log('\n')
log('*' * 80)
eval_train_batch(epoch, batch_i, loss.data[0],
flat_argument, pred, argument2idx)
log('dev:')
srl_model.eval()
score, dev_output = eval_data(srl_model,
elmo,
dev_dataset,
batch_size,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2argument,
idx2word,
False,
dev_predicate_correct,
dev_predicate_sum,
isPretrain=True)
if dev_best_score is None or score[5] > dev_best_score[5]:
dev_best_score = score
output_predict(
os.path.join(
result_path, 'dev_argument_{:.2f}.pred'.format(
dev_best_score[2] * 100)), dev_output)
torch.save(srl_model.EN_Labeler.state_dict(),
def parallel_train_1_epoc(srl_model,
criterion,
optimizer,
train_dataset,
labeled_dataset_fr,
batch_size,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2word,
shuffle=False,
lang='En',
dev_best_score=None,
test_best_score=None,
test_ood_best_score=None):
unlabeled_dataset_en, unlabeled_dataset_fr = train_dataset
unlabeled_Generator_En = inter_utils.get_batch(unlabeled_dataset_en,
batch_size,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2word,
shuffle=False,
lang="En")
for batch_i, unlabeled_data_fr in enumerate(
inter_utils.get_batch(unlabeled_dataset_fr,
batch_size,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2word,
shuffle=False,
lang='Fr')):
srl_model.train()
unlabeled_data_en = unlabeled_Generator_En.next()
predicates_1D = unlabeled_data_en['predicates_idx']
predicates_1D_fr = unlabeled_data_fr['predicates_idx']
#log(predicates_1D, predicates_1D_fr)
u_loss = srl_model((unlabeled_data_en, unlabeled_data_fr),
lang='En',
unlabeled='True')
optimizer.zero_grad()
u_loss.backward()
optimizer.step()
if batch_i % 50 == 0:
log(batch_i, u_loss)
if batch_i % 500 == 0:
log('\n')
log('*' * 80)
srl_model.eval()
# eval_train_batch(epoch, batch_i, loss.data[0], flat_argument, pred, argument2idx)
log('FR test:')
score, dev_output = eval_data(srl_model,
elmo,
labeled_dataset_fr,
30,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2argument,
idx2word,
False,
dev_predicate_correct,
dev_predicate_sum,
lang='Fr')
if dev_best_score is None or score[5] > dev_best_score[5]:
dev_best_score = score
output_predict(
os.path.join(
result_path, 'dev_argument_{:.2f}.pred'.format(
dev_best_score[2] * 100)), dev_output)
# torch.save(srl_model, os.path.join(os.path.dirname(__file__),'model/best_{:.2f}.pkl'.format(dev_best_score[2]*100)))
log('\tdev best P:{:.2f} R:{:.2f} F1:{:.2f} NP:{:.2f} NR:{:.2f} NF1:{:.2f}'
.format(dev_best_score[0] * 100, dev_best_score[1] * 100,
dev_best_score[2] * 100, dev_best_score[3] * 100,
dev_best_score[4] * 100, dev_best_score[5] * 100))
return dev_best_score
def train_1_epoc(srl_model,
criterion,
optimizer,
train_dataset,
labeled_dataset_fr,
batch_size,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2word,
shuffle=False,
lang='En',
dev_best_score=None,
test_best_score=None,
test_ood_best_score=None):
for batch_i, train_input_data in enumerate(
inter_utils.get_batch(train_dataset,
batch_size,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2word,
shuffle=shuffle,
lang=lang)):
flat_argument = train_input_data['flat_argument']
target_batch_variable = get_torch_variable_from_np(flat_argument)
out, out_word = srl_model(train_input_data, lang='En')
