items(): + if n != 'train': + info.datasets[n] = self._load(p) + + src_vocab.from_dataset( + info.datasets['train'], + field_name=input_name, + no_create_entry_dataset=[ds for n, ds in info.datasets.items() if n != 'train']) + tgt_vocab.from_dataset(info.datasets['train'], field_name=target_name) + src_vocab.index_dataset( *info.datasets.values(), field_name=input_name, new_field_name=input_name) @@ -86,10 +96,5 @@ class SSTLoader(DataSetLoader): target_name: tgt_vocab } - if src_embed_op is not None: - src_embed_op.vocab = src_vocab - init_emb = EmbedLoader.load_with_vocab(**src_embed_op) - info.embeddings[input_name] = init_emb - return info diff --git a/fastNLP/io/utils.py b/fastNLP/io/utils.py new file mode 100644 index 00000000..a7d2de85 --- /dev/null +++ b/fastNLP/io/utils.py @@ -0,0 +1,69 @@ +import os + +from typing import Union, Dict + + +def check_dataloader_paths(paths:Union[str, Dict[str, str]])->Dict[str, str]: + """ + 检查传入dataloader的文件的合法性。如果为合法路径，将返回至少包含'train'这个key的dict。类似于下面的结果 + { + 'train': '/some/path/to/', # 一定包含，建词表应该在这上面建立，剩下的其它文件应该只需要处理并index。 + 'test': 'xxx' # 可能有，也可能没有 + ... + } + 如果paths为不合法的，将直接进行raise相应的错误 + + :param paths: 路径. 可以为一个文件路径(则认为该文件就是train的文件); 可以为一个文件目录，将在该目录下寻找train(文件名 + 中包含train这个字段), test.txt, dev.txt; 可以为一个dict, 则key是用户自定义的某个文件的名称，value是这个文件的路径。 + :return: + """ + if isinstance(paths, str): + if os.path.isfile(paths): + return {'train': paths} + elif os.path.isdir(paths): + filenames = os.listdir(paths) + files = {} + for filename in filenames: + path_pair = None + if 'train' in filename: + path_pair = ('train', filename) + if 'dev' in filename: + if path_pair: + raise Exception("File:{} in {} contains bot `{}` and `dev`.".format(filename, paths, path_pair[0])) + path_pair = ('dev', filename) + if 'test' in filename: + if path_pair: + raise Exception("File:{} in {} contains bot `{}` and `test`.".format(filename, paths, path_pair[0])) + path_pair = ('test', filename) + if path_pair: + files[path_pair[0]] = os.path.join(paths, path_pair[1]) + return files + else: + raise FileNotFoundError(f"{paths} is not a valid file path.") + + elif isinstance(paths, dict): + if paths: + if 'train' not in paths: + raise KeyError("You have to include `train` in your dict.") + for key, value in paths.items(): + if isinstance(key, str) and isinstance(value, str): + if not os.path.isfile(value): + raise TypeError(f"{value} is not a valid file.") + else: + raise TypeError("All keys and values in paths should be str.") + return paths + else: + raise ValueError("Empty paths is not allowed.") + else: + raise TypeError(f"paths only supports str and dict. not {type(paths)}.") + +def get_tokenizer(): + try: + import spacy + spacy.prefer_gpu() + en = spacy.load('en') + print('use spacy tokenizer') + return lambda x: [w.text for w in en.tokenizer(x)] + except Exception as e: + print('use raw tokenizer') + return lambda x: x.split() diff --git a/fastNLP/models/star_transformer.py b/fastNLP/models/star_transformer.py index 4c944a54..1aba5a8c 100644 --- a/fastNLP/models/star_transformer.py +++ b/fastNLP/models/star_transformer.py @@ -46,7 +46,7 @@ class StarTransEnc(nn.Module): super(StarTransEnc, self).__init__() self.embedding = get_embeddings(init_embed) emb_dim = self.embedding.embedding_dim - self.emb_fc = nn.Linear(emb_dim, hidden_size) + #self.emb_fc = nn.Linear(emb_dim, hidden_size) self.emb_drop = nn.Dropout(emb_dropout) self.encoder = StarTransformer(hidden_size=hidden_size, num_layers=num_layers, @@ -65,7 +65,7 @@ class StarTransEnc(nn.Module): [batch, hidden] 全局 relay 节点, 详见论文 """ x = self.embedding(x) - x = self.emb_fc(self.emb_drop(x)) + #x = self.emb_fc(self.emb_drop(x)) nodes, relay = self.encoder(x, mask) return nodes, relay @@ -205,7 +205,7 @@ class STSeqCls(nn.Module): max_len=max_len, emb_dropout=emb_dropout, dropout=dropout) - self.cls = _Cls(hidden_size, num_cls, cls_hidden_size) + self.cls = _Cls(hidden_size, num_cls, cls_hidden_size, dropout=dropout) def forward(self, words, seq_len): """ diff --git a/fastNLP/modules/aggregator/attention.py b/fastNLP/modules/aggregator/attention.py index 4101b033..2bee7f2e 100644 --- a/fastNLP/modules/aggregator/attention.py +++ b/fastNLP/modules/aggregator/attention.py @@ -19,7 +19,7 @@ class DotAttention(nn.Module): 补上文档 """ - def __init__(self, key_size, value_size, dropout=0): + def __init__(self, key_size, value_size, dropout=0.0): super(DotAttention, self).__init__() self.key_size = key_size self.value_size = value_size @@ -37,7 +37,7 @@ class DotAttention(nn.Module): """ output = torch.matmul(Q, K.transpose(1, 2)) / self.scale if mask_out is not None: - output.masked_fill_(mask_out, -1e8) + output.masked_fill_(mask_out, -1e18) output = self.softmax(output) output = self.drop(output) return torch.matmul(output, V) @@ -67,9 +67,8 @@ class MultiHeadAttention(nn.Module): self.k_in = nn.Linear(input_size, in_size) self.v_in = nn.Linear(input_size, in_size) # follow the paper, do not apply dropout within dot-product - self.attention = DotAttention(key_size=key_size, value_size=value_size, dropout=0) + self.attention = DotAttention(key_size=key_size, value_size=value_size, dropout=dropout) self.out = nn.Linear(value_size * num_head, input_size) - self.drop = TimestepDropout(dropout) self.reset_parameters() def reset_parameters(self): @@ -105,7 +104,7 @@ class MultiHeadAttention(nn.Module): # concat all heads, do output linear atte = atte.permute(1, 2, 0, 3).contiguous().view(batch, sq, -1) - output = self.drop(self.out(atte)) + output = self.out(atte) return output diff --git a/fastNLP/modules/encoder/_elmo.py b/fastNLP/modules/encoder/_elmo.py index 4ebee819..a49634bb 100644 --- a/fastNLP/modules/encoder/_elmo.py +++ b/fastNLP/modules/encoder/_elmo.py @@ -1,12 +1,13 @@ - """ -这个页面的代码大量参考了https://github.com/HIT-SCIR/ELMoForManyLangs/tree/master/elmoformanylangs +这个页面的代码大量参考了 allenNLP """ - from typing import Optional, Tuple, List, Callable import os + +import h5py +import numpy import torch import torch.nn as nn import torch.nn.functional as F @@ -16,7 +17,6 @@ import json from ..utils import get_dropout_mask import codecs -from torch import autograd class LstmCellWithProjection(torch.nn.Module): """ @@ -58,6 +58,7 @@ class LstmCellWithProjection(torch.nn.Module): respectively. The first dimension is 1 in order to match the Pytorch API for returning stacked LSTM states. """ + def __init__(self, input_size: int, hidden_size: int, @@ -129,13 +130,13 @@ class LstmCellWithProjection(torch.nn.Module): # We have to use this '.data.new().fill_' pattern to create tensors with the correct # type - forward has no knowledge of whether these are torch.Tensors or torch.cuda.Tensors. output_accumulator = inputs.data.new(batch_size, - total_timesteps, - self.hidden_size).fill_(0) + total_timesteps, + self.hidden_size).fill_(0) if initial_state is None: full_batch_previous_memory = inputs.data.new(batch_size, - self.cell_size).fill_(0) + self.cell_size).fill_(0) full_batch_previous_state = inputs.data.new(batch_size, - self.hidden_size).fill_(0) + self.hidden_size).fill_(0) else: full_batch_previous_state = initial_state[0].squeeze(0) full_batch_previous_memory = initial_state[1].squeeze(0) @@ -169,7 +170,7 @@ class LstmCellWithProjection(torch.nn.Module): # Second conditional: Does the next shortest sequence beyond the current batch # index require computation use this timestep? while current_length_index < (len(batch_lengths) - 1) and \ - batch_lengths[current_length_index + 1] > index: + batch_lengths[current_length_index + 1] > index: current_length_index += 1 # Actually get the slices of the batch which we @@ -256,7 +257,7 @@ class LstmbiLm(nn.Module): inputs = inputs[sort_idx] inputs = nn.utils.rnn.pack_padded_sequence(inputs, sort_lens, batch_first=self.batch_first) output, hx = self.encoder(inputs, None) # -> [N,L,C] - output, _ = nn.util.rnn.pad_packed_sequence(output, batch_first=self.batch_first) + output, _ = nn.utils.rnn.pad_packed_sequence(output, batch_first=self.batch_first) _, unsort_idx = torch.sort(sort_idx, dim=0, descending=False) output = output[unsort_idx] forward, backward = output.split(self.config['encoder']['dim'], 2) @@ -316,13 +317,13 @@ class ElmobiLm(torch.nn.Module): :param seq_len: batch_size :return: torch.FloatTensor. num_layers x batch_size x max_len x hidden_size """ + max_len = inputs.size(1) sort_lens, sort_idx = torch.sort(seq_len, dim=0, descending=True) inputs = inputs[sort_idx] inputs = nn.utils.rnn.pack_padded_sequence(inputs, sort_lens, batch_first=True) output, _ = self._lstm_forward(inputs, None) _, unsort_idx = torch.sort(sort_idx, dim=0, descending=False) output = output[:, unsort_idx] - return output def _lstm_forward(self, @@ -399,7 +400,7 @@ class ElmobiLm(torch.nn.Module): torch.cat([forward_state[1], backward_state[1]], -1))) stacked_sequence_outputs: torch.FloatTensor = torch.stack(sequence_outputs) - # Stack the hidden state and memory for each layer into 2 tensors of shape + # Stack the hidden state and memory for each layer in。to 2 tensors of shape # (num_layers, batch_size, hidden_size) and (num_layers, batch_size, cell_size) # respectively. final_hidden_states, final_memory_states = zip(*final_states) @@ -408,6 +409,66 @@ class ElmobiLm(torch.nn.Module): torch.cat(final_memory_states, 0)) return stacked_sequence_outputs, final_state_tuple + def load_weights(self, weight_file: str) -> None: + """ + Load the pre-trained weights from the file. + """ + requires_grad = False + + with h5py.File(weight_file, 'r') as fin: + for i_layer, lstms in enumerate( + zip(self.forward_layers, self.backward_layers) + ): + for j_direction, lstm in enumerate(lstms): + # lstm is an instance of LSTMCellWithProjection + cell_size = lstm.cell_size + + dataset = fin['RNN_%s' % j_direction]['RNN']['MultiRNNCell']['Cell%s' % i_layer + ]['LSTMCell'] + + # tensorflow packs together both W and U matrices into one matrix, + # but pytorch maintains individual matrices. In addition, tensorflow + # packs the gates as input, memory, forget, output but pytorch + # uses input, forget, memory, output. So we need to modify the weights. + tf_weights = numpy.transpose(dataset['W_0'][...]) + torch_weights = tf_weights.copy() + + # split the W from U matrices + input_size = lstm.input_size + input_weights = torch_weights[:, :input_size] + recurrent_weights = torch_weights[:, input_size:] + tf_input_weights = tf_weights[:, :input_size] + tf_recurrent_weights = tf_weights[:, input_size:] + + # handle the different gate order convention + for torch_w, tf_w in [[input_weights, tf_input_weights], + [recurrent_weights, tf_recurrent_weights]]: + torch_w[(1 * cell_size):(2 * cell_size), :] = tf_w[(2 * cell_size):(3 * cell_size), :] + torch_w[(2 * cell_size):(3 * cell_size), :] = tf_w[(1 * cell_size):(2 * cell_size), :] + + lstm.input_linearity.weight.data.copy_(torch.FloatTensor(input_weights)) + lstm.state_linearity.weight.data.copy_(torch.FloatTensor(recurrent_weights)) + lstm.input_linearity.weight.requires_grad = requires_grad + lstm.state_linearity.weight.requires_grad = requires_grad + + # the bias weights + tf_bias = dataset['B'][...] + # tensorflow adds 1.0 to forget gate bias instead of modifying the + # parameters... + tf_bias[(2 * cell_size):(3 * cell_size)] += 1 + torch_bias = tf_bias.copy() + torch_bias[(1 * cell_size):(2 * cell_size) + ] = tf_bias[(2 * cell_size):(3 * cell_size)] + torch_bias[(2 * cell_size):(3 * cell_size) + ] = tf_bias[(1 * cell_size):(2 * cell_size)] + lstm.state_linearity.bias.data.copy_(torch.FloatTensor(torch_bias)) + lstm.state_linearity.bias.requires_grad = requires_grad + + # the projection weights + proj_weights = numpy.transpose(dataset['W_P_0'][...]) + lstm.state_projection.weight.data.copy_(torch.FloatTensor(proj_weights)) + lstm.state_projection.weight.requires_grad = requires_grad + class LstmTokenEmbedder(nn.Module): def __init__(self, config, word_emb_layer, char_emb_layer): @@ -441,7 +502,7 @@ class LstmTokenEmbedder(nn.Module): chars_emb = self.char_emb_layer(chars) # TODO 这里应该要考虑seq_len的问题 _, (chars_outputs, __) = self.char_lstm(chars_emb) - chars_outputs = chars_outputs.contiguous().view(-1, self.config['token_embedder']['char_dim'] * 2) + chars_outputs = chars_outputs.contiguous().view(-1, self.config['token_embedder']['embedding']['dim'] * 2) embs.append(chars_outputs) token_embedding = torch.cat(embs, dim=2) @@ -450,79 +511,143 @@ class LstmTokenEmbedder(nn.Module): class ConvTokenEmbedder(nn.Module): - def __init__(self, config, word_emb_layer, char_emb_layer): + def __init__(self, config, weight_file, word_emb_layer, char_emb_layer, char_vocab): super(ConvTokenEmbedder, self).__init__() - self.config = config + self.weight_file = weight_file self.word_emb_layer = word_emb_layer self.char_emb_layer = char_emb_layer self.output_dim = config['encoder']['projection_dim'] - self.emb_dim = 0 - if word_emb_layer is not None: - self.emb_dim += word_emb_layer.weight.size(1) - - if char_emb_layer is not None: - self.convolutions = [] - cnn_config = config['token_embedder'] - filters = cnn_config['filters'] - char_embed_dim = cnn_config['char_dim'] - - for i, (width, num) in enumerate(filters): - conv = torch.nn.Conv1d( - in_channels=char_embed_dim, - out_channels=num, - kernel_size=width, - bias=True - ) - self.convolutions.append(conv) - - self.convolutions = nn.ModuleList(self.convolutions) - - self.n_filters = sum(f[1] for f in filters) - self.n_highway = cnn_config['n_highway'] - - self.highways = Highway(self.n_filters, self.n_highway, activation=torch.nn.functional.relu) - self.emb_dim += self.n_filters - - self.projection = nn.Linear(self.emb_dim, self.output_dim, bias=True) + self._options = config + self.requires_grad = False + self._load_weights() + self._char_embedding_weights = char_emb_layer.weight.data + + def _load_weights(self): + self._load_cnn_weights() + self._load_highway() + self._load_projection() + + def _load_cnn_weights(self): + cnn_options = self._options['token_embedder'] + filters = cnn_options['filters'] + char_embed_dim = cnn_options['embedding']['dim'] + + convolutions = [] + for i, (width, num) in enumerate(filters): + conv = torch.nn.Conv1d( + in_channels=char_embed_dim, + out_channels=num, + kernel_size=width, + bias=True + ) + # load the weights + with h5py.File(self.weight_file, 'r') as fin: + weight = fin['CNN']['W_cnn_{}'.format(i)][...] + bias = fin['CNN']['b_cnn_{}'.format(i)][...] + + w_reshaped = numpy.transpose(weight.squeeze(axis=0), axes=(2, 1, 0)) + if w_reshaped.shape != tuple(conv.weight.data.shape): + raise ValueError("Invalid weight file") + conv.weight.data.copy_(torch.FloatTensor(w_reshaped)) + conv.bias.data.copy_(torch.FloatTensor(bias)) + + conv.weight.requires_grad = self.requires_grad + conv.bias.requires_grad = self.requires_grad + + convolutions.append(conv) + self.add_module('char_conv_{}'.format(i), conv) + + self._convolutions = convolutions + + def _load_highway(self): + # the highway layers have same dimensionality as the number of cnn filters + cnn_options = self._options['token_embedder'] + filters = cnn_options['filters'] + n_filters = sum(f[1] for f in filters) + n_highway = cnn_options['n_highway'] + + # create the layers, and load the weights + self._highways = Highway(n_filters, n_highway, activation=torch.nn.functional.relu) + for k in range(n_highway): + # The AllenNLP highway is one matrix multplication with concatenation of + # transform and carry weights. + with h5py.File(self.weight_file, 'r') as fin: + # The weights are transposed due to multiplication order assumptions in tf + # vs pytorch (tf.matmul(X, W) vs pytorch.matmul(W, X)) + w_transform = numpy.transpose(fin['CNN_high_{}'.format(k)]['W_transform'][...]) + # -1.0 since AllenNLP is g * x + (1 - g) * f(x) but tf is (1 - g) * x + g * f(x) + w_carry = -1.0 * numpy.transpose(fin['CNN_high_{}'.format(k)]['W_carry'][...]) + weight = numpy.concatenate([w_transform, w_carry], axis=0) + self._highways._layers[k].weight.data.copy_(torch.FloatTensor(weight)) + self._highways._layers[k].weight.requires_grad = self.requires_grad + + b_transform = fin['CNN_high_{}'.format(k)]['b_transform'][...] + b_carry = -1.0 * fin['CNN_high_{}'.format(k)]['b_carry'][...] + bias = numpy.concatenate([b_transform, b_carry], axis=0) + self._highways._layers[k].bias.data.copy_(torch.FloatTensor(bias)) + self._highways._layers[k].bias.requires_grad = self.requires_grad + + def _load_projection(self): + cnn_options = self._options['token_embedder'] + filters = cnn_options['filters'] + n_filters = sum(f[1] for f in filters) + + self._projection = torch.nn.Linear(n_filters, self.output_dim, bias=True) + with h5py.File(self.weight_file, 'r') as fin: + weight = fin['CNN_proj']['W_proj'][...] + bias = fin['CNN_proj']['b_proj'][...] + self._projection.weight.data.copy_(torch.FloatTensor(numpy.transpose(weight))) + self._projection.bias.data.copy_(torch.FloatTensor(bias)) + + self._projection.weight.requires_grad = self.requires_grad + self._projection.bias.requires_grad = self.requires_grad def forward(self, words, chars): - embs = [] - if self.word_emb_layer is not None: - if hasattr(self, 'words_to_words'): - words = self.words_to_words[words] - word_emb = self.word_emb_layer(words) - embs.append(word_emb) + """ + :param words: + :param chars: Tensor Shape ``(batch_size, sequence_length, 50)``: + :return Tensor Shape ``(batch_size, sequence_length + 2, embedding_dim)`` : + """ + # the character id embedding + # (batch_size * sequence_length, max_chars_per_token, embed_dim) + # character_embedding = torch.nn.functional.embedding( + # chars.view(-1, max_chars_per_token), + # self._char_embedding_weights + # ) + batch_size, sequence_length, max_char_len = chars.size() + character_embedding = self.char_emb_layer(chars).reshape(batch_size*sequence_length, max_char_len, -1) + # run convolutions + cnn_options = self._options['token_embedder'] + if cnn_options['activation'] == 'tanh': + activation = torch.tanh + elif cnn_options['activation'] == 'relu': + activation = torch.nn.functional.relu + else: + raise Exception("Unknown activation") - if self.char_emb_layer is not None: - batch_size, seq_len, _ = chars.size() - chars = chars.view(batch_size * seq_len, -1) - character_embedding = self.char_emb_layer(chars) - character_embedding = torch.transpose(character_embedding, 1, 2) - - cnn_config = self.config['token_embedder'] - if cnn_config['activation'] == 'tanh': - activation = torch.nn.functional.tanh - elif cnn_config['activation'] == 'relu': - activation = torch.nn.functional.relu - else: - raise Exception("Unknown activation") + # (batch_size * sequence_length, embed_dim, max_chars_per_token) + character_embedding = torch.transpose(character_embedding, 1, 2) + convs = [] + for i in range(len(self._convolutions)): + conv = getattr(self, 'char_conv_{}'.format(i)) + convolved = conv(character_embedding) + # (batch_size * sequence_length, n_filters for this width) + convolved, _ = torch.max(convolved, dim=-1) + convolved = activation(convolved) + convs.append(convolved) - convs = [] - for i in range(len(self.convolutions)): - convolved = self.convolutions[i](character_embedding) - # (batch_size * sequence_length, n_filters for this width) - convolved, _ = torch.max(convolved, dim=-1) - convolved = activation(convolved) - convs.append(convolved) - char_emb = torch.cat(convs, dim=-1) - char_emb = self.highways(char_emb) + # (batch_size * sequence_length, n_filters) + token_embedding = torch.cat(convs, dim=-1) - embs.append(char_emb.view(batch_size, -1, self.n_filters)) + # apply the highway layers (batch_size * sequence_length, n_filters) + token_embedding = self._highways(token_embedding) - token_embedding = torch.cat(embs, dim=2) + # final projection (batch_size * sequence_length, embedding_dim) + token_embedding = self._projection(token_embedding) - return self.projection(token_embedding) + # reshape to (batch_size, sequence_length+2, embedding_dim) + return token_embedding.view(batch_size, sequence_length, -1) class Highway(torch.nn.Module): @@ -543,6 +668,7 @@ class Highway(torch.nn.Module): activation : ``Callable[[torch.Tensor], torch.Tensor]``, optional (default=``torch.nn.functional.relu``) The non-linearity to use in the highway layers. """ + def __init__(self, input_dim: int, num_layers: int = 1, @@ -573,6 +699,7 @@ class Highway(torch.nn.Module): current_input = gate * linear_part + (1 - gate) * nonlinear_part return current_input + class _ElmoModel(nn.Module): """ 该Module是ElmoEmbedding中进行所有的heavy lifting的地方。做的工作，包括 @@ -582,11 +709,32 @@ class _ElmoModel(nn.Module): (4) 设计一个保存token的embedding，允许缓存word的表示。 """ - def __init__(self, model_dir:str, vocab:Vocabulary=None, cache_word_reprs:bool=False): + + def __init__(self, model_dir: str, vocab: Vocabulary = None, cache_word_reprs: bool = False): super(_ElmoModel, self).__init__() - config = json.load(open(os.path.join(model_dir, 'structure_config.json'), 'r')) + dir = os.walk(model_dir) + config_file = None + weight_file = None + config_count = 0 + weight_count = 0 + for path, dir_list, file_list in dir: + for file_name in file_list: + if file_name.__contains__(".json"): + config_file = file_name + config_count += 1 + elif file_name.__contains__(".hdf5"): + weight_file = file_name + weight_count += 1 + if config_count > 1 or weight_count > 1: + raise Exception(f"Multiple config files(*.json) or weight files(*.hdf5) detected in {model_dir}.") + elif config_count == 0 or weight_count == 0: + raise Exception(f"No config file or weight file found in {model_dir}") + + config = json.load(open(os.path.join(model_dir, config_file), 'r')) + self.weight_file = os.path.join(model_dir, weight_file) self.config = config + self.requires_grad = False OOV_TAG = '' PAD_TAG = '' @@ -595,48 +743,8 @@ class _ElmoModel(nn.Module): BOW_TAG = '' EOW_TAG = '' - # 将加载embedding放到这里 - token_embedder_states = torch.load(os.path.join(model_dir, 'token_embedder.pkl'), map_location='cpu') - - # For the model trained with word form word encoder. - if config['token_embedder']['word_dim'] > 0: - word_lexicon = {} - with codecs.open(os.path.join(model_dir, 'word.dic'), 'r', encoding='utf-8') as fpi: - for line in fpi: - tokens = line.strip().split('\t') - if len(tokens) == 1: - tokens.insert(0, '\u3000') - token, i = tokens - word_lexicon[token] = int(i) - # 做一些sanity check - for special_word in [PAD_TAG, OOV_TAG, BOS_TAG, EOS_TAG]: - assert special_word in word_lexicon, f"{special_word} not found in word.dic." - # 根据vocab调整word_embedding - pre_word_embedding = token_embedder_states.pop('word_emb_layer.embedding.weight') - word_emb_layer = nn.Embedding(len(vocab)+2, config['token_embedder']['word_dim']) #多增加两个是为了与 - found_word_count = 0 - for word, index in vocab: - if index == vocab.unknown_idx: # 因为fastNLP的unknow是 而在这里是所以ugly强制适配一下 - index_in_pre = word_lexicon[OOV_TAG] - found_word_count += 1 - elif index == vocab.padding_idx: # 需要pad对齐 - index_in_pre = word_lexicon[PAD_TAG] - found_word_count += 1 - elif word in word_lexicon: - index_in_pre = word_lexicon[word] - found_word_count += 1 - else: - index_in_pre = word_lexicon[OOV_TAG] - word_emb_layer.weight.data[index] = pre_word_embedding[index_in_pre] - print(f"{found_word_count} out of {len(vocab)} words were found in pretrained elmo embedding.") - word_emb_layer.weight.data[-1] = pre_word_embedding[word_lexicon[EOS_TAG]] - word_emb_layer.weight.data[-2] = pre_word_embedding[word_lexicon[BOS_TAG]] - self.word_vocab = vocab - else: - word_emb_layer = None - # For the model trained with character-based word encoder. - if config['token_embedder']['char_dim'] > 0: + if config['token_embedder']['embedding']['dim'] > 0: char_lexicon = {} with codecs.open(os.path.join(model_dir, 'char.dic'), 'r', encoding='utf-8') as fpi: for line in fpi: @@ -645,22 +753,26 @@ class _ElmoModel(nn.Module): tokens.insert(0, '\u3000') token, i = tokens char_lexicon[token] = int(i) + # 做一些sanity check for special_word in [PAD_TAG, OOV_TAG, BOW_TAG, EOW_TAG]: assert special_word in char_lexicon, f"{special_word} not found in char.dic." + # 从vocab中构建char_vocab char_vocab = Vocabulary(unknown=OOV_TAG, padding=PAD_TAG) # 需要保证与在里面 - char_vocab.add_word(BOW_TAG) - char_vocab.add_word(EOW_TAG) + char_vocab.add_word_lst([BOW_TAG, EOW_TAG, BOS_TAG, EOS_TAG]) + for word, index in vocab: char_vocab.add_word_lst(list(word)) - # 保证, 也在 - char_vocab.add_word_lst(list(BOS_TAG)) - char_vocab.add_word_lst(list(EOS_TAG)) - # 根据char_lexicon调整 - char_emb_layer = nn.Embedding(len(char_vocab), int(config['token_embedder']['char_dim'])) - pre_char_embedding = token_embedder_states.pop('char_emb_layer.embedding.weight') + + self.bos_index, self.eos_index, self._pad_index = len(vocab), len(vocab)+1, vocab.padding_idx + # 根据char_lexicon调整, 多设置一位，是预留给word padding的(该位置的char表示为全0表示) + char_emb_layer = nn.Embedding(len(char_vocab)+1, int(config['token_embedder']['embedding']['dim']), + padding_idx=len(char_vocab)) + with h5py.File(self.weight_file, 'r') as fin: + char_embed_weights = fin['char_embed'][...] + char_embed_weights = torch.from_numpy(char_embed_weights) found_char_count = 0 for char, index in char_vocab: # 调整character embedding if char in char_lexicon: @@ -668,79 +780,84 @@ class _ElmoModel(nn.Module): found_char_count += 1 else: index_in_pre = char_lexicon[OOV_TAG] - char_emb_layer.weight.data[index] = pre_char_embedding[index_in_pre] + char_emb_layer.weight.data[index] = char_embed_weights[index_in_pre] + print(f"{found_char_count} out of {len(char_vocab)} characters were found in pretrained elmo embedding.") # 生成words到chars的映射 if config['token_embedder']['name'].lower() == 'cnn': max_chars = config['token_embedder']['max_characters_per_token'] elif config['token_embedder']['name'].lower() == 'lstm': - max_chars = max(map(lambda x: len(x[0]), vocab)) + 2 # 需要补充两个与 + max_chars = max(map(lambda x: len(x[0]), vocab)) + 2 # 需要补充两个与 else: raise ValueError('Unknown token_embedder: {0}'.format(config['token_embedder']['name'])) - # 增加, 所以加2. + self.words_to_chars_embedding = nn.Parameter(torch.full((len(vocab)+2, max_chars), - fill_value=char_vocab.to_index(PAD_TAG), dtype=torch.long), + fill_value=len(char_vocab), + dtype=torch.long), requires_grad=False) - for word, index in vocab: - if len(word)+2>max_chars: - word = word[:max_chars-2] - if index==vocab.padding_idx: # 如果是pad的话，需要和给定的对齐 - word = PAD_TAG - elif index==vocab.unknown_idx: - word = OOV_TAG - char_ids = [char_vocab.to_index(BOW_TAG)] + [char_vocab.to_index(c) for c in word] + [char_vocab.to_index(EOW_TAG)] - char_ids += [char_vocab.to_index(PAD_TAG)]*(max_chars-len(char_ids)) + for word, index in list(iter(vocab)) + [(BOS_TAG, len(vocab)), (EOS_TAG, len(vocab)+1)]: + if len(word) + 2 > max_chars: + word = word[:max_chars - 2] + if index == self._pad_index: + continue + elif word == BOS_TAG or word == EOS_TAG: + char_ids = [char_vocab.to_index(BOW_TAG)] + [char_vocab.to_index(word)] + [ + char_vocab.to_index(EOW_TAG)] + char_ids += [char_vocab.to_index(PAD_TAG)] * (max_chars - len(char_ids)) + else: + char_ids = [char_vocab.to_index(BOW_TAG)] + [char_vocab.to_index(c) for c in word] + [ + char_vocab.to_index(EOW_TAG)] + char_ids += [char_vocab.to_index(PAD_TAG)] * (max_chars - len(char_ids)) self.words_to_chars_embedding[index] = torch.LongTensor(char_ids) - for index, word in enumerate([BOS_TAG, EOS_TAG]): # 加上, - if len(word)+2>max_chars: - word = word[:max_chars-2] - char_ids = [char_vocab.to_index(BOW_TAG)] + [char_vocab.to_index(c) for c in word] + [char_vocab.to_index(EOW_TAG)] - char_ids += [char_vocab.to_index(PAD_TAG)]*(max_chars-len(char_ids)) - self.words_to_chars_embedding[index+len(vocab)] = torch.LongTensor(char_ids) + self.char_vocab = char_vocab else: char_emb_layer = None if config['token_embedder']['name'].lower() == 'cnn': self.token_embedder = ConvTokenEmbedder( - config, word_emb_layer, char_emb_layer) + config, self.weight_file, None, char_emb_layer, self.char_vocab) elif config['token_embedder']['name'].lower() == 'lstm': self.token_embedder = LstmTokenEmbedder( - config, word_emb_layer, char_emb_layer) - self.token_embedder.load_state_dict(token_embedder_states, strict=False) - if config['token_embedder']['word_dim'] > 0 and vocab._no_create_word_length > 0: # 需要映射，使得来自于dev, test的idx指向unk - words_to_words = nn.Parameter(torch.arange(len(vocab)+2).long(), requires_grad=False) + config, None, char_emb_layer) + + if config['token_embedder']['word_dim'] > 0 \ + and vocab._no_create_word_length > 0: # 需要映射，使得来自于dev, test的idx指向unk + words_to_words = nn.Parameter(torch.arange(len(vocab) + 2).long(), requires_grad=False) for word, idx in vocab: if vocab._is_word_no_create_entry(word): words_to_words[idx] = vocab.unknown_idx setattr(self.token_embedder, 'words_to_words', words_to_words) self.output_dim = config['encoder']['projection_dim'] + # 暂时只考虑 elmo if config['encoder']['name'].lower() == 'elmo': self.encoder = ElmobiLm(config) elif config['encoder']['name'].lower() == 'lstm': self.encoder = LstmbiLm(config) - self.encoder.load_state_dict(torch.load(os.path.join(model_dir, 'encoder.pkl'), - map_location='cpu')) - self.bos_index = len(vocab) - self.eos_index = len(vocab) + 1 - self._pad_index = vocab.padding_idx + self.encoder.load_weights(self.weight_file) if cache_word_reprs: - if config['token_embedder']['char_dim']>0: # 只有在使用了chars的情况下有用 + if config['token_embedder']['embedding']['dim'] > 0: # 只有在使用了chars的情况下有用 print("Start to generate cache word representations.") batch_size = 320 - num_batches = self.words_to_chars_embedding.size(0)//batch_size + \ - int(self.words_to_chars_embedding.size(0)%batch_size!