def train_nfold(self, x_train, y_train, x_valid=None, y_valid=None, f_train: np.array = None, f_valid: np.array = None, fold_number=10, callbacks=None):
x_all = np.concatenate((x_train, x_valid), axis=0) if x_valid is not None else x_train
y_all = np.concatenate((y_train, y_valid), axis=0) if y_valid is not None else y_train
features_all = concatenate_or_none((f_train, f_valid), axis=0)
self.p = prepare_preprocessor(x_all, y_all, features=features_all, model_config=self.model_config)
self.model_config.char_vocab_size = len(self.p.vocab_char)
self.model_config.case_vocab_size = len(self.p.vocab_case)
self.p.return_lengths = True
if 'bert' in self.model_config.model_type.lower():
self.model = get_model(self.model_config, self.p, len(self.p.vocab_tag))
self.models = []
for k in range(0, fold_number):
model = get_model(self.model_config, self.p, len(self.p.vocab_tag))
self.models.append(model)
trainer = Trainer(self.model,
self.models,
self.embeddings,
self.model_config,
self.training_config,
checkpoint_path=self.log_dir,
preprocessor=self.p
)
trainer.train_nfold(x_train, y_train, x_valid, y_valid, f_train=f_train, f_valid=f_valid, callbacks=callbacks)
if self.embeddings.use_ELMo:
self.embeddings.clean_ELMo_cache()
if self.embeddings.use_BERT:
self.embeddings.clean_BERT_cache()
if 'bert' in self.model_config.model_type.lower():
self.save()
def load(self, dir_path='data/models/sequenceLabelling/'):
self.model_config = ModelConfig.load(
os.path.join(dir_path, self.model_config.model_name,
self.config_file))
self.p = WordPreprocessor.load(
os.path.join(dir_path, self.model_config.model_name,
self.preprocessor_file))
if self.model_config.model_type.lower().find("bert") != -1:
self.model = get_model(self.model_config,
self.p,
ntags=len(self.p.vocab_tag),
dir_path=dir_path)
self.model.load_model()
return
# load embeddings
# Do not use cache in 'production' mode
self.embeddings = Embeddings(self.model_config.embeddings_name,
use_ELMo=self.model_config.use_ELMo,
use_BERT=self.model_config.use_BERT,
use_cache=False)
self.model_config.word_embedding_size = self.embeddings.embed_size
self.model = get_model(self.model_config,
self.p,
ntags=len(self.p.vocab_tag))
self.model.load(filepath=os.path.join(
dir_path, self.model_config.model_name, self.weight_file))
def train_nfold(self, x_train, y_train, x_valid=None, y_valid=None, fold_number=10):
if x_valid is not None and y_valid is not None:
x_all = np.concatenate((x_train, x_valid), axis=0)
y_all = np.concatenate((y_train, y_valid), axis=0)
self.p = prepare_preprocessor(x_all, y_all, self.model_config)
else:
self.p = prepare_preprocessor(x_train, y_train, self.model_config)
self.model_config.char_vocab_size = len(self.p.vocab_char)
self.model_config.case_vocab_size = len(self.p.vocab_case)
self.p.return_lengths = True
#self.model = get_model(self.model_config, self.p, len(self.p.vocab_tag))
self.models = []
for k in range(0, fold_number):
model = get_model(self.model_config, self.p, len(self.p.vocab_tag))
self.models.append(model)
trainer = Trainer(self.model,
self.models,
self.embeddings,
self.model_config,
self.training_config,
checkpoint_path=self.log_dir,
preprocessor=self.p
)
trainer.train_nfold(x_train, y_train, x_valid, y_valid)
if self.embeddings.use_ELMo:
self.embeddings.clean_ELMo_cache()
def train(self, x_train, y_train, x_valid=None, y_valid=None):
# TBD if valid is None, segment train to get one
x_all = np.concatenate((x_train, x_valid), axis=0)
y_all = np.concatenate((y_train, y_valid), axis=0)
self.p = prepare_preprocessor(x_all, y_all, self.model_config)
self.model_config.char_vocab_size = len(self.p.vocab_char)
self.model_config.case_vocab_size = len(self.p.vocab_case)
"""
if self.embeddings.use_ELMo:
# dump token context independent data for the train set, done once for the training
x_train_local = x_train
if not self.training_config.early_stop:
