def evaluate(self,sess,vocab,token_indices, character_indices_padded, token_lengths, pattern ,label_indices,datatype='train'): all_predictions = [] all_y_true = [] # output_filepath = os.path.join(stats_graph_folder, '{1:03d}_{0}.txt'.format(dataset_type,epoch_number)) # output_file = codecs.open(output_filepath, 'w', 'UTF-8') # original_conll_file = codecs.open(dataset_filepaths[dataset_type], 'r', 'UTF-8') for i in range(len(token_indices)): feed_dict = { self.input_token_indices: token_indices[i], self.input_token_character_indices: character_indices_padded[i], self.input_token_lengths: token_lengths[i], self.input_token_patterns: pattern[i], self.dropout_keep_prob: 1., } unary_scores, transition_params_trained = sess.run([self.unary_scores, self.transition_parameters], feed_dict) predictions, _ = tf.contrib.crf.viterbi_decode(unary_scores, transition_params_trained) predictions = predictions[1:-1] assert (len(predictions) == len(token_indices[i])) all_predictions.extend(predictions) all_y_true.extend(label_indices[i]) label_predict = [vocab.labels[i] for i in all_predictions] label_true = [vocab.labels[i] for i in all_y_true] label_predict = utils_nlp.bioes_to_bio(label_predict) label_true = utils_nlp.bioes_to_bio(label_true) new_pre = [] new_true = [] for i in range(len(label_predict)): if label_true[i]!='O' or label_predict[i]!='O': new_pre.append(utils_nlp.remove_bio_from_label_name(label_predict[i])) new_true.append(label_true[i] if label_true[i]=='O' else label_true[i][2:]) labels = [label if label=='O' else label[2:] for label in vocab.labels] labels = list(set(labels)) report = classification_report(new_true,new_pre) print('matrix') matrix = confusion_matrix(new_true,new_pre,labels) file =codecs.open(datatype+'_evaluate.txt','w','utf-8') file.writelines(' '.join(labels)+'\n\r') for i,row in enumerate(matrix): file.writelines(' '.join([str(i) for i in row])+'\n\r') file.close() print(matrix) print(report) return report
def load_dataset(self, dataset_filepaths, parameters, token_to_vector=None): ''' dataset_filepaths : dictionary with keys 'train', 'valid', 'test', 'deploy' ''' start_time = time.time() print('Load dataset... ', end='', flush=True) if parameters['token_pretrained_embedding_filepath'] != '': if token_to_vector == None: token_to_vector = utils_nlp.load_pretrained_token_embeddings( parameters) else: token_to_vector = {} if self.verbose: print("len(token_to_vector): {0}".format(len(token_to_vector))) # Load pretraining dataset to ensure that index to label is compatible to the pretrained model, # and that token embeddings that are learned in the pretrained model are loaded properly. all_tokens_in_pretraining_dataset = [] all_characters_in_pretraining_dataset = [] if parameters['use_pretrained_model']: pretraining_dataset = pickle.load( open( os.path.join(parameters['pretrained_model_folder'], 'dataset.pickle'), 'rb')) all_tokens_in_pretraining_dataset = pretraining_dataset.index_to_token.values( ) all_characters_in_pretraining_dataset = pretraining_dataset.index_to_character.values( ) remap_to_unk_count_threshold = 1 self.UNK_TOKEN_INDEX = 0 self.PADDING_CHARACTER_INDEX = 0 self.tokens_mapped_to_unk = [] self.UNK = 'UNK' self.unique_labels = [] labels = {} tokens = {} label_count = {} token_count = {} character_count = {} for dataset_type in ['train', 'valid', 'test', 'deploy']: labels[dataset_type], tokens[dataset_type], token_count[dataset_type], label_count[dataset_type], character_count[dataset_type] \ = self._parse_dataset(dataset_filepaths.get(dataset_type, None)) if self.verbose: print("dataset_type: {0}".format(dataset_type)) if self.verbose: print("len(token_count[dataset_type]): {0}".format( len(token_count[dataset_type]))) token_count['all'] = {} for token in list(token_count['train'].keys()) + list( token_count['valid'].keys()) + list( token_count['test'].keys()) + list( token_count['deploy'].keys()): token_count['all'][token] = token_count['train'][ token] + token_count['valid'][token] + token_count['test'][ token] + token_count['deploy'][token] if parameters['load_all_pretrained_token_embeddings']: for token in token_to_vector: if token not in token_count['all']: token_count['all'][token] = -1 token_count['train'][token] = -1 for token in all_tokens_in_pretraining_dataset: if token not in token_count['all']: token_count['all'][token] = -1 token_count['train'][token] = -1 character_count['all'] = {} for character in list(character_count['train'].keys()) + list( character_count['valid'].keys()) + list( character_count['test'].keys()) + list( character_count['deploy'].keys()): character_count['all'][character] = character_count['train'][ character] + character_count['valid'][ character] + character_count['test'][ character] + character_count['deploy'][character] for character in all_characters_in_pretraining_dataset: if character not in character_count['all']: character_count['all'][character] = -1 character_count['train'][character] = -1 for dataset_type in dataset_filepaths.keys(): if self.verbose: print("dataset_type: {0}".format(dataset_type)) if self.verbose: print("len(token_count[dataset_type]): {0}".format( len(token_count[dataset_type]))) label_count['all'] = {} for character in list(label_count['train'].keys()) + list( label_count['valid'].keys()) + list( label_count['test'].keys()) + list( label_count['deploy'].keys()): label_count['all'][character] = label_count['train'][ character] + label_count['valid'][character] + label_count[ 'test'][character] + label_count['deploy'][character] token_count['all'] = utils.order_dictionary(token_count['all'], 'value_key', reverse=True) label_count['all'] = utils.order_dictionary(label_count['all'], 