Exploring Inconsistent Knowledge Distillation for Object Detection with Data Augmentation

This repository contains the official PyTorch implementation of the following paper at ACMMM 2023:

Exploring Inconsistent Knowledge Distillation for Object Detection with Data Augmentation
Jiawei Liang, Siyuan Liang, Aishan Liu, Ke Ma, Jingzhi Li, Xiaochun Cao
https://arxiv.org/abs/2209.09841

Installation

Install python (python == 3.8)
Install pytorch (pytorch == 2.0.0)
Install mmcv_full (mmcv_full == 1.6.0)

Install mmdetection (mmdetection == 2.25.0) from source code

  git clone https://github.com/JWLiang007/IKD.git
  cd IKD/
  pip install -r requirements/optional.txt
  pip install -v -e .

Download Dataset and Checkpoint

Download MS COCO2017 dataset
Unzip COCO dataset into data/coco/ in mmdetection/
Download pretrained teacher model retinanet_x101_64x4d_fpn_1x_coco_20200130-366f5af1.pth from the repository of mmdetection
Put the downloaded pretrained model into checkpoints/ in mmdetection/

Generate Adversarial Examples

# single GPU
python tools/ta_GT.py configs/retinanet/retinanet_x101_64x4d_fpn_1x_coco.py  checkpoints/retinanet_x101_64x4d_fpn_1x_coco_20200130-366f5af1.pth --method difgsm  --show-dir data/adv_rtn_coco_8_5 --gen_adv_aug --eps 8 --alpha 2 --steps 5
# multi GPU
bash tools/dist_adv.sh  configs/retinanet/retinanet_x101_64x4d_fpn_1x_coco.py  checkpoints/retinanet_x101_64x4d_fpn_1x_coco_20200130-366f5af1.pth  8  --method difgsm  --show-dir data/adv_rtn_coco_8_5 --gen_adv_aug --eps 8 --alpha 2 --steps 5

Train

#single GPU

# Step 1: train with DFA 
python tools/train.py configs/fgd/DFA_fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py
# Step 2: resume from epoch 16 and train without DFA
python tools/train.py configs/fgd/fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py --resume-from work_dirs/DFA_fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco/epoch_16.pth

#multi GPU

# Step 1: train with DFA 
bash tools/dist_train.sh configs/fgd/DFA_fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py 8
# Step 2: resume from epoch 16 and train without DFA
bash tools/dist_train.sh configs/fgd/fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py 8 --resume-from work_dirs/DFA_fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco/epoch_16.pth

Test

#single GPU
python tools/test.py configs/fgd/DFA_fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py $PATH_CHECKPOINT --eval bbox

#multi GPU
bash tools/dist_test.sh configs/fgd/DFA_fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py $PATH_CHECKPOINT 8 --eval bbox

Generalizability

Backdoor Defense

Index	Method	ASR	ASR Drop
1	Victim	96.7	-
2	NAD	82.88	13.82
3	Ours	78.26	18.44(↑33%)

For more recent progress in backdoor defense, please refers to the following repo:

https://github.com/JWLiang007/BD_DeCLIP.git

and switches to the bd branch.

Acknowledgement

Our code is based on the project MMDetection.