Agree-to-Disagree

Repository for the paper: "Agree to Disagree: Diversity through Disagreement for Better Transferability"

Reproducing our results

2D toy setup

To reproduce the 2D D-BAT results (Figure 1), see this notebook.

Results on datasets with completely spurious correlation

• To reproduce the C-MNIST results, see this notebook (notebooks/D_BAT_C_MNIST.ipynb)
• To reproduce the MF-Dominoes results, see this notebook (notebooks/D_BAT_M_F_Dominoes.ipynb)
• To reproduce the MC-Dominoes results (Table 1), see this notebook (notebooks/D_BAT_M_C_Dominoes.ipynb)

Results on "natural" datasets

Install WILDS: pip install wilds

• To reproduce the Waterbirds results, run train-waterbird.sh.
• To reproduce the Camelyon17 dataset, run train-camelyon.sh.
• To reproduce the results on the Office-Home dataset, first download the data (see datasets/README.md for link), then run train-office-home-ood_is_test.sh and train-office-home-ood_is_not_test.sh

Results on uncertainty and OOD detection

• To reproduce results on the MNIST OOD setup (Figure 3), see this notebook (notebooks/D_BAT_MNIST_Uncertainty.ipynb)
• To reproduce results on the CIFAR10 OOD setup (Figure 4), see this notebook (notebooks/D_BAT_CIFAR10_OOD_CIFAR100_MCDropout.ipynb)