run color_syncnet_train.py have a error about wav2lip_288x288

[hupu1dong]

I found that this training can be run normally without using LeakyReLU, and I believe it is caused by inconsistent data shapes during computation loss caused by LeakyReLU

[hupu1dong]

Have you verified the author before uploading the code @primepake

[carter54]

It's due to PReLU, PReLU could have negative output to audio vector and face vector, and the cosine_similarity will have negative value.
@primepake Could you share your method to solve this problem?
I saw you mentioned that using BCELogicLoss would have other problem, I think the training loss will stay near 0.6...
Did you still use cosine_loss in color_syncnet_train.py or change to other loss function?

Thanks in advance!

[DogeFlow]

For the same question, the author mentioned the use of BCELogicLoss, but in the subsequent answer, he suggested not to use it. If you do not use BCELogicLoss, you will not be able to train. So what is the correct way to deal with it? @primepake

I recommend you shouldn't you BCELogicLoss, this not correct. You should apply ReLU in the last layer to get the value in range [0,1]

[DogeFlow]

I recommend you shouldn't you BCELogicLoss, this not correct. You should apply ReLU in the last layer to get the value in range [0,1]

Thank you very much, it worked

[xiedongmingming]

I recommend you shouldn't you BCELogicLoss, this not correct. You should apply ReLU in the last layer to get the value in range [0,1]

Thank you very much, it worked

Could you share your method to solve this problem?