ibab / tensorflow-wavenet

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

Periodic debug info dump for faster training #116

Open nakosung opened 8 years ago

nakosung commented 8 years ago

Disabling tensor board summary improves training speed by ~10%. If we don't need distributions and histrograms every frame, we may dump summaries periodically except for loss.

https://wookayin.github.io/TensorflowKR-2016-talk-debugging/#37

lemonzi commented 8 years ago

Adding an option that disables all those histogram summaries completely has been in the TODO's for a while. Dumping them periodically could be useful as well.