Any success with GlowTTS -> Torchscript?

[ihshareef]

The GlowTTS implementation is excellent. Great work.
I currently have a Glow-TTS model that I trained on coqui-ai/TTS, and I'm looking to utilise it on mobile. However, I've been quite unsuccessful at converting the model to either TorchScript or ONNX. Has anyone had any such success? If so, would they mind helping me out?

Thanks

[nicolalandro]

I do not know if there is a method directly from this project, but by bring out the class of the pythorch model in a custom codebase (you only need the load model and predict) you can youse pytorch mobile to create application with Android or iOs or serve a coqi-ai/TTS server and use the api.

(I'm newest, sorry in advance if the answare is not so good)

[erogol]

We do not have ONNX yet, but in our plans. I know @synesthesiam is also interested in discovering more about it.

[synesthesiam]

As @erogol mentioned, I plan to look at ONNX export. I have it already working elsewhere, and I suspect the process will be straightforward since both our GlowTTS implementations come from the same source.

The only "trick" I needed to figure out was storing the inverse with torch.no_grad:

# Inference only
model.eval()

# Do not calcuate jacobians for fast decoding
with torch.no_grad():
    model.decoder.store_inverse()

[AIFahim]

@synesthesiam Have you implemented ONNX Export till now?

[synesthesiam]

No, I've moved to working on Piper (https://github.com/rhasspy/piper) which uses VITS and is targeted at low end machines like the RPi 4.