How to Reproduce our Results

This repository contains PyTorch implementation code for the paper MixCo: Mix-up Contrastive Learning for Visual Representation (https://arxiv.org/abs/2010.06300) that is accepted in NeurIPS 2020 Workshop on Self-Supervised Learning: Theory and Practice.

This is an instruction to reproduce our results, based on the source code we have provided.

Prerequisites

1. You should download the Tiny-ImageNet dataset. To download the images, go to https://tiny-imagenet.herokuapp.com/ and click 'Download Tiny ImageNet' button. Equivalently, try

wget http://cs231n.stanford.edu/tiny-imagenet-200.zip 

2. Unzip the file and place the folder into [your_data_path].
3. Also, download the ImageNet train dataset.
4. For linear classification, you may also need CIFAR10 and CIFAR100 dataset. If you do not have them, give the argument --download when running lincls.py. Then, it will download the dataset before training.

Structure

This repository contains python files that can train the model with mixup-based representaion learning.

pretrain.py pretrains the model in unsupervised manner, and saves the encoder part (without classification layers).

lincls.py loads and freezes the pretrained model, and then train the classifier part on the target dataset.

Experiments

1. In ./moco/experiments/ and ./simclr/scripts/, there are .sh files which include the commands that can reproduce our experimental results. Open and set the configs.

data_path="[your_data_path]"
exp_name="[experiment_name]"

2. Run the file. For example, if you want to pretrain the ResNet18 model with Tiny-ImageNet, and then see the linear classification results, run exp_mix_res18_tinyimg.sh.

bash experiments/exp_mix_res18_tinyimg.sh

3. You can find the pretraining and linear evaluation results in results/results.json file.

Model Checkpoints

You can also download the checkpoint of MixCo with ResNet-18 architecture. Link: google drive

Citing this work

@article{kim2020mixco,
  title={Mixco: Mix-up contrastive learning for visual representation},
  author={Kim, Sungnyun and Lee, Gihun and Bae, Sangmin and Yun, Se-Young},
  journal={arXiv preprint arXiv:2010.06300},
  year={2020}
}