Inference with torch and torchaudio


I’ve just set up my Jetson Nano and am trying to use it for inference on a speaker verification model (GitHub - clovaai/voxceleb_trainer: In defence of metric learning for speaker recognition). As torchaudio isn’t available as a package on the nano, I’m trying to figure out what else I can do. I’ve tried to install torchaudio from source, but that keeps failing with Illegal instruction (core dumped).

Now I’m considering to either create a docker container and run it there, or to export the model to ONNX (torchaudio is only used in the model definition files). Any advice on which of the two approaches is more likely to be successful? Or what else I can try instead?