LeGrad
Project description
LeGrad
An Explainability Method for Vision Transformers via Feature Formation Sensitivity
Walid Bousselham, Angie Boggust, Sofian Chaybouti,Hendrik Strobelt Hilde Kuehne
Vision-Language foundation models have shown remarkable performance in various zero-shot settings such as image retrieval, classification, or captioning. we propose LeGrad, an explainability method specifically designed for ViTs. We LeGrad we explore how the decision-making process of such models by leveraging their feature formation process. A by-product of understanding VL models decision-making is the ability to produce localised heatmap for any text prompt.
The following is the code for a wrapper around the OpenCLIP library to equip VL models with LeGrad.
:hammer: Installation
legrad
library can be simply installed via pip:
$ pip install legrad_torch
Demo
- Try out our web demo on HuggingFace Spaces
- Run the demo on Google Colab:
- Run
playground.py
for a usage example.
To run the gradio app locally, first install gradio and then run app.py
:
$ pip install gradio
$ python app.py
Usage
To see which pretrained models is available use the following code snippet:
import legrad
legrad.list_pretrained()
Single Image
To process an image and a text prompt use the following code snippet:
Note: the wrapper does not affect the original model, hence all the functionalities of OpenCLIP models can be used seamlessly.
import requests
from PIL import Image
import open_clip
import torch
from legrad import LeWrapper, LePreprocess
from utils import visualize
# ------- model's paramters -------
model_name = 'ViT-B-16'
pretrained = 'laion2b_s34b_b88k'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# ------- init model -------
model, _, preprocess = open_clip.create_model_and_transforms(
model_name=model_name, pretrained=pretrained, device=device)
tokenizer = open_clip.get_tokenizer(model_name=model_name)
model.eval()
# ------- Equip the model with LeGrad -------
model = LeWrapper(model)
# ___ (Optional): Wrapper for Higher-Res input image ___
preprocess = LePreprocess(preprocess=preprocess, image_size=448)
# ------- init inputs: image + text -------
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = preprocess(Image.open(requests.get(url, stream=True).raw)).unsqueeze(0).to(device)
text = tokenizer(['a photo of a cat']).to(device)
# -------
text_embedding = model.encode_text(text, normalize=True)
print(image.shape)
explainability_map = model.compute_legrad_clip(image=image, text_embedding=text_embedding)
# ___ (Optional): Visualize overlay of the image + heatmap ___
visualize(heatmaps=explainability_map, image=image)
:star: Acknowledgement
This code is build as wrapper around OpenCLIP library from LAION, visit their repo for more vision-language models. This project also takes inspiration from Transformer-MM-Explainability and the timm library, please visit their repository.
:books: Citation
If you find this repository useful, please consider citing our work :pencil: and giving a star :star2: :
@article{bousselham2024legrad,
author = {Bousselham, Walid and Boggust, Angie and Chaybouti, Sofian and Strobelt, Hendrik and Kuehne, Hilde}
title = {LeGrad: An Explainability Method for Vision Transformers via Feature Formation Sensitivity},
journal = {arXiv},
year = {2024},
}
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for legrad_torch-1.0-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 56cf378ab78c332a5f667235a591745adf40d4706c7e37990d3d01a73163847f |
|
MD5 | 61ee3478bafea7e73511db3fa5250366 |
|
BLAKE2b-256 | 9fa5117d7d280926a76e48ddb36b40fb1e5b1b15dd1e0e2d25cac88180c57c07 |