Will AWS Neuron support a PyTorch model that inferences audio to audio?

Question

The audio going into the model would be different lengths, and I saw a post saying "At the time of writing, the AWS Neuron SDK does not support dynamic shapes, which means that the input size needs to be static for compiling and inference."

Does this mean I can't use Neuron (and therefore Inferentia2) without truncating my audio going into the model to a fixed size? Or am I misunderstanding?

Answer

Variable shape tensors are normally dealt with using bucketing (to a nearest tensor size) and padding.  So the limitation currently exists, but customers can normally work around it: https://awsdocs-neuron.readthedocs-hosted.com/en/v2.11.0/general/appnotes/torch-neuron/bucketing-app-note.html?highlight=bucketing

Will AWS Neuron support a PyTorch model that inferences audio to audio?

Conteúdo relevante