SRe2L

SRe2L是一个用于数据集蒸馏的项目，旨在从大规模原始数据集中压缩信息到一个新的紧凑数据集，同时尽可能保留原始数据信息的最大程度。该项目通过匹配原始数据和蒸馏数据之间的中间统计数据（如权重轨迹、特征、梯度、BatchNorm等）来实现数据集的蒸馏。SRe2L方法在处理大型数据集时，如ImageNet-1K，能够有效地压缩数据集并保持其性能特征。此外，训练传统深度模型使用蒸馏后的图像，可以在原始验证数据上实现显著优于先前数据集蒸馏方法的测试准确性。

SRe2L is a project for dataset distillation, which aims to compress information from large-scale original datasets into a new compact dataset while retaining the maximum possible amount of original data information. This project achieves dataset distillation by matching intermediate statistics between original and distilled data, such as weight trajectories, features, gradients, BatchNorm, etc. The SRe2L method can effectively compress datasets and preserve their performance characteristics when processing large-scale datasets such as ImageNet-1K. Additionally, training conventional deep models with distilled images can achieve test accuracy on the original validation data that significantly outperforms that of previous dataset distillation methods.