lilujunai / FlatTrajectoryDistillation_FTD

The code of the paper "Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation" (CVPR2023)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation

Paper


Teaser image

This repo contains code for training expert trajectories and distilling synthetic data from our Dataset Distillation by FTD paper (CVPR 2023).

Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation
Jiawei Du*, Yidi Jiang*, Vincent Y. F. Tan, Joey tianyi Zhou, Haizhou Li
CFAR A*STAR, NUS
CVPR 2023

The task of "Dataset Distillation" is to learn a small number of synthetic images such that a model trained on this set alone will have similar test performance as a model trained on the full real dataset.

Accumulated Trajectory Error

Teaser image

State-of-the-art methods primarily rely on learning the synthetic dataset by matching the gradients obtained during training between the real and synthetic data. However, these gradient-matching methods suffer from the so-called accumulated trajectory error caused by the discrepancy between the distillation and subsequent evaluation. To mitigate the adverse impact of this accumulated trajectory error, we propose a novel approach that encourages the optimization algorithm to seek a flat trajectory.

The flat trajectory distillation (FTD) in purple line mitigates the so-called accumulated trajectory error than the baseline in blue line.


Getting Started

First, create the conda virtual enviroment

conda env create -f enviroment.yaml

You can then activate your conda environment with

conda activate distillation

Generating Expert Trajectories

Before doing any distillation, you'll need to generate some expert trajectories using .\buffer\buffer.py

The following command will train 100 ConvNet models on CIFAR-100 with ZCA whitening for 50 epochs each:

python buffer.py --dataset=CIFAR100 --model=ConvNet --train_epochs=50 --num_experts=100 --zca --buffer_path={path_to_buffer_storage} --data_path={path_to_dataset} --rho 0.01

There is an example in the .\buffer\run_buffer.sh

the default data and buffer storage path are .\data and .\buffer_storage

Distillation by Matching Training Trajectories

The following command will then use the buffers we just generated to distill CIFAR-100 down to just 10 image per class:

CUDA_VISIBLE_DEVICES=0 python distill_FTD.py --dataset=CIFAR100 --ipc=10 --syn_steps=20 --expert_epochs=2 --max_start_epoch=40 --zca \
    --lr_img=1000 --lr_lr=1e-05 --lr_teacher=0.01 --ema_decay=0.9995 --Iteration=5000 --buffer_path={path_to_buffer_storage} --data_path={path_to_dataset}

Please find a full list of hyper-parameters below:

About

The code of the paper "Minimizing the Accumulated Trajectory Error to Improve Dataset Distillation" (CVPR2023)


Languages

Language:Python 93.5%Language:Shell 5.3%Language:Perl 1.3%