Changing kernel size in wav2vec pretrained model

audio
1

I got an error that is almost the one solved here Kernel size can't greater than actual input size
the solution, as provided, is to lower the kernel size to 1 (to make it lower than the input size).

How can I do this with the pretrained wav2vec (WAV2VEC2_ASR_BASE_960H) model?

2

Changing the kernel in a pretrained model might not be the best idea as you would need to retrain the new kernel.
Instead of changing the model you could increase the spatial size of your inputs.

3

Thanks.
So, can retraining happen in certain layer(s) (at which we need to adjust the kernel size) without retraining the other layers? can you point out some references to look up this using PyTorch, if it’s possible to do so?

4

Yes, you can fine-tune specific layers only as described in e.g. this tutorial.