loss = criterion(out, target_batch_variable)
loss_word = criterion(out_word, target_batch_variable)
if batch_i % 50 == 0:
log(batch_i, loss, loss_word)
optimizer.zero_grad()
(loss + loss_word).backward()
optimizer.step()
if batch_i > 0 and batch_i % show_steps == 0:
_, pred = torch.max(out, 1)
pred = get_data(pred)
# pred = pred.reshape([bs, sl])
log('\n')
log('*' * 80)
eval_train_batch(epoch, batch_i, loss.data[0], flat_argument, pred,
argument2idx)
log('FR test:')
score, dev_output = eval_data(srl_model,
elmo,
labeled_dataset_fr,
batch_size,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2argument,
idx2word,
False,
dev_predicate_correct,
dev_predicate_sum,
lang='Fr')
if dev_best_score is None or score[5] > dev_best_score[5]:
dev_best_score = score
output_predict(
os.path.join(
result_path, 'dev_argument_{:.2f}.pred'.format(
dev_best_score[2] * 100)), dev_output)
# torch.save(srl_model, os.path.join(os.path.dirname(__file__),'model/best_{:.2f}.pkl'.format(dev_best_score[2]*100)))
log('\tdev best P:{:.2f} R:{:.2f} F1:{:.2f} NP:{:.2f} NR:{:.2f} NF1:{:.2f}'
.format(dev_best_score[0] * 100, dev_best_score[1] * 100,
dev_best_score[2] * 100, dev_best_score[3] * 100,
dev_best_score[4] * 100, dev_best_score[5] * 100))
return dev_best_score
log('\n')
log('*' * 80)
eval_train_batch(epoch, batch_i, loss.data[0],
flat_argument, pred, argument2idx)
log('FR test:')
score, dev_output = eval_data(srl_model,
elmo,
labeled_dataset_fr,
batch_size,
word2idx,
fr_word2idx,
lemma2idx,
pos2idx,
pretrain2idx,
fr_pretrain2idx,
deprel2idx,
argument2idx,
idx2argument,
idx2word,
False,
dev_predicate_correct,
dev_predicate_sum,
lang='Fr')
log('En test:')
eval_data(srl_model,
elmo,
dev_dataset,
batch_size,
word2idx,
fr_word2idx,
_, pred = torch.max(out, 1)
pred = get_data(pred)
# pred = pred.reshape([bs, sl])
print('\n')
print('*' * 80)
eval_train_batch(epoch, batch_i, loss.data[0],
flat_argument, pred, argument2idx)
print('dev:')
score, dev_output = eval_data(
srl_model, elmo, dev_dataset, batch_size, word2idx,
lemma2idx, pos2idx, pretrain2idx, deprel2idx,
argument2idx, idx2argument, False,
dev_predicate_correct, dev_predicate_sum)
if dev_best_score is None or score[2] > dev_best_score[2]:
dev_best_score = score
output_predict(
os.path.join(
result_path, 'dev_argument_{:.2f}.pred'.format(
dev_best_score[2] * 100)), dev_output)
# torch.save(srl_model, os.path.join(os.path.dirname(__file__),'model/best_{:.2f}.pkl'.format(dev_best_score[2]*100)))
print(
'\tdev best P:{:.2f} R:{:.2f} F1:{:.2f} NP:{:.2f} NR:{:.2f} NF1:{:.2f}'
.format(dev_best_score[0] * 100,
dev_best_score[1] * 100,
dev_best_score[2] * 100,
dev_best_score[3] * 100,
srl_model.eval()
_, pred = torch.max(out, 1)
pred = get_data(pred)
print('\n')
print('*' * 80)
#eval_train_batch(epoch, batch_i, loss.item(), flat_argument, pred, argument2idx)
print('Zh test:')
score, dev_output = eval_data(srl_model,
elmo,
dev_dataset_fr,
batch_size,
argument2idx,
idx2argument,
False,
dev_predicate_correct,
dev_predicate_sum,
lang='Fr',
use_bert=use_bert)
print('En test:')
eval_data(srl_model,
elmo,
dev_dataset_en,
batch_size,
argument2idx,
idx2argument,
False,
dev_predicate_correct,
dev_predicate_sum,