search

ksw0306 / ClariNet

A Pytorch Implementation of ClariNet
MIT License
289 stars 64 forks source link

Why the Synthesize (Teacher) takes very long time (take 7222 seconds to synthesis a 8 seconds wav) ? #4

Open hxs7709 opened 6 years ago

hxs7709 commented 6 years ago

The command in my test is: python synthesize.py --model_name wavenet_gaussian --num_blocks 4 --num_layers 6 --load_step 33902

I change the synthesize.py to only handle one file, by hackinging test_dataset to only include one file. The log is as below, where the 196864 is num_samples of wave. 196800/196864 7222.449482679367 seconds So, the speed is about 196864/7222 = 27 samples per second.

I know the teacher synthesis is slow, but the speed is about 180 samples per second when I test teacher synthesis in the same machine with another wavenet implementation (https://github.com/azraelkuan/parallel_wavenet_vocoder).

hxs7709 commented 6 years ago

Can anybody help to give clue about this speed issue about teacher synthesis? It's really bothering me to take so long time when evaluating.....Thank you very much.