evgeneus/graph-domain-adaptaion

PyTorch code for the paper "Curriculum Graph Co-Teaching for Multi-target Domain Adaptation" (CVPR2021)

PyTorch code for the paper "Curriculum Graph Co-Teaching for Multi-target Domain Adaptation" (CVPR2021)

This repo presents PyTorch implementation of Multi-targe Graph Domain Adaptation framework from "Curriculum Graph Co-Teaching for Multi-target Domain Adaptation" CVPR 2021. The framework is pivoted around two key concepts: graph feature aggregation and curriculum learning (see pipeline below or project web-page).

Results

Environment

Python >= 3.6
PyTorch >= 1.8.1

To install dependencies run (line 1 for pip or line 2 for conda env):

pip install -r requirements.txt
conda install --file requirements.txt

Disclaimer. This code has been tested with cuda toolkit 10.2. Please install PyTorch as supported by your machine.

Datasets

Four datasets are supported:

• Office-31 (Kate Saenko et al., 2010)
• Office-Home (Venkateswara et al., 2017)
• PACS (Li et al., 2017)
• DomainNet (Peng et al., 2019)

To run this code, one must check if the txt file names in data/<dataset_name> are matching with the downloaded domain folders. For e.g., to run OfficeHome, the domain sub-folders should be art/, clipart/, product/ and real/ corresponding to art.txt, clipart.txt, product.txt and real.txt that can be found in the data/office-home/.

Methods

• CDAN
• CDAN+E

Commands

Office-31

Run D-CGCT:

python src/main_dcgct.py \
        --method 'CDAN' \
        --encoder 'ResNet50' \
 	--dataset 'office31' \
 	--data_root [your office31 folder] \
 	--source 'webcam' \
 	--target 'dslr' 'amazon' \
 	--source_iters 200 \
 	--adapt_iters 3000 \
 	--finetune_iters 15000 \
 	--lambda_node 0.3 \
 	--output_dir 'office31-dcgct/webcam_rest/CDAN'

Run CGCT:

python src/main_cgct.py \
        --method 'CDAN' \
        --encoder 'ResNet50' \
 	--dataset 'office31' \
 	--data_root [your office31 folder] \
 	--source 'webcam' \
 	--target 'dslr' 'amazon' \
 	--source_iters 100 \
 	--adapt_iters 3000 \
 	--finetune_iters 15000 \
 	--lambda_node 0.1 \
 	--output_dir 'office31-cgct/webcam_rest/CDAN'

Office-Home

python src/main_dcgct.py \
	--method 'CDAN' \
	--encoder 'ResNet50' \
	--dataset 'office-home' \
	--data_root [your OfficeHome folder] \
	--source 'art' \
	--target 'clipart' 'product' 'real' \
	--source_iters 500 \
	--adapt_iters 10000 \
	--finetune_iters 15000 \
	--lambda_node 0.3 \
	--output_dir 'officeHome-dcgct/art_rest/CDAN' 

python src/main_cgct.py \
	--method 'CDAN' \
	--encoder 'ResNet50' \
	--dataset 'office-home' \
	--data_root [your OfficeHome folder] \
	--source 'art' \
	--target 'clipart' 'product' 'real' \
	--source_iters 500 \
	--adapt_iters 5000 \
	--finetune_iters 15000 \
	--lambda_node 0.1 \
	--output_dir 'officeHome-cgct/art_rest/CDAN' 

PACS

python src/main_dcgct.py \
	--method 'CDAN' \
	--encoder 'ResNet50' \
	--dataset 'pacs' \
	--data_root [your PACS folder] \
	--source 'photo' \
	--target 'cartoon' 'art_painting' 'sketch' \
	--source_iters 200 \
	--adapt_iters 3000 \
	--finetune_iters 15000  \
	--lambda_node 0.1 \
	--output_dir 'pacs-dcgct/photo_rest/CDAN'  

python src/main_cgct.py \
	--method 'CDAN' \
	--encoder 'ResNet50' \
	--dataset 'pacs' \
	--data_root [your PACS folder] \
	--source 'photo' \
	--target 'cartoon' 'art_painting' 'sketch' \
	--source_iters 200 \
	--adapt_iters 3000 \
	--finetune_iters 15000  \
	--lambda_node 0.1 \
	--output_dir 'pacs-cgct/photo_rest/CDAN'  

DomainNet

python src/main_dcgct.py \
	--method 'CDAN' \
	--encoder 'ResNet101' \
	--dataset 'domain-net' \
	--data_root [your DomainNet folder] \
	--source 'sketch' \
	--target 'clipart' 'infograph' 'painting' 'real' 'quickdraw' \
	--source_iters 5000 \
	--adapt_iters 50000 \
	--finetune_iters 15000  \
	--lambda_node 0.3 \
	--output_dir 'domainNet-dcgct/sketch_rest/CDAN'

python src/main_cgct.py \
	--method 'CDAN' \
	--encoder 'ResNet101' \
	--dataset 'domain-net' \
	--data_root [your DomainNet folder] \
	--source 'sketch' \
	--target 'clipart' 'infograph' 'painting' 'real' 'quickdraw' \
	--source_iters 5000 \
	--adapt_iters 50000 \
	--finetune_iters 15000  \
	--lambda_node 0.3 \
	--output_dir 'domainNet-cgct/sketch_rest/CDAN'

Citation

If you find our paper and code useful for your research, please consider citing our paper.

@inproceedings{roy2021curriculum,
  title={Curriculum Graph Co-Teaching for Multi-target Domain Adaptation},
  author={Roy, Subhankar and Krivosheev, Evgeny and Zhong, Zhun and Sebe, Nicu and Ricci, Elisa},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2021}
}

graph-domain-adaptaion's Issues

About "upgrade_source_domain"

Hi, thanks for your inspiring work! I have read the code and have some doubt about it. During the curriculum learning, why not remove the pseudo labelled target samples from the target domain after they are added to the source domain?
Expect your reply. Thanks.

How to run CGCT instead of D-CGCT?

From my understanding this code by default runs the D-CGCT variant of your method. However I am interested in comparing its performances with a MTDA method I am working on, which does not use domain labels for target samples. What should I do to run D-CGCT? Should I simply put all target data in a single txt file and use it as if it was a single target?

Thanks in advance

Multi-target settings

Thank you for the fantastic work! I realized that the "ndomain" parameter in config is always set to 2, which should be the target-combined setting in your paper.

May I know how did you train the other methods in the multi-target setting? For example, in the case of CDAN with k target domains, did you simply set the output dimension of domain classifier to be k+1?

I can't finish the 3rd step. Would you please send me model?

I am a student who are not able to afford an expensive graphics card. I have trained the data in Sagemaker Studio Lab, but the website provide little GPU time so that I can't finish the 3rd step. It will be appeciated that you can send the trained model to me via email [email protected]. Thank you.

source pretrained models

Hi,
Thank you for the awesome work! Can you provide the source pretrained model or any command to reproduce your result in Table 4 row 1 (w/o target source train) result?

DomainNet: Accuracy Information

Wanted to know that whether the accuracy reported in the paper is achieved through test split or train split.

Train	clipart_train.txt	infograph_train.txt	painting_train.txt	quickdraw_train.txt	real_train.txt	sketch_train.txt
Test	clipart_test.txt	infograph_test.txt	painting_test.txt	quickdraw_test.txt	real_test.txt	sketch_test.txt

Thanks.