=0) - self.cached_word_embedding = nn.Embedding(self.words_to_chars_embedding.size(0), + # bos eos + word_size = self.words_to_chars_embedding.size(0) + num_batches = word_size // batch_size + \ + int(word_size % batch_size != 0) + + self.cached_word_embedding = nn.Embedding(word_size, config['encoder']['projection_dim']) with torch.no_grad(): for i in range(num_batches): - words = torch.arange(i*batch_size, min((i+1)*batch_size, self.words_to_chars_embedding.size(0))).long() + words = torch.arange(i * batch_size, + min((i + 1) * batch_size, word_size)).long() chars = self.words_to_chars_embedding[words].unsqueeze(1) # batch_size x 1 x max_chars - word_reprs = self.token_embedder(words.unsqueeze(1), chars).detach() # batch_size x 1 x config['encoder']['projection_dim'] + word_reprs = self.token_embedder(words.unsqueeze(1), + chars).detach() # batch_size x 1 x config['encoder']['projection_dim'] self.cached_word_embedding.weight.data[words] = word_reprs.squeeze(1) + print("Finish generating cached word representations. Going to delete the character encoder.") del self.token_embedder, self.words_to_chars_embedding else: @@ -758,7 +875,7 @@ class _ElmoModel(nn.Module): seq_len = words.ne(self._pad_index).sum(dim=-1) expanded_words[:, 1:-1] = words expanded_words[:, 0].fill_(self.bos_index) - expanded_words[torch.arange(batch_size).to(words), seq_len+1] = self.eos_index + expanded_words[torch.arange(batch_size).to(words), seq_len + 1] = self.eos_index seq_len = seq_len + 2 if hasattr(self, 'cached_word_embedding'): token_embedding = self.cached_word_embedding(expanded_words) @@ -767,16 +884,18 @@ class _ElmoModel(nn.Module): chars = self.words_to_chars_embedding[expanded_words] else: chars = None - token_embedding = self.token_embedder(expanded_words, chars) + token_embedding = self.token_embedder(expanded_words, chars) # batch_size x max_len x embed_dim + if self.config['encoder']['name'] == 'elmo': encoder_output = self.encoder(token_embedding, seq_len) - if encoder_output.size(2) < max_len+2: - dummy_tensor = encoder_output.new_zeros(encoder_output.size(0), batch_size, - max_len + 2 - encoder_output.size(2), encoder_output.size(-1)) - encoder_output = torch.cat([encoder_output, dummy_tensor], 2) - sz = encoder_output.size() # 2, batch_size, max_len, hidden_size - token_embedding = torch.cat([token_embedding, token_embedding], dim=2).view(1, sz[1], sz[2], sz[3]) - encoder_output = torch.cat([token_embedding, encoder_output], dim=0) + if encoder_output.size(2) < max_len + 2: + num_layers, _, output_len, hidden_size = encoder_output.size() + dummy_tensor = encoder_output.new_zeros(num_layers, batch_size, + max_len + 2 - output_len, hidden_size) + encoder_output = torch.cat((encoder_output, dummy_tensor), 2) + sz = encoder_output.size() # 2, batch_size, max_len, hidden_size + token_embedding = torch.cat((token_embedding, token_embedding), dim=2).view(1, sz[1], sz[2], sz[3]) + encoder_output = torch.cat((token_embedding, encoder_output), dim=0) elif self.config['encoder']['name'] == 'lstm': encoder_output = self.encoder(token_embedding, seq_len) else: @@ -784,5 +903,4 @@ class _ElmoModel(nn.Module): # 删除, . 这里没有精确地删除，但应该也不会影响最后的结果了。 encoder_output = encoder_output[:, :, 1:-1] - return encoder_output diff --git a/fastNLP/modules/encoder/embedding.py b/fastNLP/modules/encoder/embedding.py index 005cfe75..0d6f30e3 100644 --- a/fastNLP/modules/encoder/embedding.py +++ b/fastNLP/modules/encoder/embedding.py @@ -179,16 +179,16 @@ class StaticEmbedding(TokenEmbedding): :param model_dir_or_name: 可以有两种方式调用预训练好的static embedding：第一种是传入embedding的文件名，第二种是传入embedding 的名称。目前支持的embedding包括{`en` 或者 `en-glove-840b-300` : glove.840B.300d, `en-glove-6b-50` : glove.6B.50d, `en-word2vec-300` : GoogleNews-vectors-negative300}。第二种情况将自动查看缓存中是否存在该模型，没有的话将自动下载。 - :param requires_grad: 是否需要gradient. 默认为True - :param init_method: 如何初始化没有找到的值。可以使用torch.nn.init.*中各种方法。调用该方法时传入一个tensor对象。 - :param normailize: 是否对vector进行normalize，使得每个vector的norm为1。 + :param bool requires_grad: 是否需要gradient. 默认为True + :param callable init_method: 如何初始化没有找到的值。可以使用torch.nn.init.*中各种方法。调用该方法时传入一个tensor对象。 + :param bool normailize: 是否对vector进行normalize，使得每个vector的norm为1。 + :param bool lower: 是否将vocab中的词语小写后再和预训练的词表进行匹配。如果你的词表中包含大写的词语，或者就是需要单独 + 为大写的词语开辟一个vector表示，则将lower设置为False。 """ def __init__(self, vocab: Vocabulary, model_dir_or_name: str='en', requires_grad: bool=True, init_method=None, - normalize=False): + normalize=False, lower=False): super(StaticEmbedding, self).__init__(vocab) - # 优先定义需要下载的static embedding有哪些。这里估计需要自己搞一个server， - # 得到cache_path if model_dir_or_name.lower() in PRETRAIN_STATIC_FILES: PRETRAIN_URL = _get_base_url('static') @@ -202,8 +202,40 @@ class StaticEmbedding(TokenEmbedding): raise ValueError(f"Cannot recognize {model_dir_or_name}.") # 读取embedding - embedding = self._load_with_vocab(model_path, vocab=vocab, init_method=init_method, - normalize=normalize) + if lower: + lowered_vocab = Vocabulary(padding=vocab.padding, unknown=vocab.unknown) + for word, index in vocab: + if not vocab._is_word_no_create_entry(word): + lowered_vocab.add_word(word.lower()) # 先加入需要创建entry的 + for word in vocab._no_create_word.keys(): # 不需要创建entry的 + if word in vocab: + lowered_word = word.lower() + if lowered_word not in lowered_vocab.word_count: + lowered_vocab.add_word(lowered_word) + lowered_vocab._no_create_word[lowered_word] += 1 + print(f"All word in vocab have been lowered. There are {len(vocab)} words, {len(lowered_vocab)} unique lowered " + f"words.") + embedding = self._load_with_vocab(model_path, vocab=lowered_vocab, init_method=init_method, + normalize=normalize) + # 需要适配一下 + if not hasattr(self, 'words_to_words'): + self.words_to_words = torch.arange(len(lowered_vocab, )).long() + if lowered_vocab.unknown: + unknown_idx = lowered_vocab.unknown_idx + else: + unknown_idx = embedding.size(0) - 1 # 否则是最后一个为unknow + words_to_words = nn.Parameter(torch.full((len(vocab),), fill_value=unknown_idx).long(), + requires_grad=False) + for word, index in vocab: + if word not in lowered_vocab: + word = word.lower() + if lowered_vocab._is_word_no_create_entry(word): # 如果不需要创建entry,已经默认unknown了 + continue + words_to_words[index] = self.words_to_words[lowered_vocab.to_index(word)] + self.words_to_words = words_to_words + else: + embedding = self._load_with_vocab(model_path, vocab=vocab, init_method=init_method, + normalize=normalize) self.embedding = nn.Embedding(num_embeddings=embedding.shape[0], embedding_dim=embedding.shape[1], padding_idx=vocab.padding_idx, max_norm=None, norm_type=2, scale_grad_by_freq=False, @@ -301,7 +333,7 @@ class StaticEmbedding(TokenEmbedding): if vocab._no_create_word_length>0: if vocab.unknown is None: # 创建一个专门的unknown unknown_idx = len(matrix) - vectors = torch.cat([vectors, torch.zeros(1, dim)], dim=0).contiguous() + vectors = torch.cat((vectors, torch.zeros(1, dim)), dim=0).contiguous() else: unknown_idx = vocab.unknown_idx words_to_words = nn.Parameter(torch.full((len(vocab),), fill_value=unknown_idx).long(), @@ -438,19 +470,15 @@ class ElmoEmbedding(ContextualEmbedding): :param model_dir_or_name: 可以有两种方式调用预训练好的ELMo embedding：第一种是传入ELMo权重的文件名，第二种是传入ELMo版本的名称， 目前支持的ELMo包括{`en` : 英文版本的ELMo, `cn` : 中文版本的ELMo,}。第二种情况将自动查看缓存中是否存在该模型，没有的话将自动下载 :param layers: str, 指定返回的层数, 以,隔开不同的层。如果要返回第二层的结果'2', 返回后两层的结果'1,2'。不同的层的结果 - 按照这个顺序concat起来。默认为'2'。 - :param requires_grad: bool, 该层是否需要gradient. 默认为False + 按照这个顺序concat起来。默认为'2'。'mix'会使用可学习的权重结合不同层的表示(权重是否可训练与requires_grad保持一致， + 初始化权重对三层结果进行mean-pooling, 可以通过ElmoEmbedding.set_mix_weights_requires_grad()方法只将mix weights设置为可学习。) + :param requires_grad: bool, 该层是否需要gradient, 默认为False. :param cache_word_reprs: 可以选择对word的表示进行cache; 设置为True的话，将在初始化的时候为每个word生成对应的embedding， 并删除character encoder，之后将直接使用cache的embedding。默认为False。 """ def __init__(self, vocab: Vocabulary, model_dir_or_name: str='en', layers: str='2', requires_grad: bool=False, cache_word_reprs: bool=False): super(ElmoEmbedding, self).__init__(vocab) - layers = list(map(int, layers.split(','))) - assert len(layers) > 0, "Must choose one output" - for layer in layers: - assert 0 <= layer <= 2, "Layer index should be in range [0, 2]." - self.layers = layers # 根据model_dir_or_name检查是否存在并下载 if model_dir_or_name.lower() in PRETRAINED_ELMO_MODEL_DIR: @@ -464,8 +492,49 @@ class ElmoEmbedding(ContextualEmbedding): else: raise ValueError(f"Cannot recognize {model_dir_or_name}.") self.model = _ElmoModel(model_dir, vocab, cache_word_reprs=cache_word_reprs) + + if layers=='mix': + self.layer_weights = nn.Parameter(torch.zeros(self.model.config['encoder']['n_layers']+1), + requires_grad=requires_grad) + self.gamma = nn.Parameter(torch.ones(1), requires_grad=requires_grad) + self._get_outputs = self._get_mixed_outputs + self._embed_size = self.model.config['encoder']['projection_dim'] * 2 + else: + layers = list(map(int, layers.split(','))) + assert len(layers) > 0, "Must choose one output" + for layer in layers: + assert 0 <= layer <= 2, "Layer index should be in range [0, 2]." + self.layers = layers + self._get_outputs = self._get_layer_outputs + self._embed_size = len(self.layers) * self.model.config['encoder']['projection_dim'] * 2 + self.requires_grad = requires_grad - self._embed_size = len(self.layers) * self.model.config['encoder']['projection_dim'] * 2 + + def _get_mixed_outputs(self, outputs): + # outputs: num_layers x batch_size x max_len x hidden_size + # return: batch_size x max_len x hidden_size + weights = F.softmax(self.layer_weights+1/len(outputs), dim=0).to(outputs) + outputs = torch.einsum('l,lbij->bij', weights, outputs) + return self.gamma.to(outputs)*outputs + + def set_mix_weights_requires_grad(self, flag=True): + """ + 当初始化ElmoEmbedding时layers被设置为mix时，可以通过调用该方法设置mix weights是否可训练。如果layers不是mix，调用 + 该方法没有用。 + :param bool flag: 混合不同层表示的结果是否可以训练。 + :return: + """ + if hasattr(self, 'layer_weights'): + self.layer_weights.requires_grad = flag + self.gamma.requires_grad = flag + + def _get_layer_outputs(self, outputs): + if len(self.layers) == 1: + outputs = outputs[self.layers[0]] + else: + outputs = torch.cat(tuple([*outputs[self.layers]]), dim=-1) + + return outputs def forward(self, words: torch.LongTensor): """ @@ -480,15 +549,12 @@ class ElmoEmbedding(ContextualEmbedding): if outputs is not None: return outputs outputs = self.model(words) - if len(self.layers) == 1: - outputs = outputs[self.layers[0]] - else: - outputs = torch.cat([*outputs[self.layers]], dim=-1) - - return outputs + return self._get_outputs(outputs) def _delete_model_weights(self): - del self.layers, self.model + for name in ['layers', 'model', 'layer_weights', 'gamma']: + if hasattr(self, name): + delattr(self, name) @property def requires_grad(self): @@ -892,10 +958,11 @@ class StackEmbedding(TokenEmbedding): def __init__(self, embeds: List[TokenEmbedding]): vocabs = [] for embed in embeds: - vocabs.append(embed.get_word_vocab()) + if hasattr(embed, 'get_word_vocab'): + vocabs.append(embed.get_word_vocab()) _vocab = vocabs[0] for vocab in vocabs[1:]: - assert vocab == _vocab, "All embeddings should use the same word vocabulary." + assert vocab == _vocab, "All embeddings in StackEmbedding should use the same word vocabulary." super(StackEmbedding, self).__init__(_vocab) assert isinstance(embeds, list) diff --git a/fastNLP/modules/encoder/star_transformer.py b/fastNLP/modules/encoder/star_transformer.py index 1eec7c13..76b7e922 100644 --- a/fastNLP/modules/encoder/star_transformer.py +++ b/fastNLP/modules/encoder/star_transformer.py @@ -35,11 +35,13 @@ class StarTransformer(nn.Module): self.iters = num_layers self.norm = nn.ModuleList([nn.LayerNorm(hidden_size) for _ in range(self.iters)]) + self.emb_fc = nn.Conv2d(hidden_size, hidden_size, 1) + self.emb_drop = nn.Dropout(dropout) self.ring_att = nn.ModuleList( - [_MSA1(hidden_size, nhead=num_head, head_dim=head_dim, dropout=dropout) + [_MSA1(hidden_size, nhead=num_head, head_dim=head_dim, dropout=0.0) for _ in range(self.iters)]) self.star_att = nn.ModuleList( - [_MSA2(hidden_size, nhead=num_head, head_dim=head_dim, dropout=dropout) + [_MSA2(hidden_size, nhead=num_head, head_dim=head_dim, dropout=0.0) for _ in range(self.iters)]) if max_len is not None: @@ -66,18 +68,19 @@ class StarTransformer(nn.Module): smask = torch.cat([torch.zeros(B, 1, ).byte().to(mask), mask], 1) embs = data.permute(0, 2, 1)[:, :, :, None] # B H L 1 - if self.pos_emb: + if self.pos_emb and False: P = self.pos_emb(torch.arange(L, dtype=torch.long, device=embs.device) \ .view(1, L)).permute(0, 2, 1).contiguous()[:, :, :, None] # 1 H L 1 embs = embs + P - + embs = norm_func(self.emb_drop, embs) nodes = embs relay = embs.mean(2, keepdim=True) ex_mask = mask[:, None, :, None].expand(B, H, L, 1) r_embs = embs.view(B, H, 1, L) for i in range(self.iters): ax = torch.cat([r_embs, relay.expand(B, H, 1, L)], 2) - nodes = nodes + F.leaky_relu(self.ring_att[i](norm_func(self.norm[i], nodes), ax=ax)) + nodes = F.leaky_relu(self.ring_att[i](norm_func(self.norm[i], nodes), ax=ax)) + #nodes = F.leaky_relu(self.ring_att[i](nodes, ax=ax)) relay = F.leaky_relu(self.star_att[i](relay, torch.cat([relay, nodes], 2), smask)) nodes = nodes.masked_fill_(ex_mask, 0) diff --git a/reproduction/Star_transformer/README.md b/reproduction/Star_transformer/README.md index 37c5f1e9..d07d5536 100644 --- a/reproduction/Star_transformer/README.md +++ b/reproduction/Star_transformer/README.md @@ -6,7 +6,7 @@ paper: [Star-Transformer](https://arxiv.org/abs/1902.09113) |Pos Tagging|CTB 9.0|-|ACC 92.31| |Pos Tagging|CONLL 2012|-|ACC 96.51| |Named Entity Recognition|CONLL 2012|-|F1 85.66| -|Text Classification|SST|-|49.18| +|Text Classification|SST|-|51.2| |Natural Language Inference|SNLI|-|83.76| ## Usage diff --git a/reproduction/Star_transformer/datasets.py b/reproduction/Star_transformer/datasets.py index 1532a041..1173d1a0 100644 --- a/reproduction/Star_transformer/datasets.py +++ b/reproduction/Star_transformer/datasets.py @@ -51,13 +51,15 @@ def load_sst(path, files): for sub in [True, False, False]] ds_list = [loader.load(os.path.join(path, fn)) for fn, loader in zip(files, loaders)] - word_v = Vocabulary(min_freq=2) + word_v = Vocabulary(min_freq=0) tag_v = Vocabulary(unknown=None, padding=None) for ds in ds_list: ds.apply(lambda x: [w.lower() for w in x['words']], new_field_name='words') - ds_list[0].drop(lambda x: len(x['words']) < 3) + #ds_list[0].drop(lambda x: len(x['words']) < 3) update_v(word_v, ds_list[0], 'words') + update_v(word_v, ds_list[1], 'words') + update_v(word_v, ds_list[2], 'words') ds_list[0].apply(lambda x: tag_v.add_word( x['target']), new_field_name=None) @@ -152,7 +154,10 @@ class EmbedLoader: # some words from vocab are missing in pre-trained embedding # we normally sample each dimension vocab_embed = embedding_matrix[np.where(hit_flags)] - sampled_vectors = np.random.normal(vocab_embed.mean(axis=0), vocab_embed.std(axis=0), + #sampled_vectors = np.random.normal(vocab_embed.mean(axis=0), vocab_embed.std(axis=0), + # size=(len(vocab) - np.sum(hit_flags), emb_dim)) + sampled_vectors = np.random.uniform(-0.01, 