'key', reverse=False) character_count['all'] = utils.order_dictionary(character_count['all'], 'value', reverse=True) if self.verbose: print('character_count[\'all\']: {0}'.format( character_count['all'])) token_to_index = {} token_to_index[self.UNK] = self.UNK_TOKEN_INDEX iteration_number = 0 number_of_unknown_tokens = 0 if self.verbose: print("parameters['remap_unknown_tokens_to_unk']: {0}".format( parameters['remap_unknown_tokens_to_unk'])) if self.verbose: print("len(token_count['train'].keys()): {0}".format( len(token_count['train'].keys()))) for token, count in token_count['all'].items(): if iteration_number == self.UNK_TOKEN_INDEX: iteration_number += 1 if parameters['remap_unknown_tokens_to_unk'] == 1 and \ (token_count['train'][token] == 0 or \ parameters['load_only_pretrained_token_embeddings']) and \ not utils_nlp.is_token_in_pretrained_embeddings(token, token_to_vector, parameters) and \ token not in all_tokens_in_pretraining_dataset: if self.verbose: print("token: {0}".format(token)) if self.verbose: print("token.lower(): {0}".format(token.lower())) if self.verbose: print("re.sub('\d', '0', token.lower()): {0}".format( re.sub('\d', '0', token.lower()))) token_to_index[token] = self.UNK_TOKEN_INDEX number_of_unknown_tokens += 1 self.tokens_mapped_to_unk.append(token) else: token_to_index[token] = iteration_number iteration_number += 1 if self.verbose: print("number_of_unknown_tokens: {0}".format( number_of_unknown_tokens)) infrequent_token_indices = [] for token, count in token_count['train'].items(): if 0 < count <= remap_to_unk_count_threshold: infrequent_token_indices.append(token_to_index[token]) if self.verbose: print("len(token_count['train']): {0}".format( len(token_count['train']))) if self.verbose: print("len(infrequent_token_indices): {0}".format( len(infrequent_token_indices))) # Ensure that both B- and I- versions exist for each label labels_without_bio = set() for label in label_count['all'].keys(): new_label = utils_nlp.remove_bio_from_label_name(label) labels_without_bio.add(new_label) for label in labels_without_bio: if label == 'O': continue if parameters['tagging_format'] == 'bioes': prefixes = ['B-', 'I-', 'E-', 'S-'] else: prefixes = ['B-', 'I-'] for prefix in prefixes: l = prefix + label if l not in label_count['all']: label_count['all'][l] = 0 label_count['all'] = utils.order_dictionary(label_count['all'], 'key', reverse=False) if parameters['use_pretrained_model']: self.unique_labels = sorted( list(pretraining_dataset.label_to_index.keys())) # Make sure labels are compatible with the pretraining dataset. for label in label_count['all']: if label not in pretraining_dataset.label_to_index: raise AssertionError( "The label {0} does not exist in the pretraining dataset. " .format(label) + "Please ensure that only the following labels exist in the dataset: {0}" .format(', '.join(self.unique_labels))) label_to_index = pretraining_dataset.label_to_index.copy() else: label_to_index = {} iteration_number = 0 for label, count in label_count['all'].items(): label_to_index[label] = iteration_number iteration_number += 1 self.unique_labels.append(label) if self.verbose: print('self.unique_labels: {0}'.format(self.unique_labels)) character_to_index = {} iteration_number = 0 for character, count in character_count['all'].items(): if iteration_number == self.PADDING_CHARACTER_INDEX: iteration_number += 1 character_to_index[character] = iteration_number iteration_number += 1 if self.verbose: print('token_count[\'train\'][0:10]: {0}'.format( list(token_count['train'].items())[0:10])) token_to_index = utils.order_dictionary(token_to_index, 'value', reverse=False) if self.verbose: print('token_to_index: {0}'.format(token_to_index)) index_to_token = utils.reverse_dictionary(token_to_index) if parameters['remap_unknown_tokens_to_unk'] == 1: index_to_token[self.UNK_TOKEN_INDEX] = self.UNK if self.verbose: print('index_to_token: {0}'.format(index_to_token)) if self.verbose: print('label_count[\'train\']: {0}'.format(label_count['train'])) label_to_index = utils.order_dictionary(label_to_index, 'value', reverse=False) if self.verbose: print('label_to_index: {0}'.format(label_to_index)) index_to_label = utils.reverse_dictionary(label_to_index) if self.verbose: print('index_to_label: {0}'.format(index_to_label)) character_to_index = utils.order_dictionary(character_to_index, 'value', reverse=False) index_to_character = utils.reverse_dictionary(character_to_index) if self.verbose: print('character_to_index: {0}'.format(character_to_index)) if self.verbose: print('index_to_character: {0}'.format(index_to_character)) if self.verbose: print('labels[\'train\'][0:10]: {0}'.format(labels['train'][0:10])) if self.verbose: print('tokens[\'train\'][0:10]: {0}'.format(tokens['train'][0:10])) if self.verbose: # Print sequences of length 1 in train set for token_sequence, label_sequence in zip(tokens['train'], labels['train']): if len(label_sequence) == 1 and label_sequence[0] != 'O': print("{0}\t{1}".format(token_sequence[0], label_sequence[0])) self.token_to_index = token_to_index self.index_to_token = index_to_token self.index_to_character = index_to_character self.character_to_index = character_to_index self.index_to_label = index_to_label self.label_to_index = label_to_index if self.verbose: print("len(self.token_to_index): {0}".format( len(self.token_to_index))) if self.verbose: print("len(self.index_to_token): {0}".format( len(self.index_to_token))) self.tokens = tokens self.labels = labels token_indices, label_indices, character_indices_padded, character_indices, token_lengths, characters, label_vector_indices = self._convert_to_indices( dataset_filepaths.keys()) self.token_indices = token_indices self.label_indices = label_indices self.character_indices_padded = character_indices_padded self.character_indices = character_indices self.token_lengths = token_lengths self.characters = characters self.label_vector_indices = label_vector_indices self.number_of_classes = max(self.index_to_label.keys()) + 1 self.vocabulary_size = max(self.index_to_token.keys()) + 1 self.alphabet_size = max(self.index_to_character.keys()) + 1 if self.verbose: print("self.number_of_classes: {0}".format(self.number_of_classes)) if self.verbose: print("self.alphabet_size: {0}".format(self.alphabet_size)) if self.verbose: print("self.vocabulary_size: {0}".format(self.vocabulary_size)) # unique_labels_of_interest is used to compute F1-scores. self.unique_labels_of_interest = list(self.unique_labels) self.unique_labels_of_interest.remove('O') self.unique_label_indices_of_interest = [] for lab in self.unique_labels_of_interest: self.unique_label_indices_of_interest.append(label_to_index[lab]) self.infrequent_token_indices = infrequent_token_indices if self.verbose: print('self.unique_labels_of_interest: {0}'.format( self.unique_labels_of_interest)) if self.verbose: print('self.unique_label_indices_of_interest: {0}'.format( self.unique_label_indices_of_interest)) elapsed_time = time.time() - start_time print('done ({0:.2f} seconds)'.format(elapsed_time)) return token_to_vector
def remap_labels(y_pred, y_true, dataset, evaluation_mode='bio'): ''' y_pred: list of predicted labels y_true: list of gold labels evaluation_mode: 'bio', 'token', or 'binary' Both y_pred and y_true must use label indices and names specified in the dataset # (dataset.unique_label_indices_of_interest, dataset.unique_label_indices_of_interest). ''' all_unique_labels = dataset.unique_labels if evaluation_mode == 'bio': # sort label to index new_label_names = all_unique_labels[:] new_label_names.remove('O') new_label_names.sort( key=lambda x: (utils_nlp.remove_bio_from_label_name(x), x)) new_label_names.append('O') new_label_indices = list(range(len(new_label_names))) new_label_to_index = dict(zip(new_label_names, new_label_indices)) remap_index = {} for i, label_name in enumerate(new_label_names): label_index = dataset.label_to_index[label_name] remap_index[label_index] = i elif evaluation_mode == 'token': new_label_names = set() for label_name in all_unique_labels: if label_name == 'O': continue new_label_name = utils_nlp.remove_bio_from_label_name(label_name) new_label_names.add(new_label_name) new_label_names = sorted(list(new_label_names)) + ['O'] new_label_indices = list(range(len(new_label_names))) new_label_to_index = dict(zip(new_label_names, new_label_indices)) remap_index = {} for label_name in all_unique_labels: new_label_name = utils_nlp.remove_bio_from_label_name(label_name) label_index = dataset.label_to_index[label_name] remap_index[label_index] = new_label_to_index[new_label_name] elif evaluation_mode == 'binary': new_label_names = ['NAMED_ENTITY', 'O'] new_label_indices = [0, 1] new_label_to_index = dict(zip(new_label_names, new_label_indices)) remap_index = {} for label_name in all_unique_labels: new_label_name = 'O' if label_name != 'O': new_label_name = 'NAMED_ENTITY' label_index = dataset.label_to_index[label_name] remap_index[label_index] = new_label_to_index[new_label_name] else: raise ValueError( "evaluation_mode must be either 'bio', 'token', or 'binary'.") new_y_pred = [remap_index[label_index] for label_index in y_pred] new_y_true = [remap_index[label_index] for label_index in y_true] new_label_indices_with_o = new_label_indices[:] new_label_names_with_o = new_label_names[:] new_label_names.remove('O') new_label_indices.remove(new_label_to_index['O']) return new_y_pred, new_y_true, new_label_indices, new_label_names, new_label_indices_with_o, new_label_names_with_o
def load_dataset(self, dataset_filepaths, parameters): ''' dataset_filepaths : dictionary with keys 'train', 'valid', 'test', 'deploy' ''' start_time = time.time() print('Load dataset... ', end='', flush=True) all_pretrained_tokens = [] if parameters['token_pretrained_embedding_filepath'] != '': all_pretrained_tokens = utils_nlp.load_tokens_from_pretrained_token_embeddings( parameters) if self.verbose: print("len(all_pretrained_tokens): {0}".format( len(all_pretrained_tokens))) all_tokens_in_pretraining_dataset = [] if parameters['use_pretrained_model']: pretraining_dataset = pickle.load( open( os.path.join(parameters['pretrained_model_folder'], 'dataset.pickle'), 'rb')) all_tokens_in_pretraining_dataset = pretraining_dataset.index_to_token.values( ) remap_to_unk_count_threshold = 1 self.UNK_TOKEN_INDEX = 0 self.PADDING_CHARACTER_INDEX = 0 self.tokens_mapped_to_unk = [] self.UNK = 'UNK' self.unique_labels = [] labels = {} tokens = {} if parameters['use_corrector']: labels_corrector = {} label_corrector_count = {} self.unique_labels_corrector = [] if parameters['include_pos']: labels_pos = {} label_pos_count = {} self.unique_labels_pos = [] characters = {} token_lengths = {} label_count = {} token_count = {} character_count = {} for dataset_type in ['train', 'valid', 'test', 'deploy']: # print("what am i getting?? {:s}".format(str(dataset_filepaths.get(dataset_type, None)))) if parameters['use_corrector']: if parameters['include_pos']: labels_pos[dataset_type], labels_corrector[dataset_type], labels[dataset_type], \ tokens[dataset_type], token_count[dataset_type], label_pos_count[dataset_type], \ label_corrector_count[dataset_type], label_count[dataset_type], character_count[dataset_type] \ = self._parse_dataset(dataset_filepaths.get(dataset_type, None), use_corrector=True, include_pos=True, tagging_format=parameters['tagging_format']) else: