def predict(model, test_dataloader, params, mode):
"""
预测并将结果输出至文件
Args:
mode (str): 'val' or 'test'
"""
model.eval()
# init
pre_result = pd.DataFrame()
# idx to label
cate_idx2label = {idx: value for idx, value in enumerate(params.tag_list)}
# get data
for batch in tqdm(test_dataloader, unit='Batch', ascii=True):
# to device
batch = tuple(t.to(params.device) for t in batch)
input_ids, input_mask, segment_ids, _, _, ne_cate, split_to_ori, example_ids = batch
# inference
with torch.no_grad():
start_pre, end_pre = model(input_ids, segment_ids, input_mask)
# predict label
start_label = start_pre.detach().cpu().numpy().tolist()
end_label = end_pre.detach().cpu().numpy().tolist()
# mask
input_mask = input_mask.to("cpu").detach().numpy().tolist()
ne_cate = ne_cate.to("cpu").numpy().tolist()
split_to_ori = split_to_ori.to('cpu').numpy().tolist() # (bs, max_len)
example_ids = example_ids.to('cpu').numpy().tolist() # (bs,)
# get result
for start_p, end_p, input_mask_s, ne_cate_s, s_t_o, example_id in zip(
start_label, end_label, input_mask, ne_cate, split_to_ori,
example_ids):
ne_cate_str = cate_idx2label[ne_cate_s]
# 问题长度
q_len = len(EN2QUERY[ne_cate_str])
# 有效长度
act_len = sum(input_mask_s[q_len + 2:-1])
# 转换为BIO标注
pre_bio_labels = pointer2bio(start_p[q_len + 2:q_len + 2 +
act_len],
end_p[q_len + 2:q_len + 2 + act_len],
en_cate=ne_cate_str)
# append to df
pre_result = pre_result.append(
{
'example_id': int(example_id),
'tags': pre_bio_labels,
'split_to_ori': s_t_o[q_len + 2:q_len + 2 + act_len]
},
ignore_index=True)
pre_result.to_csv(path_or_buf=params.params_path / f'{mode}_tags_pre.csv',
encoding='utf-8',
index=False)
def predict(model, test_dataloader, params, mode):
"""预测并将结果输出至文件
:param mode: 'val' or 'test'
"""
model.eval()
# init
pre_result = []
cate_result = []
mask_lst = []
# idx to label
cate_idx2label = {
idx: value
for idx, value in enumerate(params.label_list)
}
# get data
for batch in tqdm(test_dataloader, unit='Batch'):
# to device
batch = tuple(t.to(params.device) for t in batch)
input_ids, input_mask, segment_ids, _, _, ne_cate = batch
# inference
with torch.no_grad():
start_logits, end_logits = model(input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask)
# predict label
start_label = start_logits.detach().cpu().numpy().tolist()
end_label = end_logits.detach().cpu().numpy().tolist()
# mask
input_mask = input_mask.to("cpu").detach().numpy().tolist()
ne_cate = ne_cate.to("cpu").numpy().tolist()
# get result
for start_p, end_p, ne_cate_s in zip(start_label, end_label, ne_cate):
ne_cate_str = cate_idx2label[ne_cate_s]
pre_bio_labels = pointer2bio(start_p, end_p, ne_cate=ne_cate_str)
pre_result.append(pre_bio_labels)
cate_result.append(ne_cate_str)
# save mask
mask_lst += input_mask
# write to file
with open(params.data_dir / f'{mode}_tags_pre.txt', 'w',
encoding='utf-8') as file_tags:
for cate, tag, mask in zip(cate_result, pre_result, mask_lst):
# 问题长度
q_len = len(IO2QUERY[cate])
# 有效长度
act_len = sum(mask[q_len + 2:-1])
# 真实标签
file_tags.write('{}\n'.format(' '.join(tag[q_len + 2:q_len + 2 +
act_len])))
def predict(model, test_dataloader, params, mode):
"""预测并将结果输出至文件
:param mode: 'val' or 'test'
"""
model.eval()
# init
pre_result = []
# idx to label
cate_idx2label = {
idx: value
for idx, value in enumerate(params.label_list)
}
# get data
for input_ids, input_mask, segment_ids, start_pos, end_pos, ner_cate in test_dataloader:
# to device
input_ids = input_ids.to(params.device)
input_mask = input_mask.to(params.device)
segment_ids = segment_ids.to(params.device)
# inference
with torch.no_grad():
start_logits, end_logits = model(input_ids, segment_ids,
input_mask)
# predict label
start_label = start_logits.detach().cpu().numpy().tolist()
end_label = end_logits.detach().cpu().numpy().tolist()
# mask
input_mask = input_mask.to("cpu").detach().numpy().tolist()
ner_cate = ner_cate.to("cpu").numpy().tolist()
# get result
for start_p, end_p, input_mask_s, ner_cate_s in zip(
start_label, end_label, input_mask, ner_cate):
