---

Update:

Version 2 is comming soon! Faster, less parameter tuning, and better performance!

---

---

Personal update:

I am looking for a machine learning engineer/ machine learning researcher position in EU. If you know suitable opportunity, please do not hesitate to contact me! My linkedin page: Ziming Wang

---

Partial Distribution Matching via Partial Wasserstein Adversarial Networks

This is the official implement of the partial domain adaptation method proposed in Partial Distribution Matching via Partial Wasserstein Adversarial Networks. This approach applies a PWAN model to partial domain adaptation problems. Also see the application of PWAN in point set registration in our earlier work (Code)

How does it work?

A classifier is used to extract features and predict labels, while a PWAN model is used to align the source features to a fraction of the reference features.

An example of the T-SNE visualization of the reference (blue) and source features (red). Gray points represent outlier features.

without PWAN	With PWAN

Requirement

• torch=1.12
• matplotlib
• tqdm
• scikit-learn
• tensorboard

Usage

Data perparation

1. OfficeHome

   1. Download the dataset from the official website to data/OfficeHome
   2. Rename the folder of real world to Real_World

2. ImageNet:

   1. Download the ImageNet dataset (ILSVRC2012_img_train.tar) from the official website to data/ImageNetCaltech.
   2. Uncompress the ImageNet dataset to a train/ folder.

3. DomainNet and VisDa17 can be downloaded automatically. If the link expires, please also use the official websites.

Script

The main script is PWANN.py. We also provide a convinient script run_all.py for running this script with different dataset and random seeds. Commands for reproducing the results in the paper are provided in a.txt

The visualization and summary code is provided in Plot folder. For example,

cd Plot
python vis_tensorboard.py ./LOG/OfficeHome

summarize all experiments in OfficeHome folder.

Results

There will be some randomness in the results due to random initialization. We report the results with random seed 0,1,2.

OfficeHome:

seed	0	1	2
AC	0.66	0.66	0.62
AP	0.86	0.81	0.86
AR	0.89	0.90	0.90
CA	0.78	0.77	0.76
CP	0.77	0.74	0.80
CR	0.83	0.84	0.87
PA	0.79	0.75	0.79
PC	0.64	0.63	0.63
PR	0.87	0.86	0.87
RA	0.80	0.79	0.80
RC	0.66	0.66	0.68
RP	0.86	0.86	0.89
Avg	0.784	0.772	0.791

VisDa17: The accuracy ranges from 80 to 92. The large variance is due to the ambiguity of the skateboard and knife class. Visualization and discussion can be found in the appendix of the paper.

DomainNet:

seed	0	1	2
CP	0.55	0.54	0.51
CR	0.73	0.75	0.72
CS	0.57	0.61	0.56
PC	0.65	0.65	0.64
PR	0.81	0.82	0.81
PS	0.72	0.72	0.73
RC	0.77	0.78	0.77
RP	0.72	0.73	0.73
RS	0.70	0.70	0.70
SC	0.52	0.50	0.49
SP	0.56	0.54	0.54
SR	0.66	0.66	0.63
Avg	0.668	0.672	0.657

ImageNetCaltech:

seed	0	1	2	Avg
IC	0.858	0.864	0.857	0.860

Reference

If you find the code useful, please cite the following papers.

@misc{wang2024partialdistributionmatchingpartial,
      title={Partial Distribution Matching via Partial Wasserstein Adversarial Networks}, 
      author={Zi-Ming Wang and Nan Xue and Ling Lei and Rebecka Jörnsten and Gui-Song Xia},
      year={2024},
      eprint={2409.10499},
      url={https://arxiv.org/abs/2409.10499}, 
}

@inproceedings{wang2022partial,
    title={Partial Wasserstein Adversarial Network for Non-rigid Point Set Registration},
    author={Zi-Ming Wang and Nan Xue and Ling Lei and Gui-Song Xia},
    booktitle={International Conference on Learning Representations (ICLR)},
    year={2022}
}

For any question, welcome to open an issue or contact me.

Acknowledgement

The datasets module was adopted from the Transfer Learning Library. We thank the authors of this repository and other authors in the community for their code.

LICENSE

The code is available under a MIT license.