parser.add_argument('--gpu-id',
default=None,
type=int,
help='Use CUDA on the listed devices')
args = parser.parse_args()
args.lazy_loading = True
args.embeddings_format = None
args.vocab_size = 99999999
args.vocab_min_frequency = 0
set_seed(42)
corpus_cls = available_corpora[args.corpus]
fields_tuples = corpus_cls.create_fields_tuples()
print('Building dataset...')
text_dataset = dataset.build(args.corpus_path, fields_tuples, args)
print('Building iterator...')
dataset_iter = iterator.build(text_dataset,
args.gpu_id,
args.batch_size,
is_train=False,
lazy=True)
print('Building vocabularies...')
fields.build_vocabs(fields_tuples, text_dataset, [text_dataset], args)
print('Calculating nonzeros...')
lens_src = []
lens_hyp = []
lens_trg = []
def run(options):
fields_tuples = available_corpora[options.corpus].create_fields_tuples()
# 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
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] != 'target', 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,
lazy=options.lazy_loading)
if options.text is not None and options.test_path is None:
logger.info('Preparing text...')
test_tuples = list(filter(lambda x: x[0] != 'target', 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,
lazy=options.lazy_loading)
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':
target_field = dict(fields_tuples)['target']
prediction_target = transform_classes_to_target(
target_field, predictions)
predictions_str = transform_predictions_to_text(prediction_target)
else:
predictions_str = transform_predictions_to_text(predictions)
if options.test_path is not None:
save_predictions(
options.output_dir,
predictions_str,
)
else:
logger.info(options.text)
logger.info(predictions_str)
return predictions
def run(options):
logger.info('Running with options: {}'.format(options))
fields_tuples = available_corpora[options.corpus].create_fields_tuples()
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,
lazy=options.lazy_loading)
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=True,
lazy=options.lazy_loading)
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=True,
lazy=options.lazy_loading)
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('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)))
for name, field in fields_tuples:
if field.use_vocab:
logger.info('{} vocab size: {}'.format(name, len(field.vocab)))
if name == 'target':
logger.info('target vocab: {}'.format(field.vocab.stoi))
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
def run(options):
logger.info('Running with options: {}'.format(options))
fields_tuples = available_corpora[options.corpus].create_fields_tuples()
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,
lazy=options.lazy_loading)
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,
lazy=options.lazy_loading)
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,
lazy=options.lazy_loading)
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)
# OR LOAD
else:
logger.info('Loading vocabularies...')
fields.load_vocabs(options.load, fields_tuples)
logger.info('Loading vectors...')
vectors = fields.load_vectors(options)
if vectors is not None:
train_dataset.fields['words'].vocab.load_vectors(vectors)
# STATS
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)))
for name, field in fields_tuples:
if field.use_vocab:
logger.info('{} vocab size: {}'.format(name, len(field.vocab)))
# BUILD COMMUNICATION
if not options.load_communication:
logger.info('Building explainer...')
explainer = explainers.build(options, fields_tuples, None)
logger.info('Building explainer optimizer...')
explainer_optim = optimizer.build(options, explainer.parameters())
logger.info('Building layman...')
msg_size = explainer.get_output_size()
layman = laymen.build(options, fields_tuples, msg_size, None)
logger.info('Building layman optimizer...')
layman_optim = optimizer.build(options, layman.parameters())
# OR LOAD COMMUNICATION
else:
logger.info('Loading explainer...')
explainer = explainers.load(options.load_communication, fields_tuples,
options.gpu_id)
logger.info('Loading explainer optimizer...')
explainer_optim = optimizer.load(
options.load_communication,
explainer.parameters(),
name=constants.EXPLAINER_OPTIMIZER,
config_name=constants.COMMUNICATION_CONFIG)
logger.info('Loading layman...')
msg_size = explainer.get_output_size()
layman = laymen.load(options.load_communication, fields_tuples,
msg_size, options.gpu_id)
logger.info('Loading layman optimizer...')
layman_optim = optimizer.load(
options.load_communication,
layman.parameters(),
name=constants.LAYMAN_OPTIMIZER,
config_name=constants.COMMUNICATION_CONFIG)
logger.info('Explainer info: ')
logger.info(str(explainer))
logger.info('Explainer optimizer info: ')
logger.info(str(explainer_optim))
logger.info('Layman info: ')
logger.info(str(layman))
logger.info('Layman optimizer info: ')
logger.info(str(layman_optim))
if options.freeze_explainer_params:
logger.info('Freezing explainer params...')
freeze_all_module_params(explainer)
# TRAIN
logger.info('Building trainer...')
communicator = CommunicatorTranslation(train_iter,
explainer,
layman,
explainer_optim,
layman_optim,
options,
dev_iter=dev_iter,
test_iter=test_iter)
# resume training from a checkpoint
if options.resume_epoch and options.load is None:
logger.info('Resuming communication...')
communicator.resume(options.resume_epoch)
# train the communication
communicator.train()
if options.save_explanations:
logger.info('Saving explanations to {}'.format(
options.save_explanations))
# save explanations (run with 0 epochs to ignore the communication)
ds_iterator = test_iter if test_iter is not None else dev_iter
communicator.save_explanations(
options.save_explanations,
ds_iterator,
max_explanations=options.max_explanations)
# SAVE
if options.save:
logger.info('Saving path: {}'.format(options.save))
config_path = Path(options.save)
config_path.mkdir(parents=True, exist_ok=True)
# save communication modules
logger.info('Saving communication config options...')
opts.save(config_path, options, name=constants.COMMUNICATION_CONFIG)
logger.info('Saving explainer...')
explainers.save(config_path, explainer)
logger.info('Saving layman...')
laymen.save(config_path, layman)
logger.info('Saving explainer optimizer...')
optimizer.save(config_path,
explainer_optim,
name=constants.EXPLAINER_OPTIMIZER)
logger.info('Saving layman optimizer...')
optimizer.save(config_path,
layman_optim,
name=constants.LAYMAN_OPTIMIZER)