search

chrisdonahue / wavegan

WaveGAN: Learn to synthesize raw audio with generative adversarial networks
MIT License
1.33k stars 281 forks source link

errata #3

Closed physhik closed 6 years ago

physhik commented 6 years ago

Hi Chris,

I have checked the beautiful generation of piano sound using wavegan model based on your tfrecord piano. BTW, in the code of colab, you miswrote the name of the tensor. 'G_z_spec:0' should be modified into 'global_step:0'.

chrisdonahue commented 6 years ago

What do you mean? G_z_spec:0 is the tensor that computes the spectrogram of the generated audio. Why do you think it should be global_step:0?

chrisdonahue commented 6 years ago

Closing this for now. Feel free to reopen and clarify the issue

physhik commented 6 years ago

Hi sorry for late comment. Was a bit busy. I could not find "G_z_spec:0" in any file. The tensorial multiplication is not matching. I also tried to test tfrecord file from custom data and your uploaded voice wav files but it did not work. Are you sure if you git push all the codes?

chrisdonahue commented 6 years ago

Hi there. I see the confusion now! Yes, I do not have the computational graph to produce the spectrogram in the code in this repository. You can get it from the metagraphs on the colab and then load weights generated by this repository.

polisen commented 5 years ago

Hey, I too encountered this problem and I'm not sure how to solve it. I understand that this is one of those "go figure it out" situations but I've tried and would love some general pointers on how to resolve this. The error I get is G_z_spec:0 cannot be found.

chrisdonahue commented 5 years ago

The G_z_spec:0 tensor is used to create a preview spectrogram for visualization purposes. The reason it is not found in the default metagraph from WaveGAN is that it requires parameters representing the average spectrogram from a particular dataset, which the WaveGAN script does not know by default (I manually added them to create the colab). You shouldn't need this tensor unless you really want to look at spectrograms for your generated waveforms. If you just want to hear the generated waveforms, use G_z:0 instead.

polisen commented 5 years ago

Thanks - I tried it and of course it worked.