Can't load fine-tuned model

[PavelPanjaya]

I've fine-tuned a model using tune.py script and got a model_save_path.jit file

Trying to load it:

import torch
torch.set_num_threads(1)

model, utils = torch.hub.load(
    repo_or_dir='.............../silero_training/silero-vad/',
    model='model_save_path.jit',
    source="local")
(get_speech_timestamps, _, read_audio, _, _) = utils

But get this error:

Traceback (most recent call last):
  File ".................../silero-vad/main.py", line 3, in <module>
    model, utils = torch.hub.load(
                   ^^^^^^^^^^^^^^^
  File ".................../lib/python3.12/site-packages/torch/hub.py", line 647, in load
    model = _load_local(repo_or_dir, model, *args, **kwargs)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".................../lib/python3.12/site-packages/torch/hub.py", line 679, in _load_local
    entry = _load_entry_from_hubconf(hub_module, model)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".................../lib/python3.12/site-packages/torch/hub.py", line 387, in _load_entry_from_hubconf
    raise RuntimeError(f"Cannot find callable {model} in hubconf")
RuntimeError: Cannot find callable model_save_path.jit in hubconf

Please advise how do I supposed to load the fine-tuned model.
Thanks!

[snakers4]

Will your custom model load, if you use the direct loading function?

silero-vad/src/silero_vad/model.py

Lines 33 to 36 in 63fe03a

	else:
	model = init_jit_model(model_file_path)
	return model

[PavelPanjaya]

Basically yes - And I could run the model like this:

model = torch.jit.load("............/model_save_path.jit")
model.eval()

_, utils = torch.hub.load(
        repo_or_dir='snakers4/silero-vad:v6.0',
        model='silero_vad',
        force_reload=False,
        onnx=True,
        verbose=False
    )
(get_speech_timestamps, _, read_audio, _, _) = utils
wav = read_audio(audio_path)
speech_timestamps = get_speech_timestamps(wav, model, return_seconds=True)

It works, but I'd like to load it more... gracefully :)

[snakers4]

Without modifications, this looks a bit more polished, if you need to use the utils from the library:

  from silero_vad import (read_audio,
                          get_speech_timestamps)
  model = torch.jit.load(local_path)
  model.eval()

[banjamn]

@PavelPanjaya Thanks, is the performance improvement of the fine-tuned model significant on your dataset?
How do you label the data? (The author stated that the labeling error should be within 30 milliseconds.)

[PavelPanjaya]

performance improvement of the fine-tuned model significant on your dataset

Yes

How do you label the data?

Manual labeling in Audacity of few thousands samples