Bor-Shiun Wang, Chien-Yi Wang*, Wei-Chen Chiu*

^{*Equal Advising}

Official PyTorch implementation of CVPR 2024 paper "MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes".

[Paper] [Supplementary] [Website] [arXiv] [Youtube] [BibTeX]

For business inquiries, please visit our website and submit the form: NVIDIA Research Licensing.

Recent advancements in post-hoc and inherently interpretable methods have markedly enhanced the explanations of black box classifier models. These methods operate either through post-analysis or by integrating concept learning during model training. Although being effective in bridging the semantic gap between a model's latent space and human interpretation, these explanation methods only partially reveal the model's decision-making process. The outcome is typically limited to high-level semantics derived from the last feature map. We argue that the explanations lacking insights into the decision processes at low and mid-level features are neither fully faithful nor useful. Addressing this gap, we introduce the Multi-Level Concept Prototypes Classifier (MCPNet), an inherently interpretable model. MCPNet autonomously learns meaningful concept prototypes across multiple feature map levels using Centered Kernel Alignment (CKA) loss and an energy-based weighted PCA mechanism, and it does so without reliance on predefined concept labels. Further, we propose a novel classifier paradigm that learns and aligns multi-level concept prototype distributions for classification purposes via Class-aware Concept Distribution (CCD) loss. Our experiments reveal that our proposed MCPNet while being adaptable to various model architectures, offers comprehensive multi-level explanations while maintaining classification accuracy. Additionally, its concept distribution-based classification approach shows improved generalization capabilities in few-shot classification scenarios.

• Pytorch(includa torchvision, test with Pytorch 1.13.1)
• Matplotlib
• OpenCV
• NumPy
• tqdm
• argparse
• easydict
• importlib

You can also simply build the environment via .yml file.

conda env create -f ./environment.yml

The code can be applied to any imaging classification dataset, structured according to the Imagefolder format:

root/class1/xxx.png
root/class1/xxy.png
root/class2/xyy.png
root/class2/yyy.png

Add or update the paths to your dataset in arg_reader.py.

python -m torch.distributed.launch --nproc_per_node=1 --master_port 9560 train.py --index AWA2_test --model ResNet --basic_model resnet50_relu --device 1 --dataset_name AWA2 --margin 0.01 --concept_cha 32 32 32 32 --concept_per_layer 8 16 32 64 --optimizer adam

To visualize the top-k response images of each concept, we first have to calculate the weighted covariance matrix and weighted means, which are used to calculate the concept prototypes via weighted PCA.

python extract_prototypes.py --case_name AWA2_test --device 0 --model ResNet --basic_model resnet50_relu --concept_per_layer 8 16 32 64 --cha 32 32 32 32

The above code will calculate each concept prototype's weighted covariance matrix and mean.

Next, the following code will scan the whole training set to find the highest k point for each prototype from the training set. There might be multi-selected points from the same image. Select multiple candidates for each prototype to prevent the visualization from showing the same image. **The code will store the image index list indexed by ImageFolder, so make sure the image set will be the same in this and next step; also, set the shuffle=False for these two steps.

python ./vis_utils/find_topk_response.py --case_name AWA2_test --model ResNet --basic_model resnet50_relu --concept_per_layer 8 16 32 64 --cha 32 32 32 32 --device 0 --eigen_topk 1

The final step stores the top-k concept prototype result. Each image will only be selected once to present for each prototype. Precisely, the image of the top responses won't be duplicated for each prototype.

python ./vis_utils/find_topk_area.py --case_name AWA2_test --model ResNet --basic_model resnet50_relu --concept_per_layer 8 16 32 64 --cha 32 32 32 32 --topk 5 --device 0 --eigen_topk 1 --masked --heatmap --individually

To evaluate the performance of the trained MCPNet.

(Optional) Firstly, extract the concept prototypes (If the concept prototypes have already been extracted, this step can be passed).

python extract_prototypes.py --case_name AWA2_test --device 0 --model ResNet --basic_model resnet50_relu --concept_per_layer 8 16 32 64 --cha 32 32 32 32

Next, the class Multi-level Concept Prototypes distribution (MCP distribution) was calculated.

python ./classify_utils/cal_class_MCP.py --case_name AWA2_test --device 0 --model ResNet --basic_model resnet50_relu --concept_mode pca --concept_per_layer 8 16 32 64 --cha 32 32 32 32 --all_class

Finally, classify the image by matching the images' MCP distribution to the closest class MCP distribution.

python ./classify_utils/cal_acc_MCP.py --case_name AWA2_test --model ResNet --basic_model resnet50_relu --device 0 --concept_per_layer 8 16 32 64 --cha 32 32 32 32 --all_class

Bor-Shiun Wang: [email protected]

Chien-Yi Wang: [email protected]

@InProceedings{Wang2024MCPNet,
  author    = {Wang, Bor-Shiun and Wang, Chien-Yi and Chiu, Wei-Chen},
  title     = {MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2024},
  pages     = {10885-10894}
}

Copyright © 2024, NVIDIA Corporation. All rights reserved.

This work is made available under the NVIDIA Source Code License-NC. Click here to view a copy of this license.