Issue with evaluating during training using weighted layer sum

#10
by kimedaka - opened

Hello @sanchit-gandhi ,

When config.use_weighted_layer_sum = True, eval_pred.predictions in compute metric method in Trainer object returns two dimensional tuple/array instead of one dimensional at evaluation step during training. Consequently, it throws the following error:

setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (2,) + inhomogeneous part.

Also, it takes more than 40 GB or GPU RAM at evaluation step during training with Whisper base model when config.use_weighted_layer_sum = True.
Is it a bug or am I mistaking something.

Regards,
K

Sign up or log in to comment