CausalVLR is a python open-source framework for causal relation discovery, causal inference that implements state-of-the-art causal learning algorithms for various visual-linguistic reasoning tasks, such as VQA, Image/Video Captioning, Model Generalization and Robustness, Medical Report Generation, etc.
📄 Table of Contents
📄 Table of Contents📚 Introduction🚀 What's New👨🏫 Get Started👀 Model Zoo🎫 License🖊️ Citation🙌 Contributing🤝 Acknowledgement🏗️ Projects in HCPLab
📚 Introduction 🔝
Major features
-
Modular Design
We decompose the causal framework of visual-linguistic tasks into different components and one can easily construct a customized causal-reasoning framework by combining different modules.
-
Support of multiple tasks
The toolbox directly supports multiple visual-linguistic reasoning tasks such as VQA, Image/Video Caption, Medical Report Generation, Model Generalization and Robustness and so on.
-
State of the art
The toolbox stems from the codebase developed by the HCPLab team, who dedicated to solving a variety of complex logic tasks through causal reasoning, and we keep pushing it forward.
🚀 What's New 🔝
🔥 2023.6.29.
- v0.0.1 was released in 6/30/2023
- Support VLCI for Medical Report Generation task
- Support CMCIR (T-PAMI 2023) for Event-Level Visual Question Answering task
- Support VCSR for Visual Causal Scene Discovery task
- Support Robust Fine-tuning (CVPR 2023) for Model Generalization and Robustness
✨ VLCI-Visual Causal Intervention for Radiology Report Generation
| Dataset | B@1 | B@2 | B@3 | B@4 | Meteor | Rough-L | CIDEr |
|---|---|---|---|---|---|---|---|
| IU-Xray | 50.5 | 33.4 | 24.5 | 18.9 | 20.4 | 39.7 | 45.6 |
| MIMIC-CXR | 39.6 | 24.3 | 16.3 | 11.7 | 14.9 | 28.1 | 15.7 |
✨ CMCIR-Cross-modal Causal Intervention for Event-level Video Question Answering
| Method | Basic | Attribution | Introspection | Counterfactual | Forecasting | Reverse | All |
|---|---|---|---|---|---|---|---|
| VQAC | 34.02 | 49.43 | 34.44 | 39.74 | 38.55 | 49.73 | 36.00 |
| MASN | 33.83 | 50.86 | 34.23 | 41.06 | 41.57 | 50.80 | 36.03 |
| DualVGR | 33.91 | 50.57 | 33.40 | 41.39 | 41.57 | 50.62 | 36.07 |
| HCRN | 34.17 | 50.29 | 33.40 | 40.73 | 44.58 | 50.09 | 36.26 |
| CMCIR | 36.10 (+1.93) | 52.59 (+1.73) | 38.38 (+3.94) | 46.03 (+4.64) | 48.80 (+4.22) | 52.21 (+1.41) | 38.58 (+1.53) |
👨🏫 Getting Started 🔝
Please see Overview for the general introduction of Causal-VLR.
For detailed user guides and advanced guides, please refer to our documentation, and here is the code structure of toolbox.
Installation
Please refer to Installation for installation instructions in documentation.
Briefly, to use Causal-VLR, we could install it using pip:
pip install Causal-VLR
Running examples
For causal discovery, there are various running examples in the ‘tests’ directory.
For the implemented modules, we provide unit tests for the convenience of developing your own methods.
👀 Model Zoo 🔝
Please feel free to let us know if you have any recommendation regarding datasets with high-quality. We are grateful for any effort that benefits the development of causality community.
| Task | Model | Benchmark |
|---|---|---|
| Medical Report Generation | VLCI | IU-Xray, MIMIC-CXR |
| VQA | CMCIR | SUTD-TrafficQA, TGIF-QA, MSVD-QA, MSRVTT-QA |
| Visual Causal Scene Discovery | VCSR | NExT-QA, Causal-VidQA, and MSRVTT-QA |
| Model Generalization and Robustness | Robust Fine-tuning | ImageNet-V2, ImageNet-R, ImageNet-Sketch, ObjectNet, ImageNet-A |
🎫 License 🔝
This project is released under the Apache 2.0 license.
🖊️ Citation🔝
If you find this project useful in your research, please consider cite:
@misc{liu2023causalvlr,
title={CausalVLR: A Toolbox and Benchmark for Visual-Linguistic Causal Reasoning},
author={Yang Liu and Weixing Chen and Guanbin Li and Liang Lin},
year={2023},
eprint={2306.17462},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
🙌 Contribution 🔝
Please feel free to open an issue if you find anything unexpected. We are always targeting to make our community better!
🤝 Acknowledgement 🔝
Causal-VLR is an open-source project and We appreciate all the contributors who implement their methods or add new features and users who give valuable feedback. We wish that the toolbox and benchmark could serve the growing research community by providing a flexible toolkit to reimplement existing methods and develop their new models.
🪐 The review paper here can provide some help
Causal Reasoning Meets Visual Representation Learning: A Prospective Study
Machine Intelligence Research (MIR) 2022
A Review paper for causal reasoning and visual representation learning
@article{liu2022causal,
title={Causal Reasoning Meets Visual Representation Learning: A Prospective Study},
author={Liu, Yang and Wei, Yu-Shen and Yan, Hong and Li, Guan-Bin and Lin, Liang},
journal={Machine Intelligence Research},
pages={1--27},
year={2022},
publisher={Springer}
