Unsupervised Coreset Selection(UCS)

PyTorch Code for the paper:
"Extending Contrastive Learning to Unsupervised Coreset Selection"
Jeongwoo Ju, Heechul Jung, Yoonju Oh and Junmo Kim

Original code is SVP from Stanford.
Based on the original code, we implemented our unsupervised coreset algorithm.

@article{ju2021extending,
  title={Extending Contrastive Learning to Unsupervised Coreset Selection},
  author={Ju, Jeongwoo and Jung, Heechul and Oh, Yoonju and Kim, Junmo},
  journal={arXiv preprint arXiv:2103.03574},
  year={2021}
}

Installation

Prerequisites

• Linux or macOS (Windows is in experimental support)
• Python 3.6 +
• PyTorch 0.4.1
• TorchVision 0.2.1
• CUDA 9.1

Shell Script File Description

File Name	Description
run_sim_core_svhn.sh	UCS w\ SimCLR on SVHN
run_sim_core_qmnist.sh	UCS w\ SimCLR on QMNIST
run_sim_core_cifar.sh	UCS w\ SimCLR on CIFAR10
run_moco_core_svhn.sh	UCS w\ MoCo on SVHN
run_moco_core_qmnist.sh	UCS w\ MoCo on QMNIST
run_moco_core_cifar.sh	UCS w\ MoCo on CIFAR10

Folder Description

Folder Name	Description
loss	coreset score for SimCLR and MoCo
index	example indices for each dataset

Training Examples

see each shell script file in main branch For example, ./run_sim_core_svhn.sh is as follows

for subsize in 22500 30000 37500 45000 52500
do
    for run in 1 2 3 4 5
    do
        CUDA_VISIBLE_DEVICES=2 python3 -m svp.svhn active \
            --run-dir ./run/svhn/resnet18/simclr \
            --dataset svhn \
            --datasets-dir './data' \
            --arch resnet18 \
            --num-workers 4 \
            --weighted-loss False \
            --coreset-path ./index/simclr_coreset_svhn_run$run.index \
            --coreset-loss-path ./loss/simclr_coreset_svhn_run$run.loss \
            --runs $run \
            --initial-subset $subsize \
            --eval-target-at $subsize 2>&1 | tee "./log_svhn_test_simclr_coreset_resnet18_subsize"$subsize"_run$run.txt"
     done
done

Dataset

CIFAR10

Link: https://www.cs.toronto.edu/~kriz/cifar.html

SVHN

Download train_32x32.mat and test_32x32.mat from the web http://ufldl.stanford.edu/housenumbers/

QMNIST

Download QMNIST dataset using torchvision

Experimental Results

Coreset selection performance on CIFAR10

Coreset selection performance on SVHN

Coreset selection performance on QMNIST