search

mostafaelaraby / wavegan-pytorch

Pytorch Implementation of wavegan model to generate audio
https://arxiv.org/abs/1802.04208
Apache License 2.0
162 stars 43 forks source link

Are these codes available on sc09 dataset? #2

Open north-redwing opened 3 years ago

north-redwing commented 3 years ago

Dear mostafaelaraby,

Hi, thank you for your pytorch implementation.

I've done your codes on sc09 speech dataset with some hyperparameters, but I couldn't get any plausible sound results (no sound).

Are these codes available on sc09 dataset? If you have succeeded on sc09, please tell me the hyperparameters.

Regards,

fjiang9 commented 3 years ago

@north-redwing Hi, I couldn't get plausible sound results either. Have you figured it out? Thank you in advance! Best, Fei

north-redwing commented 3 years ago

@fjiang9 Hi, Sorry for late replay but I have not trained on sc09 dataset anymore. Maybe WaveGAN is somewhat unstable as is usual with other GANs, I think. However, on other datasets I could get plausible sounds with decreasing lr_g to 1e-5, so please try with the params as possible. I suppose generator has fell into a local optimum solution, that is silence...?

Anyway, if you succeed or figure it out why we fail in training on sc09, please tell me!

yuji

north-redwing commented 3 years ago

@fjiang9 I have done with lr_g=1e-5 on sc09 and I counld get no silent but very noisy sounds. So maybe with better parameters we will get plausible sounds.

mostafaelaraby commented 2 years ago

@north-redwing hope you managed to make it work, Gans in general are sensitive to hyper-parameters (specifically lr and batch size ) and Gans that generates wavelets are even harder to tune the model would easily reach local minima and collapse causing silent waves .

Rongjiehuang commented 2 years ago

@mostafaelaraby @fjiang9 @north-redwing. Hi, I have generated silence waves in SC09 datasets, have you succeeded in getting plausible sounds?