# in case we want to train with the validation set too, we dump also
# the ELMo embeddings for the token of the valid set
x_train_local = np.concatenate((x_train, x_valid), axis=0)
self.embeddings.dump_ELMo_token_embeddings(x_train_local)
"""
self.model = get_model(self.model_config, self.p, len(self.p.vocab_tag))
trainer = Trainer(self.model,
self.models,
self.embeddings,
self.model_config,
self.training_config,
checkpoint_path=self.log_dir,
preprocessor=self.p
)
trainer.train(x_train, y_train, x_valid, y_valid)
if self.embeddings.use_ELMo:
self.embeddings.clean_ELMo_cache()
def load(self,
dir_path='data/models/sequenceLabelling/',
weight_file=DEFAULT_WEIGHT_FILE_NAME):
model_path = os.path.join(dir_path, self.model_config.model_name)
self.model_config = ModelConfig.load(
os.path.join(model_path, CONFIG_FILE_NAME))
if self.model_config.embeddings_name is not None:
# load embeddings
# Do not use cache in 'prediction/production' mode
self.embeddings = Embeddings(self.model_config.embeddings_name,
resource_registry=self.registry,
use_ELMo=self.model_config.use_ELMo,
use_cache=False)
self.model_config.word_embedding_size = self.embeddings.embed_size
else:
self.embeddings = None
self.model_config.word_embedding_size = 0
self.p = Preprocessor.load(
os.path.join(dir_path, self.model_config.model_name,
PROCESSOR_FILE_NAME))
self.model = get_model(self.model_config,
self.p,
ntags=len(self.p.vocab_tag),
load_pretrained_weights=False,
local_path=os.path.join(
dir_path, self.model_config.model_name))
print(
"load weights from",
os.path.join(dir_path, self.model_config.model_name, weight_file))
self.model.load(filepath=os.path.join(
dir_path, self.model_config.model_name, weight_file))
self.model.print_summary()
def train(self,
x_train,
y_train,
f_train=None,
x_valid=None,
y_valid=None,
f_valid=None,
callbacks=None):
# TBD if valid is None, segment train to get one if early_stop is True
# we concatenate all the training+validation data to create the model vocabulary
if not x_valid is None:
x_all = np.concatenate((x_train, x_valid), axis=0)
else:
x_all = x_train
if not y_valid is None:
y_all = np.concatenate((y_train, y_valid), axis=0)
else:
y_all = y_train
features_all = concatenate_or_none((f_train, f_valid), axis=0)
self.p = prepare_preprocessor(x_all,
y_all,
features=features_all,
model_config=self.model_config)
self.model_config.char_vocab_size = len(self.p.vocab_char)
self.model_config.case_vocab_size = len(self.p.vocab_case)
self.model = get_model(self.model_config,
self.p,
len(self.p.vocab_tag),
load_pretrained_weights=True)
print_parameters(self.model_config, self.training_config)
self.model.print_summary()
# uncomment to plot graph
#plot_model(self.model,
# to_file='data/models/textClassification/'+self.model_config.model_name+'_'+self.model_config.architecture+'.png')
trainer = Trainer(
self.model,
self.models,
self.embeddings,
self.model_config,
self.training_config,
checkpoint_path=self.log_dir,
preprocessor=self.p,
transformer_preprocessor=self.model.transformer_preprocessor)
trainer.train(x_train,
y_train,
x_valid,
y_valid,
features_train=f_train,
features_valid=f_valid,
callbacks=callbacks)
if self.embeddings and self.embeddings.use_ELMo:
self.embeddings.clean_ELMo_cache()
def train(self, x_train, y_train, f_train: np.array = None, x_valid=None, y_valid=None, f_valid: np.array = None, callbacks=None):
# TBD if valid is None, segment train to get one
x_all = np.concatenate((x_train, x_valid), axis=0) if x_valid is not None else x_train
y_all = np.concatenate((y_train, y_valid), axis=0) if y_valid is not None else y_train
features_all = concatenate_or_none((f_train, f_valid), axis=0)
self.p = prepare_preprocessor(x_all, y_all, features=features_all, model_config=self.model_config)
self.model_config.char_vocab_size = len(self.p.vocab_char)
self.model_config.case_vocab_size = len(self.p.vocab_case)