labels_corrector[dataset_type], labels[dataset_type], tokens[dataset_type], \ token_count[dataset_type], label_corrector_count[dataset_type], label_count[dataset_type], \ character_count[dataset_type] \ = self._parse_dataset(dataset_filepaths.get(dataset_type, None), use_corrector=True, tagging_format=parameters['tagging_format']) else: labels[dataset_type], tokens[dataset_type], token_count[dataset_type], label_count[dataset_type], \ character_count[dataset_type] \ = self._parse_dataset(dataset_filepaths.get(dataset_type, None), tagging_format=parameters['tagging_format']) if self.verbose: print("len(token_count[{1}]): {0}".format( len(token_count[dataset_type]), dataset_type)) # sys.exit(0) token_count['all'] = {} for token in list(token_count['train'].keys()) + list( token_count['valid'].keys()) + list( token_count['test'].keys()) + list( token_count['deploy'].keys()): token_count['all'][token] = token_count['train'][ token] + token_count['valid'][token] + token_count['test'][ token] + token_count['deploy'][token] if self.verbose: print("len(token_count[all]): {0}".format(len(token_count['all']))) for dataset_type in dataset_filepaths.keys(): if self.verbose: print("len(token_count[{1}]): {0}".format( len(token_count[dataset_type]), dataset_type)) character_count['all'] = {} for character in list(character_count['train'].keys()) + list( character_count['valid'].keys()) + list( character_count['test'].keys()) + list( character_count['deploy'].keys()): character_count['all'][character] = character_count['train'][ character] + character_count['valid'][ character] + character_count['test'][ character] + character_count['deploy'][character] label_count['all'] = {} for character in list(label_count['train'].keys()) + list( label_count['valid'].keys()) + list( label_count['test'].keys()) + list( label_count['deploy'].keys()): label_count['all'][character] = label_count['train'][character] + label_count['valid'][character] + \ label_count['test'][character] + label_count['deploy'][character] if parameters['use_corrector']: label_corrector_count['all'] = {} for label in list(label_corrector_count['train'].keys()) + list( label_corrector_count['valid'].keys()) + list( label_corrector_count['test'].keys()) + list( label_corrector_count['deploy'].keys()): label_corrector_count['all'][label] = label_corrector_count['train'][label] + label_corrector_count['valid'][label] + \ label_corrector_count['test'][label] + label_corrector_count['deploy'][label] label_corrector_count['all'] = utils.order_dictionary( label_corrector_count['all'], 'key', reverse=False) if parameters['include_pos']: label_pos_count['all'] = {} for label in list(label_pos_count['train'].keys()) + list( label_pos_count['valid'].keys()) + list( label_pos_count['test'].keys()) + list( label_pos_count['deploy'].keys()): label_pos_count['all'][label] = label_pos_count['train'][label] + label_pos_count['valid'][label] + \ label_pos_count['test'][label] + label_pos_count['deploy'][label] label_pos_count['all'] = utils.order_dictionary( label_pos_count['all'], 'key', reverse=False) token_count['all'] = utils.order_dictionary(token_count['all'], 'value_key', reverse=True) label_count['all'] = utils.order_dictionary(label_count['all'], 'key', reverse=False) character_count['all'] = utils.order_dictionary(character_count['all'], 'value', reverse=True) if self.verbose: print('character_count[\'all\']: {0}'.format( character_count['all'])) token_to_index = {} token_to_index[self.UNK] = self.UNK_TOKEN_INDEX iteration_number = 0 number_of_unknown_tokens = 0 if self.verbose: print("parameters['remap_unknown_tokens_to_unk']: {0}".format( parameters['remap_unknown_tokens_to_unk'])) if self.verbose: print("len(token_count['train'].keys()): {0}".format( len(token_count['train'].keys()))) for token, count in token_count['all'].items(): if iteration_number == self.UNK_TOKEN_INDEX: iteration_number += 1 if parameters['remap_unknown_tokens_to_unk'] == 1 and \ (token_count['train'][token] == 0 or \ parameters['load_only_pretrained_token_embeddings']) and \ not utils_nlp.is_token_in_pretrained_embeddings(token, all_pretrained_tokens, parameters) and \ token not in all_tokens_in_pretraining_dataset: token_to_index[token] = self.UNK_TOKEN_INDEX number_of_unknown_tokens += 1 self.tokens_mapped_to_unk.append(token) else: token_to_index[token] = iteration_number iteration_number += 1 if self.verbose: print("number_of_unknown_tokens: {0}".format( number_of_unknown_tokens)) infrequent_token_indices = [] for token, count in token_count['train'].items(): if 0 < count <= remap_to_unk_count_threshold: infrequent_token_indices.append(token_to_index[token]) if self.verbose: print("len(token_count['train']): {0}".format( len(token_count['train']))) if self.verbose: print("len(infrequent_token_indices): {0}".format( len(infrequent_token_indices))) # Ensure that both B- and I- versions exist for each label labels_without_bio = set() for label in label_count['all'].keys(): new_label = utils_nlp.remove_bio_from_label_name(label) labels_without_bio.add(new_label) for label in labels_without_bio: if label == 'O': continue if parameters['tagging_format'] == 'bioes': prefixes = ['B-', 'I-', 'E-', 'S-'] else: prefixes = ['B-', 'I-'] for prefix in prefixes: l = prefix + label if l not in label_count['all']: label_count['all'][l] = 0 label_count['all'] = utils.order_dictionary(label_count['all'], 'key', reverse=False) if parameters['use_pretrained_model'] and not parameters['add_class']: self.unique_labels = sorted( list(pretraining_dataset.label_to_index.keys())) # Make sure labels are compatible with the pretraining dataset. for label in label_count['all']: