def main():
arguments = load_arguments()
config = load_config(arguments.config)
train_labeled_dataloader, train_unlabeled_dataloader, validation_dataloader = load_train_data(config)
model, ema_model, optimizer, ema_optimizer, semi_supervised, semi_supervised_loss = load(config)
metrics = {
'train_loss': 0,
'train_accuracy': 0,
'train_steps': 0,
'validation_accuracy': 0,
'validation_steps': 0
}
best_validation_accuracy = config.best_test_accuracy
for epoch_step in range(config.epoch_step + 1, config.epochs + 1):
train_labeled_dataloader_iterator = iter(train_labeled_dataloader)
train_unlabeled_dataloader_iterator = iter(train_unlabeled_dataloader) if train_unlabeled_dataloader else None
train_progress_bar = tqdm(range(config.iterations))
for batch_step in train_progress_bar:
inputs_x, targets_x, ub, train_labeled_dataloader_iterator, train_unlabeled_dataloader_iterator = on_train_batch_start(
train_labeled_dataloader, train_unlabeled_dataloader, train_labeled_dataloader_iterator,
train_unlabeled_dataloader_iterator, config)
train_step(epoch_step, batch_step, inputs_x, targets_x, ub,
semi_supervised, semi_supervised_loss, model,
ema_model, optimizer, ema_optimizer, metrics, config)
on_train_batch_end(epoch_step, inputs_x, targets_x, ema_model, metrics, train_progress_bar)
test_progress_bar = tqdm(enumerate(validation_dataloader), total=len(validation_dataloader))
for batch_step, batch in test_progress_bar:
validation_step(ema_model, batch, metrics, config)
on_validation_batch_end(epoch_step, metrics, test_progress_bar)
best_validation_accuracy = on_epoch_end(epoch_step, best_validation_accuracy, model, ema_model, optimizer,
config, metrics)
metrics = {
'train_loss': 0,
'train_accuracy': 0,
'train_steps': 0,
'validation_accuracy': 0,
'validation_steps': 0
}
# elimiate the first variable scope, and restore the classifier from the path
def restore_classifier_by_path(classifier, classifier_path, scope):
new_vars = {}
for var in classifier.params:
pos = var.name.find('/')
# eliminate the first variable scope, e.g., target, source
new_vars[var.name[pos + 1:-2]] = var
saver = tf.train.Saver(new_vars)
saver.restore(sess, os.path.join(classifier_path, 'model'))
logger.info("-----%s classifier model loading from %s successfully!-----" %
(scope, classifier_path))
if __name__ == '__main__':
args = load_arguments()
assert args.domain_adapt, "domain_adapt arg should be True."
if not os.path.isfile(args.multi_vocab):
build_unify_vocab([args.target_train_path, args.source_train_path],
args.multi_vocab)
multi_vocab = Vocabulary(args.multi_vocab)
logger.info('vocabulary size: %d' % multi_vocab.size)
# use tensorboard
if args.suffix:
tensorboard_dir = os.path.join(args.logDir, 'tensorboard', args.suffix)
else:
tensorboard_dir = os.path.join(args.logDir, 'tensorboard')
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
