From MNIST to ImageNet and Back: Benchmarking Continual Curriculum Learning

Introduction

This repository contains a code for carrying out experiments for the paper ''From MNIST to ImageNet and Back: Benchmarking Continual Curriculum Learning''. The details of the algorithm are described here (preprint): https://arxiv.org/abs/2303.11076

Environment

1. The project was developed and tested leveraging Python 3.8
2. All python dependencies are listed in requirements.txt file. You can install them using pip install -r requirements.txt.
3. The project was tested on NVIDIA GPUs, including NVIDIA A100.

How to run experiments?

1. The main file that can be used to run the experiments is experiment.py.
2. The experiments can be configured using the configuration files in the config folder.
3. The configuration files are in the YAML format. There is a separate configuration file for each strategy.
4. You can run the experiments using the following command: python3 experiment.py --config config/class_incremental/config_file_name.yml
5. The results will be put in logs/out directory.

The structure of the repository

• config - folder with configuration files for experiments
• models - folder with code for models
• scenarios - folder with code for scenarios and datasets
• strategies - folder with code for strategies