Time warp fails on short spectrogram

[tann9949]

Hi, I have a question whether the time warping can sometimes cause a spectrogram to be too distorted or not.

I applied Time Warping with W=2 and it sometimes results in ugly padding on the right side of the spectrogram. Is there any way to alleviate this problem?

[shangeth]

Hey did you find what's wrong? I am using Time Warp with W=2(tried with W=1,...10) I am getting different features at each time.

run-1

run-2

[dtchantao]

Hey did you find what's wrong? I am using Time Warp with W=2(tried with W=1,...10) I am getting different features at each time.

run-1

run-2

same here, any solutions?

[PrathamSoni]

https://github.com/DemisEom/SpecAugment/blob/7f1435963b37ac8f9e4de9e44d754ecc41eaba85/SpecAugment/spec_augment_tensorflow.py#L46

It seems the implementation in this source code differs from the approach taken from in the above tensorflow implementation. This repo uses 2 points for fitting the transform, whereas the tf code uses multiple points

[zcaceres]

Great catch @PrathamSoni . Please feel free to PR. I'm afraid I have too much on my plate to fix this module.

[PrathamSoni]

@zcaceres Will open the PR within the next couple of days.

This line in the bilinear interpolation is also deprecated (copying tensor from construct). I can patch that as well. alpha = torch.tensor((queries - floor), dtype=grid_type, device=grid_device)

[keshawnhsieh]

Got same situation. This BUG seriously affects the model performance. Any updates? @PrathamSoni

[PrathamSoni]

@keshawnhsieh just put up the pr. Could you let me know if it works better?

[keshawnhsieh]

@keshawnhsieh just put up the pr. Could you let me know if it works better?

Many thanks. I will test it after Chinese New Year.

[zcaceres]

Fantastic @PrathamSoni . Thanks so much.

[PrathamSoni]

@thread especially @keshawnhsieh, I've made some fixes to the pull request. I think it should work perfectly now. LMK if you try it out!