Awesome

HINT

PyTorch implementation for the paper:

HINT: Hierarchical Neuron Concept Explainer, CVPR 2022.

We propose HIerarchical Neuron concepT explainer (HINT) to effectively build bidirectional associations between neurons and hierarchical concepts in a low-cost and scalable manner. HINT enables us to systematically and quantitatively study whether and how the implicit hierarchical relationships of concepts are embedded into neurons, such as identifying collaborative neurons responsible to one concept and multimodal neurons for different concepts, at different semantic levels from concrete concepts (e.g., dog) to more abstract ones (e.g., animal).

Pipeline

Dependencies

• Check the required python packages in requirements.txt.

Run

• Execute Get_feature_maps_and_saliency_maps_of_dif_concepts.py to get feature maps and saliency maps (the images are from ImageNet train set) of the chosen layer(s) of the chosen model(s) (pre-trained on ImageNet).

• Based on previous step, execute Get_responsible_regions.py to get the responsible regions of different concepts.

• Run Get_concept_classifiers_and_shap_values.ipynb to train concept classifiers and calculate Shapley Values of neurons (which indicates neuron contribution to concept).

• Run Object_localization_using_concept_classifier_ILSVRC.ipynb, Object_localization_using_concept_classifier_PASCAL_VOC.ipynb, and Object_localization_using_concept_classifier_CUB-200-2011.ipynb to see the performance of object localization using concept classifiers (see folder output/concept_clf) on different datasets.

Datasets

• ImageNet

I put ImageNet dataset under /data/ImageNet_ILSVRC2012/

/data/ImageNet_ILSVRC2012/
    /ILSVRC2012_train
        /n07715103
            n07715103_8433.JPEG
            ...
        ...
    /val
        ILSVRC2012_val_00040001.JPEG
        ILSVRC2012_val_00040001.xml
        ...
    ...

• ImageNet sample data (with password 2333)

I put it under /data/imagenet_sample_5000/

• Pascal VOC

I put Pascal VOC dataset under /data/Pascal_VOC_dataset/VOCdevkit/VOC2007$ (the path name may seem wired...)

/data/Pascal_VOC_dataset/VOCdevkit/VOC2007
    /Annotations
        009279.xml
        ...
    /ImageSets
        /Layout
        /Main
        /Segmentation
    /JPEGImages
        009279.jpg
        ...
    /SegmentationClass
        001457.png
        ...
    /SegmentationObject
        009654.png
        ...

• CUB-200-2011

I put CUB-200-2011 dataset under /data/CUB-200-2011$

/data/CUB-200-2011
    /CUB_200_2011
        /attributes
        bounding_boxes.txt
        classes.txt
        image_class_labels.txt
        /images
            /001.Black_footed_Albatross
                Black_Footed_Albatross_0001_796111.jpg
                ...
            ...
        images.txt
        /parts
        README
        train_test_split.txt
    /segmentations
        /001.Black_footed_Albatross
            Black_Footed_Albatross_0001_796111.png
            ...
        ...
    attributes.txt
    README.txt

Results

Responsible neurons for hierarchical concepts

Object localization

Contact

If you have any questions, please contact wangad@connect.hku.hk.