-
Hi, It should be very easy code/config change but I couldn't find the proper way to pass the top-k param from PTL(I see that Nemo is extending PTL ModelCheckpoint class a litle bit). So the question is - How to save more than 3 best checkpoints(and the last one)? I would like to validate my FastPitch model during finetuning because val_loss doesn't usually make sense in this case. Thanks a lot for your help folks! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
Hi, you may add save_top_k to your exp_manager like here: https://github.com/NVIDIA/NeMo/blob/main/examples/asr/conf/conformer/conformer_transducer_bpe.yaml#L238 Also you may set always_save_nemo=true so that it stores the checkpoints in nemo format too so that you can evaluate them easily. |
Beta Was this translation helpful? Give feedback.
Hi, you may add save_top_k to your exp_manager like here: https://github.com/NVIDIA/NeMo/blob/main/examples/asr/conf/conformer/conformer_transducer_bpe.yaml#L238
Also you may set always_save_nemo=true so that it stores the checkpoints in nemo format too so that you can evaluate them easily.