Bottom-Up Feature Restoration

This repository contains the code accompanying the paper Source-Free Adaptation to Measurement Shift via Bottom-Up Feature Restoration (ICLR 2022, Spotlight).

Prerequisites:

• python 3
• pytorch 1.7.0+
• WILDS (for Camelyon-17)
• numpy, scipy, sklearn, matplotlib, PIL, torchvision

Data

By default datasets are assumed to be located in a ./datasets/ directory, this can be changed with the data-root flag in the shell scripts used to run experiments. The table below shows where the datasets can be downloaded from and the path where they should be stored. To start running the code only EMNIST-DA should be required.

Dataset	Download	Location
EMNIST-DA	Data	<data-root>/EMNIST/
MNIST-C	Data	<data-root>/mnist_c/
MNIST-M	Set download=True in mnistm.py or here	<data-root>/MNISTM/
CIFAR-10-C	Data	<data-root>/CIFAR-10-C/
CIFAR-100-C	Data	<data-root>/CIFAR-100-C/
CAMELYON-17	Downloads automatically (download=True) from WILDS	<data-root>/CAMELYON17/

Simple Usage

To run Bottom-Up Feature Restoration there are three steps:

1. Pretrain the source model

./scripts/pretrain.sh

2. Save the source statistics

./scripts/save_train_stats.sh

3. Adapt on the new data

./scripts/bufr.sh

Other (not bottom-up) methods, including source-only performance, AdaBN and feature restoration can be run after steps 1 & 2 by runnning

./scripts/adapt.sh

Further Information

Choosing what to run

In the configs directory are yaml files that set hyperparameters for the different methods for each dataset

• configs/<DATASET_NAME>/dataset.yml sets arguments used for all methods across a dataset, namely the network architecture and which shifts to train on.

• configs/<DATASET_NAME>/<method>.yml sets the specific hyperparameters for the different methods (learning rate, epochs etc.)

The shell scripts in the scripts directory set which method to run, which dataset to use and some global arguments. To choose a specific dataset the following arguments should be set where applicable

• --alg-config to ./configs/<MDATASET_NAME>/pretrain.yml
• --alg-configs-dir to ./configs/<DATASET_NAME>/
• --data-config to` ./configs/<DATASET_NAME>/dataset.yml

Possible values for <DATASET_NAME> are: EMNIST-DA, MNIST-C, CIFAR-10-C, CIFAR-100-C, CAMELYON17. (The directory names inside configs.)

To choose a specific method in scripts/adapt.sh set --alg-name to the name of the method. Possible values for --alg-name are:

• all - trains all methods for the dataset
• fewshot - for EMNIST-DA and CAMELYON17 trains the few-shot methods
• Any one of: adabn, bnm, bnm_im, fr, im, jg, label, pl, shot, source_only. (The names of the yaml files inside configs/<DATASET_NAME>/. Not all methods are implemented for all datasets.)

The same random seed should be used for all steps (pretraining, saving statistics and adapting), this can be set with the --seed flag.

SHOT baseline

SHOT uses different pretraining techniques, to run this baseline on EMNIST-DA or MNIST-C requires first rerunning ./scripts/pretrain.sh on with the shot_pretrain argument set to True in pretrain.yml.

Analyses

The analysis directory contains files for creating analysis plots and calculating the feature restoration score. These analyses require an adapted models to be saved in advance, this can be done by adding --save-adapted-model to the adapt.sh and bufr.sh scripts during training.

Analysis files can be run using the scripts in ./scripts/analysis and changing the PYTHONPATH appropriately. The files calibration.py, max_patches.py and tsne.py contain hardcoded paths at the start of the files that you may wish to change. The only exception is vis_per_unit_scores.py which has no such run script in ./scripts/analysis and is quite messy. If you would like to use this to recreate Figure 17 of Appendix J.3 in our paper, please let us know and we can provide some run scripts!

Results

The results we achieved from running this code can be found here.

Citation

@inproceedings{eastwood2022sourcefree,
     title={Source-Free Adaptation to Measurement Shift via Bottom-Up Feature Restoration}, 
     author={Cian Eastwood and Ian Mason and Christopher K. I. Williams and Bernhard Sch\"olkopf},
     booktitle={International Conference on Learning Representations},
     year={2022}
}

Readers may also be interested in our related study on Unit-level surprise in neural networks.