Nazukixv
commented
1 year ago Basically the code of subsampling module implementation was borrowed from espnet. It was used to subsample the 83-dimensional feature (80 fbank + 3 pitch) to a 256-dimensional tensor and it works fine in my previous ASR training tasks.
I'm not quite clear on why that problem occurred during your training process after adding the module. Therefore, I suggest trying out additional optimization settings to see if the problem still persists.
I tried to Implement Subsampling in my code, but just after adding the subsampling class to my TDNN model, the model stops learning (both training and testing accuracies doesn't change for all epochs ). Without sampling, The model is learning properly Attach is my code for your reference. Note : all other parameters are unchanged in both cases
Link of the code : Link