Official PyTorch Implementation for the "Distilling Datasets Into Less Than One Image" paper.

🌐 Project | 📃 Paper

Poster Dataset Distillation (PoDD): We propose PoDD, a new dataset distillation setting for a tiny, under 1 image-per-class (IPC) budget. In this example, the standard method attains an accuracy of 35.5% on CIFAR-100 with approximately 100k pixels, PoDD achieves an accuracy of 35.7% with less than half the pixels (roughly 40k)

Distilling Datasets Into Less Than One Image
Asaf Shul*, Eliahu Horwitz*, Yedid Hoshen
*Equal contribution
https://arxiv.org/abs/2403.12040

Abstract: Dataset distillation aims to compress a dataset into a much smaller one so that a model trained on the distilled dataset achieves high accuracy. Current methods frame this as maximizing the distilled classification accuracy for a budget of K distilled images-per-class, where K is a positive integer. In this paper, we push the boundaries of dataset distillation, compressing the dataset into less than an image-per-class. It is important to realize that the meaningful quantity is not the number of distilled images-per-class but the number of distilled pixels-per-dataset. We therefore, propose Poster Dataset Distillation (PoDD), a new approach that distills the entire original dataset into a single poster. The poster approach motivates new technical solutions for creating training images and learnable labels. Our method can achieve comparable or better performance with less than an image-per-class compared to existing methods that use one image-per-class. Specifically, our method establishes a new state-of-the-art performance on CIFAR-10, CIFAR-100, and CUB200 using as little as 0.3 images-per-class.

Poster distillation progress over time followed by a semantic visualization of the distilled classes using a poster of CIFAR-10 with 1 IPC

This project consists of:

• main.py - Main entry point (handles user run arguments).
• src/base.py - Main worker for the distillation process.
• src/PoDD.py - PoDD implementation using RaT-BPTT as the underlying dataset distillation algorithm.
• src/PoCO.py - PoCO class ordering strategy implementation, using CLIP text embeddings.
• src/PoDDL.py - PoDDL soft labeling strategy implementation.
• src/PoDD_utils.py - Utility functions for PoDD.
• src/data_utils.py - Utility functions for data handling.
• src/util.py - General utility functions.
• src/convnet.py - ConvNet model for the distillation process.

1. Clone the repo:

git clone https://github.com/AsafShul/PoDD
cd PoDD

2. Create a new environment with needed libraries from the environment.yml file, then activate it:

conda env create -f environment.yml
conda activate podd

This implementation supports the following 4 datasets:

• CIFAR-10
• CIFAR-100
• CUB200
• Tiny ImageNet

Both the CIFAR-10 and CIFAR-100 datasets are built-in and will be downloaded automatically.

1. Download the data from here
2. Extract the dataset into ./datasets/CUB200

1. Download the dataset by running wget http://cs231n.stanford.edu/tiny-imagenet-200.zip
2. Extract the dataset into ./tiny-imagenet-200/tiny-imagenet-200
3. Preprocess the validation split of the dataset to fit torchvision's ImageFolder structure. This can be done by running the function format_tiny_imagenet_val located in ./src/data_utils.py

The main.py script is the main script in this project.

Below are examples for running PoDD on CIFAR-10, CIFAR100, CUB200 and Tiny ImageNet datasets for 0.9 IPC.

python main.py --name=PoDD-CIFAR10-LT1-90 --distill_batch_size=96 --patch_num_x=16 --patch_num_y=6 --dataset=cifar10 --num_train_eval=8 --update_steps=1 --batch_size=5000 --ddtype=curriculum --cctype=2 --epoch=10000 --test_freq=10 --print_freq=10 --arch=convnet --window=60 --minwindow=0 --totwindow=200 --inner_optim=Adam --outer_optim=Adam --inner_lr=0.001 --lr=0.001 --syn_strategy=flip_rotate --real_strategy=flip_rotate --seed=0 --zca --comp_ipc=1 --class_area_width=32 --class_area_height=32 --poster_width=153 --poster_height=60 --poster_class_num_x=5 --poster_class_num_y=2

python main.py --name=PoDD-CIFAR100-LT1-90 --distill_batch_size=50 --patch_num_x=20 --patch_num_y=20 --dataset=cifar100 --num_train_eval=8 --update_steps=1 --batch_size=2000 --ddtype=curriculum --cctype=2 --epoch=10000 --test_freq=10 --print_freq=10 --arch=convnet --window=100 --minwindow=0 --totwindow=300 --inner_optim=Adam --outer_optim=Adam --inner_lr=0.001 --lr=0.001 --syn_strategy=flip_rotate --real_strategy=flip_rotate --seed=0 --zca --comp_ipc=1 --class_area_width=32 --class_area_height=32 --poster_width=303 --poster_height=303 --poster_class_num_x=10 --poster_class_num_y=10 --train_y

python main.py --name=PoDD-CUB200-LT1-90 --distill_batch_size=200 --patch_num_x=60 --patch_num_y=30 --dataset=cub-200 --num_train_eval=8 --update_steps=1 --batch_size=3000 --ddtype=curriculum --cctype=2 --epoch=10000 --test_freq=25 --print_freq=10 --arch=convnet --window=60 --minwindow=0 --totwindow=200 --inner_optim=Adam --outer_optim=Adam --inner_lr=0.001 --lr=0.001 --syn_strategy=flip_rotate --real_strategy=flip_rotate --seed=1 --zca --comp_ipc=1 --class_area_width=32 --class_area_height=32 --poster_width=610 --poster_height=302 --poster_class_num_x=20 --poster_class_num_y=10 --train_y

python main.py --name=PoDD_TinyImageNet-LT1-90 --distill_batch_size=30 --patch_num_x=40 --patch_num_y=20 --dataset=tiny-imagenet-200 --num_train_eval=8 --update_steps=1 --batch_size=500 --ddtype=curriculum --cctype=2 --epoch=10000 --test_freq=5 --print_freq=1 --arch=convnet --window=100 --minwindow=0 --totwindow=300 --inner_optim=Adam --outer_optim=Adam --inner_lr=0.0005 --lr=0.0005 --syn_strategy=flip_rotate --real_strategy=flip_rotate --seed=0 --zca --comp_ipc=1 --class_area_width=64 --class_area_height=64 --poster_width=1211 --poster_height=608 --poster_class_num_x=20 --poster_class_num_y=10 --train_y

• --patch_num_x and --patch_num_y - The number of extracted overlapping patches in the x and y axis of the poster.
• --poster_width and --poster_height - The width and height of the poster (controls the distillation data budget).
• --poster_class_num_x and --poster_class_num_y - The class layout dimensions within the poster as a 2d array (e.g., 10X10 or 20X5), (the product must be equal to the number of classes).
• --train_y - If set, the model will also optimize a set of learnable labels for the poster.

Although we use RaT-BPTT as the underlying distillation algorithm, using PoDD with other dataset distillation algorithms should be straight forward. The main change is replacing the distillation functionality in src/base.py and src/PoDD.py with the desired distillation algorithm.

If you find this useful for your research, please use the following.

@article{shul2024distilling,
  title={Distilling Datasets Into Less Than One Image},
  author={Shul, Asaf and Horwitz, Eliahu and Hoshen, Yedid},
  journal={arXiv preprint arXiv:2403.12040},
  year={2024}
}

• This repo uses RaT-BPTT as the underlying distillation algorithm, the implementation of RaT-BPTT is based on the following code found in their supplementary materials.