ner_cate_str = cate_idx2label[ner_cate_s]
# 问题长度
q_len = len(IO2QUERY[ner_cate_str])
# 有效长度
act_len = sum(input_mask_s[q_len + 2:-1])
# 转换为BIO标注
pre_bio_labels = pointer2bio(start_p[q_len + 2:q_len + 2 +
act_len],
end_p[q_len + 2:q_len + 2 + act_len],
ne_cate=ner_cate_str)
pre_result.append(pre_bio_labels)
# write to file
with open(params.data_dir / f'{mode}_tags_pre.txt', 'w',
encoding='utf-8') as file_tags:
for tag in pre_result:
# 真实标签
file_tags.write('{}\n'.format(' '.join(tag)))
def predict(model, test_dataloader, params, mode):
"""预测并将结果输出至文件
:param mode: 'val' or 'test'
"""
model.eval()
# init
pre_result = []
mask_lst = []
# idx to label
cate_idx2label = {
idx: int(idx + 1)
for idx, _ in enumerate(params.label_list)
}
# get data
for batch in test_dataloader:
# to device
batch = tuple(t.to(params.device) for t in batch)
input_ids, input_mask, start_pos, end_pos = batch
# inference
with torch.no_grad():
start_logits, end_logits = model(input_ids,
attention_mask=input_mask)
# predict label
start_label = start_logits.detach().cpu().numpy().transpose(
(0, 2, 1)).tolist()
end_label = end_logits.detach().cpu().numpy().transpose(
(0, 2, 1)).tolist()
# mask
input_mask = input_mask.to("cpu").detach().numpy().tolist()
# get result
for start_p_s, end_p_s, input_mask_s in zip(start_label, end_label,
input_mask):
# 有效长度
act_len = sum(input_mask_s)
for idx, (start_p, end_p) in enumerate(zip(start_p_s, end_p_s)):
pre_bio_labels = pointer2bio(start_p[:act_len],
end_p[:act_len],
ne_cate=cate_idx2label[idx])
pre_result.append(pre_bio_labels)
# write to file
with open(params.data_dir / f'{mode}_tags_pre.txt', 'w',
encoding='utf-8') as file_tags:
for idx, tag in enumerate(pre_result):
# 真实标签
file_tags.write('{}\n'.format(' '.join(tag)))
def evaluate(args, model, eval_dataloader, params):
model.eval()
# 记录平均损失
loss_avg = utils.RunningAverage()
# init
pre_result = []
gold_result = []
# get data
for batch in tqdm(eval_dataloader, unit='Batch'):
# to device
batch = tuple(t.to(params.device) for t in batch)
input_ids, input_mask, segment_ids, start_pos, end_pos, ne_cate = batch
with torch.no_grad():
# get loss
loss = model(input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask,
start_positions=start_pos,
end_positions=end_pos)
if params.n_gpu > 1 and args.multi_gpu:
loss = loss.mean() # mean() to average on multi-gpu.
# update the average loss
loss_avg.update(loss.item())
# inference
start_logits, end_logits = model(input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask)
# gold label
start_pos = start_pos.to("cpu").numpy().tolist()
end_pos = end_pos.to("cpu").numpy().tolist()
input_mask = input_mask.to('cpu').numpy().tolist()
ne_cate = ne_cate.to("cpu").numpy().tolist()
# predict label
start_label = start_logits.detach().cpu().numpy().tolist()
end_label = end_logits.detach().cpu().numpy().tolist()
# idx to label
cate_idx2label = {
idx: value
for idx, value in enumerate(params.label_list)
}
# get bio result
for start_p, end_p, start_g, end_g, input_mask_s, ne_cate_s in zip(
start_label, end_label, start_pos, end_pos, input_mask,
ne_cate):
ne_cate_str = cate_idx2label[ne_cate_s]
# 问题长度
q_len = len(IO2QUERY[ne_cate_str])
# 有效长度
act_len = sum(input_mask_s[q_len + 2:-1])
# get BIO labels
pre_bio_labels = pointer2bio(start_p[q_len + 2:q_len + 2 +
act_len],
end_p[q_len + 2:q_len + 2 + act_len],
ne_cate=ne_cate_str)
gold_bio_labels = pointer2bio(start_g[q_len + 2:q_len + 2 +
act_len],
end_g[q_len + 2:q_len + 2 + act_len],
ne_cate=ne_cate_str)
pre_result.append(pre_bio_labels)
gold_result.append(gold_bio_labels)
# metrics
f1 = f1_score(y_true=gold_result, y_pred=pre_result)
acc = accuracy_score(y_true=gold_result, y_pred=pre_result)
# f1, acc
metrics = {'loss': loss_avg(), 'f1': f1, 'acc': acc}
metrics_str = "; ".join("{}: {:05.2f}".format(k, v)
for k, v in metrics.items())
logging.info("- {} metrics: ".format('Val') + metrics_str)
# f1 classification report
report = classification_report(y_true=gold_result, y_pred=pre_result)
logging.info(report)
return metrics