self.model = get_model(self.model_config, self.p, len(self.p.vocab_tag))
if self.p.return_features is not False:
print('x_train.shape: ', x_train.shape)
print('features_train.shape: ', f_train.shape)
sample_transformed_features = self.p.transform_features(f_train)
self.model_config.max_feature_size = np.asarray(sample_transformed_features).shape[-1]
print('max_feature_size: ', self.model_config.max_feature_size)
trainer = Trainer(self.model,
self.models,
self.embeddings,
self.model_config,
self.training_config,
checkpoint_path=self.log_dir,
preprocessor=self.p
)
trainer.train(x_train, y_train, x_valid, y_valid, features_train=f_train, features_valid=f_valid, callbacks=callbacks)
if self.embeddings.use_ELMo:
self.embeddings.clean_ELMo_cache()
if self.embeddings.use_BERT:
self.embeddings.clean_BERT_cache()
def load(self, dir_path='data/models/sequenceLabelling/'):
self.p = WordPreprocessor.load(os.path.join(dir_path, self.model_config.model_name, self.preprocessor_file))
self.model_config = ModelConfig.load(os.path.join(dir_path, self.model_config.model_name, self.config_file))
# load embeddings
self.embeddings = Embeddings(self.model_config.embeddings_name, use_ELMo=self.model_config.use_ELMo)
self.model_config.word_embedding_size = self.embeddings.embed_size
self.model = get_model(self.model_config, self.p, ntags=len(self.p.vocab_tag))
self.model.load(filepath=os.path.join(dir_path, self.model_config.model_name, self.weight_file))
def train(self, x_train, y_train, x_valid=None, y_valid=None):
# TBD if valid is None, segment train to get one
x_all = np.concatenate((x_train, x_valid), axis=0)
y_all = np.concatenate((y_train, y_valid), axis=0)
self.p = prepare_preprocessor(x_all, y_all, self.model_config)
self.model_config.char_vocab_size = len(self.p.vocab_char)
self.model_config.case_vocab_size = len(self.p.vocab_case)
self.model = get_model(self.model_config, self.p,
len(self.p.vocab_tag))
trainer = Trainer(self.model,
self.models,
self.embeddings,
self.model_config,
self.training_config,
checkpoint_path=self.log_dir,
preprocessor=self.p)
trainer.train(x_train, y_train, x_valid, y_valid)
if self.embeddings.use_ELMo:
self.embeddings.clean_ELMo_cache()
if self.embeddings.use_BERT:
self.embeddings.clean_BERT_cache()
def eval_nfold(self, x_test, y_test, features=None):
if self.models is not None:
total_f1 = 0
best_f1 = 0
best_index = 0
worst_f1 = 1
worst_index = 0
reports = []
reports_as_map = []
total_precision = 0
total_recall = 0
for i in range(self.model_config.fold_number):
print('\n------------------------ fold ' + str(i) +
' --------------------------------------')
if 'bert' not in self.model_config.model_type.lower():
# Prepare test data(steps, generator)
test_generator = DataGenerator(
x_test,
y_test,
batch_size=self.model_config.batch_size,
preprocessor=self.p,
char_embed_size=self.model_config.char_embedding_size,
max_sequence_length=self.model_config.
max_sequence_length,
embeddings=self.embeddings,
shuffle=False,
features=features)
# Build the evaluator and evaluate the model
scorer = Scorer(test_generator, self.p, evaluation=True)
scorer.model = self.models[i]
scorer.on_epoch_end(epoch=-1)
f1 = scorer.f1
precision = scorer.precision
recall = scorer.recall
reports.append(scorer.report)
reports_as_map.append(scorer.report_as_map)
else:
# BERT architecture model
dir_path = 'data/models/sequenceLabelling/'
self.model_config = ModelConfig.load(
os.path.join(dir_path, self.model_config.model_name,
self.config_file))
self.p = WordPreprocessor.load(
os.path.join(dir_path, self.model_config.model_name,
self.preprocessor_file))
self.model = get_model(self.model_config,
self.p,
ntags=len(self.p.vocab_tag))
self.model.load_model(i)
y_pred = self.model.predict(x_test, fold_id=i)
nb_alignment_issues = 0
for j in range(len(y_test)):
if len(y_test[i]) != len(y_pred[j]):
