These directories contain code for reproducing the MEMO results for the CIFAR-10 and ImageNet distribution shift test sets.

Please note: this code has been modified from the version that generated the results in the paper for the purpose of cleaning the code. Though it is likely that the code is correct and should produce the same results, it is possible that there may be some discrepancies that were not caught. Though minor results differences may arise from stochasticity, please report any major differences or bugs by submitting an issue, and we will aim to resolve these promptly.

First, create a Anaconda environment with requirements.txt, e.g.,

conda create -n memo python=3.8 -y -q --file requirements.txt
conda activate memo

After doing so, you will need to pip install tqdm. For the robust vision transformer models, you will also need to pip install timm einops.

The cifar-10-exps directory contains code for the CIFAR-10 experiments. You can run bash script_c10.sh for the full set of experiments. Alternatively, you can run python script_test_c10.py directly with the experiment you wish to run (see script_c10.sh for more details).

For convenience, we provide the ResNet26 model that we trained in results/cifar10_rn26_gn/ckpt.pth. We do not provide the datasets themselves, though you can download the non standard test sets here:

• CIFAR-10.1
• CIFAR-10-C

After downloading and setting up the datasets, make sure to modify the dataroot variable on line 8 of script_test_c10.py.

The imagenet-exps directory contains code for the ImageNet experiments. You can run bash script_in.sh for the full set of experiments, though this is very slow. You can again run python script_test_in.py directly with the experiment you wish to run. For the corrupted image datasets, you may wish to slightly modify the code to only run one corruption-level pair (and then parallelize).

As an example, we provide the baseline ResNet-50 model from torchvision in results/imagenet_rn50/ckpt.pth. Including all of the pretrained model weights would be prohibitively large. We did not train our own models for ImageNet, and all other models we used can be downloaded:

• ResNet-50 w/ DeepAugment + AugMix
• ResNet-50 w/ moment exchange + CutMix
• RVT*-small
• ResNext-101 (32x8d) w/ WSL

We also experimented with a baseline ResNext-101 (32x8d) model which we obtained from torchvision.

Please note: some of these models provide the weights in slightly different conventions, thus loading the downloaded state_dict may not directly work, and the keys in the state_dict may need to be modified to match with the code. We have done this modification already for the baseline ResNet-50 model, and thus this ckpt.pth can be used as a template for modifying other model checkpoints.

We again do not provide the datasets themselves, though you can download the test sets here:

• ImageNet-C
• ImageNet-R
• ImageNet-A

After downloading and setting up the datasets, again make sure to modify the dataroot variable on line 8 of script_test_in.py.

Please use the following citation:

@article{memo,
    author={Zhang, M. and Levine, S. and Finn, C.},
    title={{MEMO}: Test Time Robustness via Adaptation and Augmentation},
    article={arXiv preprint arXiv:2110.09506},
    year={2021},
}

The paper can be found on arXiv here.

The design of this code was adapted from the TTT codebases. Other parts of the code that were adapted from third party sources are credited via comments in the code itself.