Trick to Upsampling to High sampling rates using VITS model #1456
Conversation
Edresson
force-pushed
the
VITS-upsample
branch
3 times, most recently
from
c583a8a to
a5f5eba
Compare
|
To Fix the zoo-tests we need to merge this coqpit PR. |
| assert batch["spec"].shape[2] == batch["mel"].shape[2], f"{batch['spec'].shape[2]}, {batch['mel'].shape[2]}" | ||
|
|
||
| if not self.args.TTS_part_sample_rate: | ||
| assert batch["spec"].shape[2] == batch["mel"].shape[2], f"{batch['spec'].shape[2]}, {batch['mel'].shape[2]}" |
There was a problem hiding this comment.
can you also implement the corresponding assert when upsampling is used?
| @@ -104,6 +104,7 @@ class HifiganConfig(BaseGANVocoderConfig): | |||
| "resblock_type": "1", | |||
| } | |||
| ) | |||
| discriminator_model_params: dict = field(default_factory=lambda: {"periods": [2, 3, 5, 7, 11]}) | |||
There was a problem hiding this comment.
HifiGAN changes must be a separate PR
There was a problem hiding this comment.
Yeah, my bad, I removed these commits, it is in the PR: #1526
| from TTS.vocoder.datasets.preprocess import load_wav_data | ||
| from TTS.vocoder.models.gan import GAN | ||
|
|
||
| output_path = "/home/julian/workspace/train" |
There was a problem hiding this comment.
Why it is pointin Julian's workspace?
Why there is HiFiGan recipe update in VITS PR?
There was a problem hiding this comment.
My bad, I removed these commits, it is in the PR: #1526
| @@ -0,0 +1,90 @@ | |||
| import glob | |||
There was a problem hiding this comment.
We don't need these costly tests.
Just write unittests in test_vits.py
There was a problem hiding this comment.
Indeed, good catch, Done :)
| @@ -0,0 +1,90 @@ | |||
| import glob | |||
There was a problem hiding this comment.
We don't need this one too if you tests things in test_vits.py
|
Awesome PR |
No description provided.