def load(self, dir_path):
# load options from the json file
self.options = opts.load(dir_path)
# load vocabularies for each field
fields.load_vocabs(dir_path, self.fields_tuples)
# set the current gpu
self.options.gpu_id = self.gpu_id
# load model, optimizer and scheduler
self.model = models.load(dir_path, self.fields_tuples, self.gpu_id)
self.optimizer = optimizer.load(dir_path, self.model.parameters())
self.scheduler = scheduler.load(dir_path, self.optimizer)
# now we have a loaded tagger
self._loaded = True
def detect(self, text=None, test_path=None):
self.options.text = text
self.options.test_path = test_path
words_field = fields.WordsField()
tags_field = fields.TagsField()
fields_tuples = [('words', words_field), ('tags', tags_field)]
dataset_iter = None
save_dir_path = None
if self.options.test_path is None and self.options.text is None:
raise Exception('You should inform a path to test data or a text.')
if self.options.test_path is not None and self.options.text is not None:
raise Exception(
'You cant inform both a path to test data and a text.')
if self.options.test_path is not None and self.options.text is None:
logger.info('Building test dataset: {}'.format(
self.options.test_path))
test_tuples = list(filter(lambda x: x[0] != 'tags', fields_tuples))
test_dataset = dataset.build(self.options.test_path, test_tuples,
self.options)
logger.info('Building test iterator...')
dataset_iter = iterator.build(test_dataset,
self.options.gpu_id,
self.options.dev_batch_size,
is_train=False)
save_dir_path = self.options.test_path
if self.options.text is not None and self.options.test_path is None:
logger.info('Preparing text...')
test_tuples = list(filter(lambda x: x[0] != 'tags', fields_tuples))
test_dataset = dataset.build_texts(self.options.text, test_tuples,
self.options)
logger.info('Building iterator...')
dataset_iter = iterator.build(test_dataset,
self.options.gpu_id,
self.options.dev_batch_size,
is_train=False)
save_dir_path = None
logger.info('Loading vocabularies...')
fields.load_vocabs(self.options.load, fields_tuples)
logger.info('Loading model...')
model = models.load(self.options.load, fields_tuples,
self.options.gpu_id)
logger.info('Predicting...')
predicter = Predicter(dataset_iter, model)
predictions = predicter.predict(self.options.prediction_type)
logger.info('Preparing to save...')
if self.options.prediction_type == 'classes':
prediction_tags = transform_classes_to_tags(
tags_field, predictions)
predictions_str = transform_predictions_to_text(prediction_tags)
else:
predictions_str = transform_predictions_to_text(predictions)
words_labels = None
if self.options.text is not None:
orig_words = self.options.text.split()
labels = predictions_str.split()
print(orig_words, labels)
words_labels = join_words_and_labels(orig_words, labels)
return predictions, predictions_str, words_labels
def run(options):
words_field = fields.WordsField()
tags_field = fields.TagsField()
fields_tuples = [('words', words_field), ('tags', tags_field)]
# fields_tuples += features.load(options.load)
if options.test_path is None and options.text is None:
raise Exception('You should inform a path to test data or a text.')
if options.test_path is not None and options.text is not None:
raise Exception('You cant inform both a path to test data and a text.')
dataset_iter = None
save_dir_path = None
if options.test_path is not None and options.text is None:
logger.info('Building test dataset: {}'.format(options.test_path))
test_tuples = list(filter(lambda x: x[0] != 'tags', fields_tuples))
test_dataset = dataset.build(options.test_path, test_tuples, options)
logger.info('Building test iterator...')
dataset_iter = iterator.build(test_dataset,
options.gpu_id,
options.dev_batch_size,
is_train=False)
save_dir_path = options.test_path
if options.text is not None and options.test_path is None:
logger.info('Preparing text...')
test_tuples = list(filter(lambda x: x[0] != 'tags', fields_tuples))
test_dataset = dataset.build_texts(options.text, test_tuples, options)
logger.info('Building iterator...')
dataset_iter = iterator.build(test_dataset,
options.gpu_id,
options.dev_batch_size,
is_train=False)
save_dir_path = None
logger.info('Loading vocabularies...')
fields.load_vocabs(options.load, fields_tuples)
logger.info('Loading model...')
model = models.load(options.load, fields_tuples, options.gpu_id)
logger.info('Predicting...')
predicter = Predicter(dataset_iter, model)
predictions = predicter.predict(options.prediction_type)
logger.info('Preparing to save...')
