thunderboom / text_similarity Goto Github PK

原：
if labels is not None:
if i == 0:
loss = compute_loss(out, labels, loss_method=self.loss_method) / self.multi_drop
else:
loss += loss / self.multi_drop
改成这样？
if labels is not None:
loss += compute_loss(out, labels, loss_method=self.loss_method) / self.multi_drop

这里的train加上dev数据集不会使得 训练的时候看到dev的信息，导致准确度有问题吗？
elif pattern == 'full_train':
train_examples.extend(dev_examples)
model_example, _, _ = train_dev_test(
config=config,
model=model,
tokenizer=tokenizer,
train_data=train_examples,
train_enhancement=train_enhancement,
dev_data=dev_examples,
test_examples=None)
return model_example, None, None