Infer mode

[snykral]

It would be interesting to add single audio infer/classification scripts for models. It requires some code tracking to remount it from the loss calculation.