Attention Consistency on Visual Corruptions for Single-Source Domain Generalization
Ilke Cugu, Massimiliano Mancini, Yanbei Chen, Zeynep Akata
IEEE Computer Vision and Pattern Recognition Workshops (CVPRW), 2022
The official PyTorch implementation of the CVPR 2022, L3D-IVU Workshop paper titled "Attention Consistency on Visual Corruptions for Single-Source Domain Generalization". This repository contains: (1) our single-source domain generalization benchmark that aims at generalizing from natural images to other domains such as paintings, cliparts and skethces, (2) our adaptation/version of well-known advanced data augmentation techniques in the literaure, and (3) our final model ACVC which fuses visual corruptions with an attention consistency loss.
torch~=1.5.1+cu101
numpy~=1.19.5
torchvision~=0.6.1+cu101
Pillow~=8.3.1
matplotlib~=3.1.1
sklearn~=0.0
scikit-learn~=0.24.1
scipy~=1.6.1
imagecorruptions~=1.1.2
tqdm~=4.58.0
pycocotools~=2.0.0
-
We also include a YAML script
acvc-pytorch.yml
that is prepared for an easy Anaconda environment setup. -
One can also use the
requirements.txt
if one knows one's craft.
Training is done via run.py
. To get the up-to-date list of commands:
python run.py --help
We include a sample script run_experiments.sh
for a quick start.
The benchmark results are prepared by analysis/GeneralizationExpProcessor.py
, which outputs LaTeX tables of the cumulative results in a .tex file.
- For example:
python GeneralizationExpProcessor.py --path generalization.json --to_dir ./results --image_format pdf
- You can also run distributed experiments, and merge the results later on:
python GeneralizationExpProcessor.py --merge_logs generalization_gpu0.json generalization_gpu1.json
COCO benchmark is especially useful for further studies on ACVC since it includes segmentation masks per image.
Here are the steps to make it work:
- For this benchmark you only need 10 classes:
airplane
bicycle
bus
car
horse
knife
motorcycle
skateboard
train
truck
-
Download COCO 2017 trainset, valset, and annotations
-
Extract the annotations zip file into a folder named
COCO
inside your choice ofdata_dir
(For example:datasets/COCO
) -
Extract train and val set zip files into a subfolder named
downloads
(For example:datasets/COCO/downloads
) -
Download DomainNet (clean version)
-
Create a new
DomainNet
folder next to yourCOCO
folder -
Extract each domain's zip file under its respective subfolder (For example:
datasets/DomainNet/clipart
) -
Back to the project, use
--first_run
argument once while running the training script:
python run.py --loss CrossEntropy --epochs 1 --corruption_mode None --data_dir datasets --first_run --train_dataset COCO --test_datasets DomainNet:Real --print_config
-
If everything works fine, you will see
train2017
andval2017
folders underCOCO
-
Both folders must contain 10 subfolders that belong to shared classes between COCO and DomainNet
-
Now, try running ACVC as well:
python run.py --loss CrossEntropy AttentionConsistency --epochs 1 --corruption_mode acvc --data_dir datasets --train_dataset COCO --test_datasets DomainNet:Real --print_config
-
All good? Then, you are good to go with the COCO section of
run_experiments.sh
to run multiple experiments -
That's it!
If you use these codes in your research, please cite:
@InProceedings{Cugu_2022_CVPR,
author = {Cugu, Ilke and Mancini, Massimiliano and Chen, Yanbei and Akata, Zeynep},
title = {Attention Consistency on Visual Corruptions for Single-Source Domain Generalization},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2022},
pages = {4165-4174}
}
We indicate if a function or script is borrowed externally inside each file. Specifically for visual corruption implementations we benefit from:
- The imagecorruptions library of Autonomous Driving when Winter is Coming.
Consider citing this work as well if you use it in your project.