Causality for Machine Learning

This repo accompanies the prototype code from our report, specifically the Invariant Risk Minimization (IRM) approach discussed in chapters 3 & 4.

Setup environment

conda create --name irm_env python=3.7 ipykernel
conda activate irm_env
conda install pip
pip install -r requirements.txt

Setup data

Steps to generate the WildCam dataset used for this experiment.

Step 1 : Download the Camera Traps (or Wild Cams) dataset - iWildCam 2019 using Kaggle.
- Use Kaggle account to download data. This involves first creating a new Kaggle API from your account and then downloading the kaggle.json file in [user-home]/.kaggle folder. If there is no .kaggle folder yet, create it and then move the kaggle.json to the .kaggle folder.
- Install kaggle package, conda install -c conda-forge kaggle
- chmod 600 ~/.kaggle/kaggle.json
- Download data, kaggle competitions download -c iwildcam-2019-fgvc6. This is a 44GB file!
- unzip iwildcam-2019-fgvc6.zip -d ./iWildCam
- unzip -q train_images.zip -d ./train. NOTE: The test set images are unlabeled so are being ignored from our experiments. Instead, we create a test set from the training data in the next step.
Step 2 :
- Run python create_denoised_data.py - This creates a new directory ./data/wildcam_denoised consisting of the images we used for training and testing both the IRM and ERM models. The list of images are available in ./data/train_test_filenames.json.

Repo structure

.
├── ./create_denoised_data.py
├── ./data
│   ├── ./data/train_test_filenames.json
│   └── ./data/wildcam_denoised
│       ├── ./data/wildcam_denoised/test
│       ├── ./data/wildcam_denoised/train_43
│       └── ./data/wildcam_denoised/train_46
├── ./dataset.py
├── ./ERM_results.out
├── ./IRM_results.out
├── ./main.py
├── ./models
│   └── ./models/wildcam_denoised_121_0.001_0_0.0_ERM.pth
│   └── ./models/wildcam_denoised_121_0.001_40_10000.0_IRM.pth
├── ./models.py
├── ./README.md
├── ./requirements.txt
└── ./train.py

Training

Simply run python main.py to train an ERM/ IRM model. The arguments penalty_anneal_iters and penalty_weight when set to 0 trains an ERM model.
The model is saved in the ./models folder

Results

Results from training both IRM and ERM are available in ERM_results.out and IRM_results.out files

Interpretability

The model_explanation_irm.ipynb notebook provides LIME explanations for a sample image based on the IRM model. For a deeper dive into explanations and comparison between all images based on both the ERM and IRM models look at our prototype - Scene

References

Leveraged source code from the paper:

@article{InvariantRiskMinimization,
    title={Invariant Risk Minimization},
    author={Arjovsky, Martin and Bottou, L{\'e}on and Gulrajani, Ishaan and Lopez-Paz, David},
    journal={arXiv},
    year={2019}
}

fastforwardlabs/causality-for-ml