Anyone has tried XLA to accelerate Tacotron2

Rayhane-mamah / Tacotron-2

DeepMind's Tacotron-2 Tensorflow implementation

MIT License

2.28k stars 905 forks source link

Anyone has tried XLA to accelerate Tacotron2 #411

Open lee-bin opened 5 years ago

lee-bin commented 5 years ago

Before using XLA, the Tacotron 2 runs about 0.5s, however, after using XLA, it increases to about 0.7s which is much slower.

config = tf.ConfigProto()
config.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1              
self.session = tf.Session(config=config)

I use XLA like above. Any thoughts would be appreciated. Thanks.

superhg2012 commented 5 years ago

Hi, when did you use XLA ? train or inference phrase?

lee-bin commented 5 years ago

Just for inference.