0.01, size=(len(vocab) - np.sum(hit_flags), emb_dim)) + embedding_matrix[np.where(1 - hit_flags)] = sampled_vectors return embedding_matrix diff --git a/reproduction/Star_transformer/run.sh b/reproduction/Star_transformer/run.sh index 0972c662..5cd6954b 100644 --- a/reproduction/Star_transformer/run.sh +++ b/reproduction/Star_transformer/run.sh @@ -1,5 +1,5 @@ #python -u train.py --task pos --ds conll --mode train --gpu 1 --lr 3e-4 --w_decay 2e-5 --lr_decay .95 --drop 0.3 --ep 25 --bsz 64 > conll_pos102.log 2>&1 & #python -u train.py --task pos --ds ctb --mode train --gpu 1 --lr 3e-4 --w_decay 2e-5 --lr_decay .95 --drop 0.3 --ep 25 --bsz 64 > ctb_pos101.log 2>&1 & -#python -u train.py --task cls --ds sst --mode train --gpu 2 --lr 1e-4 --w_decay 1e-5 --lr_decay 0.9 --drop 0.5 --ep 50 --bsz 128 > sst_cls201.log & +python -u train.py --task cls --ds sst --mode train --gpu 0 --lr 1e-4 --w_decay 5e-5 --lr_decay 1.0 --drop 0.4 --ep 20 --bsz 64 > sst_cls.log & #python -u train.py --task nli --ds snli --mode train --gpu 1 --lr 1e-4 --w_decay 1e-5 --lr_decay 0.9 --drop 0.4 --ep 120 --bsz 128 > snli_nli201.log & -python -u train.py --task ner --ds conll --mode train --gpu 0 --lr 1e-4 --w_decay 1e-5 --lr_decay 0.9 --drop 0.4 --ep 120 --bsz 64 > conll_ner201.log & +#python -u train.py --task ner --ds conll --mode train --gpu 0 --lr 1e-4 --w_decay 1e-5 --lr_decay 0.9 --drop 0.4 --ep 120 --bsz 64 > conll_ner201.log & diff --git a/reproduction/Star_transformer/train.py b/reproduction/Star_transformer/train.py index 6fb58daf..480748df 100644 --- a/reproduction/Star_transformer/train.py +++ b/reproduction/Star_transformer/train.py @@ -1,4 +1,6 @@ from util import get_argparser, set_gpu, set_rng_seeds, add_model_args +seed = set_rng_seeds(15360) +print('RNG SEED {}'.format(seed)) from datasets import load_seqtag, load_sst, load_snli, EmbedLoader, MAX_LEN import torch.nn as nn import torch @@ -7,8 +9,8 @@ import fastNLP as FN from fastNLP.models.star_transformer import STSeqLabel, STSeqCls, STNLICls from fastNLP.core.const import Const as C import sys -sys.path.append('/remote-home/yfshao/workdir/dev_fastnlp/') - +#sys.path.append('/remote-home/yfshao/workdir/dev_fastnlp/') +pre_dir = '/home/ec2-user/fast_data/' g_model_select = { 'pos': STSeqLabel, @@ -17,8 +19,8 @@ g_model_select = { 'nli': STNLICls, } -g_emb_file_path = {'en': '/remote-home/yfshao/workdir/datasets/word_vector/glove.840B.300d.txt', - 'zh': '/remote-home/yfshao/workdir/datasets/word_vector/cc.zh.300.vec'} +g_emb_file_path = {'en': pre_dir + 'glove.840B.300d.txt', + 'zh': pre_dir + 'cc.zh.300.vec'} g_args = None g_model_cfg = None @@ -53,7 +55,7 @@ def get_conll2012_ner(): def get_sst(): - path = '/remote-home/yfshao/workdir/datasets/SST' + path = pre_dir + 'sst' files = ['train.txt', 'dev.txt', 'test.txt'] return load_sst(path, files) @@ -94,6 +96,7 @@ class MyCallback(FN.core.callback.Callback): nn.utils.clip_grad.clip_grad_norm_(self.model.parameters(), 5.0) def on_step_end(self): + return warm_steps = 6000 # learning rate warm-up & decay if self.step <= warm_steps: @@ -108,12 +111,11 @@ class MyCallback(FN.core.callback.Callback): def train(): - seed = set_rng_seeds(1234) - print('RNG SEED {}'.format(seed)) print('loading data') ds_list, word_v, tag_v = g_datasets['{}-{}'.format( g_args.ds, g_args.task)]() print(ds_list[0][:2]) + print(len(ds_list[0]), len(ds_list[1]), len(ds_list[2])) embed = load_pretrain_emb(word_v, lang='zh' if g_args.ds == 'ctb' else 'en') g_model_cfg['num_cls'] = len(tag_v) print(g_model_cfg) @@ -123,11 +125,14 @@ def train(): def init_model(model): for p in model.parameters(): if p.size(0) != len(word_v): - nn.init.normal_(p, 0.0, 0.05) + if len(p.size())<2: + nn.init.constant_(p, 0.0) + else: + nn.init.normal_(p, 0.0, 0.05) init_model(model) train_data = ds_list[0] - dev_data = ds_list[2] - test_data = ds_list[1] + dev_data = ds_list[1] + test_data = ds_list[2] print(tag_v.word2idx) if g_args.task in ['pos', 'ner']: @@ -145,14 +150,26 @@ def train(): } metric_key, metric = metrics[g_args.task] device = 'cuda' if torch.cuda.is_available() else 'cpu' - ex_param = [x for x in model.parameters( - ) if x.requires_grad and x.size(0) != len(word_v)] - optim_cfg = [{'params': model.enc.embedding.parameters(), 'lr': g_args.lr*0.1}, - {'params': ex_param, 'lr': g_args.lr, 'weight_decay': g_args.w_decay}, ] - trainer = FN.Trainer(train_data=train_data, model=model, optimizer=torch.optim.Adam(optim_cfg), loss=loss, - batch_size=g_args.bsz, n_epochs=g_args.ep, print_every=10, dev_data=dev_data, metrics=metric, - metric_key=metric_key, validate_every=3000, save_path=g_args.log, use_tqdm=False, - device=device, callbacks=[MyCallback()]) + + params = [(x,y) for x,y in list(model.named_parameters()) if y.requires_grad and y.size(0) != len(word_v)] + no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] + print([n for n,p in params]) + optim_cfg = [ + #{'params': model.enc.embedding.parameters(), 'lr': g_args.lr*0.1}, + {'params': [p for n, p in params if not any(nd in n for nd in no_decay)], 'lr': g_args.lr, 'weight_decay': 1.0*g_args.w_decay}, + {'params': [p for n, p in params if any(nd in n for nd in no_decay)], 'lr': g_args.lr, 'weight_decay': 0.0*g_args.w_decay} + ] + + print(model) + trainer = FN.Trainer(model=model, train_data=train_data, dev_data=dev_data, + loss=loss, metrics=metric, metric_key=metric_key, + optimizer=torch.optim.Adam(optim_cfg), + n_epochs=g_args.ep, batch_size=g_args.bsz, print_every=100, validate_every=1000, + device=device, + use_tqdm=False, prefetch=False, + save_path=g_args.log, + sampler=FN.BucketSampler(100, g_args.bsz, C.INPUT_LEN), + callbacks=[MyCallback()]) trainer.train() tester = FN.Tester(data=test_data, model=model, metrics=metric, @@ -195,12 +212,12 @@ def main(): 'init_embed': (None, 300), 'num_cls': None, 'hidden_size': g_args.hidden, - 'num_layers': 4, + 'num_layers': 2, 'num_head': g_args.nhead, 'head_dim': g_args.hdim, 'max_len': MAX_LEN, - 'cls_hidden_size': 600, - 'emb_dropout': 0.3, + 'cls_hidden_size': 200, + 'emb_dropout': g_args.drop, 'dropout': g_args.drop, } run_select[g_args.mode.lower()]() diff --git a/reproduction/coreference_resolution/__init__.py b/reproduction/coreference_resolution/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/reproduction/coreference_resolution/data_load/__init__.py b/reproduction/coreference_resolution/data_load/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/reproduction/coreference_resolution/data_load/cr_loader.py b/reproduction/coreference_resolution/data_load/cr_loader.py new file mode 100644 index 00000000..986afcd5 --- /dev/null +++ b/reproduction/coreference_resolution/data_load/cr_loader.py @@ -0,0 +1,68 @@ +from fastNLP.io.dataset_loader import JsonLoader,DataSet,Instance +from fastNLP.io.file_reader import _read_json +from fastNLP.core.vocabulary import Vocabulary +from fastNLP.io.base_loader import DataInfo +from reproduction.coreference_resolution.model.config import Config +import reproduction.coreference_resolution.model.preprocess as preprocess + + +class CRLoader(JsonLoader): + def __init__(self, fields=None, dropna=False): + super().__init__(fields, dropna) + + def _load(self, path): + """ + 加载数据 + :param path: + :return: + """ + dataset = DataSet() + for idx, d in _read_json(path, fields=self.fields_list, dropna=self.dropna): + if self.fields: + ins = {self.fields[k]: v for k, v in d.items()} + else: + ins = d + dataset.append(Instance(**ins)) + return dataset + + def process(self, paths, **kwargs): + data_info = DataInfo() + for name in ['train', 'test', 'dev']: + data_info.datasets[name] = self.load(paths[name]) + + config = Config() + vocab = Vocabulary().from_dataset(*data_info.datasets.values(), field_name='sentences') + vocab.build_vocab() + word2id = vocab.word2idx + + char_dict = preprocess.get_char_dict(config.char_path) + data_info.vocabs = vocab + + genres = {g: i for i, g in enumerate(["bc", "bn", "mz", "nw", "pt", "tc", "wb"])} + + for name, ds in data_info.datasets.items(): + ds.apply(lambda x: preprocess.doc2numpy(x['sentences'], word2id, char_dict, max(config.filter), + config.max_sentences, is_train=name=='train')[0], + new_field_name='doc_np') + ds.apply(lambda x: preprocess.doc2numpy(x['sentences'], word2id, char_dict, max(config.filter), + config.max_sentences, is_train=name=='train')[1], + new_field_name='char_index') + ds.apply(lambda x: preprocess.doc2numpy(x['sentences'], word2id, char_dict, max(config.filter), + config.max_sentences, is_train=name=='train')[2], + new_field_name='seq_len') + ds.apply(lambda x: preprocess.speaker2numpy(x["speakers"], config.max_sentences, is_train=name=='train'), + new_field_name='speaker_ids_np') + ds.apply(lambda x: genres[x["doc_key"][:2]], new_field_name='genre') + + ds.set_ignore_type('clusters') + ds.set_padder('clusters', None) + ds.set_input("sentences", "doc_np", "speaker_ids_np", "genre", "char_index", "seq_len") + ds.set_target("clusters") + + # train_dev, test = self.ds.split(348 / (2802 + 343 + 348), shuffle=False) + # train, dev = train_dev.split(343 / (2802 + 343), shuffle=False) + + return data_info + + + diff --git a/reproduction/coreference_resolution/model/__init__.py b/reproduction/coreference_resolution/model/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/reproduction/coreference_resolution/model/config.py b/reproduction/coreference_resolution/model/config.py new file mode 100644 index 00000000..6011257b --- /dev/null +++ b/reproduction/coreference_resolution/model/config.py @@ -0,0 +1,54 @@ +class Config(): + def __init__(self): + self.is_training = True + # path + self.glove = 'data/glove.840B.300d.txt.filtered' + self.turian = 'data/turian.50d.txt' + self.train_path = "data/train.english.jsonlines" + self.dev_path = "data/dev.english.jsonlines" + self.test_path = "data/test.english.jsonlines" + self.char_path = "data/char_vocab.english.txt" + + self.cuda = "0" + self.max_word = 1500 + self.epoch = 200 + + # config + # self.use_glove = True + # self.use_turian = True #No + self.use_elmo = False + self.use_CNN = True + self.model_heads = True #Yes + self.use_width = True # Yes + self.use_distance = True #Yes + self.use_metadata = True #Yes + + self.mention_ratio = 0.4 + self.max_sentences = 50 + self.span_width = 10 + self.feature_size = 20 #宽度信息emb的size + self.lr = 0.001 + self.lr_decay = 1e-3 + self.max_antecedents = 100 # 这个参数在mention detection中没有用 + self.atten_hidden_size = 150 + self.mention_hidden_size = 150 + self.sa_hidden_size = 150 + + self.char_emb_size = 8 + self.filter = [3,4,5] + + + # decay = 1e-5 + + def __str__(self): + d = self.__dict__ + out = 'config==============\n' + for i in list(d): + out += i+":" + out += str(d[i])+"\n" + out+="config==============\n" + return out + +if __name__=="__main__": + config = Config() + print(config) diff --git a/reproduction/coreference_resolution/model/metric.py b/reproduction/coreference_resolution/model/metric.py new file mode 100644 index 00000000..2c924660 --- /dev/null +++ b/reproduction/coreference_resolution/model/metric.py @@ -0,0 +1,163 @@ +from fastNLP.core.metrics import MetricBase + +import numpy as np + +from collections import Counter +from sklearn.utils.linear_assignment_ import linear_assignment + +""" +Mostly borrowed from https://github.com/clarkkev/deep-coref/blob/master/evaluation.py +""" + + + +class CRMetric(MetricBase): + def __init__(self): + super().__init__() + self.evaluators = [Evaluator(m) for m in (muc, b_cubed, ceafe)] + + # TODO 改名为evaluate，输入也 + def evaluate(self, predicted, mention_to_predicted,clusters): + for e in self.evaluators: + e.update(predicted,mention_to_predicted, clusters) + + def get_f1(self): + return sum(e.get_f1() for e in self.evaluators) / len(self.evaluators) + + def get_recall(self): + return sum(e.get_recall() for e in self.evaluators) / len(self.evaluators) + + def get_precision(self): + return sum(e.get_precision() for e in self.evaluators) / len(self.evaluators) + + # TODO 原本的getprf + def get_metric(self,reset=False): + res = {"pre":self.get_precision(), "rec":self.get_recall(), "f":self.get_f1()} + self.evaluators = [Evaluator(m) for m in (muc, b_cubed, ceafe)] + return res + + + + + + +class Evaluator(): + def __init__(self, metric, beta=1): + self.p_num = 0 + self.p_den = 0 + self.r_num = 0 + self.r_den = 0 + self.metric = metric + self.beta = beta + + def update(self, predicted,mention_to_predicted,gold): + gold = gold[0].tolist() + gold = [tuple(tuple(m) for m in gc) for gc in gold] + mention_to_gold = {} + for gc in gold: + for mention in gc: + mention_to_gold[mention] = gc + + if self.metric == ceafe: + pn, pd, rn, rd = self.metric(predicted, gold) + else: + pn, pd = self.metric(predicted, mention_to_gold) + rn, rd = self.metric(gold, mention_to_predicted) + self.p_num += pn + self.p_den += pd + self.r_num += rn + self.r_den += rd + + def get_f1(self): + return f1(self.p_num, self.p_den, self.r_num, self.r_den, beta=self.beta) + + def get_recall(self): + return 0 if self.r_num == 0 else self.r_num / float(self.r_den) + + def get_precision(self): + return 0 if self.p_num == 0 else self.p_num / float(self.p_den) + + def get_prf(self): + return self.get_precision(), self.get_recall(), self.get_f1() + + def get_counts(self): + return self.p_num, self.p_den, self.r_num, self.r_den + + + +def b_cubed(clusters, mention_to_gold): + num, dem = 0, 0 + + for c in clusters: + if len(c) == 1: + continue + + gold_counts = Counter() + correct = 0 + for m in c: + if m in mention_to_gold: + gold_counts[tuple(mention_to_gold[m])] += 1 + for c2, count in gold_counts.items(): + if len(c2) != 1: + correct += count * count + + num += correct / float(len(c)) + dem += len(c) + + return num, dem + + +def muc(clusters, mention_to_gold): + tp, p = 0, 0 + for c in clusters: + p += len(c) - 1 + tp += len(c) + linked = set() + for m in c: + if m in mention_to_gold: + linked.add(mention_to_gold[m]) + else: + tp -= 1 + tp -= len(linked) + return tp, p + + +def phi4(c1, c2): + return 2 * len([m for m in c1 if m in c2]) / float(len(c1) + len(c2)) + + +def ceafe(clusters, gold_clusters): + clusters = [c for c in clusters if len(c) != 1] + scores = np.zeros((len(gold_clusters), len(clusters))) + for i in range(len(gold_clusters)): + for j in range(len(clusters)): + scores[i, j] = phi4(gold_clusters[i], clusters[j]) + matching = linear_assignment(-scores) + similarity = sum(scores[matching[:, 0], matching[:, 1]]) + return similarity, len(clusters), similarity, len(gold_clusters) + + +def lea(clusters, mention_to_gold): + num, dem = 0, 0 + + for c in clusters: + if len(c) == 1: + continue + + common_links = 0 + all_links = len(c) * (len(c) - 1) / 2.0 + for i, m in enumerate(c): + if m in mention_to_gold: + for m2 in c[i + 1:]: + if m2 in mention_to_gold and mention_to_gold[m] == mention_to_gold[m2]: + common_links += 1 + + num += len(c) * common_links / float(all_links) + dem += len(c) + + return num, dem + +def f1(p_num, p_den, r_num, r_den, beta=1): + p = 0 if p_den == 0 else p_num / float(p_den) + r = 0 if r_den == 0 else r_num / float(r_den) + return 0 if p + r == 0 else (1 + beta * beta) * p * r / (beta * beta * p + r) diff --git a/reproduction/coreference_resolution/model/model_re.py b/reproduction/coreference_resolution/model/model_re.py new file mode 100644 index 00000000..9dd90ec4 --- /dev/null +++ b/reproduction/coreference_resolution/model/model_re.py @@ -0,0 +1,576 @@ +import torch +import numpy as np +import torch.nn as nn +import torch.nn.functional as F + +from allennlp.commands.elmo import ElmoEmbedder +from fastNLP.models.base_model import BaseModel +from fastNLP.modules.encoder.variational_rnn import VarLSTM +from reproduction.coreference_resolution.model import preprocess +from fastNLP.io.embed_loader import EmbedLoader +import random + +# 设置seed +torch.manual_seed(0) # cpu +torch.cuda.manual_seed(0) # gpu +np.random.seed(0) # numpy +random.seed(0) + + +class ffnn(nn.Module): + def __init__(self, input_size, hidden_size, output_size): + super(ffnn, self).