Vocoder Project

Based on this template

Installation guide

pip install -r ./requirements.txt
mkdir default_test_model

Usage guide

You can choose between 2 pipelines: "the clown" and "the normal one".

"The clown" is trained using 2 different configs and fixing dataset bug in between.
"The normal one" is trained using one config and no bugs, but its checkpoint is undertrained at the moment.

To train

To train "the clown":

python3 train.py --config-name hifigan-bug
python3 train.py --config-name hifigan2 +resume=saved/hifigan-mk1/path2checkoint.pth

To train "the normal one":

python3 train.py --config-name hifigan

To synthesize audio:

"The clown":

gdown https://drive.google.com/file/d/1SoDH__65dA808Eh5EVEJTgcMLgLsldyP/view?usp=sharing -O default_test_model/checkpoint.pth --fuzzy
python3 train.py --config-name hifigan_bug +resume=default_test_model/checkpoint.pth ++trainer.len_epoch=1 ++g_optimizer.lr=0 ++trainer.epochs=77 +data.val.dataset.limit=1

"The normal one":

gdown https://drive.google.com/file/d/1LmNT_XSEHxnd6IgAhX7CULXZEWgFI7Sw/view?usp=sharing -O default_test_model/checkpoint.pth --fuzzy
python3 train.py --config-name hifigan +resume=default_test_model/checkpoint.pth ++trainer.len_epoch=1 ++g_optimizer.lr=0 ++trainer.epochs=77 +data.val.dataset.limit=1

The data will be logged to wandb. You can change wavs in test_data if you want.