Guided-AbsoluteGrad

This is the official implementation for Guided AbsoluteGrad. For more details, please refer to the paper Guided AbsoluteGrad: Magnitude of GradientsMatters to Explanation’s Localization and Saliency.

Quick Navigator

• cv_exp/gradient/_guided_absolute_grad.py: implementation of the Guided AbsoluteGrad;
• cv_exp/eval/rcap_eval.py: implementation of RCAP evaluation for saliency maps;

Data Preparation

Please run data_preparation.ipynb.

If everything is correct, you should have the following structure in the data:

The isic_exp/weights folder should contain:

The places365_exp/pretrained folder should contain:

Quick Demo for each Case

Please run imagenet_demo.ipynb, isic_demo.ipynb, places365_demo.ipynb.

Figures in the paper can be partially reproduced from the demo.

Experiment Reproduction

To reproduce the experiment, run the command:

python cv_exp/exp.py -t sa -mk [ModelKey] -dk [DatasetKey] -m [Mode] -c [CaseName] [SettingKeys]

1. ModelKey: The explained model key. It should be the key defined in the settings.py of each case.

   For instance, resnet50 is defined in the get_model() function of cases/imagenet_exp/settings.py.

2. DatasetKey: Similar to the ModelKey. It should be the key defined in the get_dataset() function.

3. Mode: should be one of:

   1. x: run xai explanation to get all saliency map results;
   2. e: run RCAP evaluation;
   3. a: run AUC evaluation;
   4. s: run the x and the e at the same time;
   5. sn: run sanity check evaluation;

4. CaseName: the name of cases. It is just for naming the experiment result dir. It could be anything.

5. SettingKeys: experiment keys defined in each case's settings.py file. If not provided, it will run for all experiments.

Example:

Run all experiments of the ImageNet case for XAI and the RCAP evaluation.

python main.py -mk resnet50 -dk imagenet_seg -t sa -m s -c imagenet

Run only the 2.a_places365 experiments of the Places365 case.

python main.py -mk densenet161 -dk places365 -t sa -m x -c places365 2.a_places365 

Experiment Evaluation & Result Observation

The RCAP should be calculated finally by the batch_rcap(...) function in the cv_exp/eval/rcap_eval.py

The evaluation_result.ipynb can run this function and plot all case results.

If all experiments are reproduced, one can verify the experiment results with the file evaluation_result_author_side.ipynb

Cite

@article{Huang2024Guided,
	author = {Huang, Jun and Liu, Yan},
	journal = {Proceedings of the Canadian Conference on Artificial Intelligence},
	year = {2024},
	month = {may 27},
	note = {https://caiac.pubpub.org/pub/8rzaolsn},
	publisher = {Canadian Artificial Intelligence Association (CAIAC)},
	title = {Guided {AbsoluteGrad} : Magnitude of {GradientsMatters} to {Explanation}\textquoteright{}s {Localization} and {Saliency}},
}