A Fully Configurable Pipeline For Image Classifier

Introduction

This goal of this repository was to minimize the number of code edits by enabling easy configuration of the Image Classifier pipeline using Hydra, Timm & Lightning. In the /config/dataset directory, we are able to configure the data transformation without having to edit the code. In the /config/pytorch-lightning directory, we are able to configure over 700 State-of-the-art CNN model & 10 optimizers supported by Timm. In the /config/training directory, variables like max_epochs can be set.

Structure

The structure of the configuration folder is shown below. The main configuration file can be found at /config/config.yaml.

├── config
│   ├── dataset
│   ├── pytorch-lightning     <- models & optimizers
│   ├── testing
│   ├── training
│   |
│   ├── configs.yaml          <- main config

Usage

1. Install packages

   pip install -r requirements.txt

2. Change the configs at config/dataset/dataset.yaml, etc.

   data_loader:
       train_dataset_path: PATH_TO_YOUR_DATASET
       val_dataset_path: PATH_TO_YOUR_DATASET

3. Launch Training

   python train.py

Contact

Chau Yuan Qi - @chauyuanqi - yuanqichau@gmail.com