Visual Representation Learning over Latent Domains

This repository contains code for training sparse latent adapters on latent domain benchmarks. To select a dataset simply pass --dataset pacs or --dataset office_home, respectively.

To download the datasets, run scripts/download.py --data_dir data from where datasets will be read by default. The models require pretrained weights, which will be downloaded automatically using gdown.

Other configuration settings can be changed in util/parser.py or providing matching commands to train.py, e.g. to modify the learning rate python3 train.py --lr_sgd 0.001.

If you find this code useful in your research, please cite our work as:

@inproceedings{deecke22,
    author       = "Deecke, Lucas and Hospedales, Timothy and Bilen, Hakan",
    title        = "Visual Representation Learning over Latent Domains",
    booktitle    = "International Conference on Learning Representations",
    year         = "2022"
}