Dabi Ahn
Dabi Ahn
voxceleb_norm is the processed dataset. The dataset is structured to directories for each celeb. Each directory contains each celeb's wav files which have sample rate 16,000 and format is 'wav'....
Keep the training at least a few days because voxceleb is huge. I kept training the model a few days using 8 gpu to get over 90% eval accuracy.
@chikiuso I used a server which has 8 Nvidia Tesla P40 GPU and 200GB memory. But single GPU core was enough.
@HudsonHuang Please check the paths(data_path or something) in hparam.py again. If you set the paths incorrectly, the problem you mentioned could happen.
@tbfly I think the queue runner is not working properly. To debug, set queue=False in train1.py and then run and see what message is up.
@HudsonHuang It occurs when the number of wavfiles is smaller than the batch size.
@roger865477 I also got 65~75% accuracy in train1. I found train2 worked quite well even when train1 accuracy of around 70%. But I'm still researching to improve the train1 accuracy....
It's possible and actually I was working for the same purpose. Check https://github.com/andabi/parallel-wavenet-vocoder
@niccanicci Please add some case name like below. python train1.py default
@coice I trained the model over 20K steps with Adam and learning rate 0.0005. It took more a day.