IntroductionΒΆ
Training deep networks in any field can be a challenging endeavor. It requires patience, luck, and a lot of debugging. The goal of this section is to provide an overview of both the theory and the practice of training deep networks for audio source separation. We will cover the following topics:
Gradient-descent based optimization
Inputs and outputs for audio separation networks
The building blocks of modern deep nets for separation
Training a model with a straightforward script based on Scaper and nussl.
How to diagnose and fix common bugs
As we proceed through this section, we will gradually build up a script that can be used and modified to train and evaluate a deep audio source separation network.