VITS checkpoint trained on Hi-Fi TTS

#1
Amphion org
No description provided.
Amphion org

Hi Tze Ying,

  1. Please remove the private paths that are not used in the inference process.
  2. Have you tested the inference code?
Amphion org
edited Feb 15

Yes, I have tested the inference code according to the instructions written in README.md of this PR. There's one problem in the model loading part of accelerator, it appears that installing accelerate==0.24.1 as specified here (https://github.com/open-mmlab/Amphion/blob/main/env.sh) will result in the following error during inference, since pytorch_model.bin does not exist in our trained checkpoint folder:
Capture.PNG

This problem could be solved by editing code of models/tts/base/tts_inferece.py to load the model.safetensors file in a similar way as suggested here (https://github.com/open-mmlab/Amphion/issues/67#issuecomment-1869897191) or installing accelerate >=0.25.

Amphion org

Please write this information in the README so that uers can use this checkpoint successfully.

lmxue changed pull request status to merged

Sign up or log in to comment