if label not in pretraining_dataset.label_to_index: raise AssertionError( "The label {0} does not exist in the pretraining dataset. " .format(label) + "Please ensure that only the following labels exist in the dataset: {0}" .format(', '.join(self.unique_labels))) label_to_index = pretraining_dataset.label_to_index.copy() elif parameters['use_pretrained_model'] and parameters['add_class']: # make sure that the added labels are mapped to the end of the dectionary print('Adding new label-index pair to label_to_index dictionary') old_label_to_index = pretraining_dataset.label_to_index.copy() for label, count in label_count['all'].items(): if label not in old_label_to_index.keys(): old_label_to_index[label] = len(old_label_to_index.keys()) label_to_index = old_label_to_index.copy() self.unique_labels = list(label_to_index.keys()) else: label_to_index = {} iteration_number = 0 for label, count in label_count['all'].items(): label_to_index[label] = iteration_number iteration_number += 1 self.unique_labels.append(label) if parameters['use_corrector']: label_corrector_to_index = {} self.unique_labels_corrector = list( label_corrector_count['all'].keys()) for n, label in enumerate(self.unique_labels_corrector): label_corrector_to_index[label] = n if parameters['include_pos']: label_pos_to_index = {} self.unique_labels_pos = list(label_pos_count['all'].keys()) for n, pos in enumerate(self.unique_labels_pos): label_pos_to_index[pos] = n if self.verbose: print('self.unique_labels: {0}'.format(self.unique_labels)) character_to_index = {} iteration_number = 0 for character, count in character_count['all'].items(): if iteration_number == self.PADDING_CHARACTER_INDEX: iteration_number += 1 character_to_index[character] = iteration_number iteration_number += 1 if self.verbose: print('token_count[\'train\'][0:10]: {0}'.format( list(token_count['train'].items())[0:10])) token_to_index = utils.order_dictionary(token_to_index, 'value', reverse=False) index_to_token = utils.reverse_dictionary(token_to_index) if parameters['remap_unknown_tokens_to_unk'] == 1: index_to_token[self.UNK_TOKEN_INDEX] = self.UNK if self.verbose: print('label_count[\'train\']: {0}'.format(label_count['train'])) label_to_index = utils.order_dictionary(label_to_index, 'value', reverse=False) if self.verbose: print('label_to_index: {0}'.format(label_to_index)) index_to_label = utils.reverse_dictionary(label_to_index) if self.verbose: print('index_to_label: {0}'.format(index_to_label)) if parameters['use_corrector']: label_corrector_to_index = utils.order_dictionary( label_corrector_to_index, 'value', reverse=False) index_to_label_corrector = utils.reverse_dictionary( label_corrector_to_index) if parameters['include_pos']: label_pos_to_index = utils.order_dictionary(label_pos_to_index, 'value', reverse=False) index_to_label_pos = utils.reverse_dictionary( label_pos_to_index) character_to_index = utils.order_dictionary(character_to_index, 'value', reverse=False) index_to_character = utils.reverse_dictionary(character_to_index) if self.verbose: print('character_to_index: {0}'.format(character_to_index)) if self.verbose: print('index_to_character: {0}'.format(index_to_character)) if self.verbose: print('labels[\'train\'][0:10]: {0}'.format(labels['train'][0:10])) if self.verbose: print('tokens[\'train\'][0:10]: {0}'.format(tokens['train'][0:10])) if self.verbose: # Print sequences of length 1 in train set for token_sequence, label_sequence in zip(tokens['train'], labels['train']): if len(label_sequence) == 1 and label_sequence[0] != 'O': print("{0}\t{1}".format(token_sequence[0], label_sequence[0])) token_indices = {} label_indices = {} if parameters['use_corrector']: label_indices_corrector = {} if parameters['include_pos']: label_indices_pos = {} character_indices = {} character_indices_padded = {} for dataset_type in dataset_filepaths.keys(): # print("dataset_type: {:s}".format(dataset_type)) token_indices[dataset_type] = [] characters[dataset_type] = [] character_indices[dataset_type] = [] token_lengths[dataset_type] = [] character_indices_padded[dataset_type] = [] for token_sequence in tokens[dataset_type]: token_indices[dataset_type].append( [token_to_index[token] for token in token_sequence]) characters[dataset_type].append( [list(token) for token in token_sequence]) character_indices[dataset_type].append( [[character_to_index[character] for character in token] for token in token_sequence]) token_lengths[dataset_type].append( [len(token) for token in token_sequence]) longest_token_length_in_sequence = max( token_lengths[dataset_type][-1]) character_indices_padded[dataset_type].append([ utils.pad_list(temp_token_indices, longest_token_length_in_sequence, self.PADDING_CHARACTER_INDEX) for temp_token_indices in character_indices[dataset_type][-1] ]) label_indices[dataset_type] = [] for label_sequence in labels[dataset_type]: label_indices[dataset_type].append( [label_to_index[label] for label in label_sequence]) if parameters['use_corrector']: label_indices_corrector[dataset_type] = [] for label_sequence_corrector in labels_corrector[dataset_type]: label_indices_corrector[dataset_type].append([ label_corrector_to_index[label] for label in label_sequence_corrector ]) if parameters['include_pos']: label_indices_pos[dataset_type] = [] for label_sequence_pos in labels_pos[dataset_type]: label_indices_pos[dataset_type].append([ label_pos_to_index[label] for label in label_sequence_pos ]) if self.verbose: print('token_lengths[\'train\'][0][0:10]: {0}'.format( token_lengths['train'][0][0:10])) if self.verbose: print('characters[\'train\'][0][0:10]: {0}'.format( characters['train'][0][0:10])) if self.verbose: print('token_indices[\'train\'][0:10]: {0}'.format( token_indices['train'][0:10])) if self.verbose: print('label_indices[\'train\'][0:10]: {0}'.format( label_indices['train'][0:10])) if self.verbose: print('character_indices[\'train\'][0][0:10]: {0}'.format( character_indices['train'][0][0:10])) if self.verbose: print('character_indices_padded[\'train\'][0][0:10]: {0}'.format( character_indices_padded['train'][0][0:10])) label_binarizer = sklearn.preprocessing.LabelBinarizer() label_binarizer.fit(range(max(index_to_label.keys()) + 1)) label_vector_indices = {} for dataset_type in dataset_filepaths.keys(): label_vector_indices[dataset_type] = [] for label_indices_sequence in label_indices[dataset_type]: label_vector_indices[dataset_type].append( label_binarizer.transform(label_indices_sequence)) if parameters['use_corrector']: label_binarizer_corrector = sklearn.preprocessing.LabelBinarizer() label_binarizer_corrector.fit( range(max(index_to_label_corrector.keys()) + 1)) label_corrector_vector_indices = {} for dataset_type in dataset_filepaths.keys(): label_corrector_vector_indices[dataset_type] = [] for label_indices_sequence in label_indices_corrector[ dataset_type]: label_corrector_vector_indices[dataset_type].append( label_binarizer_corrector.transform( label_indices_sequence)) if parameters['include_pos']: label_binarizer_pos = sklearn.preprocessing.LabelBinarizer() label_binarizer_pos.fit( range(max(index_to_label_pos.keys()) + 1)) label_pos_vector_indices = {} for dataset_type in dataset_filepaths.keys(): label_pos_vector_indices[dataset_type] = [] for label_indices_sequence in label_indices_pos[ dataset_type]: label_pos_vector_indices[dataset_type].append( label_binarizer_pos.transform( label_indices_sequence)) if self.verbose: print('label_vector_indices[\'train\'][0:2]: {0}'.format( label_vector_indices['train'][0:2])) if self.verbose: print('len(label_vector_indices[\'train\']): {0}'.format( len(label_vector_indices['train']))) self.token_to_index = token_to_index self.index_to_token = index_to_token self.token_indices = token_indices self.label_indices = label_indices self.character_indices_padded = character_indices_padded self.index_to_character = index_to_character self.character_to_index = character_to_index self.character_indices = character_indices self.token_lengths = token_lengths self.characters = characters self.tokens = tokens self.labels = labels self.label_vector_indices = label_vector_indices self.index_to_label = index_to_label self.label_to_index = label_to_index if parameters['use_corrector']: self.index_to_label_corrector = index_to_label_corrector self.label_corrector_to_index = label_corrector_to_index self.label_indices_corrector = label_indices_corrector self.label_corrector_vector_indices = label_corrector_vector_indices if parameters['include_pos']: self.index_to_label_pos = index_to_label_pos self.label_pos_to_index = label_pos_to_index self.label_indices_pos = label_indices_pos self.label_pos_vector_indices = label_pos_vector_indices if self.verbose: print("len(self.token_to_index): {0}".format( len(self.token_to_index))) if self.verbose: print("len(self.index_to_token): {0}".format( len(self.index_to_token))) if parameters['add_class'] and parameters[ 'tagging_format'] == 'bioes' and len( self.index_to_label) > 100: self.number_of_classes = max(self.index_to_label.keys()) + 1 - 8 elif parameters['add_class'] and parameters[ 'tagging_format'] == 'bioes': print('here') self.number_of_classes = max(self.index_to_label.keys()) + 1 - 4 elif parameters['add_class'] and parameters['tagging_format'] == 'bio': print('here2') self.number_of_classes = max(self.index_to_label.keys()) + 1 - 2 else: self.number_of_classes = max( self.index_to_label.keys()) + 1 # 1 is for O label print('max(self.index_to_label.keys()) : {:d}'.format( max(self.index_to_label.keys()))) print(self.index_to_label.keys()) print(self.number_of_classes) self.vocabulary_size = max(self.index_to_token.keys()) + 1 self.alphabet_size = max(self.index_to_character.keys()) + 1 if self.verbose: print("self.number_of_classes: {0}".format(self.number_of_classes)) if self.verbose: print("self.alphabet_size: {0}".format(self.alphabet_size)) if self.verbose: print("self.vocabulary_size: {0}".format(self.vocabulary_size)) self.unique_labels_of_interest = list(self.unique_labels) self.unique_labels_of_interest.remove('O') self.unique_label_indices_of_interest = [] for lab in self.unique_labels_of_interest: self.unique_label_indices_of_interest.append(label_to_index[lab]) self.infrequent_token_indices = infrequent_token_indices if self.verbose: print('self.unique_labels_of_interest: {0}'.format( self.unique_labels_of_interest)) if self.verbose: print('self.unique_label_indices_of_interest: {0}'.format( self.unique_label_indices_of_interest)) elapsed_time = time.time() - start_time print('done ({0:.2f} seconds)'.format(elapsed_time))
