Home

Awesome

Sequential Subset Matching for Dataset Distillation

Paper

The code is for training expert trajectories and distilling synthetic data from our Sequential Subset Matching for Dataset Distillation paper (NIPS 2023). We provide the Sequential Subset Matching methods in MTT model and IDC model

Getting Started

Download

git clone https://github.com/shqii1j/seqmatch.git
cd seqmatch

The requirement for SeqMatch in MTT model

If you have an RTX 30XX GPU (or newer), run

conda env create -f requirements_11_3.yaml

If you have an RTX 20XX GPU (or older), run

conda env create -f requirements_10_2.yaml

You can then activate your conda environment with

conda activate distillation

The requirement for SeqMatch in IDC model

If you have created distillation environment, run it

conda activate distillation

If not, install the packages pytorch and efficientnet_pytorch

Sequential Subset Matching in MTT

There is an example .sh file to use our code. This command will generate 3 subsets to distill CIFAR-10 50 image per class:

cd seqmatch-mtt
python run.sh

Using buffer.py, you can generate some expert trajectories for the first subset. The following command will train 100 ConvNet models on CIFAR-10 with ZCA whitening for 50 epochs each:

python buffer.py --dataset=CIFAR10 --model=ConvNet --train_epochs=50 --num_experts=100 --zca --buffer_path={path_to_buffer_storage} --data_path={path_to_dataset}

Using distill_eval_new.py, you can generate the first subset via the buffers. The following command will generate the first distilled subset for CIFAR-10 down to just 1 image per class:

python distill.py --dataset=CIFAR10 --ipc=1 --syn_steps=30 --expert_epochs=2 --max_start_epoch=20 --zca --lr_img=100 --lr_lr=1e-05 --lr_teacher=0.01 --buffer_path={path_to_buffer_storage} --data_path={path_to_dataset} --run_name={path to the task} --name={path to the subset}

For the following subsets, you need to use --pre_names and --reparam_syn flags. The following command will generate the expert trajectories based on the previous subsets and the new distilled subset for CIFAR-10:

python buffer.py --dataset=CIFAR10 --model=ConvNet --train_epochs=20 --num_experts=100 --zca --image_path=logged_files --data_path={path_to_dataset} --run_name={path to the task} --pre_names={paths to the previous subsets} --reparam_syn

python distill_eval_new.py --dataset=CIFAR10 --ipc=1 --syn_steps=30 --expert_epochs=2 --zca --image_path=logged_files --data_path={path_to_dataset} --buffer_path=./logged_files/CIFAR10/{path to the task}/{paths to the last subset}/buffer --intervals=0-20 --lr_img=100 --lr_lr=1e-05 --lr_teacher=0.01 --run_name={path to the task} --pre_names={paths to the previous subsets} --name={path to the new subset} --reparam_syn

Please find a full list of hyper-parameters in our paper (https://arxiv.org/abs/2311.01570).

ImageNet

When generating expert trajectories with buffer.py or distilling the dataset with distill.py for ImageNet, you must designate a named subset of ImageNet with the --subset flag.

Sequential Subset Matching in IDC

There is an example .sh file to use our code. This command will generate 2 subsets to distill CIFAR-10 50 image per class (25 image per class in one subset):

cd seqmatch-idc
python condense_new.py --reproduce -d cifar10 -f 2 --ipcs=[25,25] --inner_loop=[50,100] --niters=[2000,4000] --lrs_img_ori=[5e-3,5e-3] --it_log=100 --it_eval=100 --seed=2023 --fix_iter=50

You can get more robust result (repeat test 5 times) via th following code:

python test.py -d cifar10 -n convnet -f 2 --reproduce --ipcs=[5,5] --repeat=5 --seed=2023 --data_path={path to the results} --test_paths={paths to the subsets}

Reference

If you find our code useful for your research, please cite our paper.

@inproceedings{
du2023sequential,
title={Sequential Subset Matching for Dataset Distillation},
author={Jiawei Du and Qin Shi and Joey Tianyi Zhou},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023}
}