nb_alignment_issues += 1
# BERT tokenizer appears to introduce some additional tokens without ## prefix,
# but this is normally handled when predicting.
# To be very conservative, the following ensure the number of tokens always
# match, but it should never be used in practice.
if len(y_test[j]) < len(y_pred[j]):
y_test[j] = y_test[j] + ["O"] * (
len(y_pred[j]) - len(y_test[j]))
if len(y_test[j]) > len(y_pred[j]):
y_pred[j] = y_pred[j] + ["O"] * (
len(y_test[j]) - len(y_pred[j]))
if nb_alignment_issues > 0:
print("number of alignment issues with test set:",
nb_alignment_issues)
f1 = f1_score(y_test, y_pred)
precision = precision_score(y_test, y_pred)
recall = recall_score(y_test, y_pred)
print("\tf1: {:04.2f}".format(f1 * 100))
print("\tprecision: {:04.2f}".format(precision * 100))
print("\trecall: {:04.2f}".format(recall * 100))
report, report_as_map = classification_report(y_test,
y_pred,
digits=4)
reports.append(report)
reports_as_map.append(report_as_map)
if best_f1 < f1:
best_f1 = f1
best_index = i
if worst_f1 > f1:
worst_f1 = f1
worst_index = i
total_f1 += f1
total_precision += precision
total_recall += recall
fold_average_evaluation = {'labels': {}, 'micro': {}, 'macro': {}}
micro_f1 = total_f1 / self.model_config.fold_number
micro_precision = total_precision / self.model_config.fold_number
micro_recall = total_recall / self.model_config.fold_number
micro_eval_block = {
'f1': micro_f1,
'precision': micro_precision,
'recall': micro_recall
}
fold_average_evaluation['micro'] = micro_eval_block
# field-level average over the n folds
labels = []
for label in sorted(self.p.vocab_tag):
if label == 'O' or label == '<PAD>':
continue
if label.startswith("B-") or label.startswith(
"S-") or label.startswith("I-") or label.startswith(
"E-"):
label = label[2:]
if label in labels:
continue
labels.append(label)
sum_p = 0
sum_r = 0
sum_f1 = 0
sum_support = 0
for j in range(0, self.model_config.fold_number):
if not label in reports_as_map[j]['labels']:
continue
report_as_map = reports_as_map[j]['labels'][label]
sum_p += report_as_map["precision"]
sum_r += report_as_map["recall"]
sum_f1 += report_as_map["f1"]
sum_support += report_as_map["support"]
avg_p = sum_p / self.model_config.fold_number
avg_r = sum_r / self.model_config.fold_number
avg_f1 = sum_f1 / self.model_config.fold_number
avg_support = sum_support / self.model_config.fold_number
avg_support_dec = str(avg_support - int(avg_support))[1:]
if avg_support_dec != '0':
avg_support = math.floor(avg_support)
block_label = {
'precision': avg_p,
'recall': avg_r,
'support': avg_support,
'f1': avg_f1
}
fold_average_evaluation['labels'][label] = block_label
print(
"----------------------------------------------------------------------"
)
print("\n** Worst ** model scores - run", str(worst_index))
print(reports[worst_index])
print("\n** Best ** model scores - run", str(best_index))
print(reports[best_index])
if 'bert' not in self.model_config.model_type.lower():
self.model = self.models[best_index]
else:
# copy best BERT model fold_number
best_model_dir = 'data/models/sequenceLabelling/' + self.model_config.model_name + str(
best_index)
new_model_dir = 'data/models/sequenceLabelling/' + self.model_config.model_name
# update new_model_dir if it already exists, keep its existing config content
merge_folders(best_model_dir, new_model_dir)
# clean other fold directory
for i in range(self.model_config.fold_number):
shutil.rmtree('data/models/sequenceLabelling/' +
self.model_config.model_name + str(i))
print(
"----------------------------------------------------------------------"
)
print("\nAverage over", self.model_config.fold_number, "folds")
print(
get_report(fold_average_evaluation,
digits=4,
include_avgs=['micro']))
def eval_nfold(self, x_test, y_test, features=None):
