Silence on all iterations

[luisarandas]

Hello.

I'm currently training wavegan-pytorch with a set of files (<300) and I'm getting silence files every iteration - it generates 10 .wav but it's silence, 0 values all duration. During the installation, I had some problems with librosa in my machine and changed line 172 on utils.py to:

    #librosa.output.write_wav(output_path, sample, sampling_rate)
    sf.write(output_path, sample, sampling_rate) #'PCM_24')

sf as of soundfile library. Might this be the problem? Is there a script here to navigate the latent space and debug this?

Thanks in advance.
Luis

[ALI7861111]

your solution might have felt into local optima . This has happened to me while training . But still i am trying to get the proper drum sound but getting very different sound after 150 epoch not even similar to drum just big booms

[luisarandas]

What does that mean? The problem is not writing files, the thing is that they are all silent. I do 200 epochs and it is always silent, except for the 0 folder which are an inaudible noise with very little amplitude.

[ALI7861111]

Introduce momentum in the optimizer that might help u to escape local minima

[luisarandas]

That's a good point. @ALI7861111 do you have any example/fork that I can look at/use?

[ALI7861111]

I did not apply the strategy to introduce momentum. It is a very logical way to overcome this problem. I was being lazy and i planned to changed the normal distribution of random initialization of the weights in utils.py that resulted in producing sounds approx similar to drums. i asked a very expert guy(professor) he recommended me to train for 1000 epochs currently my algo is still training i hope to get good sounds near to drums . I have to wait 46 hours

[ALI7861111]

u can avoid local minima by many ways i did the shortcut but i dont recommend u that

[jsphweid]

@luisarandas did you ever resolve this? This is happening to me as well.