This is an implementation of the Waveglow model based on Tensorflow. The model architecture follows the NVIDIA public Waveglow model. It includes the inference part only for now. The main purpose is to benchmark its inference performance on GPU. You may use it freely.
File an issue if you have any questions.
This repo has the following files:
glow.py
,wn.py
andupsample.py
: waveglow modelconfig.py
: model parametersbenchmark.py
: benchmark utilities
To run waveglow inference:
python ./benchmark.py --gpu=0
On a NVidia V100 32GB NVLink GPU, we measure the inference RTF to be 0.024 in fp16
mode with n_channels = 512
.
The docker image we used is nvcr.io/nvidia/tensorflow:19.06-py3