CurML: A Curriculum Machine Learning Library

A library & toolkit for Curriculum Learning.

The code of the ACM MM 2022 Open Source Software paper CurML: A Curriculum Machine Learning Library.

The paper can be downloaded from the official website or in the docs directory.

Target

To reproduce, evaluate and compare existing Curriculum Learning algorithms
To provide baselines and benchmarks for Curriculum Learning study.
To utilize Curriculum Learning as a plugin to improve the performance of existing works.

Introduction

The figure below shows the overall framework of the library. The main part is a CL Trainer class, composed of a CL Algorithm class and a Model Trainer class, which interact with each other through five APIs of the CL Algorithm class to implement the curriculum.

An general workflow of curriculum machine learning is illustrated below.

Environment

python >= 3.6
pytorch >= 1.9.0

Quick Start

# 1. clone from the repository
git clone https://github.com/THUMNLab/CurML
cd CurML

# 2. pip install local module: curriculum
pip install -e .

# 3. run the example code
python examples/base.py

Development & Extension

Every CL method implemented has been aranged in /curriculum/algorithms, each consists of a Trainer class and an Algorithm class. Their roles and how they interact with each other are illustrated in the figure above. Image_classifier.py in /curriculum/trainer deals with the whole learning process. It embeds the five APIs from existing CL methods, initializing datasets and models, then completes the whole training procedure(forwarding and back propagation). Examples in /examples show how you can call the implemented methods, and how to pass the hyperparameters and start training in an easy manner.

For extension, more latest CL algorithms will be added, or you can add your own CL method in /curriculum/algorithms, which should be consistent with the overall framework. For convenience, /curriculum/backbones incorporates some standard implementations of popular neural networks, including resnet, vgg, etc. For other customized datasets, you can add your dataset processing scripts in /curriculum/datasets, following the example in cifar10.py(only support cifar10 at present). Trainers for tasks other than image classification in the field of CV, NLP, etc., will be implemented in the future.

Cite

Please cite our paper as follows if you find our code useful:

@inproceedings{zhou2022curml,
  title={CurML: A Curriculum Machine Learning Library},
  author={Zhou, Yuwei and Chen, Hong and Pan, Zirui and Yan, Chuanhao and Lin, Fanqi and Wang, Xin and Zhu, Wenwu},
  booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
  pages={7359--7363},
  year={2022}
}

You may also find our survey paper helpful:

@article{wang2021survey,
  title={A survey on curriculum learning},
  author={Wang, Xin and Chen, Yudong and Zhu, Wenwu},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2021},
  publisher={IEEE}
}

License

We follow Apache license across the entire codebase from v0.2.

THUMNLab / CurML

readme