Giters

ibab / tensorflow-wavenet

A TensorFlow implementation of DeepMind's WaveNet paper

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Possible to reduce time/step ? + gc_channels explanations

opened this issue · comments

commented

Hello,

I had some questions:

  1. Is it possible to reduce the time needed per steps? Currently it takes my computer about 40 sec/steps. I have an i7 Skylake and 16 Gb RAM. Is 40 sec/steps slow ?

  2. My laptop has a bad GPU (Intel HD 520). Can I still use it alongside the CPU or is it exclusive (use CPU only or GPU only)? No it works only with NVidia Cuda compatible GPU. Tensorflow-gpu enable both CPU and GPU use.

  3. Does increasing batch_size reduce the time needed to train the model? No. It processes multiple files at once, so if batch_size = 3, it will take 3 times more time per step than batch_size=1.

  4. gc_channels makes it possible to learn different languages ? for example: I have 2 folders: 01 (English voices) and 02 (Chinese mandarin voices). Does use gc_channels=2 makes the program learn English and Chinese ?

  5. Since I let the variable SAMPLES = 16000, does that means I have to go to the step n°16000 ? Or is it NUM_STEPS (that default to 1e5)?

Thank you for your attention!