This repository contains the code for experiments in following paper
Zhi Chen, Yijie Bei, Cynthia Rudin
Concept Whitening for Interpretable Image Recognition
Nature Machine Intelligence 2, 772–782 (2020). https://doi.org/10.1038/s42256-020-00265-z
arXiv pre-publication version
We have adapted skeleton code from https://github.com/Jongchan/attention-module in train_places.py, and adapted IterNrom implementation from https://github.com/huangleiBuaa/IterNorm as part of the implementation
Most of our contributions can be found in the implementation of IterNormRotation class from /MODELS/iterative_normalization.py. Useful code for producing experimental results and visualizations are found in /train_imagenet.py and /plot_functions.py. /scripts folder contains shell scripts using which experiments are done. Also under the /scripts folder, there are example checkpoints we saved during experiments.
PyTorch (1.1.0), torchvision (0.3.0), NumPy (1.18.1), sklearn (0.20.3), matplotlib (3.1.3), PIL (6.2.1), Seaborn (0.9.0), skimage (0.15.0).
NVIDIA Tesla P-100 GPUs or NVIDIA Tesla K-80 GPUs
data_256
├── concept_train
│ ├── airplane
│ │ ├── airplane
│ ├── bed
│ │ ├── bed
│ ├── desk
│ │ ├── desk
│ ├── fridge
│ │ ├── fridge
│ ├── lamp
│ │ ├── lamp
│ ├── person
│ │ ├── person
│ ├── sofa
│ │ ├── sofa
│ └── ......
├── concept_test
│ ├── airplane
│ ├── bed
│ ├── desk
│ ├── fridge
│ ├── lamp
│ ├── person
│ ├── sofa
│ └── ......
├── test
│ ├── airfield
│ ├── airplane_cabin
│ ├── airport_terminal
│ ├── alcove
│ ├── alley
│ ├── amphitheater
│ ├── amusement_arcade
│ ├── amusement_park
│ ├── apartment_building_outdoor
│ ├── aquarium
│ ├── yard
│ ├── youth_hostel
│ └── ......
├── train
│ ├── airfield
│ ├── airplane_cabin
│ ├── airport_terminal
│ ├── alcove
│ ├── alley
│ ├── amphitheater
│ ├── amusement_arcade
│ ├── amusement_park
│ ├── apartment_building_outdoor_outdoor
│ ├── aquarium
│ ├── aqueduct
│ ├── arcade
│ ├── arch
│ ├── archaelogical_excavation
│ ├── archive
│ ├── arena_hockey
│ ├── arena_performance
│ ├── bamboo_forest
│ ├── bank_vault
│ ├── banquet_hall
│ ├── bar
│ ├── barn
│ ├── barndoor
│ ├── baseball_field
│ ├── basement
│ ├── basketball_court_indoor
│ ├── bathroom
│ ├── bazaar_indoor
│ ├── bazaar_outdoor
│ ├── youth_hostel
│ └── ......
└── val
├── airfield
├── airplane_cabin
├── airport_terminal
├── alcove
├── alley
├── amphitheater
├── amusement_arcade
├── amusement_park
├── apartment_building_outdoor_outdoor
├── aquarium
├── aqueduct
├── arcade
├── arch
├── archaelogical_excavation
├── archive
├── arena_hockey
├── arena_performance
├── arena_rodeo
├── army_base
├── art_gallery
├── artists_loft
├── art_school
├── ......
An example dataset folder structure looks like the above, where in each bottom level folder, there are images. Notice that during training, a sampler will randomly sample from dataset folders for images. Notice that in the above folder structure, the list of subfolders has been cut short and the complete dataset is more extensive.
There generally two types of dataset, main objective dataset and auxiliary concept dataset.
Main dataset: We mainly use Places365 as the main dataset, and it can be downloaded from Here (alt). It should be divided into train, test, validation sets and stored in corresponding folders shown by the example dataset folder structure above.
Concept dataset: We mainly use objects in MS COCO as our auxiliary concept dataset, and it can be downloaded from Here. Each annotation, e.g., “person” in MS COCO, was used as one concept, and we selected all the images with this annotation (images having “person” in it), cropped them using bounding boxes and used the cropped images as the data representing the concept. The preprocessing code of the COCO dataset is provided in cropping_images_COCO.py. After downloading the 2017 COCO dataset, one can extract the concept images by running
python3 cropping_images_COCO.py -coco-path <coco_dataset_folder> -concept-path <folder_containing_concept_datasets>
. However, our model generalize to various concepts as well. It should be divided into train and test and stored in concept_train/ and /concept_test as is shown by the example dataset folder structure above. Note that in order to load data easily, the structures of the two folders are different: /concept_train allows loading images of one concept while /concept_test allows loading images of all concepts.
We also use the ISIC dataset in the experiments, and it can be downloaded from Here. One can download the entire dataset using this downloader code. The attributes of the lesion images, stored as metadata, such as "age", are used to define the concepts. After downloading the ISIC dataset, one can run
python3 extract_ISIC_concepts.py
to create the dataset used in the paper and extract the concepts (age<20 and size>=10mm) from the raw ISIC images.
To accelerate training process of our model, we utilize the pretrained weights of standard CNNs trained on Places365. The weights of those models can be downloaded from Here
Once the weights are downloaded, please put them under the folder checkpoints/
The weights of models trained with CW can be downloaded from Here. The folders are named by concepts trained in CW. For example airplane_bed_person/ contains weights when CW is trained with concepts "airplane", "bed" and "person".
Once the weights are downloaded, please put the folders of weights under checkpoints/
train_places.py has many arguments that one can use, here are some interesting ones:
--arch: resnet_cw, resnet_original, densnet_cw, densnet_original, vgg16_cw, vgg16_bn_original
--whitened_layers: refers to where the concept whitening module is added in the architecture (see explanation in example)
--concepts: comma delimited list of concepts that needs to be disentangled (see example below)
--act_mode: mean, max, pos_mean, pool_max (refer to paper for explanation)
Example training invocation scripts are inside the /scripts folder. A typical configuration looks like:
python3 ../train_places.py --ngpu 1 --workers 4 --arch resnet_cw --depth 18 --epochs 200 --batch-size 64 --lr 0.05 --whitened_layers 5 --concepts airplane,bed,person --prefix RESNET18_PLACES365_CPT_WHITEN_TRANSFER /data_256
This will start a training that adds concept whitening module to only the fifth residual block of the of the resnet 18. Note that a resnet 18 structure only has 8 residual blocks, and therefore the valid configuration numbers for resnet18 architecure are only 1 through 8. Note that when configuration number is 8, it corresponds to the 16th layer of the ResNet18 instead of 8th layer. Similarly for resnet 50, the valid configuration numbers are only 1 through 16. Currently we only support adding concept whitening to one layer.
The concepts to be disentangled will be specified through
--concepts airplane,bed,person
Note that these corresponds to directories in the /dataset_256/concept_train and /dataset_256/concept_test folder.
Similarly, example testing invocation scripts are inside the /scripts folder, and testing is done by invoking:
python train_places.py --ngpu 1 --workers 2 --arch resnet_cw --depth 18 --epochs 200 --batch-size 64 --lr 0.1 --whitened_layers 5 --concepts airplane,bed,person --prefix RESNET18_PLACES365_CPT_WHITEN_TRANSFER --resume ./checkpoints/RESNET18_PLACES365_CPT_WHITEN_TRANSFER_model_best.pth.tar ./data_256 --evaluate plot_top50
which plot the 50 top activated images along the concept axes.