akira-l / ELP

Code for A Simple Episodic Linear Probe Improves Visual Recognition in the Wild

ELP

Introduction

This project is an implementation of [A Simple Episodic Linear Probe Improves Visual Recognition in the Wild]

Insight

• A simple online linear probe can boost recognition performances. The simple regularization leads to better performances without complex network designs or additional data.

Requirements

Python 3 & Pytorch >= 0.4.0

Datasets Orgnization

Similar to DCL.

Training

Run train.py to train ELP.

For CUB / STCAR / AIR

python train.py --data $DATASET --epoch 360 --backbone resnet50 \
                    --tb 16 --tnw 16 --vb 512 --vnw 16 \
                    --lr 0.0008 --lr_step 60 \
                    --cls_lr_ratio 10 --start_epoch 0 \
                    --detail training_descibe --size 512 \
                    --crop 448 

For ImageNet

python train.py --data CUB --epoch 100 --backbone resnet50 \
                    --tb 1024 --tnw 16 --vb 2048 --vnw 16 \
                    --lr 0.01 --lr_step 10 \
                    --cls_lr_ratio 10 --start_epoch $LAST_EPOCH \
                    --detail training_descibe4checkpoint --size 256 \
                    --crop 224 

You can rewrite line 98-125 in utils/train_model.py for your own codebase.

Citation

Please cite ELP paper if you find ELP is helpful in your work:

@InProceedings{liang2022elp,
title={A Simple Episodic Linear Probe Improves Visual Recognition in the Wild},
author={Liang, Yuanzhi and Zhu, Linchao and Wang, Xiaohan and Yang, Yi},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022}
}