A simplified PyTorch implementation of GANsynth from magenta. Some of the codes are borrowed from magenta orginal repo.
This repo only support the best setting in the GANsynth paper which is a simple version comparing to orginal Tensorflow version by magenta team. So if you want to test other frequency setting, you may need to modify the code.
Use Make Training Data notebook to generate HDF5 file for training.
You have to make a directory to save model checkpint and output spectrum first.
python3 train.pyUse Inference notebook to load the model and generate audio.
https://drive.google.com/open?id=1tNnOtcqCpgTTXGmkHJBA4K6MalBjdsPC