__init__() + + self.f = nn.Sequential( + # 多少层数 + nn.Linear(input_size, hidden_size), + nn.ReLU(inplace=True), + nn.Dropout(p=0.2), + nn.Linear(hidden_size, hidden_size), + nn.ReLU(inplace=True), + nn.Dropout(p=0.2), + nn.Linear(hidden_size, output_size) + ) + self.reset_param() + + def reset_param(self): + for name, param in self.named_parameters(): + if param.dim() > 1: + nn.init.xavier_normal_(param) + # param.data = torch.tensor(np.random.randn(*param.shape)).float() + else: + nn.init.zeros_(param) + + def forward(self, input): + return self.f(input).squeeze() + + +class Model(BaseModel): + def __init__(self, vocab, config): + word2id = vocab.word2idx + super(Model, self).__init__() + vocab_num = len(word2id) + self.word2id = word2id + self.config = config + self.char_dict = preprocess.get_char_dict('data/char_vocab.english.txt') + self.genres = {g: i for i, g in enumerate(["bc", "bn", "mz", "nw", "pt", "tc", "wb"])} + self.device = torch.device("cuda:" + config.cuda) + + self.emb = nn.Embedding(vocab_num, 350) + + emb1 = EmbedLoader().load_with_vocab(config.glove, vocab,normalize=False) + emb2 = EmbedLoader().load_with_vocab(config.turian, vocab ,normalize=False) + pre_emb = np.concatenate((emb1, emb2), axis=1) + pre_emb /= (np.linalg.norm(pre_emb, axis=1, keepdims=True) + 1e-12) + + if pre_emb is not None: + self.emb.weight = nn.Parameter(torch.from_numpy(pre_emb).float()) + for param in self.emb.parameters(): + param.requires_grad = False + self.emb_dropout = nn.Dropout(inplace=True) + + + if config.use_elmo: + self.elmo = ElmoEmbedder(options_file='data/elmo/elmo_2x4096_512_2048cnn_2xhighway_options.json', + weight_file='data/elmo/elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5', + cuda_device=int(config.cuda)) + print("elmo load over.") + self.elmo_args = torch.randn((3), requires_grad=True).to(self.device) + + self.char_emb = nn.Embedding(len(self.char_dict), config.char_emb_size) + self.conv1 = nn.Conv1d(config.char_emb_size, 50, 3) + self.conv2 = nn.Conv1d(config.char_emb_size, 50, 4) + self.conv3 = nn.Conv1d(config.char_emb_size, 50, 5) + + self.feature_emb = nn.Embedding(config.span_width, config.feature_size) + self.feature_emb_dropout = nn.Dropout(p=0.2, inplace=True) + + self.mention_distance_emb = nn.Embedding(10, config.feature_size) + self.distance_drop = nn.Dropout(p=0.2, inplace=True) + + self.genre_emb = nn.Embedding(7, config.feature_size) + self.speaker_emb = nn.Embedding(2, config.feature_size) + + self.bilstm = VarLSTM(input_size=350+150*config.use_CNN+config.use_elmo*1024,hidden_size=200,bidirectional=True,batch_first=True,hidden_dropout=0.2) + # self.bilstm = nn.LSTM(input_size=500, hidden_size=200, bidirectional=True, batch_first=True) + self.h0 = nn.init.orthogonal_(torch.empty(2, 1, 200)).to(self.device) + self.c0 = nn.init.orthogonal_(torch.empty(2, 1, 200)).to(self.device) + self.bilstm_drop = nn.Dropout(p=0.2, inplace=True) + + self.atten = ffnn(input_size=400, hidden_size=config.atten_hidden_size, output_size=1) + self.mention_score = ffnn(input_size=1320, hidden_size=config.mention_hidden_size, output_size=1) + self.sa = ffnn(input_size=3980+40*config.use_metadata, hidden_size=config.sa_hidden_size, output_size=1) + self.mention_start_np = None + self.mention_end_np = None + + def _reorder_lstm(self, word_emb, seq_lens): + sort_ind = sorted(range(len(seq_lens)), key=lambda i: seq_lens[i], reverse=True) + seq_lens_re = [seq_lens[i] for i in sort_ind] + emb_seq = self.reorder_sequence(word_emb, sort_ind, batch_first=True) + packed_seq = nn.utils.rnn.pack_padded_sequence(emb_seq, seq_lens_re, batch_first=True) + + h0 = self.h0.repeat(1, len(seq_lens), 1) + c0 = self.c0.repeat(1, len(seq_lens), 1) + packed_out, final_states = self.bilstm(packed_seq, (h0, c0)) + + lstm_out, _ = nn.utils.rnn.pad_packed_sequence(packed_out, batch_first=True) + back_map = {ind: i for i, ind in enumerate(sort_ind)} + reorder_ind = [back_map[i] for i in range(len(seq_lens_re))] + lstm_out = self.reorder_sequence(lstm_out, reorder_ind, batch_first=True) + return lstm_out + + def reorder_sequence(self, sequence_emb, order, batch_first=True): + """ + sequence_emb: [T, B, D] if not batch_first + order: list of sequence length + """ + batch_dim = 0 if batch_first else 1 + assert len(order) == sequence_emb.size()[batch_dim] + + order = torch.LongTensor(order) + order = order.to(sequence_emb).long() + + sorted_ = sequence_emb.index_select(index=order, dim=batch_dim) + + del order + return sorted_ + + def flat_lstm(self, lstm_out, seq_lens): + batch = lstm_out.shape[0] + seq = lstm_out.shape[1] + dim = lstm_out.shape[2] + l = [j + i * seq for i, seq_len in enumerate(seq_lens) for j in range(seq_len)] + flatted = torch.index_select(lstm_out.view(batch * seq, dim), 0, torch.LongTensor(l).to(self.device)) + return flatted + + def potential_mention_index(self, word_index, max_sent_len): + # get mention index [3,2]:the first sentence is 3 and secend 2 + # [0,0,0,1,1] --> [[0, 0], [0, 1], [1, 1], [1, 2], [2, 2], [3, 3], [3, 4], [4, 4]] (max =2) + potential_mention = [] + for i in range(len(word_index)): + for j in range(i, i + max_sent_len): + if (j < len(word_index) and word_index[i] == word_index[j]): + potential_mention.append([i, j]) + return potential_mention + + def get_mention_start_end(self, seq_lens): + # 序列长度转换成mention + # [3,2] --> [0,0,0,1,1] + word_index = [0] * sum(seq_lens) + sent_index = 0 + index = 0 + for length in seq_lens: + for l in range(length): + word_index[index] = sent_index + index += 1 + sent_index += 1 + + # [0,0,0,1,1]-->[[0,0],[0,1],[0,2]....] + mention_id = self.potential_mention_index(word_index, self.config.span_width) + mention_start = np.array(mention_id, dtype=int)[:, 0] + mention_end = np.array(mention_id, dtype=int)[:, 1] + return mention_start, mention_end + + def get_mention_emb(self, flatten_lstm, mention_start, mention_end): + mention_start_tensor = torch.from_numpy(mention_start).to(self.device) + mention_end_tensor = torch.from_numpy(mention_end).to(self.device) + emb_start = flatten_lstm.index_select(dim=0, index=mention_start_tensor) # [mention_num,embed] + emb_end = flatten_lstm.index_select(dim=0, index=mention_end_tensor) # [mention_num,embed] + return emb_start, emb_end + + def get_mask(self, mention_start, mention_end): + # big mask for attention + mention_num = mention_start.shape[0] + mask = np.zeros((mention_num, self.config.span_width)) # [mention_num,span_width] + for i in range(mention_num): + start = mention_start[i] + end = mention_end[i] + # 实际上是宽度 + for j in range(end - start + 1): + mask[i][j] = 1 + mask = torch.from_numpy(mask) # [mention_num,max_mention] + # 0-->-inf 1-->0 + log_mask = torch.log(mask) + return log_mask + + def get_mention_index(self, mention_start, max_mention): + # TODO 后面可能要改 + assert len(mention_start.shape) == 1 + mention_start_tensor = torch.from_numpy(mention_start) + num_mention = mention_start_tensor.shape[0] + mention_index = mention_start_tensor.expand(max_mention, num_mention).transpose(0, + 1) # [num_mention,max_mention] + assert mention_index.shape[0] == num_mention + assert mention_index.shape[1] == max_mention + range_add = torch.arange(0, max_mention).expand(num_mention, max_mention).long() # [num_mention,max_mention] + mention_index = mention_index + range_add + mention_index = torch.min(mention_index, torch.LongTensor([mention_start[-1]]).expand(num_mention, max_mention)) + return mention_index.to(self.device) + + def sort_mention(self, mention_start, mention_end, candidate_mention_emb, candidate_mention_score, seq_lens): + # 排序记录,高分段在前面 + mention_score, mention_ids = torch.sort(candidate_mention_score, descending=True) + preserve_mention_num = int(self.config.mention_ratio * sum(seq_lens)) + mention_ids = mention_ids[0:preserve_mention_num] + mention_score = mention_score[0:preserve_mention_num] + + mention_start_tensor = torch.from_numpy(mention_start).to(self.device).index_select(dim=0, + index=mention_ids) # [lamda*word_num] + mention_end_tensor = torch.from_numpy(mention_end).to(self.device).index_select(dim=0, + index=mention_ids) # [lamda*word_num] + mention_emb = candidate_mention_emb.index_select(index=mention_ids, dim=0) # [lamda*word_num,emb] + assert mention_score.shape[0] == preserve_mention_num + assert mention_start_tensor.shape[0] == preserve_mention_num + assert mention_end_tensor.shape[0] == preserve_mention_num + assert mention_emb.shape[0] == preserve_mention_num + # TODO 不交叉没做处理 + + # 对start进行再排序，实际位置在前面 + # TODO 这里只考虑了start没有考虑end + mention_start_tensor, temp_index = torch.sort(mention_start_tensor) + mention_end_tensor = mention_end_tensor.index_select(dim=0, index=temp_index) + mention_emb = mention_emb.index_select(dim=0, index=temp_index) + mention_score = mention_score.index_select(dim=0, index=temp_index) + return mention_start_tensor, mention_end_tensor, mention_score, mention_emb + + def get_antecedents(self, mention_starts, max_antecedents): + num_mention = mention_starts.shape[0] + max_antecedents = min(max_antecedents, num_mention) + # mention和它是第几个mention之间的对应关系 + antecedents = np.zeros((num_mention, max_antecedents), dtype=int) # [num_mention,max_an] + # 记录长度 + antecedents_len = [0] * num_mention + for i in range(num_mention): + ante_count = 0 + for j in range(max(0, i - max_antecedents), i): + antecedents[i, ante_count] = j + ante_count += 1 + # 补位操作 + for j in range(ante_count, max_antecedents): + antecedents[i, j] = 0 + antecedents_len[i] = ante_count + assert antecedents.shape[1] == max_antecedents + return antecedents, antecedents_len + + def get_antecedents_score(self, span_represent, mention_score, antecedents, antecedents_len, mention_speakers_ids, + genre): + num_mention = mention_score.shape[0] + max_antecedent = antecedents.shape[1] + + pair_emb = self.get_pair_emb(span_represent, antecedents, mention_speakers_ids, genre) # [span_num,max_ant,emb] + antecedent_scores = self.sa(pair_emb) + mask01 = self.sequence_mask(antecedents_len, max_antecedent) + maskinf = torch.log(mask01).to(self.device) + assert maskinf.shape[1] <= max_antecedent + assert antecedent_scores.shape[0] == num_mention + antecedent_scores = antecedent_scores + maskinf + antecedents = torch.from_numpy(antecedents).to(self.device) + mention_scoreij = mention_score.unsqueeze(1) + torch.gather( + mention_score.unsqueeze(0).expand(num_mention, num_mention), dim=1, index=antecedents) + antecedent_scores += mention_scoreij + + antecedent_scores = torch.cat([torch.zeros([mention_score.shape[0], 1]).to(self.device), antecedent_scores], + 1) # [num_mentions, max_ant + 1] + return antecedent_scores + + ############################## + def distance_bin(self, mention_distance): + bins = torch.zeros(mention_distance.size()).byte().to(self.device) + rg = [[1, 1], [2, 2], [3, 3], [4, 4], [5, 7], [8, 15], [16, 31], [32, 63], [64, 300]] + for t, k in enumerate(rg): + i, j = k[0], k[1] + b = torch.LongTensor([i]).unsqueeze(-1).expand(mention_distance.size()).to(self.device) + m1 = torch.ge(mention_distance, b) + e = torch.LongTensor([j]).unsqueeze(-1).expand(mention_distance.size()).to(self.device) + m2 = torch.le(mention_distance, e) + bins = bins + (t + 1) * (m1 & m2) + return bins.long() + + def get_distance_emb(self, antecedents_tensor): + num_mention = antecedents_tensor.shape[0] + max_ant = antecedents_tensor.shape[1] + + assert max_ant <= self.config.max_antecedents + source = torch.arange(0, num_mention).expand(max_ant, num_mention).transpose(0,1).to(self.device) # [num_mention,max_ant] + mention_distance = source - antecedents_tensor + mention_distance_bin = self.distance_bin(mention_distance) + distance_emb = self.mention_distance_emb(mention_distance_bin) + distance_emb = self.distance_drop(distance_emb) + return distance_emb + + def get_pair_emb(self, span_emb, antecedents, mention_speakers_ids, genre): + emb_dim = span_emb.shape[1] + num_span = span_emb.shape[0] + max_ant = antecedents.shape[1] + assert span_emb.shape[0] == antecedents.shape[0] + antecedents = torch.from_numpy(antecedents).to(self.device) + + # [num_span,max_ant,emb] + antecedent_emb = torch.gather(span_emb.unsqueeze(0).expand(num_span, num_span, emb_dim), dim=1, + index=antecedents.unsqueeze(2).expand(num_span, max_ant, emb_dim)) + # [num_span,max_ant,emb] + target_emb_tiled = span_emb.expand((max_ant, num_span, emb_dim)) + target_emb_tiled = target_emb_tiled.transpose(0, 1) + + similarity_emb = antecedent_emb * target_emb_tiled + + pair_emb_list = [target_emb_tiled, antecedent_emb, similarity_emb] + + # get speakers and genre + if self.config.use_metadata: + antecedent_speaker_ids = mention_speakers_ids.unsqueeze(0).expand(num_span, num_span).gather(dim=1, + index=antecedents) + same_speaker = torch.eq(mention_speakers_ids.unsqueeze(1).expand(num_span, max_ant), + antecedent_speaker_ids) # [num_mention,max_ant] + speaker_embedding = self.speaker_emb(same_speaker.long().to(self.device)) # [mention_num.max_ant,emb] + genre_embedding = self.genre_emb( + torch.LongTensor([genre]).expand(num_span, max_ant).to(self.device)) # [mention_num,max_ant,emb] + pair_emb_list.append(speaker_embedding) + pair_emb_list.append(genre_embedding) + + # get distance emb + if self.config.use_distance: + distance_emb = self.get_distance_emb(antecedents) + pair_emb_list.append(distance_emb) + + pair_emb = torch.cat(pair_emb_list, 2) + return pair_emb + + def sequence_mask(self, len_list, max_len): + x = np.zeros((len(len_list), max_len)) + for i in range(len(len_list)): + l = len_list[i] + for j in range(l): + x[i][j] = 1 + return torch.from_numpy(x).float() + + def logsumexp(self, value, dim=None, keepdim=False): + """Numerically stable implementation of the operation + + value.exp().sum(dim, keepdim).log() + """ + # TODO: torch.max(value, dim=None) threw an error at time of writing + if dim is not None: + m, _ = torch.max(value, dim=dim, keepdim=True) + value0 = value - m + if keepdim is False: + m = m.squeeze(dim) + return m + torch.log(torch.sum(torch.exp(value0), + dim=dim, keepdim=keepdim)) + else: + m = torch.max(value) + sum_exp = torch.sum(torch.exp(value - m)) + + return m + torch.log(sum_exp) + + def softmax_loss(self, antecedent_scores, antecedent_labels): + antecedent_labels = torch.from_numpy(antecedent_labels * 1).to(self.device) + gold_scores = antecedent_scores + torch.log(antecedent_labels.float()) # [num_mentions, max_ant + 1] + marginalized_gold_scores = self.logsumexp(gold_scores, 1) # [num_mentions] + log_norm = self.logsumexp(antecedent_scores, 1) # [num_mentions] + return torch.sum(log_norm - marginalized_gold_scores) # [num_mentions]reduce_logsumexp + + def get_predicted_antecedents(self, antecedents, antecedent_scores): + predicted_antecedents = [] + for i, index in enumerate(np.argmax(antecedent_scores.detach(), axis=1) - 1): + if index < 0: + predicted_antecedents.append(-1) + else: + predicted_antecedents.append(antecedents[i, index]) + return predicted_antecedents + + def get_predicted_clusters(self, mention_starts, mention_ends, predicted_antecedents): + mention_to_predicted = {} + predicted_clusters = [] + for i, predicted_index in enumerate(predicted_antecedents): + if predicted_index < 0: + continue + assert i > predicted_index + predicted_antecedent = (int(mention_starts[predicted_index]), int(mention_ends[predicted_index])) + if predicted_antecedent in mention_to_predicted: + predicted_cluster = mention_to_predicted[predicted_antecedent] + else: + predicted_cluster = len(predicted_clusters) + predicted_clusters.append([predicted_antecedent]) + mention_to_predicted[predicted_antecedent] = predicted_cluster + + mention = (int(mention_starts[i]), int(mention_ends[i])) + predicted_clusters[predicted_cluster].append(mention) + mention_to_predicted[mention] = predicted_cluster + + predicted_clusters = [tuple(pc) for pc in predicted_clusters] + mention_to_predicted = {m: predicted_clusters[i] for m, i in mention_to_predicted.items()} + + return predicted_clusters, mention_to_predicted + + def evaluate_coref(self, mention_starts, mention_ends, predicted_antecedents, gold_clusters, evaluator): + gold_clusters = [tuple(tuple(m) for m in gc) for gc in gold_clusters] + mention_to_gold = {} + for gc in gold_clusters: + for mention in gc: + mention_to_gold[mention] = gc + predicted_clusters, mention_to_predicted = self.get_predicted_clusters(mention_starts, mention_ends, + predicted_antecedents) + evaluator.update(predicted_clusters, gold_clusters, mention_to_predicted, mention_to_gold) + return predicted_clusters + + + def forward(self, sentences, doc_np, speaker_ids_np, genre, char_index, seq_len): + """ + 实际输入都是tensor + :param sentences: 句子，被fastNLP转化成了numpy， + :param doc_np: 被fastNLP转化成了Tensor + :param speaker_ids_np: 被fastNLP转化成了Tensor + :param genre: 被fastNLP转化成了Tensor + :param char_index: 被fastNLP转化成了Tensor + :param seq_len: 被fastNLP转化成了Tensor + :return: + """ + # change for fastNLP + sentences = sentences[0].tolist() + doc_tensor = doc_np[0] + speakers_tensor = speaker_ids_np[0] + genre = genre[0].item() + char_index = char_index[0] + seq_len = seq_len[0].cpu().numpy() + + # 类型 + + # doc_tensor = torch.from_numpy(doc_np).to(self.device) + # speakers_tensor = torch.from_numpy(speaker_ids_np).to(self.device) + mention_emb_list = [] + + word_emb = self.emb(doc_tensor) + word_emb_list = [word_emb] + if self.config.use_CNN: + # [batch, length, char_length, char_dim] + char = self.char_emb(char_index) + char_size = char.size() + # first transform to [batch *length, char_length, char_dim] + # then transpose to [batch * length, char_dim, char_length] + char = char.view(char_size[0] * char_size[1], char_size[2], char_size[3]).transpose(1, 2) + + # put into cnn [batch*length, char_filters, char_length] + # then put into maxpooling [batch * length, char_filters] + char_over_cnn, _ = self.conv1(char).max(dim=2) + # reshape to [batch, length, char_filters] + char_over_cnn = torch.tanh(char_over_cnn).view(char_size[0], char_size[1], -1) + word_emb_list.append(char_over_cnn) + + char_over_cnn, _ = self.conv2(char).max(dim=2) + char_over_cnn = torch.tanh(char_over_cnn).view(char_size[0], char_size[1], -1) + word_emb_list.append(char_over_cnn) + + char_over_cnn, _ = self.conv3(char).max(dim=2) + char_over_cnn = torch.tanh(char_over_cnn).view(char_size[0], char_size[1], -1) + word_emb_list.append(char_over_cnn) + + # word_emb = torch.cat(word_emb_list, dim=2) + + # use elmo or not + if self.config.use_elmo: + # 如果确实被截断了 + if doc_tensor.shape[0] == 50 and len(sentences) > 50: + sentences = sentences[0:50] + elmo_embedding, elmo_mask = self.elmo.batch_to_embeddings(sentences) + elmo_embedding = elmo_embedding.to( + self.device) # [sentence_num,max_sent_len,3,1024]--[sentence_num,max_sent,1024] + elmo_embedding = elmo_embedding[:, 0, :, :] * self.elmo_args[0] + elmo_embedding[:, 1, :, :] * \ + self.elmo_args[1] + elmo_embedding[:, 2, :, :] * self.elmo_args[2] + word_emb_list.append(elmo_embedding) + # print(word_emb_list[0].shape) + # print(word_emb_list[1].shape) + # print(word_emb_list[2].shape) + # print(word_emb_list[3].shape) + # print(word_emb_list[4].shape) + + word_emb = torch.cat(word_emb_list, dim=2) + + word_emb = self.emb_dropout(word_emb) + # word_emb_elmo = self.emb_dropout(word_emb_elmo) + lstm_out = self._reorder_lstm(word_emb, seq_len) + flatten_lstm = self.flat_lstm(lstm_out, seq_len) # [word_num,emb] + flatten_lstm = self.bilstm_drop(flatten_lstm) + # TODO 没有按照论文写 + flatten_word_emb = self.flat_lstm(word_emb, seq_len) # [word_num,emb] + + mention_start, mention_end = self.get_mention_start_end(seq_len) # [mention_num] + self.mention_start_np = mention_start # [mention_num] np + self.mention_end_np = mention_end + mention_num = mention_start.shape[0] + emb_start, emb_end = self.get_mention_emb(flatten_lstm, mention_start, mention_end) # [mention_num,emb] + + # list + mention_emb_list.append(emb_start) + mention_emb_list.append(emb_end) + + if self.config.use_width: + mention_width_index = mention_end - mention_start + mention_width_tensor = torch.from_numpy(mention_width_index).to(self.device) # [mention_num] + mention_width_emb = self.feature_emb(mention_width_tensor) + mention_width_emb = self.feature_emb_dropout(mention_width_emb) + mention_emb_list.append(mention_width_emb) + + if self.config.model_heads: + mention_index = self.get_mention_index(mention_start, self.config.span_width) # [mention_num,max_mention] + log_mask_tensor = self.get_mask(mention_start, mention_end).float().to( + self.device) # [mention_num,max_mention] + alpha = self.atten(flatten_lstm).to(self.device) # [word_num] + + # 得到attention + mention_head_score = torch.gather(alpha.expand(mention_num, -1), 1, + mention_index).float().to(self.device) # [mention_num,max_mention] + mention_attention = F.softmax(mention_head_score + log_mask_tensor, dim=1) # [mention_num,max_mention] + + # TODO flatte lstm + word_num = flatten_lstm.shape[0] + lstm_emb = flatten_lstm.shape[1] + emb_num = flatten_word_emb.shape[1] + + # [num_mentions, max_mention_width, emb] + mention_text_emb = torch.gather( + flatten_word_emb.unsqueeze(1).expand(word_num, self.config.span_width, emb_num), + 0, mention_index.unsqueeze(2).expand(mention_num, self.config.span_width, + emb_num)) + # [mention_num,emb] + mention_head_emb = torch.sum( + mention_attention.unsqueeze(2).expand(mention_num, self.config.span_width, emb_num) * mention_text_emb, + dim=1) + mention_emb_list.append(mention_head_emb) + + candidate_mention_emb = torch.cat(mention_emb_list, 1) # [candidate_mention_num,emb] + candidate_mention_score = self.mention_score(candidate_mention_emb) # [candidate_mention_num] + + antecedent_scores, antecedents, mention_start_tensor, mention_end_tensor = (None, None, None, None) + mention_start_tensor, mention_end_tensor, mention_score, mention_emb = \ + self.sort_mention(mention_start, mention_end, candidate_mention_emb, candidate_mention_score, seq_len) + mention_speakers_ids = speakers_tensor.index_select(dim=0, index=mention_start_tensor) # num_mention + + antecedents, antecedents_len = self.get_antecedents(mention_start_tensor, self.config.max_antecedents) + antecedent_scores = self.get_antecedents_score(mention_emb, mention_score, antecedents, antecedents_len, + mention_speakers_ids, genre) + + ans = {"candidate_mention_score": candidate_mention_score, "antecedent_scores": antecedent_scores, + "antecedents": antecedents, "mention_start_tensor": mention_start_tensor, + "mention_end_tensor": mention_end_tensor} + + return ans + + def predict(self, sentences, doc_np, speaker_ids_np, genre, char_index, seq_len): + ans = self(sentences, + doc_np, + speaker_ids_np, + genre, + char_index, + seq_len) + + predicted_antecedents = self.get_predicted_antecedents(ans["antecedents"], ans["antecedent_scores"]) + predicted_clusters, mention_to_predicted = self.get_predicted_clusters(ans["mention_start_tensor"], + ans["mention_end_tensor"], + predicted_antecedents) + + return {'predicted':predicted_clusters,"mention_to_predicted":mention_to_predicted} + + +if __name__ == '__main__': + pass diff --git a/reproduction/coreference_resolution/model/preprocess.py b/reproduction/coreference_resolution/model/preprocess.py new file mode 100644 index 00000000..d97fcb4d --- /dev/null +++ b/reproduction/coreference_resolution/model/preprocess.py @@ -0,0 +1,225 @@ +import json +import numpy as np +from . import util +import collections + +def load(path): + """ + load the file from jsonline + :param path: + :return: examples with many example(dict): {"clusters":[[[mention],[mention]],[another cluster]], + "doc_key":"str","speakers":[[,,,],[]...],"sentence":[[][]]} + """ + with open(path) as f: + train_examples = [json.loads(jsonline) for jsonline in f.readlines()] + return train_examples + +def get_vocab(): + """ + 从所有的句子中得到最终的字典,被main调用,不止是train，还有dev和test + :param examples: + :return: word2id & id2word + """ + word2id = {'PAD':0,'UNK':1} + id2word = {0:'PAD',1:'UNK'} + index = 2 + data = [load("../data/train.english.jsonlines"),load("../data/dev.english.jsonlines"),load("../data/test.english.jsonlines")] + for examples in data: + for example in examples: + for sent in example["sentences"]: + for word in sent: + if(word not in word2id): + word2id[word]=index + id2word[index] = word + index += 1 + return word2id,id2word + +def normalize(v): + norm = np.linalg.norm(v) + if norm > 0: + return v / norm + else: + return v + +# 加载glove得到embedding +def get_emb(id2word,embedding_size): + glove_oov = 0 + turian_oov = 0 + both = 0 + glove_emb_path = "../data/glove.840B.300d.txt.filtered" + turian_emb_path = "../data/turian.50d.txt" + word_num = len(id2word) + emb = np.zeros((word_num,embedding_size)) + glove_emb_dict = util.load_embedding_dict(glove_emb_path,300,"txt") + turian_emb_dict = util.load_embedding_dict(turian_emb_path,50,"txt") + for i in range(word_num): + if id2word[i] in glove_emb_dict: + word_embedding = glove_emb_dict.get(id2word[i]) + emb[i][0:300] = np.array(word_embedding) + else: + # print(id2word[i]) + glove_oov += 1 + if id2word[i] in turian_emb_dict: + word_embedding = turian_emb_dict.get(id2word[i]) + emb[i][300:350] = np.array(word_embedding) + else: + # print(id2word[i]) + turian_oov += 1 + if id2word[i] not in glove_emb_dict and id2word[i] not in turian_emb_dict: + both += 1 + emb[i] = normalize(emb[i]) + print("embedding num:"+str(word_num)) + print("glove num:"+str(glove_oov)) + print("glove oov rate:"+str(glove_oov/word_num)) + print("turian num:"+str(turian_oov)) + print("turian oov rate:"+str(turian_oov/word_num)) + print("both num:"+str(both)) + return emb + + +def _doc2vec(doc,word2id,char_dict,max_filter,max_sentences,is_train): + max_len = 0 + max_word_length = 0 + docvex = [] + length = [] + if is_train: + sent_num = min(max_sentences,len(doc)) + else: + sent_num = len(doc) + + for i in range(sent_num): + sent = doc[i] + length.append(len(sent)) + if (len(sent) > max_len): + max_len = len(sent) + sent_vec =[] + for j,word in enumerate(sent): + if len(word)>max_word_length: + max_word_length = len(word) + if word in word2id: + sent_vec.append(word2id[word]) + else: + sent_vec.append(word2id["UNK"]) + docvex.append(sent_vec) + + char_index = np.zeros((sent_num, max_len, max_word_length),dtype=int) + for i in range(sent_num): + sent = doc[i] + for j,word in enumerate(sent): + char_index[i, j, :len(word)] = [char_dict[c] for c in word] + + return docvex,char_index,length,max_len + +# TODO 修改了接口，确认所有该修改的地方都修改好 +def doc2numpy(doc,word2id,chardict,max_filter,max_sentences,is_train): + docvec, char_index, length, max_len = _doc2vec(doc,word2id,chardict,max_filter,max_sentences,is_train) + assert max(length) == max_len + assert char_index.shape[0]==len(length) + assert char_index.shape[1]==max_len + doc_np = np.zeros((len(docvec), max_len), int) + for i in range(len(docvec)): + for j in range(len(docvec[i])): + doc_np[i][j] = docvec[i][j] + return doc_np,char_index,length + +# TODO 没有测试 +def speaker2numpy(speakers_raw,max_sentences,is_train): + if is_train and len(speakers_raw)> max_sentences: + speakers_raw = speakers_raw[0:max_sentences] + speakers = flatten(speakers_raw) + speaker_dict = {s: i for i, s in enumerate(set(speakers))} + speaker_ids = np.array([speaker_dict[s] for s in speakers]) + return speaker_ids + + +def flat_cluster(clusters): + flatted = [] + for cluster in clusters: + for item in cluster: + flatted.append(item) + return flatted + +def get_right_mention(clusters,mention_start_np,mention_end_np): + flatted = flat_cluster(clusters) + cluster_num = len(flatted) + mention_num = mention_start_np.shape[0] + right_mention = np.zeros(mention_num,dtype=int) + for i in range(mention_num): + if [mention_start_np[i],mention_end_np[i]] in flatted: + right_mention[i]=1 + return right_mention,cluster_num + +def handle_cluster(clusters): + gold_mentions = sorted(tuple(m) for m in flatten(clusters)) + gold_mention_map = {m: i for i, m in enumerate(gold_mentions)} + cluster_ids = np.zeros(len(gold_mentions), dtype=int) + for cluster_id, cluster in enumerate(clusters): + for mention in cluster: + cluster_ids[gold_mention_map[tuple(mention)]] = cluster_id + gold_starts, gold_ends = tensorize_mentions(gold_mentions) + return cluster_ids, gold_starts, gold_ends + +# 展平 +def flatten(l): + return [item for sublist in l for item in sublist] + +# 把mention分成start end +def tensorize_mentions(mentions): + if len(mentions) > 0: + starts, ends = zip(*mentions) + else: + starts, ends = [], [] + return np.array(starts), np.array(ends) + +def get_char_dict(path): + vocab = [""] + with open(path) as f: + vocab.extend(c.strip() for c in