Number of seconds per step about tensorflow-wavenet

[r-zemblys]

Using default params I get loss of ~2 quite fast, in about 300 steps. You just need to lower learning rate, otherwise loss does not get lower that ~2.5.
I get 0.6-0.9 sec/step on TitanX GPU, using batch_size of 2

[ibab]

I get about 0.5 sec / step on a K40 using batch_size = 1.

[Zeta36]

"I get 0.6-0.9 sec/step on TitanX GPU, using batch_size of 2"

Wow!! I would love to have enough money to buy one of those :P.

Thank you for the info, @r-zemblys

[Zeta36]

0.5 sec/step!! You can run more than 7k steps per hour!! I can run only 4k per day :P

Definitely, I have to buy a GPU for my computer.

By hte way, @ibab. Right now, it's only developed the non-conditional WaveNet form, isn't it? I mean, we are just reading all the raw wav file from a folder (the Corpus in this case), but we are not yet using any conditional one-hot vector (h) as described in the paper, isn't it?

[ibab]

@Zeta36: Right, there's an issue on it now: #112.

[aclaussen1]

How do you lower the learning rate?

[nakosung]

In my case 3 sec per step. Learning rate was annealed by scheduled shell script.

[yarodevuci]

I have 0.1 sec/step is that good?

[cgberlin]

yeah thats pretty good, i'm averaging 0.8 sec/step on a v100 using a custom dataset