This repository presents the PyTorch code for PIP-Net (Patch-based Intuitive Prototypes Network).

Main Paper at CVPR: "PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification" introduces PIP-Net for natural images.
Medical applications, data quality inspection and manual corrections: Interpreting and Correcting Medical Image Classification with PIP-Net, applies PIP-Net to X-rays and skin lesion images where biases can be fixed by (manually) disabling prototypes.
Evaluation of part-prototype models like PIP-Net: The Co-12 Recipe for Evaluating Interpretable Part-Prototype Image Classifiers, presented at the XAI World Conference in July 2023.

PIP-Net is an interpretable and intuitive deep learning method for image classification. PIP-Net learns prototypical parts: interpretable concepts visualized as image patches. PIP-Net classifies an image with a sparse scoring sheet where the presence of a prototypical part in an image adds evidence for a class. PIP-Net is globally interpretable since the set of learned prototypes shows the entire reasoning of the model. A smaller local explanation locates the relevant prototypes in a test image. The model can also abstain from a decision for out-of-distribution data by saying “I haven’t seen this before”. The model only uses image-level labels and does not rely on any part annotations.

• PyTorch (incl torchvision, tested with PyTorch 1.13)
• tqdm
• scikit-learn
• openCV (optional, used to generate heatmaps)
• pandas
• matplotlib

PIP-Net can be trained by running main.py with arguments. Run main.py --help to see all the argument options. Recommended parameters per dataset are present in the used_arguments.txt file (usually corresponds to the default options).

Want to train PIP-Net on another dataset? Add your dataset in util/data.py by creating a function get_yourdata with the desired data augmentation (that captures human perception of similarity), add it to the existing get_data function in util/data.py and give your dataset a name. Use --dataset your_dataset_name as argument to run PIP-Net on your dataset.

Other relevant arguments are for example --weighted_loss which is useful when your data is imbalanced. In case of a 2-class task with presence/absence reasoning, you could consider using --bias to include a traininable bias term in the linear classification layer (which could decrease the OoD abilities) such that PIP-Net does not necessarily need to find evidence for the absence-class.

Check your --log_dir to keep track of the training progress. This directory contains log_epoch_overview.csv which prints statistics per epoch. File tqdm.txt prints updates per iteration and potential errors. File out.txt includes all print statements such as additional info. See the Interpreting the Results section for further details.

Visualizations of prototypes are included in your --log_dir / --dir_for_saving_images.

Various trained versions of PIP-Net are made available:

• PIP-Net with the ConvNext backbone (recommended) trained on the birds CUB-200-2011 dataset is available for download here (320MB). Download the CUB dataset (see instructions in this README) and run the following command to generate the prototypes and evaluate the model: python3 main.py --dataset CUB-200-2011 --epochs_pretrain 0 --batch_size 64 --freeze_epochs 10 --epochs 0 --log_dir ./runs/pipnet_cub --state_dict_dir_net ./pipnet_cub_trained. Update the path of --state_dict_dir_net to the checkpoint if needed.
• PIP-Net with the ResNet50 backbone trained on the birds CUB-200-2011 dataset is available for download here (280MB). Use --net resnet50.
• PIP-Net with the ConvNext backbone (recommended) trained on the CARS dataset is available for download here (320MB). Use --dataset CARS.
• PIP-Net with the ResNet50 backbone trained on the CARS dataset is available for download here (280MB).

The code can be applied to any imaging classification data set, 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 util/data.py.

For preparing CUB-200-2011 with 200 bird species, use util/preprocess_cub.py. For Stanford Cars with 196 car types, use the Instructions of ProtoTree.

During training, various files will be created in your --log_dir:

• log_epoch_overview.csv keeps track of the training progress per epoch. It contains accuracies, the number of prototypes, loss values etc. In case of a 2-class task, the third value is F1-score, otherwise this is top5-accuracy.
• out.txt collects the standard output from print statements. Its most relevant content is:
  • More performance metrics are printed, such as sparsity ratio. In case of a 2-class task, it also shows the sensitivity, specificity, confusion matrix, etc.
  • At the end of the file, after training, the relevant prototypes per class are printed. E.g., ``Class 0 has 5 relevant prototypes: [(prototype_id, class weight), ...]''. This information thus shows the learned scoring sheet of PIP-Net.
• tqdm.txt contains the progress via progress bar package tqdm. Useful to see how long one epoch will take, and how the losses evolve. Errors are also printed here.
• metadata folder logs the provided arguments.
• checkpoints folder contains state_dicts of the saved models.
• Prototype visualisations After training, various folders are created to visualise the learned reasoning of PIP-Net.
  • visualised_pretrained_prototypes_topk visualises the top-10 most similar image patches per prototype after the pretraining phase. Each row in grid_topk_all corresponds to one prototype. The number corresponds with the index of the prototype node, starting at 0.
  • visualised_prototypes_topk visualises the top-10 most similar image patches after the full (first and second stage) training. Prototypes that are not relevant to any class (all weights are zero) are excluded.
  • visualised_prototypes is a more extensive visualisation of the prototypes learned after training PIP-Net. The grid_xxx.png images show all image patches that are similar to prototype with index xxx. The number of image patches (or the size of the png file) already gives an indication how often this prototype is found in the training set. If you want to know where these image patches come from (to see some more context), you can open the corresponding folder prototype_xxx. Each image contains a yellow square indicating where prototype xxx was found, coresponding with an image patch in grid_xxx.png. The file name is pxxx_imageid_similarityscore_imagename_rect.png.
  • visualization_results (or other --dir_for_saving_images) contains predictions including local explanations for test images. A subfolder corresponding to a test image contains the test image itself, and folders with predicted classes: classname_outputscore. In such a class folder, it is visualised where which prototypes are detected: muliplicationofsimilarityandweight_prototypeindex_similarityscore_classweight_rect_or_patch.png.

• What is the best number of epochs for my dataset? The right number of epochs (--epochs and --epochs_pretrain) will depend on the data set size and difficulty of the classification task. Hence, tuning the parameters might require some trial-and-error. You can start with the default values. For datasets of different sizes, we recommend to set the number of epochs such that the number of iterations (i.e., weight updates) during the second training state is around 10,000 (rule of thumb). Hence, epochs = 10000 / (num_images_in_trainingset / batch_size). The number of iterations for one epoch is easily found in tqdm.txt. Similarly, the number of pretraining epochs --epochs_pretrain can be set such that there are 2000 weight updates.

• I have CUDA memory issues, what can I do? PIP-Net is designed to fit onto one GPU. If your GPU has less CUDA memory, you have the following options: 1) reduce your batch size --batch_size or --batch_size_pretrain. Set it as large as possible to still fit in CUDA memory. 2) freeze more layers of the CNN backbone. Rather than optimizing the whole CNN backbone from --freeze_epochs onwards, you could keep the first layers frozen during the whole training process. Adapt the code around line 200 in util/args.py as indicated in the comments there. Alternatively, set --freeze_epochs equal to --epochs. 3) Use --net convnext_tiny_13 instead of the default convnext_tiny_26 to make training faster and more efficient. The potential downside is that the latent output grid is less fine-grained and could therefore impact prototype localization, but the impact will depend on your data and classification task.

Please refer to our work when using or discussing PIP-Net:

Meike Nauta, Jörg Schlötterer, Maurice van Keulen, Christin Seifert (2023). “PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification.” IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

BibTex citation:

@article{nauta2023pipnet,
  title={PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification},
  author={Nauta, Meike and Schlötterer, Jörg and van Keulen, Maurice and Seifert, Christin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023},
}