f.readlines()) + char_dict = collections.defaultdict(int) + char_dict.update({c: i for i, c in enumerate(vocab)}) + return char_dict + +def get_labels(clusters,mention_starts,mention_ends,max_antecedents): + cluster_ids, gold_starts, gold_ends = handle_cluster(clusters) + num_mention = mention_starts.shape[0] + num_gold = gold_starts.shape[0] + max_antecedents = min(max_antecedents, num_mention) + mention_indices = {} + + for i in range(num_mention): + mention_indices[(mention_starts[i].detach().item(), mention_ends[i].detach().item())] = i + # 用来记录哪些mention是对的，-1表示错误，正数代表这个mention实际上对应哪个gold cluster的id + mention_cluster_ids = [-1] * num_mention + # test + right_mention_count = 0 + for i in range(num_gold): + right_mention = mention_indices.get((gold_starts[i], gold_ends[i])) + if (right_mention != None): + right_mention_count += 1 + mention_cluster_ids[right_mention] = cluster_ids[i] + + # i j 是否属于同一个cluster + labels = np.zeros((num_mention, max_antecedents + 1), dtype=bool) # [num_mention,max_an+1] + for i in range(num_mention): + ante_count = 0 + null_label = True + for j in range(max(0, i - max_antecedents), i): + if (mention_cluster_ids[i] >= 0 and mention_cluster_ids[i] == mention_cluster_ids[j]): + labels[i, ante_count + 1] = True + null_label = False + else: + labels[i, ante_count + 1] = False + ante_count += 1 + for j in range(ante_count, max_antecedents): + labels[i, j + 1] = False + labels[i, 0] = null_label + return labels + +# test=========================== + + +if __name__=="__main__": + word2id,id2word = get_vocab() + get_emb(id2word,350) + + diff --git a/reproduction/coreference_resolution/model/softmax_loss.py b/reproduction/coreference_resolution/model/softmax_loss.py new file mode 100644 index 00000000..c75a31d6 --- /dev/null +++ b/reproduction/coreference_resolution/model/softmax_loss.py @@ -0,0 +1,32 @@ +from fastNLP.core.losses import LossBase + +from reproduction.coreference_resolution.model.preprocess import get_labels +from reproduction.coreference_resolution.model.config import Config +import torch + + +class SoftmaxLoss(LossBase): + """ + 交叉熵loss + 允许多标签分类 + """ + + def __init__(self, antecedent_scores=None, clusters=None, mention_start_tensor=None, mention_end_tensor=None): + """ + + :param pred: + :param target: + """ + super().__init__() + self._init_param_map(antecedent_scores=antecedent_scores, clusters=clusters, + mention_start_tensor=mention_start_tensor, mention_end_tensor=mention_end_tensor) + + def get_loss(self, antecedent_scores, clusters, mention_start_tensor, mention_end_tensor): + antecedent_labels = get_labels(clusters[0], mention_start_tensor, mention_end_tensor, + Config().max_antecedents) + + antecedent_labels = torch.from_numpy(antecedent_labels*1).to(torch.device("cuda:" + Config().cuda)) + gold_scores = antecedent_scores + torch.log(antecedent_labels.float()).to(torch.device("cuda:" + Config().cuda)) # [num_mentions, max_ant + 1] + marginalized_gold_scores = gold_scores.logsumexp(dim=1) # [num_mentions] + log_norm = antecedent_scores.logsumexp(dim=1) # [num_mentions] + return torch.sum(log_norm - marginalized_gold_scores) diff --git a/reproduction/coreference_resolution/model/util.py b/reproduction/coreference_resolution/model/util.py new file mode 100644 index 00000000..42cd09fe --- /dev/null +++ b/reproduction/coreference_resolution/model/util.py @@ -0,0 +1,101 @@ +import os +import errno +import collections +import torch +import numpy as np +import pyhocon + + + +# flatten the list +def flatten(l): + return [item for sublist in l for item in sublist] + + +def get_config(filename): + return pyhocon.ConfigFactory.parse_file(filename) + + +# safe make directions +def mkdirs(path): + try: + os.makedirs(path) + except OSError as exception: + if exception.errno != errno.EEXIST: + raise + return path + + +def load_char_dict(char_vocab_path): + vocab = [""] + with open(char_vocab_path) as f: + vocab.extend(c.strip() for c in f.readlines()) + char_dict = collections.defaultdict(int) + char_dict.update({c: i for i, c in enumerate(vocab)}) + return char_dict + +# 加载embedding +def load_embedding_dict(embedding_path, embedding_size, embedding_format): + print("Loading word embeddings from {}...".format(embedding_path)) + default_embedding = np.zeros(embedding_size) + embedding_dict = collections.defaultdict(lambda: default_embedding) + skip_first = embedding_format == "vec" + with open(embedding_path) as f: + for i, line in enumerate(f.readlines()): + if skip_first and i == 0: + continue + splits = line.split() + assert len(splits) == embedding_size + 1 + word = splits[0] + embedding = np.array([float(s) for s in splits[1:]]) + embedding_dict[word] = embedding + print("Done loading word embeddings.") + return embedding_dict + + +# safe devide +def maybe_divide(x, y): + return 0 if y == 0 else x / float(y) + + +def shape(x, dim): + return x.get_shape()[dim].value or torch.shape(x)[dim] + + +def normalize(v): + norm = np.linalg.norm(v) + if norm > 0: + return v / norm + else: + return v + + +class RetrievalEvaluator(object): + def __init__(self): + self._num_correct = 0 + self._num_gold = 0 + self._num_predicted = 0 + + def update(self, gold_set, predicted_set): + self._num_correct += len(gold_set & predicted_set) + self._num_gold += len(gold_set) + self._num_predicted += len(predicted_set) + + def recall(self): + return maybe_divide(self._num_correct, self._num_gold) + + def precision(self): + return maybe_divide(self._num_correct, self._num_predicted) + + def metrics(self): + recall = self.recall() + precision = self.precision() + f1 = maybe_divide(2 * recall * precision, precision + recall) + return recall, precision, f1 + + + +if __name__=="__main__": + print(load_char_dict("../data/char_vocab.english.txt")) + embedding_dict = load_embedding_dict("../data/glove.840B.300d.txt.filtered",300,"txt") + print("hello") diff --git a/reproduction/coreference_resolution/readme.md b/reproduction/coreference_resolution/readme.md new file mode 100644 index 00000000..67d8cdc7 --- /dev/null +++ b/reproduction/coreference_resolution/readme.md @@ -0,0 +1,49 @@ +# 共指消解复现 +## 介绍 +Coreference resolution是查找文本中指向同一现实实体的所有表达式的任务。 +对于涉及自然语言理解的许多更高级别的NLP任务来说， +这是一个重要的步骤，例如文档摘要，问题回答和信息提取。 +代码的实现主要基于[ End-to-End Coreference Resolution (Lee et al, 2017)](https://arxiv.org/pdf/1707.07045). + + +## 数据获取与预处理 +论文在[OntoNote5.0](https://allennlp.org/models)数据集上取得了当时的sota结果。 +由于版权问题，本文无法提供数据集的下载，请自行下载。 +原始数据集的格式为conll格式，详细介绍参考数据集给出的官方介绍页面。 + +代码实现采用了论文作者Lee的预处理方法，具体细节参加[链接](https://github.com/kentonl/e2e-coref/blob/e2e/setup_training.sh)。 +处理之后的数据集为json格式，例子： +``` +{ + "clusters": [], + "doc_key": "nw", + "sentences": [["This", "is", "the", "first", "sentence", "."], ["This", "is", "the", "second", "."]], + "speakers": [["spk1", "spk1", "spk1", "spk1", "spk1", "spk1"], ["spk2", "spk2", "spk2", "spk2", "spk2"]] +} +``` + +### embedding 数据集下载 +[turian emdedding](https://lil.cs.washington.edu/coref/turian.50d.txt) + +[glove embedding]( https://nlp.stanford.edu/data/glove.840B.300d.zip) + + + +## 运行 +```python +# 训练代码 +CUDA_VISIBLE_DEVICES=0 python train.py +# 测试代码 +CUDA_VISIBLE_DEVICES=0 python valid.py +``` + +## 结果 +原论文作者在测试集上取得了67.2%的结果，AllenNLP复现的结果为 [63.0%](https://allennlp.org/models)。 +其中allenNLP训练时没有加入speaker信息，没有variational dropout以及只使用了100的antecedents而不是250。 + +在与allenNLP使用同样的超参和配置时，本代码复现取得了63.6%的F1值。 + + +## 问题 +如果您有什么问题或者反馈，请提issue或者邮件联系我： +yexu_i@qq.com diff --git a/reproduction/coreference_resolution/test/__init__.py b/reproduction/coreference_resolution/test/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/reproduction/coreference_resolution/test/test_dataloader.py b/reproduction/coreference_resolution/test/test_dataloader.py new file mode 100644 index 00000000..0d9dae52 --- /dev/null +++ b/reproduction/coreference_resolution/test/test_dataloader.py @@ -0,0 +1,14 @@ +import unittest +from ..data_load.cr_loader import CRLoader + +class Test_CRLoader(unittest.TestCase): + def test_cr_loader(self): + train_path = 'data/train.english.jsonlines.mini' + dev_path = 'data/dev.english.jsonlines.minid' + test_path = 'data/test.english.jsonlines' + cr = CRLoader() + data_info = cr.process({'train':train_path,'dev':dev_path,'test':test_path}) + + print(data_info.datasets['train'][0]) + print(data_info.datasets['dev'][0]) + print(data_info.datasets['test'][0]) diff --git a/reproduction/coreference_resolution/train.py b/reproduction/coreference_resolution/train.py new file mode 100644 index 00000000..a231a575 --- /dev/null +++ b/reproduction/coreference_resolution/train.py @@ -0,0 +1,69 @@ +import sys +sys.path.append('../..') + +import torch +from torch.optim import Adam + +from fastNLP.core.callback import Callback, GradientClipCallback +from fastNLP.core.trainer import Trainer + +from reproduction.coreference_resolution.data_load.cr_loader import CRLoader +from reproduction.coreference_resolution.model.config import Config +from reproduction.coreference_resolution.model.model_re import Model +from reproduction.coreference_resolution.model.softmax_loss import SoftmaxLoss +from reproduction.coreference_resolution.model.metric import CRMetric +from fastNLP import SequentialSampler +from fastNLP import cache_results + + +# torch.backends.cudnn.benchmark = False +# torch.backends.cudnn.deterministic = True + +class LRCallback(Callback): + def __init__(self, parameters, decay_rate=1e-3): + super().__init__() + self.paras = parameters + self.decay_rate = decay_rate + + def on_step_end(self): + if self.step % 100 == 0: + for para in self.paras: + para['lr'] = para['lr'] * (1 - self.decay_rate) + + +if __name__ == "__main__": + config = Config() + + print(config) + + @cache_results('cache.pkl') + def cache(): + cr_train_dev_test = CRLoader() + + data_info = cr_train_dev_test.process({'train': config.train_path, 'dev': config.dev_path, + 'test': config.test_path}) + return data_info + data_info = cache() + print("数据集划分：\ntrain:", str(len(data_info.datasets["train"])), + "\ndev:" + str(len(data_info.datasets["dev"])) + "\ntest:" + str(len(data_info.datasets["test"]))) + # print(data_info) + model = Model(data_info.vocabs, config) + print(model) + + loss = SoftmaxLoss() + + metric = CRMetric() + + optim = Adam(model.parameters(), lr=config.lr) + + lr_decay_callback = LRCallback(optim.param_groups, config.lr_decay) + + trainer = Trainer(model=model, train_data=data_info.datasets["train"], dev_data=data_info.datasets["dev"], + loss=loss, metrics=metric, check_code_level=-1,sampler=None, + batch_size=1, device=torch.device("cuda:" + config.cuda), metric_key='f', n_epochs=config.epoch, + optimizer=optim, + save_path='/remote-home/xxliu/pycharm/fastNLP/fastNLP/reproduction/coreference_resolution/save', + callbacks=[lr_decay_callback, GradientClipCallback(clip_value=5)]) + print() + + trainer.train() diff --git a/reproduction/coreference_resolution/valid.py b/reproduction/coreference_resolution/valid.py new file mode 100644 index 00000000..826332c6 --- /dev/null +++ b/reproduction/coreference_resolution/valid.py @@ -0,0 +1,24 @@ +import torch +from reproduction.coreference_resolution.model.config import Config +from reproduction.coreference_resolution.model.metric import CRMetric +from reproduction.coreference_resolution.data_load.cr_loader import CRLoader +from fastNLP import Tester +import argparse + + +if __name__=='__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--path') + args = parser.parse_args() + + cr_loader = CRLoader() + config = Config() + data_info = cr_loader.process({'train': config.train_path, 'dev': config.dev_path, + 'test': config.test_path}) + metirc = CRMetric() + model = torch.load(args.path) + tester = Tester(data_info.datasets['test'],model,metirc,batch_size=1,device="cuda:0") + tester.test() + print('test over') + + diff --git a/reproduction/matching/matching_cntn.py b/reproduction/matching/matching_cntn.py new file mode 100644 index 00000000..d813164d --- /dev/null +++ b/reproduction/matching/matching_cntn.py @@ -0,0 +1,105 @@ +import argparse +import torch +import os + +from fastNLP.core import Trainer, Tester, Adam, AccuracyMetric, Const +from fastNLP.modules.encoder.embedding import StaticEmbedding + +from reproduction.matching.data.MatchingDataLoader import QNLILoader, RTELoader, SNLILoader, MNLILoader +from reproduction.matching.model.cntn import CNTNModel + +# define hyper-parameters +argument = argparse.ArgumentParser() +argument.add_argument('--embedding', choices=['glove', 'word2vec'], default='glove') +argument.add_argument('--batch-size-per-gpu', type=int, default=256) +argument.add_argument('--n-epochs', type=int, default=200) +argument.add_argument('--lr', type=float, default=1e-5) +argument.add_argument('--seq-len-type', choices=['mask', 'seq_len'], default='mask') +argument.add_argument('--save-dir', type=str, default=None) +argument.add_argument('--cntn-depth', type=int, default=1) +argument.add_argument('--cntn-ns', type=int, default=200) +argument.add_argument('--cntn-k-top', type=int, default=10) +argument.add_argument('--cntn-r', type=int, default=5) +argument.add_argument('--dataset', choices=['qnli', 'rte', 'snli', 'mnli'], default='qnli') +argument.add_argument('--max-len', type=int, default=50) +arg = argument.parse_args() + +# dataset dict +dev_dict = { + 'qnli': 'dev', + 'rte': 'dev', + 'snli': 'dev', + 'mnli': 'dev_matched', +} + +test_dict = { + 'qnli': 'dev', + 'rte': 'dev', + 'snli': 'test', + 'mnli': 'dev_matched', +} + +# set num_labels +if arg.dataset == 'qnli' or arg.dataset == 'rte': + num_labels = 2 +else: + num_labels = 3 + +# load data set +if arg.dataset == 'qnli': + data_info = QNLILoader().process( + paths='path/to/qnli/data', to_lower=True, seq_len_type=arg.seq_len_type, bert_tokenizer=None, + get_index=True, concat=False, auto_pad_length=arg.max_len) +elif arg.dataset == 'rte': + data_info = RTELoader().process( + paths='path/to/rte/data', to_lower=True, seq_len_type=arg.seq_len_type, bert_tokenizer=None, + get_index=True, concat=False, auto_pad_length=arg.max_len) +elif arg.dataset == 'snli': + data_info = SNLILoader().process( + paths='path/to/snli/data', to_lower=True, seq_len_type=arg.seq_len_type, bert_tokenizer=None, + get_index=True, concat=False, auto_pad_length=arg.max_len) +elif arg.dataset == 'mnli': + data_info = MNLILoader().process( + paths='path/to/mnli/data', to_lower=True, seq_len_type=arg.seq_len_type, bert_tokenizer=None, + get_index=True, concat=False, auto_pad_length=arg.max_len) +else: + raise ValueError(f'now we only support [qnli,rte,snli,mnli] dataset for cntn model!') + +# load embedding +if arg.embedding == 'word2vec': + embedding = StaticEmbedding(data_info.vocabs[Const.INPUT], model_dir_or_name='en-word2vec-300', requires_grad=True) +elif arg.embedding == 'glove': + embedding = StaticEmbedding(data_info.vocabs[Const.INPUT], model_dir_or_name='en-glove-840b-300', + requires_grad=True) +else: + raise ValueError(f'now we only support word2vec or glove embedding for cntn model!') + +# define model +model = CNTNModel(embedding, ns=arg.cntn_ns, k_top=arg.cntn_k_top, num_labels=num_labels, depth=arg.cntn_depth, + r=arg.cntn_r) +print(model) + +# define trainer +trainer = Trainer(train_data=data_info.datasets['train'], model=model, + optimizer=Adam(lr=arg.lr, model_params=model.parameters()), + batch_size=torch.cuda.device_count() * arg.batch_size_per_gpu, + n_epochs=arg.n_epochs, print_every=-1, + dev_data=data_info.datasets[dev_dict[arg.dataset]], + metrics=AccuracyMetric(), metric_key='acc', + device=[i for i in range(torch.cuda.device_count())], + check_code_level=-1) + +# train model +trainer.train(load_best_model=True) + +# define tester +tester = Tester( + data=data_info.datasets[test_dict[arg.dataset]], + model=model, + metrics=AccuracyMetric(), + batch_size=torch.cuda.device_count() * arg.batch_size_per_gpu, + device=[i for i in range(torch.cuda.device_count())] +) + +# test model +tester.test() diff --git a/reproduction/matching/model/cntn.py b/reproduction/matching/model/cntn.py new file mode 100644 index 00000000..0b4803fa --- /dev/null +++ b/reproduction/matching/model/cntn.py @@ -0,0 +1,120 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +import numpy as np + +from torch.nn import CrossEntropyLoss + +from fastNLP.models import BaseModel +from fastNLP.modules.encoder.embedding import TokenEmbedding +from fastNLP.core.const import Const + + +class DynamicKMaxPooling(nn.Module): + """ + :param k_top: Fixed number of pooling output features for the topmost convolutional layer. + :param l: Number of convolutional layers. + """ + + def __init__(self, k_top, l): + super(DynamicKMaxPooling, self).