train error

[SFGHRT]

Hello author, if you set --num_quantizers=1 when training vq, then at the beginning of the final res transformer training, you will get RuntimeError: Trying to create tensor with negative dimension -1: [-1, 8193, 32]? Is it a code error or you have written the wrong setting here? @exitudio

[SFGHRT]

Can the training script be the same as momask, and then train it, or should I follow the script in your experiment?

[exitudio]

I didn't clean the VQVQAE training code. Not sure if it has any errors. But the training for VQVAE is the same as MoMask. You can use their script.

[SFGHRT]

Thank you for your reply, but when I looked at your res_transformer, I noticed that your loss = '' #cal_loss(pred, labels, ignore_index, smoothing=smoothing) does not calculate the loss. Is that the part that is commented afterward?

[exitudio]

I calculated the cross-entropy loss averaged per sample, instead of per token as in MoMask. I implemented the loss function here Then, I commented out the loss in that function.