Implementation plan

[Yeongtae]

Data feeding

• [ ] using parameters in hparam.py to reduce independent parameters (3rd)
• [ ] muLaw quantization to preserve more important information (4th)

Model

• modifying the wavernn model to reduce matrix multiflications (1st, ~ 11.04) (Fail)
• test half precision and holding on network parameters on half precision register. (extended 1st, ~11.11 11.14)
• [ ] weight sparsification to reduce computation times (2nd, ~11.16)

[m-toman]

Thanks.
I've updated to the latest version of Tacotron-2 and forked this version. Currently training the LJ taco model.

[bjtommychen]

Thanks you both. I am looking forward to your progress.

[Yeongtae]

Intro

WaveRNN use GRU operation optimization techniques.
To figure out efficiency of operation optimization, I had completed the GRU opreation optimization experiment.

I had compared Pythorch's built-in GRU module with self implementation of GRU operation optimization

Experimental setup

I made a simple two-dimensional coordinate prediction problem, f(x_i) = y = x_(i+1),
The dataset was generated by using r x rotationmatrix(theta)．[1, 0 ]^T.

Device: k80 GPU

Implementation

Built-in GRU module

Self implementation of original GRU

Self implementation of GRU operation optimization

Experimental result

I performed full batch inferences, 10000 times each.

Built-in GRU module: 6.02 seconds
Self implementation of original GRU: 16.0 seconds
Self implementation of GRU operation optimization: 9.70 seconds

The pytorch build-in GRU module is well optimized.
Because it use cudnn_gru,

check rossumai/OCkRE#2

Conclusion

In my opinion, it's faster to use built-in GRU in pytorch.

[m-toman]

Thanks for the experiment. Unfortunately I'm in hospital and couldn't continue to integrate the original wavernn implementation for now, but it's still on my agenda.

[Yeongtae]

I'm trying to use faster wave generator, which is waveglow.
Sorry about stopping implementation.

[m-toman]

Perhaps this post is interesting for you: mozilla/TTS#9 (comment)
I'm generally still watching the Mozilla TTS repo to check their progress with WaveRNN.

[m-toman]

I'll close this for now.