This is the official PyTorch implementation of the Iterated Integrated Attributrions (IIA) method.
We introduce a novel method that enables visualization of predictions made by vision models, as well as visualization of explanations for a specific class. In this method, we present the concept of iterated integrated attributrions.
Images should be stored in the data\ILSVRC2012_img_val directory.
The information on the images being evaluated and their target class can be found in the data\pics.txt file, with each line formatted as follows: <file_name> target_class_number (e.g. ILSVRC2012_val_00002214.JPEG 153).
To generate saliency maps using our method on CNN, run the following command:
python cnn_saliency_map_generator.py
And to produce maps for ViT, run:
python vit_saliency_map_generator.py
By default, saliency maps for CNNs are generated using the Resnet101 network on the last layer. You can change the selected network and layer by modifying the model_name and FEATURE_LAYER_NUMBER variables in the cnn_saliency_map_generator.py class. For ViT, the default is ViT-Base, which can also be configured using the model_name variable in the vit_saliency_map_generator.py class.
The generated saliency maps will be stored in the qualitive_results directory.
- ViT-B Link to download
- ViT-S Link to download
- Download imagenet_dataset Link to download dataset
- Download the COCO_Val2017 Link to download dataset
- Download Pascal_val_2012 Link to download dataset
To run the image segmentation, download the VOC and COCO datasets to data/VOC and data/COCO respectively.
For Imagenet segmentation, use the dataset provided in the link (Link to download dataset).
To run segmentation using ViT, configure the chosen_dataset variable in the seg_vit_datasets class (the default is COCO) and run the following command:
python seg_vit_datasets.py
To run segmentation using CNN, specify the desired dataset {dataset} (which can be imagenet, coco, or voc) in the following command and run it:
python segmentation_cnn_{dataset}.py
For comparison, we used the following implementations of code from git repositories: