[BUG]: Issue with the vocoder calling during inference

[ArkhamImp]

Describe the bug

In models/base/new_inference.py

vocoder_cfg, vocoder_ckpt = self._parse_vocoder(self.args.vocoder_dir)

If the vocoder checkpoint is saved as '*.pt' it works fine, but there will be error when I using Hifigan checkpoint which is save as '*.bin'.

[VocodexElysium]

Could you tell me which task you tried to run and encountered this BUG?

Currently, regarding loading neural vocoder for other tasks, both a .pt checkpoint or a statedict folder that contains a series of .bin files are supported as in models/vocoders/vocoder_inference.py, the function load_nn_vocoder. The function parse_vocoder should have been discarded.

[ArkhamImp]

So I should not use the inference function in models/base/new_inference.py and write my own?

Or maybe copy the load model function from models/vocoders/vocoder_inference.py is more straightforward.

[VocodexElysium]

So I should not use the inference function in models/base/new_inference.py and write my own?

Or maybe copy the load model function from models/vocoders/vocoder_inference.py is more straightforward.

Yes, you can just write your own inference function with just base names aligned with the new_inference.py

Thanks for your reporting, I will check the compatibility of loading both old .pt vocoder file and new vocoder state dict folder of each downstream tasks this week.

[mysxs]

So I should not use the inference function in models/base/new_inference.py and write my own?

Or maybe copy the load model function from models/vocoders/vocoder_inference.py is more straightforward.

Hi @VocodexElysium~, excuse me! I would like to ask which .bin file do you use with higfigan? Or how did you change your code?

[VocodexElysium]

So I should not use the inference function in models/base/new_inference.py and write my own? Or maybe copy the load model function from models/vocoders/vocoder_inference.py is more straightforward.

Hi @VocodexElysium~, excuse me! I would like to ask which .bin file do you use with higfigan? Or how did you change your code?

Hi @mysxs.

We use the whole folder as a statedict for accelerate loading, so all the .bin files should be used.

I have checked out the issue#142. You encountered a misuse of a deprecated file. For your case, update your related code snippet with the following can solve your problem:

def _parse_vocoder(vocoder_dir):
        r"""Parse vocoder config"""
        vocoder_dir = os.path.abspath(vocoder_dir)
        ckpt_list = [ckpt for ckpt in Path(vocoder_dir).glob("*.pt")]
        if len(ckpt_list) == 0:
            ckpt_path = vocoder_dir
        else:
            ckpt_list.sort(key=lambda x: int(x.stem), reverse=True)
            ckpt_path = str(ckpt_list[0])
        vocoder_cfg = load_config(
            os.path.join(vocoder_dir, "args.json"), lowercase=True
        )
        return vocoder_cfg, ckpt_path

Thanks for your report! We will fix this issue ASAP.