Skip to content

SOTA Open Source TTS

Update vqgan config & baseline

Lengyuee596deff1fd8
diff --git a/fish_speech/configs/vqgan_pretrain.yaml b/fish_speech/configs/vqgan_pretrain.yaml
index 372e9ef..1fb568e 100644
--- a/fish_speech/configs/vqgan_pretrain.yaml
+++ b/fish_speech/configs/vqgan_pretrain.yaml
@@ -126,3 +126,4 @@ callbacks:
- encoder
- decoder
- quantizer
+ - discriminator
diff --git a/fish_speech/models/vqgan/lit_module.py b/fish_speech/models/vqgan/lit_module.py
index 8c1ed7d..0744fc5 100644
--- a/fish_speech/models/vqgan/lit_module.py
+++ b/fish_speech/models/vqgan/lit_module.py
@@ -213,6 +213,14 @@ class VQGAN(L.LightningModule):
prog_bar=False,
logger=True,
)
+ self.log(
+ "train/generator/loss_adv",
+ loss_adv,
+ on_step=True,
+ on_epoch=False,
+ prog_bar=False,
+ logger=True,
+ )
# Generator backward
optim_g.zero_grad()
diff --git a/tools/vqgan/inference.py b/tools/vqgan/inference.py
index 8e8aaf1..a8a8076 100644
--- a/tools/vqgan/inference.py
+++ b/tools/vqgan/inference.py
@@ -26,7 +26,7 @@ OmegaConf.register_new_resolver("eval", eval)
@click.option(
"--input-path",
"-i",
- default="data/sft/Rail_ZH/三月七/1fe0cc6fc3fe3e6d.wav",
+ default="test.wav",
type=click.Path(exists=True, path_type=Path),
)
@click.option(
@@ -36,7 +36,7 @@ OmegaConf.register_new_resolver("eval", eval)
@click.option(
"--checkpoint-path",
"-ckpt",
- default="results/vq-group-fsq-8x1024-wn-20x512-cond-e009/checkpoints/step_000355000.ckpt",
+ default="checkpoints/vq-gan-group-fsq-8x1024-wn-20x768-30kh.pth",
)
def main(input_path, output_path, config_name, checkpoint_path):
with initialize(version_base="1.3", config_path="../../fish_speech/configs"):