search

JayYip / m3tl

BERT for Multitask Learning
https://jayyip.github.io/m3tl/
Apache License 2.0
545 stars 125 forks source link

bug of label set in classification task #34

Closed chiyuzhang94 closed 4 years ago

chiyuzhang94 commented 4 years ago

I find a bug of generating the label set of classification task.

The label set always has one more label: [PAD]. I find in utils.fit() (line 33 and 35), the [PAD] will be add to label set. Could you please check this issue?

https://github.com/JayYip/bert-multitask-learning/blob/9fe97739194f801e539efbadbaaf97a9c945eaaa/bert_multitask_learning/utils.py#L33

JayYip commented 4 years ago

Hi, you're right and this is intentional. The reason behind this is that I want to make sure the label of [PAD] is 0 in the sequence labeling scenario so that various metric calculation of t2t could work. Is it causing problems?

chiyuzhang94 commented 4 years ago

@JayYip Thanks for your reply. I am thinking whether this affects the model performance some how.

JayYip commented 4 years ago

I don't think it will. Actually, in daily work, sometimes we cannot set a clear number of classes(I know it's weird, but that's business). We will set a large number and it's proved that the performance is quite similar. I think the situation is somewhat similar here.

It actually make sense because classes are independent in the classification layer and the redundant classes will not create any loss.

chiyuzhang94 commented 4 years ago

Sure. Thanks for your explanations. I will change this personally.