ner_model.to(device)
ner_model.eval()
output_list = list()
fout = open(args.output_text, 'w')
iterator = data_loader.get_tqdm(device)
max_score = -float('inf')
min_score = float('inf')
for word_t, char_t, chunk_mask, chunk_index, chunk_surface in iterator:
output = ner_model(word_t, char_t, chunk_mask)
chunk_score = ner_model.chunking(output)
tmp_min = utils.to_scalar(chunk_score.min())
tmp_max = utils.to_scalar(chunk_score.max())
max_score = max(max_score, tmp_max)
min_score = min(min_score, tmp_min)
pred_chunk = (chunk_score < args.threshold)
chunk_index = chunk_index.masked_select(pred_chunk).data.cpu()
output = ner_model.typing(output, pred_chunk)
output = output.data.cpu()
offset = chunk_index[0]
for ind in range(0, output.size(0)):
st, ed = chunk_index[ind].item(), chunk_index[ind + 1].item()
surface = ' '.join(chunk_surface[st - offset:ed - offset])
chunk_score = ner_model.chunking(output)
chunk_loss = crit_chunk(chunk_score, chunk_label)
type_score = ner_model.typing(output, type_mask)
type_loss = crit_type(type_score, type_label)
loss = type_loss + chunk_loss
loss.backward()
torch.nn.utils.clip_grad_norm_(ner_model.parameters(), args.clip)
optimizer.step()
batch_index += 1
if 0 == batch_index % args.interval:
pw.add_loss_vs_batch({'loss_chunk': utils.to_scalar(chunk_loss), 'loss_type': utils.to_scalar(type_loss)}, batch_index, use_logger = False)
if 0 == batch_index % args.check:
# NER evaluation
pre_dev, rec_dev, f1_dev, type2pre_dev, type2rec_dev, type2f1_dev = utils.evaluate_ner(dev_loader.get_tqdm(device), ner_model, tl_map['None'], id2label)
pw.add_loss_vs_batch({'dev_pre': pre_dev, 'dev_rec': rec_dev}, batch_index, use_logger = False)
pw.add_loss_vs_batch({'dev_f1': f1_dev}, batch_index, use_logger = True)
pw.save_checkpoint(model = ner_model, is_best = f1_dev > best_eval, s_dict = {'config': ner_config, 'w_map': w_map, 'c_map': c_map, 'tl_map': tl_map, 'cl_map': cl_map})
if f1_dev > best_eval:
best_eval = f1_dev
# best_f1, best_pre, best_rec, best_type2pre, best_type2rec, best_type2f1 = utils.evaluate_ner(test_loader.get_tqdm(device), ner_model, tl_map['None'])
best_pre, best_rec, best_f1, best_type2pre, best_type2rec, best_type2f1 = utils.evaluate_ner(test_loader.get_tqdm(device), ner_model, tl_map['None'], id2label)
type_score = ner_model.typing(output, type_mask)
type_loss = crit_type(type_score, type_label)
loss = type_loss + chunk_loss
loss.backward()
torch.nn.utils.clip_grad_norm_(ner_model.parameters(),
args.clip)
optimizer.step()
batch_index += 1
if 0 == batch_index % args.interval:
pw.add_loss_vs_batch(
{
'loss_chunk': utils.to_scalar(chunk_loss),
'loss_type': utils.to_scalar(type_loss)
},
batch_index,
use_logger=False)
if 0 == batch_index % args.check:
# NER evaluation
pre_dev, rec_dev, f1_dev, type2pre_dev, type2rec_dev, type2f1_dev = utils.evaluate_ner(
dev_loader.get_tqdm(device), ner_model, tl_map['None'],
id2label)
pw.add_loss_vs_batch(
{
'dev_pre': pre_dev,
'dev_rec': rec_dev