Domain Generalization via Gradient Surgery

This repository contains the source code corresponding to the paper "Domain Generalization via Gradient Surgery" (Accepted for publication at ICCV 2021). To learn more about our work, please visit this link: https://arxiv.org/abs/2108.01621

Instructions

This project uses Python 3.6.9 and PyTorch 1.8.1.

Data:

1. Download PACS (Li et al., 2017), VLCS (Fang et al., 2013) and Office-Home (Venkateswara et al., 2017) datasets.
2. Put the data of each dataset into a folder named orig, inside a parent folder named data. For example: /path/to/DGvGS/data/orig/PACS.
3. Run the run_pre.sh script to resize the images and generate the training, validation and test splits.

Python environment:

1. Create the virtual environment: python -m venv /path/to/myenv.
2. Activate it: source /path/to/myenv/bin/activate.
3. Install the required packages: pip install -r /path/to/DGvGS/requirements.txt.

Simulations:

To run the simulations across all datasets (PACS, VLCS and Office-Home) and methods (Deep-All, Agr-Sum, Agr-Rand and PCGrad), execute the run_all.sh script.

Otherwise, if you want to run a particular combination of dataset and method, use the train.py script. For example, the following instruction:

python train_all.py --data_dir=data/pre --output_dir=result/train_all --dataset=PACS --method=deep-all

will run Deep-All on PACS and save the results in /path/to/DGvGS/result/train_all.

License

MIT