for whisper_large: generation_max_length not respected inside evaluation step of training loop
#3
by sitandon - opened
I have set generation_max_length as 225 but the input of compute_metrics function has shape [batch_size, 448] (both for predictions and labels). I know whisper's max_length is 448 but why generation_max_length is not respected. As a workaround I have to explicitly truncate the response before passing to batch_decode otherwise my memory requirements (RAM) shoot up beyond 100 GB
sitandon changed discussion title from generation_max_length not respected inside evaluation step of training loop to for whisper_large: generation_max_length not respected inside evaluation step of training loop