Kiril Murschel

seems to be the same case even with Google Colab Pro and GPU activated? > Warning: multi_tensor_applier fused unscale kernel is unavailable, possibly because apex was installed without --cuda_ext --cpp_ext....

https://github.com/jupyter/notebook/issues/2527 its CMD+Click on macOS.

I do not see any explicit calculation or logging of a validation loss in the code, can you maybe share corresponding code? I am also looking into the 1D version....

in class Trainer, def train() you can for example import wandb, init and after pbar.set_description(f'loss: {total_loss:.4f}') # Log the total_loss to wandb wandb.log({"total_loss": total_loss}, step=self.step)