Introduction

This PyTorch template helps you to start a new machine learning project quickly. It has the following structure:

pytorch_template
├── config
│     ├─ASR
│     │     ├─experiment.yaml
│     │     ├─model
│     │     │   ├─ simpleLSTM.yaml
│     │     │   └─ simpleLinear.yaml
│     │     ├─spec_layer
│     │     │   ├─ Mel.yaml
│     │     │   └─ STFT.yaml
│     └─AMT
├── models
│     ├─Models.py
│     └─Tasks.py
├── utils

Tasks.py contains the tasks you want to work on, such as Automatic Speech Recognition (ASR).

Models.py contains the model architectures you want to experiment with.

config contains the .yaml configuration files.

If you want to try a new model on the existing ASR task, simply add it to models/Models.py, and then add a new configuration file for this model as config/ASR/model/YourNewModel.yaml.

python train_ASR.py model=YourNewModel will train an ASR using your newly defined model YourNewModel.

Usage

A. Quick Start

Step1: Download the source code

git clone https://github.com/KinWaiCheuk/pytorch_template.git

Step2: Install the dependencies

pip install -r requirements.txt

Step3: Train a model

python train_ASR.py

This command trains a simple LSTM model using the TIMIT dataset.

By default, train_ASR.py uses the configuration file at config/ASR/experiment.yaml. i.e. This model is using Melspectrograms (spec_layer: Mel) as the inputs of the LSTM model (model: simpleLSTM).

B. Configuration changes

Due to the usage of hydra in this framework, there are two ways to change the configurations.

1. Modifying config/ASR/experiment.yaml directly
2. Overwritting the arguments on a command line interface (CLI)

The following section explains how to change the configurations using CLI.

Change model type

python train_ASR.py model=simpleLinear This line overwrites model: simpleLSTM with model: simpleLinear and then runs train_ASR.py using the overwritten configuration

Change input type

python train_ASR.py spec_layer=STFT This line overwrites spec_layer: Mel with spec_layer: STFT and then runs train_ASR.py using the overwritten configuration.