tensorflow-wavenet
tensorflow-wavenet copied to clipboard
ibab
Pre Trained Model
Does anyone have a pre trained model they could share? Specifically one trained on music (piano) data?
This is a model trained with 72000+ steps with original configuration using VCTK corpus. Still far from fully trained, but could save sometime. only changed SILENCE_TRESHOLD = 0.1 other than original 0.3 2017-12-04T13-48-11.zip
copy the model to the path "logdir/train/" and run it by type
python3 train.py --data_dir=corpus --gc_channels=32 --restore_from logdir/train/2017-12-04T13-48-11
@DiyuanLu what was the loss you were getting for this model? Just curious - I'm around 40k steps and its between 1.7 and 2.5. I think I need to adjust some parameters.
@DiyuanLu How is your result of the final trained model? Does SILENCE_TRESHOLD = 0.1 work better than the original 0.3?
Hi, for me it worked better. The threshold is basically a control of how much audio data you want to use to train your model. The higher, maybe the data is less but the quality maybe better. The lower, the more training data but a lot of them maybe non-speech noise. You have to find a trade-off, I guess.
@DiyuanLu How is your result of the final trained model? Does SILENCE_TRESHOLD = 0.1 work better than the original 0.3?
-- You are receiving this because you were mentioned. Reply to this email directly or view it on GitHub: https://github.com/ibab/tensorflow-wavenet/issues/307#issuecomment-371141095
@skywolf829 , I think the loss I got is similar. later around 60K, I think I got around 1.2.