This repository is the official implementation of Domain Adaptation under Open Set Label Shift. We also release implementation of relevant baselines for Open Set Domain Adaptation and alternative methods adapted from PU learning. If you find this repository useful or use this code in your research, please cite the following paper:
Garg, S., Balakrishnan, S., Lipton, Z. (2022). Domain Adaptation under Open Set Label Shift. arxiv preprint arXiv:2207.13048.
@article{garg2022OSLS,
title={Domain Adaptation under Open Set Label Shift},
author={Garg, Saurabh and Balakrishnan, Sivaraman and Lipton, Zachary},
year={2022},
journal={arXiv preprint arXiv:2207.13048},
}
The code is written in Python and uses PyTorch. To install requirements, setup a conda enviornment using the following command:
conda create --file enviornment.yml
We use CIFAR10, CIFAR100 from pytorch torchvision library. For Newsgroups20, we use sklearn datasets. Hence, we do not need any manual setup for them. For other datasets, our repository includes the code to setup an OSLS problem after the datasets have been manually downloaded. To download these datasets follow the steps below:
- Entity30: We use BREEDs benchmark library provided here. This mainly involves: (i) installing the robustness library; (ii) downloading Imagenet dataset; and (iii) downloading the Imagenet heirarchy from here.
- BreakHis: We use BreakHis dataset provided here. Use the following command to download the dataset with Kaggle API:
kaggle datasets download -d ambarish/breakhis && mkdir -p ./data/ && unzip breakhis.zip -d ./data/ && python dataset_cleaning/Breakhis_dataset.py ./data/BreaKHis_v1/ - DermNet: We use Dermnet NZ dataset provided here. Use the following command to download the dataset with Kaggle API:
kaggle datasets download -d shubhamgoel27/dermnet/ && mkdir -p ./data/dermnet && unzip dermnet.zip -d ./data/dermnet/ - Tabula Muris: Download the dataset following instructions from this repository. In particular, use the script here or simply run the following:
wget http://snap.stanford.edu/comet/data/tabula-muris-comet.zip && unzip tabula-muris-comet.zip - UTKFace: We use UTKFace dataset. Use the following command to download and setup the dataset:
gdown 0BxYys69jI14kYVM3aVhKS1VhRUk && mkdir -p ./data/UTKDataset && tar -xvf UTKFace.tar.gz -C ./data/UTKDataset/ && python ./dataset_cleaning/UTK_dataset.py ./data/UTKDataset/
For pretrained models, download and store the following files in the folder ./pretrained_models/: For CIFAR100 we use the pretrained models provided here and for entity30 we use the pretrained models provided here. For newsgroups20, we use glove vectors from here.
run.py file is the main entry point for training the model and run the code with the following command:
CUDA_VISIBLE_DEVICES=0 python run.py -m models=trainPU_labelshift.yaml datamodule=random_split_module.yaml seed=42 dataset=breakhis arch=Resnet50 num_source_classes=6 fraction_ood_class=0.25 max_epochs=80 batch_size=32 pretrained=True learning_rate=0.0001 separate=True datamodule.seed=1
Change the parameters to your liking and run the experiment. For example, change dataset by changing dataset to one of [CIFAR10, CIFAR100, entity30, breakhis, dermnet, tabula_munis, newsgroups20] and vary algorithm with varying models to any yaml file in configs/models. We implement our PULSE framework in src/algorithm/trainPU_labelshift.py.
We provide a set of scripts to run the main experiments. See scripts folder for details.
We provide code to aggregate and plot results in aggregate_nums.py, plot_acc.py and plot_mpe.py.
This repository is licensed under the terms of the Apache-2.0 License.
The code uses hydra to manage config files and pytorch lightening to manage training. Algorithm specific config files
are in configs/models folder. We use random datamodule configs/datamodule/random_split_module.yaml to generate an OSLS setup.
For more details, refer to the accompanying the paper: Domain Adaptation under Open Set Label Shift. If you have questions, please feel free to reach us at sgarg2@andrew.cmu.edu or open an issue.