Skip to content

Latest commit

 

History

History
111 lines (90 loc) · 4.46 KB

README.md

File metadata and controls

111 lines (90 loc) · 4.46 KB
Raw

Integrated-Gradient Optimized Saliency Maps (iGOS++/I-GOS)

PyTorch implementation for a family of salieny (attribution) map methods that optimize for an explanation mask using integrated gradient.

  • Mingqi Jiang, Saeed Khorram, Li Fuxin. "Diverse Explanations for Object Detectors with Nesterov-Accelerated iGOS++", BMVC 2023.

Contributors:

This repository is published by @Lawsonty, @khorrams and @mingqiJ.

Dependencies

First install and activate a python3.6 virtual environment:

$ python3.6 -m venv env
$ source env/bin/activate

You can update the pip and install the dependencies using:

(env) $ pip install --upgrade pip
(env) $ pip install -r req.txt

Detector Preparation

Download the Mask R-CNN, Faster R-CNN and YOLOv3-SPP, then put them in the weight/:

weight/
  maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth
  fasterrcnn_resnet50_fpn_coco-258fb6c6.pth
  yolov3-spp-ultralytics-512.pt

Quick Start

To generate explanations, you can simply run:

(env) $ python main.py --method <I-GOS/iGOS++> --opt <LS/NAG> --data <path to images>

where I-GOS and iGOS++ are the explanations methods, LS (Line Search) and NAG (Nesterov Accelerated Gradient) are the optimization methods, and the --data defines the path to the images.

The hyperparameters of our method can be directly passed as arguments when running the code:

Classification

(env) $ python main.py --method iGOS++ --opt NAG --data samples/ --dataset imagenet --input_size 224
--size 224 --model vgg19 --L1 10 --L2 20 --ig_iter 20 --iterations 20 --alpha 1000

Object Detection and Instance Segmentation

(env) $ python main.py --method iGOS++ --opt NAG --data samples/ --dataset coco --input_size 800
--size 100 --model m-rcnn --model_file ./weight/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth --L1 0.1 
--L2 20 --ig_iter 5 --iterations 5 --alpha 10

Diverse Initialization

To generate explanations with different initializations, you can run:

(env) $ python main.py --method iGOS++ --opt NAG --data samples/ --dataset coco --input_size 800
--size 100 --model m-rcnn --model_file ./weight/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth --L1 0.1 
--L2 20 --ig_iter 5 --iterations 5 --alpha 10 --diverse_k 2 --init_val 0.2 --init_posi 2

Citation

If you use this code for your research, please consider citing our papers:

@inproceedings{jiang2023nagigos++,
  title={Diverse Explanations for Object Detectors with Nesterov-Accelerated iGOS++},
  author={Jiang, Mingqi and Khorram, Saeed and Fuxin, Li},
  booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
  year={2023}
}

@inproceedings{khorram2021igos++,
author = {Khorram, Saeed and Lawson, Tyler and Fuxin, Li},
title = {IGOS++: Integrated Gradient Optimized Saliency by Bilateral Perturbations},
year = {2021},
isbn = {9781450383592},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3450439.3451865},
doi = {10.1145/3450439.3451865},
booktitle = {Proceedings of the Conference on Health, Inference, and Learning},
pages = {174–182},
keywords = {COVID-19, chest X-ray, saliency methods, medical imaging, integrated-gradient, explainable AI},
series = {CHIL '21}
}

@inproceedings{qi2020visualizing,
  title={Visualizing Deep Networks by Optimizing with Integrated Gradients},
  author={Qi, Zhongang and Khorram, Saeed and Fuxin, Li},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={34},
  number={07},
  pages={11890--11898},
  year={2020}
}

Acknowledgement

Some parts of the code are borrowed from TorchVision and yolov3_spp_gradcam. We thank the authors for their excellent work.