Add torch.package.PackageImporter(local_file).load_pickle("tts_models", "model") and model.save_wav()

[mizantrop777]

I need to port this code:

[NiklasGustafsson]

Hello @mizantrop777,

Since PyTorch packages involve loading Python code, it is outside the scope of TorchSharp. It is a firm design principle of TorchSharp that it will provide access to the native libraries underneath PyTorch, i.e. "libtorch" but not require a Python environment in order to run.

torch.package.PackageImporter could possibly be implemented outside of TorchSharp in some form of extension library, but dealing with pickling, which is a Pyton-specific serialization format, will have some challenges in .NET.

Regarding 'save_wav,' I've been searching for the documentation in PyTorch as well as the torchaudio documentation, but didn't find it. Could you point me at the docs? Maybe @kaiidams knows?

Thanks,

Niklas

[kaiidams]

@mizantrop777 @NiklasGustafsson
We have TorchSharp.torchaudio.save() interface, but it is a stub function. You need to register an IO backend by calling TorchSharp.backend.utils.set_audio_backend(AudioBackend backend).
There is a sample code to do it in

TorchSharp/src/Examples/SpeechCommands.cs

Line 50 in 8a08803

torchaudio.backend.utils.set_audio_backend(new WaveAudioBackend());

, but it only supports load().

[kaiidams]

@mizantrop777
If you want to port Silere TTS to train models yourself, you could convert their model from PyTorch to C#. If you have to only port inference code to C#, probably you could convert their models into TorchScript module or ONNX model and then either use TorchSharp to run TorchScript or use ONNX runtime to run ONNX model.