search

krasserm / fairseq-image-captioning

Transformer-based image captioning extension for pytorch/fairseq
Apache License 2.0
312 stars 55 forks source link

teacher forcing ratio #24

Closed goodluck110706112 closed 3 years ago

goodluck110706112 commented 3 years ago

Thank you for your high quality code! Is there any method to achieve teacher forcing？ Teacher forcing means in the decoding phase, we use the ground truth or prediction as prev_output_token according to probability. Thank you!

krasserm commented 3 years ago

Teacher forcing is used during training, why do you think it is not?

goodluck110706112 commented 3 years ago

Thank you for your reply! Yes, teacher forcing is used during training. But I think the fairseq's teacher forcing ratio always equals to one, which means model always use ground truth as prev_output_tokens, can I set teacher forcing ratio as a hyperparameter? Such as teacher forcing ratio=0.5 or teacher forcing ratio = 0.7?

krasserm commented 3 years ago

Did you already ask that question at the fairseq project directly? fairseq-image-captioning doesn't add anything special to make that possible.

goodluck110706112 commented 3 years ago

I didn't ask this question at the fairseq before. Yes, this is a question about fairseq, not fairseq-image-captioning. I'm sorry for ask the question in a wrong place. But if we consider different teacher forcing ratio, we may have a better score in image-caption.

krasserm commented 3 years ago

Yes, would be interesting too see. Let us know if you have any updates.