search

ibab / tensorflow-wavenet

A TensorFlow implementation of DeepMind's WaveNet paper
MIT License
5.41k stars 1.29k forks source link

Pre Trained Model #307

Open brappazzo opened 6 years ago

brappazzo commented 6 years ago

Does anyone have a pre trained model they could share? Specifically one trained on music (piano) data?

DiyuanLu commented 6 years ago

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

skywolf829 commented 6 years ago

@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.

dreamibor commented 6 years ago

@DiyuanLu How is your result of the final trained model? Does SILENCE_TRESHOLD = 0.1 work better than the original 0.3?

DiyuanLu commented 6 years ago

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

DiyuanLu commented 6 years ago

@skywolf829 , I think the loss I got is similar. later around 60K, I think I got around 1.2.