__init__() + self.k_top = k_top + self.L = l + + def forward(self, x, l): + """ + :param x: Input sequence. + :param l: Current convolutional layers. + """ + s = x.size()[3] + k_ll = ((self.L - l) / self.L) * s + k_l = int(round(max(self.k_top, np.ceil(k_ll)))) + out = F.adaptive_max_pool2d(x, (x.size()[2], k_l)) + return out + + +class CNTNModel(BaseModel): + """ + 使用CNN进行问答匹配的模型 + 'Qiu, Xipeng, and Xuanjing Huang. + Convolutional neural tensor network architecture for community-based question answering. + Twenty-Fourth International Joint Conference on Artificial Intelligence. 2015.' + + :param init_embedding: Embedding. + :param ns: Sentence embedding size. + :param k_top: Fixed number of pooling output features for the topmost convolutional layer. + :param num_labels: Number of labels. + :param depth: Number of convolutional layers. + :param r: Number of weight tensor slices. + :param drop_rate: Dropout rate. + """ + + def __init__(self, init_embedding: TokenEmbedding, ns=200, k_top=10, num_labels=2, depth=2, r=5, + dropout_rate=0.3): + super(CNTNModel, self).__init__() + self.embedding = init_embedding + self.depth = depth + self.kmaxpooling = DynamicKMaxPooling(k_top, depth) + self.conv_q = nn.ModuleList() + self.conv_a = nn.ModuleList() + width = self.embedding.embed_size + for i in range(depth): + self.conv_q.append(nn.Sequential( + nn.Dropout(p=dropout_rate), + nn.Conv2d( + in_channels=1, + out_channels=width // 2, + kernel_size=(width, 3), + padding=(0, 2)) + )) + self.conv_a.append(nn.Sequential( + nn.Dropout(p=dropout_rate), + nn.Conv2d( + in_channels=1, + out_channels=width // 2, + kernel_size=(width, 3), + padding=(0, 2)) + )) + width = width // 2 + + self.fc_q = nn.Sequential(nn.Dropout(p=dropout_rate), nn.Linear(width * k_top, ns)) + self.fc_a = nn.Sequential(nn.Dropout(p=dropout_rate), nn.Linear(width * k_top, ns)) + self.weight_M = nn.Bilinear(ns, ns, r) + self.weight_V = nn.Linear(2 * ns, r) + self.weight_u = nn.Sequential(nn.Dropout(p=dropout_rate), nn.Linear(r, num_labels)) + + def forward(self, words1, words2, seq_len1, seq_len2, target=None): + """ + :param words1: [batch, seq_len, emb_size] Question. + :param words2: [batch, seq_len, emb_size] Answer. + :param seq_len1: [batch] + :param seq_len2: [batch] + :param target: [batch] Glod labels. + :return: + """ + in_q = self.embedding(words1) + in_a = self.embedding(words2) + in_q = in_q.permute(0, 2, 1).unsqueeze(1) + in_a = in_a.permute(0, 2, 1).unsqueeze(1) + + for i in range(self.depth): + in_q = F.relu(self.conv_q[i](in_q)) + in_q = in_q.squeeze().unsqueeze(1) + in_q = self.kmaxpooling(in_q, i + 1) + in_a = F.relu(self.conv_a[i](in_a)) + in_a = in_a.squeeze().unsqueeze(1) + in_a = self.kmaxpooling(in_a, i + 1) + + in_q = self.fc_q(in_q.view(in_q.size(0), -1)) + in_a = self.fc_q(in_a.view(in_a.size(0), -1)) + score = torch.tanh(self.weight_u(self.weight_M(in_q, in_a) + self.weight_V(torch.cat((in_q, in_a), -1)))) + + if target is not None: + loss_fct = CrossEntropyLoss() + loss = loss_fct(score, target) + return {Const.LOSS: loss, Const.OUTPUT: score} + else: + return {Const.OUTPUT: score} + + def predict(self, **kwargs): + return self.forward(**kwargs) diff --git a/reproduction/seqence_labelling/ner/data/Conll2003Loader.py b/reproduction/seqence_labelling/ner/data/Conll2003Loader.py deleted file mode 100644 index 577987c6..00000000 --- a/reproduction/seqence_labelling/ner/data/Conll2003Loader.py +++ /dev/null @@ -1,93 +0,0 @@ - -from fastNLP.core.vocabulary import VocabularyOption -from fastNLP.io.base_loader import DataSetLoader, DataInfo -from typing import Union, Dict -from fastNLP import Vocabulary -from fastNLP import Const -from reproduction.utils import check_dataloader_paths - -from fastNLP.io.dataset_loader import ConllLoader -from reproduction.seqence_labelling.ner.data.utils import iob2bioes, iob2 - - -class Conll2003DataLoader(DataSetLoader): - def __init__(self, task:str='ner', encoding_type:str='bioes'): - """ - 加载Conll2003格式的英语语料，该数据集的信息可以在https://www.clips.uantwerpen.be/conll2003/ner/找到。当task为pos - 时，返回的DataSet中target取值于第2列; 当task为chunk时，返回的DataSet中target取值于第3列;当task为ner时，返回 - 的DataSet中target取值于第4列。所有"-DOCSTART- -X- O O"将被忽略，这会导致数据的数量少于很多文献报道的值，但 - 鉴于"-DOCSTART- -X- O O"只是用于文档分割的符号，并不应该作为预测对象，所以我们忽略了数据中的-DOCTSTART-开头的行 - ner与chunk任务读取后的数据的target将为encoding_type类型。pos任务读取后就是pos列的数据。 - - :param task: 指定需要标注任务。可选ner, pos, chunk - """ - assert task in ('ner', 'pos', 'chunk') - index = {'ner':3, 'pos':1, 'chunk':2}[task] - self._loader = ConllLoader(headers=['raw_words', 'target'], indexes=[0, index]) - self._tag_converters = None - if task in ('ner', 'chunk'): - self._tag_converters = [iob2] - if encoding_type == 'bioes': - self._tag_converters.append(iob2bioes) - - def load(self, path: str): - dataset = self._loader.load(path) - def convert_tag_schema(tags): - for converter in self._tag_converters: - tags = converter(tags) - return tags - if self._tag_converters: - dataset.apply_field(convert_tag_schema, field_name=Const.TARGET, new_field_name=Const.TARGET) - return dataset - - def process(self, paths: Union[str, Dict[str, str]], word_vocab_opt:VocabularyOption=None, lower:bool=True): - """ - 读取并处理数据。数据中的'-DOCSTART-'开头的行会被忽略 - - :param paths: - :param word_vocab_opt: vocabulary的初始化值 - :param lower: 是否将所有字母转为小写 - :return: - """ - # 读取数据 - paths = check_dataloader_paths(paths) - data = DataInfo() - input_fields = [Const.TARGET, Const.INPUT, Const.INPUT_LEN] - target_fields = [Const.TARGET, Const.INPUT_LEN] - for name, path in paths.items(): - dataset = self.load(path) - dataset.apply_field(lambda words: words, field_name='raw_words', new_field_name=Const.INPUT) - if lower: - dataset.words.lower() - data.datasets[name] = dataset - - # 对construct vocab - word_vocab = Vocabulary(min_freq=2) if word_vocab_opt is None else Vocabulary(**word_vocab_opt) - word_vocab.from_dataset(data.datasets['train'], field_name=Const.INPUT, - no_create_entry_dataset=[dataset for name, dataset in data.datasets.items() if name!='train']) - word_vocab.index_dataset(*data.datasets.values(), field_name=Const.INPUT, new_field_name=Const.INPUT) - data.vocabs[Const.INPUT] = word_vocab - - # cap words - cap_word_vocab = Vocabulary() - cap_word_vocab.from_dataset(data.datasets['train'], field_name='raw_words', - no_create_entry_dataset=[dataset for name, dataset in data.datasets.items() if name!='train']) - cap_word_vocab.index_dataset(*data.datasets.values(), field_name='raw_words', new_field_name='cap_words') - input_fields.append('cap_words') - data.vocabs['cap_words'] = cap_word_vocab - - # 对target建vocab - target_vocab = Vocabulary(unknown=None, padding=None) - target_vocab.from_dataset(*data.datasets.values(), field_name=Const.TARGET) - target_vocab.index_dataset(*data.datasets.values(), field_name=Const.TARGET) - data.vocabs[Const.TARGET] = target_vocab - - for name, dataset in data.datasets.items(): - dataset.add_seq_len(Const.INPUT, new_field_name=Const.INPUT_LEN) - dataset.set_input(*input_fields) - dataset.set_target(*target_fields) - - return data - -if __name__ == '__main__': - pass \ No newline at end of file diff --git a/reproduction/seqence_labelling/ner/data/OntoNoteLoader.py b/reproduction/seqence_labelling/ner/data/OntoNoteLoader.py deleted file mode 100644 index 8a2c567d..00000000 --- a/reproduction/seqence_labelling/ner/data/OntoNoteLoader.py +++ /dev/null @@ -1,152 +0,0 @@ -from fastNLP.core.vocabulary import VocabularyOption -from fastNLP.io.base_loader import DataSetLoader, DataInfo -from typing import Union, Dict -from fastNLP import DataSet -from fastNLP import Vocabulary -from fastNLP import Const -from reproduction.utils import check_dataloader_paths - -from fastNLP.io.dataset_loader import ConllLoader -from reproduction.seqence_labelling.ner.data.utils import iob2bioes, iob2 - -class OntoNoteNERDataLoader(DataSetLoader): - """ - 用于读取处理为Conll格式后的OntoNote数据。将OntoNote数据处理为conll格式的过程可以参考https://github.com/yhcc/OntoNotes-5.0-NER。 - - """ - def __init__(self, encoding_type:str='bioes'): - assert encoding_type in ('bioes', 'bio') - self.encoding_type = encoding_type - if encoding_type=='bioes': - self.encoding_method = iob2bioes - else: - self.encoding_method = iob2 - - def load(self, path:str)->DataSet: - """ - 给定一个文件路径，读取数据。返回的DataSet包含以下的field - raw_words: List[str] - target: List[str] - - :param path: - :return: - """ - dataset = ConllLoader(headers=['raw_words', 'target'], indexes=[3, 10]).load(path) - def convert_to_bio(tags): - bio_tags = [] - flag = None - for tag in tags: - label = tag.strip("()*") - if '(' in tag: - bio_label = 'B-' + label - flag = label - elif flag: - bio_label = 'I-' + flag - else: - bio_label = 'O' - if ')' in tag: - flag = None - bio_tags.append(bio_label) - return self.encoding_method(bio_tags) - - def convert_word(words): - converted_words = [] - for word in words: - word = word.replace('/.', '.') # 有些结尾的.是/.形式的 - if not word.startswith('-'): - converted_words.append(word) - continue - # 以下是由于这些符号被转义了，再转回来 - tfrs = {'-LRB-':'(', - '-RRB-': ')', - '-LSB-': '[', - '-RSB-': ']', - '-LCB-': '{', - '-RCB-': '}' - } - if word in tfrs: - converted_words.append(tfrs[word]) - else: - converted_words.append(word) - return converted_words - - dataset.apply_field(convert_word, field_name='raw_words', new_field_name='raw_words') - dataset.apply_field(convert_to_bio, field_name='target', new_field_name='target') - - return dataset - - def process(self, paths: Union[str, Dict[str, str]], word_vocab_opt:VocabularyOption=None, - lower:bool=True)->DataInfo: - """ - 读取并处理数据。返回的DataInfo包含以下的内容 - vocabs: - word: Vocabulary - target: Vocabulary - datasets: - train: DataSet - words: List[int], 被设置为input - target: int. label，被同时设置为input和target - seq_len: int. 句子的长度，被同时设置为input和target - raw_words: List[str] - xxx(根据传入的paths可能有所变化) - - :param paths: - :param word_vocab_opt: vocabulary的初始化值 - :param lower: 是否使用小写 - :return: - """ - paths = check_dataloader_paths(paths) - data = DataInfo() - input_fields = [Const.TARGET, Const.INPUT, Const.INPUT_LEN] - target_fields = [Const.TARGET, Const.INPUT_LEN] - for name, path in paths.items(): - dataset = self.load(path) - dataset.apply_field(lambda words: words, field_name='raw_words', new_field_name=Const.INPUT) - if lower: - dataset.words.lower() - data.datasets[name] = dataset - - # 对construct vocab - word_vocab = Vocabulary(min_freq=2) if word_vocab_opt is None else Vocabulary(**word_vocab_opt) - word_vocab.from_dataset(data.datasets['train'], field_name=Const.INPUT, - no_create_entry_dataset=[dataset for name, dataset in data.datasets.items() if name!='train']) - word_vocab.index_dataset(*data.datasets.values(), field_name=Const.INPUT, new_field_name=Const.INPUT) - data.vocabs[Const.INPUT] = word_vocab - - # cap words - cap_word_vocab = Vocabulary() - cap_word_vocab.from_dataset(*data.datasets.values(), field_name='raw_words') - cap_word_vocab.index_dataset(*data.datasets.values(), field_name='raw_words', new_field_name='cap_words') - input_fields.append('cap_words') - data.vocabs['cap_words'] = cap_word_vocab - - # 对target建vocab - target_vocab = Vocabulary(unknown=None, padding=None) - target_vocab.from_dataset(*data.datasets.values(), field_name=Const.TARGET) - target_vocab.index_dataset(*data.datasets.values(), field_name=Const.TARGET) - data.vocabs[Const.TARGET] = target_vocab - - for name, dataset in data.datasets.items(): - dataset.add_seq_len(Const.INPUT, new_field_name=Const.INPUT_LEN) - dataset.set_input(*input_fields) - dataset.set_target(*target_fields) - - return data - - -if __name__ == '__main__': - loader = OntoNoteNERDataLoader() - dataset = loader.load('/hdd/fudanNLP/fastNLP/others/data/v4/english/test.txt') - print(dataset.target.value_count()) - print(dataset[:4]) - - -""" -train 115812 2200752 -development 15680 304684 -test 12217 230111 - -train 92403 1901772 -valid 13606 279180 -test 10258 204135 -""" \ No newline at end of file diff --git a/reproduction/seqence_labelling/ner/data/utils.py b/reproduction/seqence_labelling/ner/data/utils.py deleted file mode 100644 index 8f7af792..00000000 --- a/reproduction/seqence_labelling/ner/data/utils.py +++ /dev/null @@ -1,49 +0,0 @@ -from typing import List - -def iob2(tags:List[str])->List[str]: - """ - 检查数据是否是合法的IOB数据，如果是IOB1会被自动转换为IOB2。 - - :param tags: 需要转换的tags - """ - for i, tag in enumerate(tags): - if tag == "O": - continue - split = tag.split("-") - if len(split) != 2 or split[0] not in ["I", "B"]: - raise TypeError("The encoding schema is not a valid IOB type.") - if split[0] == "B": - continue - elif i == 0 or tags[i - 1] == "O": # conversion IOB1 to IOB2 - tags[i] = "B" + tag[1:] - elif tags[i - 1][1:] == tag[1:]: - continue - else: # conversion IOB1 to IOB2 - tags[i] = "B" + tag[1:] - return tags - -def iob2bioes(tags:List[str])->List[str]: - """ - 将iob的tag转换为bmeso编码 - :param tags: - :return: - """ - new_tags = [] - for i, tag in enumerate(tags): - if tag == 'O': - new_tags.append(tag) - else: - split = tag.split('-')[0] - if split == 'B': - if i+1!=len(tags) and tags[i+1].split('-')[0] == 'I': - new_tags.append(tag) - else: - new_tags.append(tag.replace('B-', 'S-')) - elif split == 'I': - if i + 1