Why does this library outputs to ONNX and tensorflow if Pytorch is still needed?

[DavidGOrtega]

👋 Having looked into the examples I have seen that despite that the model is converted into ONNX or TF you still need the Decoder in Pytorch. What is the benefit of the conversion then?

[snakers4]

If the question is about the stt model, the decoder does not rely on PyTorch.

[DavidGOrtega]

Thanks for the reply @snakers4 🙏
What I meant is that despite the effort of transforming the model into ONNX or TF some parts like the Decoder still relies in Pytorch. Or at least that what I can infer.

Within the current onnx example:

_, decoder, utils = torch.hub.load(repo_or_dir='snakers4/silero-models', model='silero_stt', language='en')

Then de Decoder class has Torch dependency like i.e.:

argm = torch.argmax(probs, axis=1)

Ideally the model should work completely with ONNX alone? If not why should I run the stt model in ONNX and not in Pytorch?

[snakers4]

Well, I guess it could depend on numpy's argmax. We just did not really invest into this, because the users most likely will rewrite the decoder anyaway.

[DavidGOrtega]

because the users most likely will rewrite the decoder anyaway.

Why? 🤔