Robustness to Spurious Correlations Improves Semantic Out-of-distribution Detection

Overview

Set Up

All experiments were conducted on CentOS 8.2.2004 with Python 3.8.5 and Pytorch 1.10.2. See requirements.txt for further details.

Data

To begin, first download:

1. Caltech-UCSD Birds-200-2011
2. PlacesBG
3. CelebA

See scripts/dataset_creation for scripts generating Waterbirds in-distribution and shared-nuisance out-of-distribution datasets. Place the data in the location corresponding to root_dir or save_dir in each of the dataset files. Otherwise, data is downloaded automatically from torchvision in the location specified by root_dir or save_dir.

Quickstart

Experiments were run using Weights and Biases. To use, simply create an account, enter your API key locally where experiments will be run, initialize a sweep via wandb sweep <path/to/config>, and launch agents via wandb agent <username/project_name/sweep_id>. See Weights and Biases documentation for further details on sweeps.