if self.models is not None:
total_f1 = 0
best_f1 = 0
best_index = 0
worst_f1 = 1
worst_index = 0
reports = []
reports_as_map = []
total_precision = 0
total_recall = 0
for i in range(self.model_config.fold_number):
if self.model_config.transformer_name is None:
the_model = self.models[i]
bert_preprocessor = None
else:
# the architecture model uses a transformer layer, it is large and needs to be loaded from disk
dir_path = 'data/models/sequenceLabelling/'
weight_file = DEFAULT_WEIGHT_FILE_NAME.replace(
".hdf5",
str(i) + ".hdf5")
self.model = get_model(self.model_config,
self.p,
ntags=len(self.p.vocab_tag),
load_pretrained_weights=False,
local_path=os.path.join(
dir_path,
self.model_config.model_name))
self.model.load(filepath=os.path.join(
dir_path, self.model_config.model_name, weight_file))
the_model = self.model
bert_preprocessor = self.model.transformer_preprocessor
if i == 0:
the_model.print_summary()
print_parameters(self.model_config, self.training_config)
print('\n------------------------ fold ' + str(i) +
' --------------------------------------')
# we can use a data generator for evaluation
# Prepare test data(steps, generator)
generator = the_model.get_generator()
test_generator = generator(
x_test,
y_test,
batch_size=self.model_config.batch_size,
preprocessor=self.p,
bert_preprocessor=bert_preprocessor,
char_embed_size=self.model_config.char_embedding_size,
max_sequence_length=self.model_config.max_sequence_length,
embeddings=self.embeddings,
shuffle=False,
features=features,
output_input_offsets=True,
use_chain_crf=self.model_config.use_chain_crf)
# Build the evaluator and evaluate the model
scorer = Scorer(test_generator,
self.p,
evaluation=True,
use_crf=self.model_config.use_crf,
use_chain_crf=self.model_config.use_chain_crf)
scorer.model = the_model
scorer.on_epoch_end(epoch=-1)
f1 = scorer.f1
precision = scorer.precision
recall = scorer.recall
reports.append(scorer.report)
reports_as_map.append(scorer.report_as_map)
if best_f1 < f1:
best_f1 = f1
best_index = i
if worst_f1 > f1:
worst_f1 = f1
worst_index = i
total_f1 += f1
total_precision += precision
total_recall += recall
fold_average_evaluation = {'labels': {}, 'micro': {}, 'macro': {}}
micro_f1 = total_f1 / self.model_config.fold_number
micro_precision = total_precision / self.model_config.fold_number
micro_recall = total_recall / self.model_config.fold_number
micro_eval_block = {
'f1': micro_f1,
'precision': micro_precision,
'recall': micro_recall
}
fold_average_evaluation['micro'] = micro_eval_block
# field-level average over the n folds
labels = []
for label in sorted(self.p.vocab_tag):
if label == 'O' or label == '<PAD>':
continue
if label.startswith("B-") or label.startswith(
"S-") or label.startswith("I-") or label.startswith(
"E-"):
label = label[2:]
if label in labels:
continue
labels.append(label)
sum_p = 0
sum_r = 0
sum_f1 = 0
sum_support = 0
for j in range(0, self.model_config.fold_number):
if label not in reports_as_map[j]['labels']:
continue
report_as_map = reports_as_map[j]['labels'][label]
sum_p += report_as_map["precision"]
sum_r += report_as_map["recall"]
sum_f1 += report_as_map["f1"]
sum_support += report_as_map["support"]
avg_p = sum_p / self.model_config.fold_number
avg_r = sum_r / self.model_config.fold_number
avg_f1 = sum_f1 / self.model_config.fold_number
avg_support = sum_support / self.model_config.fold_number
avg_support_dec = str(avg_support - int(avg_support))[1:]
if avg_support_dec != '0':
avg_support = math.floor(avg_support)
block_label = {
'precision': avg_p,
'recall': avg_r,
'support': avg_support,
'f1': avg_f1
}
fold_average_evaluation['labels'][label] = block_label
print(
"----------------------------------------------------------------------"
)
print("\n** Worst ** model scores - run", str(worst_index))