def load_dataset(self, dataset_filepaths, parameters, token_to_vector=None): ''' dataset_filepaths : dictionary with keys 'train', 'valid', 'test', 'deploy' Load word vectors từ file đã chuẩn bị sẵn ''' start_time = time.time() print('Load dataset... ', end='', flush=True) if parameters['token_pretrained_embedding_filepath'] != '': if token_to_vector == None: token_to_vector = utils_nlp.load_pretrained_token_embeddings( parameters) else: token_to_vector = {} if self.verbose: print("len(token_to_vector): {0}".format(len(token_to_vector))) # Load pretraining dataset to ensure that index to label is compatible to the pretrained model, # and that token embeddings that are learned in the pretrained model are loaded properly. all_tokens_in_pretraining_dataset = [] all_characters_in_pretraining_dataset = [] if parameters['use_pretrained_model']: pretraining_dataset = pickle.load( open( os.path.join(parameters['pretrained_model_folder'], 'dataset.pickle'), 'rb')) all_tokens_in_pretraining_dataset = pretraining_dataset.index_to_token.values( ) # Những token lưu ở đợt train trước all_characters_in_pretraining_dataset = pretraining_dataset.index_to_character.values( ) # Những character lưu ở đợt train trước remap_to_unk_count_threshold = 1 self.UNK_TOKEN_INDEX = 0 # Index của những unknow token self.PADDING_CHARACTER_INDEX = 0 self.tokens_mapped_to_unk = [] # những unknown token self.UNK = 'UNK' self.unique_labels = [] # Các nhãn tồn tại trong dataset labels = {} # nhãn {all: ...., train: ..., test: ...} tokens = {} # token {all: ...., train: ..., test: ...} label_count = {} # Đếm số nhãn {all: ...., train: ..., test: ...} token_count = {} # Đếm số token {all: ...., train: ..., test: ...} character_count = {} # Đếm số ký tự {all: ...., train: ..., test: ...} for dataset_type in ['train', 'valid', 'test', 'deploy']: labels[dataset_type], tokens[dataset_type], token_count[dataset_type], label_count[dataset_type], character_count[dataset_type] \ = self._parse_dataset(dataset_filepaths.get(dataset_type, None)) if self.verbose: print("dataset_type: {0}".format(dataset_type)) if self.verbose: print("len(token_count[dataset_type]): {0}".format( len(token_count[dataset_type]))) # Tính tổng hợp lại cho tất cả các dataset token_count['all'] = {} for token in list(token_count['train'].keys()) + list( token_count['valid'].keys()) + list( token_count['test'].keys()) + list( token_count['deploy'].keys()): token_count['all'][token] = token_count['train'][ token] + token_count['valid'][token] + token_count['test'][ token] + token_count['deploy'][token] # Thêm những token ở pretrained trước với giá trị -1 if parameters['load_all_pretrained_token_embeddings']: for token in token_to_vector: if token not in token_count['all']: token_count['all'][token] = -1 token_count['train'][token] = -1 for token in all_tokens_in_pretraining_dataset: if token not in token_count['all']: token_count['all'][token] = -1 token_count['train'][token] = -1 # Tính tổng hợp lại cho tất cả các dataset character_count['all'] = {} for character in list(character_count['train'].keys()) + list( character_count['valid'].keys()) + list( character_count['test'].keys()) + list( character_count['deploy'].keys()): character_count['all'][character] = character_count['train'][ character] + character_count['valid'][ character] + character_count['test'][ character] + character_count['deploy'][character] # Thêm những token ở pretrained trước với giá trị -1 for character in all_characters_in_pretraining_dataset: if character not in character_count['all']: character_count['all'][character] = -1 character_count['train'][character] = -1 for dataset_type in dataset_filepaths.keys(): if self.verbose: print("dataset_type: {0}".format(dataset_type)) if self.verbose: print("len(token_count[dataset_type]): {0}".format( len(token_count[dataset_type]))) # Tính tổng hợp lại các nhãn ở đợt train trước label_count['all'] = {} for character in list(label_count['train'].keys()) + list( label_count['valid'].keys()) + list( label_count['test'].keys()) + list( label_count['deploy'].keys()): label_count['all'][character] = label_count['train'][ character] + label_count['valid'][character] + label_count[ 'test'][character] + label_count['deploy'][character] token_count['all'] = utils.order_dictionary( token_count['all'], 'value_key', reverse=True ) # Sort token count theo các token có freq cao đến thấp, token desc label_count['all'] = utils.order_dictionary( label_count['all'], 'key', reverse=False) # Sort label count theo label asc character_count['all'] = utils.order_dictionary( character_count['all'], 'value', reverse=True ) # Sort character count theo các character có freq cao đến thấp if self.verbose: print('character_count[\'all\']: {0}'.format( character_count['all'])) token_to_index = {} token_to_index[self.UNK] = self.UNK_TOKEN_INDEX iteration_number = 0 number_of_unknown_tokens = 0 if self.verbose: print("parameters['remap_unknown_tokens_to_unk']: {0}".format( parameters['remap_unknown_tokens_to_unk'])) if self.verbose: print("len(token_count['train'].keys()): {0}".format( len(token_count['train'].keys()))) for token, count in token_count['all'].items(): if iteration_number == self.UNK_TOKEN_INDEX: iteration_number += 1 ''' UNK_TOKEN: token không xuất hiện trong pretraining_dataset và trong word vectors ''' if parameters['remap_unknown_tokens_to_unk'] == 1 and \ (token_count['train'][token] == 0 or \ parameters['load_only_pretrained_token_embeddings']) and \ not utils_nlp.is_token_in_pretrained_embeddings(token, token_to_vector, parameters) and \ token not in all_tokens_in_pretraining_dataset: if self.verbose: print("token: {0}".format(token)) if self.verbose: print("token.lower(): {0}".format(token.lower())) if self.verbose: print("re.sub('\d', '0', token.lower()): {0}".format( re.sub('\d', '0', token.lower()))) token_to_index[token] = self.UNK_TOKEN_INDEX number_of_unknown_tokens += 1 self.tokens_mapped_to_unk.append(token) else: token_to_index[token] = iteration_number iteration_number += 1 if self.verbose: print("number_of_unknown_tokens: {0}".format( number_of_unknown_tokens)) infrequent_token_indices = [ ] # Các token xuất hiện thấp trong train dataset for token, count in token_count['train'].items(): if 0 < count <= remap_to_unk_count_threshold: infrequent_token_indices.append(token_to_index[token]) if self.verbose: print("len(token_count['train']): {0}".format( len(token_count['train']))) if self.verbose: