Auxiliary loss is counted twice in the loss

microsoft / protein-frame-flow

Fast protein backbone generation with SE(3) flow matching.

MIT License

200 stars 13 forks source link

Auxiliary loss is counted twice in the loss #23

Closed SizhuangHe closed 1 month ago

SizhuangHe commented 2 months ago

Hi, In model_step, we have se3_vf_loss += auxiliary_loss. Yet in training_step, we also have `train_loss = ( total_losses[self._exp_cfg.training.loss]

total_losses['auxiliary_loss'] ) The default training loss isse3_vf_loss` so in this case, auxiliary loss is being counted twice in the loss. Is this expected or a bug?

Thanks!

chaitjo commented 2 months ago

I think it may be a bug. We removed and trained and attained very similar results.

jasonkyuyim commented 1 month ago

This has been fixed.