print(reports[worst_index])
print("\n** Best ** model scores - run", str(best_index))
print(reports[best_index])
fold_nb = self.model_config.fold_number
self.model_config.fold_number = 1
if self.model_config.transformer_name is None:
self.model = self.models[best_index]
else:
dir_path = 'data/models/sequenceLabelling/'
weight_file = DEFAULT_WEIGHT_FILE_NAME.replace(
".hdf5",
str(best_index) + ".hdf5")
# saved config file must be updated to single fold
self.model.load(filepath=os.path.join(
dir_path, self.model_config.model_name, weight_file))
print(
"----------------------------------------------------------------------"
)
print("\nAverage over", str(int(fold_nb)), "folds")
print(
get_report(fold_average_evaluation,
digits=4,
include_avgs=['micro']))
def train_nfold(self,
x_train,
y_train,
x_valid=None,
y_valid=None,
f_train=None,
f_valid=None,
callbacks=None):
"""
n-fold training for the instance model
for RNN models:
-> the n models are stored in self.models, and self.model left unset at this stage
fold number is available with self.model_config.fold_number
for models with transformer layer:
-> fold models are saved on disk (because too large) and self.models is not used, we identify the usage
of folds with self.model_config.fold_number
"""
fold_count = self.model_config.fold_number
fold_size = len(x_train) // fold_count
dir_path = 'data/models/sequenceLabelling/'
output_directory = os.path.join(dir_path, self.model_config.model_name)
print("Output directory:", output_directory)
if not os.path.exists(output_directory):
os.makedirs(output_directory)
if self.model_config.transformer_name is not None:
# save the config, preprocessor and transformer layer config on disk
self.model_config.save(
os.path.join(output_directory, CONFIG_FILE_NAME))
self.preprocessor.save(
os.path.join(output_directory, PROCESSOR_FILE_NAME))
for fold_id in range(0, fold_count):
if x_valid is None:
# segment train and valid
fold_start = fold_size * fold_id
fold_end = fold_start + fold_size
if fold_id == fold_size - 1:
fold_end = len(x_train)
train_x = np.concatenate(
[x_train[:fold_start], x_train[fold_end:]])
train_y = np.concatenate(
[y_train[:fold_start], y_train[fold_end:]])
train_f = np.concatenate(
[f_train[:fold_start], f_train[fold_end:]])
val_x = x_train[fold_start:fold_end]
val_y = y_train[fold_start:fold_end]
val_f = f_train[fold_start:fold_end]
else:
# reuse given segmentation
train_x = x_train
train_y = y_train
train_f = f_train
val_x = x_valid
val_y = y_valid
val_f = f_valid
foldModel = get_model(self.model_config,
self.preprocessor,
ntags=len(self.preprocessor.vocab_tag),
load_pretrained_weights=True)
if fold_id == 0:
print_parameters(self.model_config, self.training_config)
foldModel.print_summary()
print('\n------------------------ fold ' + str(fold_id) +
'--------------------------------------')
self.transformer_preprocessor = foldModel.transformer_preprocessor
foldModel = self.compile_model(foldModel, len(train_x))
foldModel = self.train_model(
foldModel,
train_x,
train_y,
x_valid=val_x,
y_valid=val_y,
f_train=train_f,
f_valid=val_f,
max_epoch=self.training_config.max_epoch,
callbacks=callbacks)
if self.model_config.transformer_name is None:
self.models.append(foldModel)
else:
# save the model with transformer layer on disk
weight_file = DEFAULT_WEIGHT_FILE_NAME.replace(
".hdf5",
str(fold_id) + ".hdf5")
foldModel.save(os.path.join(output_directory, weight_file))
if fold_id == 0:
foldModel.transformer_config.to_json_file(
os.path.join(output_directory,
TRANSFORMER_CONFIG_FILE_NAME))
if self.model_config.transformer_name is not None:
transformer_preprocessor = foldModel.transformer_preprocessor
transformer_preprocessor.tokenizer.save_pretrained(
os.path.join(output_directory,
DEFAULT_TRANSFORMER_TOKENIZER_DIR))