kenJPG / ExplainAnything

Explain Anything 🎩

Explaining any model through user-defined concepts

---

Setup

Dockerfile and run scripts are provided. Simply build with

docker build . -t explain_anything

and run the container via

sudo bash run_docker.sh

Please note that weights for the examples (only for the example ResNet and YOLO model, SegGPT is provided via Dockerfile) are not available in this repo and will have to be re-trained.

Additionally, the datasets are removed however, the classification dataset CelebAMask-HQ is kept here for demonstration purposes (not full size, roughly 400 images).

Usage

For best demonstration and understanding of how to use, refer to the notebooks found in examples.

Explaining classification scores can be found in examples/classification. The dataset used is CelebA. Explaining object detection through detection scores can be found in examples/detection. The dataset used here is NEA Rodent dataset.

Here are some high level API snippets:

# We have some hand labelled segmentations. Let's use these to generate some more!
from explain_anything.segmenters import FeatureSegmenter

fs = FeatureSegmenter(
	seggpt_checkpoint = '/seggpt_vit_large.pth'
	feature_count = 3
)

support_images = ['img1.jpg', 'img2.jpg']
support_masks = ['img1_mask.png', 'img2_mask.png']
for s_img, s_mask in zip(support_images, support_masks):
	fs.add_support_data(
		s_img, s_mask
	)

for image in ['img3.jpg', 'img4.jpg', 'img5.jpg']:
	fs.generate_segmentation(
		input_image_path = image,
		output_dir = '/output_masks'
	)

Once we have gotten more samples and features, we can evaluate our model

from explain_anything.explain import ExA
model = resnet50() # Note this model should only return a confidence score

exa = ExA()
features = ['leaves', 'trunk', 'roots']

importances = exa.ablate(
	model,
	['img3.jpg', 'img4.jpg', 'img5.jpg'],
	'/output_masks',
	features
)

for importance, feature_name in list(sorted(zip(importances, ['background] + features)))[::-1]:
	print(f'{feature_name}:', importance)