if options.prediction_type == 'classes':
prediction_tags = transform_classes_to_tags(tags_field, predictions)
predictions_str = transform_predictions_to_text(prediction_tags)
else:
predictions_str = transform_predictions_to_text(predictions)
if options.test_path is not None:
save_predictions(
options.output_dir,
predictions_str,
save_dir_path=save_dir_path,
)
else:
logger.info(options.text)
logger.info(predictions_str)
return predictions
def run(options):
logger.info('Running with options: {}'.format(options))
words_field = fields.WordsField()
tags_field = fields.TagsField()
fields_tuples = [('words', words_field), ('tags', tags_field)]
logger.info('Building train corpus: {}'.format(options.train_path))
train_dataset = dataset.build(options.train_path, fields_tuples, options)
logger.info('Building train iterator...')
train_iter = iterator.build(train_dataset,
options.gpu_id,
options.train_batch_size,
is_train=True)
dev_dataset = None
dev_iter = None
if options.dev_path is not None:
logger.info('Building dev dataset: {}'.format(options.dev_path))
dev_dataset = dataset.build(options.dev_path, fields_tuples, options)
logger.info('Building dev iterator...')
dev_iter = iterator.build(dev_dataset,
options.gpu_id,
options.dev_batch_size,
is_train=False)
test_dataset = None
test_iter = None
if options.test_path is not None:
logger.info('Building test dataset: {}'.format(options.test_path))
test_dataset = dataset.build(options.test_path, fields_tuples, options)
logger.info('Building test iterator...')
test_iter = iterator.build(test_dataset,
options.gpu_id,
options.dev_batch_size,
is_train=False)
datasets = [train_dataset, dev_dataset, test_dataset]
datasets = list(filter(lambda x: x is not None, datasets))
# BUILD
if not options.load:
logger.info('Building vocabulary...')
fields.build_vocabs(fields_tuples, train_dataset, datasets, options)
loss_weights = None
if options.loss_weights == 'balanced':
loss_weights = train_dataset.get_loss_weights()
logger.info('Building model...')
model = models.build(options, fields_tuples, loss_weights)
logger.info('Building optimizer...')
optim = optimizer.build(options, model.parameters())
logger.info('Building scheduler...')
sched = scheduler.build(options, optim)
# OR LOAD
else:
logger.info('Loading vocabularies...')
fields.load_vocabs(options.load, fields_tuples)
logger.info('Loading model...')
model = models.load(options.load, fields_tuples, options.gpu_id)
logger.info('Loading optimizer...')
optim = optimizer.load(options.load, model.parameters())
logger.info('Loading scheduler...')
sched = scheduler.load(options.load, optim)
# STATS
logger.info('Word vocab size: {}'.format(len(words_field.vocab)))
logger.info('Tag vocab size: {}'.format(len(tags_field.vocab) - 1))
logger.info('Number of training examples: {}'.format(len(train_dataset)))
if dev_dataset:
logger.info('Number of dev examples: {}'.format(len(dev_dataset)))
if test_dataset:
logger.info('Number of test examples: {}'.format(len(test_dataset)))
logger.info('Model info: ')
logger.info(str(model))
logger.info('Optimizer info: ')
logger.info(str(optim))
logger.info('Scheduler info: ')
logger.info(str(sched))
nb_trainable_params = 0
for p_name, p_tensor in model.named_parameters():
if p_tensor.requires_grad:
if options.print_parameters_per_layer:
logger.info('{} {}: {}'.format(p_name, tuple(p_tensor.size()),
p_tensor.size().numel()))
nb_trainable_params += p_tensor.size().numel()
logger.info('Nb of trainable parameters: {}'.format(nb_trainable_params))
# TRAIN
logger.info('Building trainer...')
trainer = Trainer(train_iter,
model,
optim,
sched,
options,
dev_iter=dev_iter,
test_iter=test_iter)
if options.resume_epoch and options.load is None:
logger.info('Resuming training...')
trainer.resume(options.resume_epoch)
trainer.train()
# SAVE
if options.save:
logger.info('Saving path: {}'.format(options.save))
config_path = Path(options.save)
config_path.mkdir(parents=True, exist_ok=True)
logger.info('Saving config options...')
opts.save(config_path, options)
logger.info('Saving vocabularies...')
fields.save_vocabs(config_path, fields_tuples)
logger.info('Saving model...')
models.save(config_path, model)
logger.info('Saving optimizer...')
optimizer.save(config_path, optim)
logger.info('Saving scheduler...')
scheduler.save(config_path, sched)
return fields_tuples, model, optim, sched