mediabrain-sjtu / fact

FACT

This repo provides a demo for the CVPR 2021 paper "A Fourier-based Framework for Domain Generalization" on the PACS dataset.

To cite, please use:

@InProceedings{Xu_2021_CVPR,
    author    = {Xu, Qinwei and Zhang, Ruipeng and Zhang, Ya and Wang, Yanfeng and Tian, Qi},
    title     = {A Fourier-Based Framework for Domain Generalization},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {14383-14392}
}

Requirements

• Python 3.6
• Pytorch 1.1.0

Evaluation

Firstly create directory ckpt/ and drag your model under it. For running the evaluation code, please download the PACS dataset from http://www.eecs.qmul.ac.uk/~dl307/project_iccv2017. Then update the files with suffix _test.txt in data/datalists for each domain, following styles below:

/home/user/data/images/PACS/kfold/art_painting/dog/pic_001.jpg 0
/home/user/data/images/PACS/kfold/art_painting/dog/pic_002.jpg 0
/home/user/data/images/PACS/kfold/art_painting/dog/pic_003.jpg 0
...

Once the data is prepared, remember to update the path of test files and output logs in shell_test.py:

input_dir = 'path/to/test/files'
output_dir = 'path/to/output/logs'

then simply run:

 python shell_test.py -d=art_painting

You can use the argument -d to specify the held-out target domain.

Training from scratch

After downloading the dataset, update the files with suffix _train.txt and _val.txt in data/datalists for each domain, following the same styles as the test files above. Please make sure you are using the official train-val-split. Then update the the path of train&val files and output logs in shell_train.py:

input_dir = 'path/to/train/files'
output_dir = 'path/to/output/logs'

Then running the code:

python shell_train.py -d=art_painting

Use the argument -d to specify the held-out target domain.

Acknowledgement

Part of our code is borrowed from the following repositories.

• JigenDG: "Domain Generalization by Solving Jigsaw Puzzles", CVPR 2019
• DDAIG: "Deep Domain-Adversarial Image Generation for Domain Generalisation", AAAI 2020

We thank to the authors for releasing their codes. Please also consider citing their works.

fact's Issues

About the division of OfficeHome dataset

Thanks for your work. There is no official division of the OfficeHome dataset. How do you divide OfficeHome into a training set and a validation set. Is the result in your paper the result of the test set corresponding to the highest accuracy of the validation set? Thanks

Issue on the PACS dataset

First of all, thank you for your wonderful work. I found that the link to the PACS dataset (http://www.eecs.qmul.ac.uk/~dl307/project_iccv2017) is now unavailable but can be accessed at (https://dali-dl.github.io/project_iccv2017.html).

The config file of ResNet18

Hi, could you share the config file of ResNet18?

On the model selection

In your paper, you write: "we conduct the leave-one-domain-out evaluation. We train our model on the training splits and select the best model on the validation splits of all source domains"

But, in my opinion and according to the description in "https://github.com/facebookresearch/DomainBed", "select the best model on the validation splits of all source domains" is the "Training-domain validation" or "IIDAccuracySelectionMethod", rather than the " leave-one-domain-out evaluation".

What's your opiniton? Thanks for your reply.

Code for generating phase-only reconstruction

In the paper, you mentioned "we first generate phase-only reconstructed images by setting the amplitude component as a constant.".
However, I cannot understand how you set the amplitude component as a constant.
Whenever I tried that way, it gives me the black image.
Can you provide implementations for generating phase-only recon images?