print("len(infrequent_token_indices): {0}".format( len(infrequent_token_indices))) # Ensure that both B- and I- versions exist for each label # Bỏ các tiền tố B-, O-, I-... labels_without_bio = set() for label in label_count['all'].keys(): new_label = utils_nlp.remove_bio_from_label_name(label) labels_without_bio.add(new_label) # Kết hợp các ENTITY vs các tiền tố B-, I-,... và thêm vào label count for label in labels_without_bio: if label == 'O': continue if parameters['tagging_format'] == 'bioes': prefixes = ['B-', 'I-', 'E-', 'S-'] else: prefixes = ['B-', 'I-'] for prefix in prefixes: l = prefix + label if l not in label_count['all']: label_count['all'][l] = 0 # Sắp xếp label_count theo label asc label_count['all'] = utils.order_dictionary(label_count['all'], 'key', reverse=False) if parameters['use_pretrained_model']: self.unique_labels = sorted( list(pretraining_dataset.label_to_index.keys())) # Make sure labels are compatible with the pretraining dataset. for label in label_count['all']: if label not in pretraining_dataset.label_to_index: raise AssertionError( "The label {0} does not exist in the pretraining dataset. " .format(label) + "Please ensure that only the following labels exist in the dataset: {0}" .format(', '.join(self.unique_labels))) label_to_index = pretraining_dataset.label_to_index.copy() else: label_to_index = {} iteration_number = 0 for label, count in label_count['all'].items(): label_to_index[label] = iteration_number iteration_number += 1 self.unique_labels.append(label) if self.verbose: print('self.unique_labels: {0}'.format(self.unique_labels)) character_to_index = {} iteration_number = 0 for character, count in character_count['all'].items(): if iteration_number == self.PADDING_CHARACTER_INDEX: iteration_number += 1 character_to_index[character] = iteration_number iteration_number += 1 if self.verbose: print('token_count[\'train\'][0:10]: {0}'.format( list(token_count['train'].items())[0:10])) token_to_index = utils.order_dictionary(token_to_index, 'value', reverse=False) if self.verbose: print('token_to_index: {0}'.format(token_to_index)) index_to_token = utils.reverse_dictionary(token_to_index) if parameters['remap_unknown_tokens_to_unk'] == 1: index_to_token[self.UNK_TOKEN_INDEX] = self.UNK if self.verbose: print('index_to_token: {0}'.format(index_to_token)) if self.verbose: print('label_count[\'train\']: {0}'.format(label_count['train'])) label_to_index = utils.order_dictionary(label_to_index, 'value', reverse=False) if self.verbose: print('label_to_index: {0}'.format(label_to_index)) index_to_label = utils.reverse_dictionary(label_to_index) if self.verbose: print('index_to_label: {0}'.format(index_to_label)) character_to_index = utils.order_dictionary(character_to_index, 'value', reverse=False) index_to_character = utils.reverse_dictionary(character_to_index) if self.verbose: print('character_to_index: {0}'.format(character_to_index)) if self.verbose: print('index_to_character: {0}'.format(index_to_character)) if self.verbose: print('labels[\'train\'][0:10]: {0}'.format(labels['train'][0:10])) if self.verbose: print('tokens[\'train\'][0:10]: {0}'.format(tokens['train'][0:10])) if self.verbose: # Print sequences of length 1 in train set for token_sequence, label_sequence in zip(tokens['train'], labels['train']): if len(label_sequence) == 1 and label_sequence[0] != 'O': print("{0}\t{1}".format(token_sequence[0], label_sequence[0])) self.token_to_index = token_to_index # {token: index sau khi sắp xếp theo freq từ cao đến thấp, 0 nếu là unk token} self.index_to_token = index_to_token # Ngược token_to_index self.index_to_character = index_to_character # Ngược character_to_index self.character_to_index = character_to_index # { character: index sau khi sắp xếp freq từ cao đến thấp} self.index_to_label = index_to_label # Ngược label_to_index self.label_to_index = label_to_index # {label: index sau khi sắp xếp asc} if self.verbose: print("len(self.token_to_index): {0}".format( len(self.token_to_index))) if self.verbose: print("len(self.index_to_token): {0}".format( len(self.index_to_token))) self.tokens = tokens self.labels = labels token_indices, label_indices, character_indices_padded, character_indices, token_lengths, characters, label_vector_indices = self._convert_to_indices( dataset_filepaths.keys()) self.token_indices = token_indices self.label_indices = label_indices self.character_indices_padded = character_indices_padded self.character_indices = character_indices self.token_lengths = token_lengths self.characters = characters self.label_vector_indices = label_vector_indices self.number_of_classes = max(self.index_to_label.keys()) + 1 self.vocabulary_size = max(self.index_to_token.keys()) + 1 self.alphabet_size = max(self.index_to_character.keys()) + 1 if self.verbose: print("self.number_of_classes: {0}".format(self.number_of_classes)) if self.verbose: print("self.alphabet_size: {0}".format(self.alphabet_size)) if self.verbose: print("self.vocabulary_size: {0}".format(self.vocabulary_size)) # unique_labels_of_interest is used to compute F1-scores. self.unique_labels_of_interest = list(self.unique_labels) self.unique_labels_of_interest.remove('O') self.unique_label_indices_of_interest = [] for lab in self.unique_labels_of_interest: self.unique_label_indices_of_interest.append(label_to_index[lab]) self.infrequent_token_indices = infrequent_token_indices if self.verbose: print('self.unique_labels_of_interest: {0}'.format( self.unique_labels_of_interest)) if self.verbose: print('self.unique_label_indices_of_interest: {0}'.format( self.unique_label_indices_of_interest)) elapsed_time = time.time() - start_time print('done ({0:.2f} seconds)'.format(elapsed_time)) return token_to_vector