bit-da/cirl

[CVPR 2022 Oral] Code release for "Causality Inspired Representation Learning for Domain Generalization"

CIRL

This repo provides a demo for the CVPR 2022 paper "Causality Inspired Representation Learning for Domain Generalization" on the PACS dataset.

Requirements

• Python 3.6
• Pytorch 1.1.0

Training from scratch

Please first download the PACS dataset from http://www.eecs.qmul.ac.uk/~dl307/project_iccv2017 or from https://pan.baidu.com/s/1KxMA6SiQX1jdRxwkeKMqOw (password：pacs). Then update the files with suffix _train.txt and _val.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
...

Please make sure you are using the official train-val-split. Once the data is prepared, then remember to update 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.

Evaluation

After training the model, firstly create directory ckpt/ and drag your model under it. For running the evaluation code, please update the files with suffix _test.txt in data/datalists for each domain, following the same styles as the train/val files above.

Then 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.

Acknowledgements

Some codes are adapted from FACT. We thank them for their excellent projects.

Contact

If you have any problem about our code, feel free to contact [email protected] or describe your problem in Issues.

cirl's Issues

On the issue of model selection

CIRL/train.py

Line 192 in 1b50cd6

if self.results['test'][self.current_epoch] >= self.best_acc:

Choose the best model in your training code based on the correctness of the test set rather than the best correctness of the validation set.

希望可以提供requirements.txt

目前安装依赖需要根据报错手动一个一个装很麻烦，并且t使用orch==1.1.0和py==3.6根本无法运行

What kind of question is that

res = func(*args, **kwrds)
File "mkl_fft_pydfti.pyx", line 1105, in mkl_fft._pydfti.fftn
File "mkl_fft_pydfti.pyx", line 1083, in mkl_fft._pydfti._fftnd_impl
File "mkl_fft_pydfti.pyx", line 952, in mkl_fft._pydfti.iter_complementary
File "mkl_fft_pydfti.pyx", line 985, in mkl_fft._pydfti._direct_fftnd
AssertionError

Reproduced experiment results

Is there anyone who has reproduced the reported results? Does the code need to be modified?

Only 48% acc on: Leave-one-domain-out(photo), PACS official split, default parameters of ResNet50.
Test result:

Part of the training log:

Office-home dataset official split

thanks for your works，can you upload Office-home dataset splitting，thank you

Official train-val split for the PACS dataset

Hi, Thanks for your great work. I enjoyed reading your paper! I wonder where to get the official train-val split for the PACS dataset, can you point me to that?

About the Fourier transform of causality intervention module

Hi author, thank you for your excellent work.
I have a question about the paper and code. The Causal Intervention Module in the paper says that $x^a$ is obtained by using Fourier transform, but in your code, I see the data transformation in the below function, which seems to be not Fourier transform. I am not an expert in the image field, so I have some doubts about this. In addition, I have not seen the paper in the data transformation λ~U(0,η) either.

def get_pre_transform(image_size=224, crop=False, jitter=0):
    if crop:
        img_transform = [transforms.RandomResizedCrop(image_size, scale=[0.8, 1.0])]
    else:
        img_transform = [transforms.Resize((image_size, image_size))]
    if jitter > 0:
        img_transform.append(transforms.ColorJitter(brightness=jitter,
                                                    contrast=jitter,
                                                    saturation=jitter,
                                                    hue=min(0.5, jitter)))
    img_transform += [transforms.RandomHorizontalFlip(), lambda x: np.asarray(x)]
    img_transform = transforms.Compose(img_transform)
    return img_transform

Looking forward to your reply.

Why do we set the mask in the following way rather than directly using the Gumbel Softmax?

https://github.com/BIT-DA/CIRL/blob/1b50cd605fa1779b4adc739569371d0f19a6bcce/models/classifier.py#L42C8-L44C37

Is the environment configured? Versions do not correspond

Unresolved reference 'get_copy_dataset'

Thanks for your excellent work, I found a problem: the get_copy_dataset() function seems to be undefined in DGDataLoader.py - Unresolved reference 'get_copy_dataset' ,Looking forward to your response