The "nnet3" setup

Introduction

This documentation covers the latest, "nnet3", DNN setup in Kaldi. For an overview of all deep neural network code in Kaldi, explaining Karel's version, see Deep Neural Networks in Kaldi.

The nnet3 setup is intended to support more general kinds of networks than simple feedforward networks (e.g. things like RNNs and LSTMs) in a natural way that should not require any actual coding. Like the nnet2 setup, it supports parallel training across GPUs on multiple machines (using an approach based on natural gradient-stabilized SGD with model averaging, see this paper .

The documentation has been broken up into multiple pages: see

• Data types in the "nnet3" setup.
• Compilation in the "nnet3" setup
• Optimization in the "nnet3" setup
• Context and chunk-size in the "nnet3" setup