Attract, Perturb, and Explore: Learning a Feature Alignment Network for Semi-supervised Domain Adaptation (ECCV 2020)
The implementation is built on the pytorch implementation of SSDA_MME and we refer a specific module in DTA.
- CUDA 10.0 or 10.1
- Python 3.7 (or 3.6)
- Pytorch 1.0.1
conda install pytorch==1.0.1 torchvision==0.2.2 cudatoolkit=10.0 -c pytorch
- Pillow, numpy, tqdm
- You can easily install dependencies through
pip install -r requirements.txt
You can download the datasets by following the instructions in SSDA_MME.
data---
|
multi---
| |
| Real
| Clipart
| Product
| Real
office_home---
| |
| Art
| Clipart
| Product
| Real
office---
| |
| amazon
| dslr
| webcam
txt---
|
multi---
| |
| labeled_source_images_real.txt
| unlabeled_target_images_real_3.txt
| labeled_target_images_real_3.txt
| unlabeled_source_images_sketch.txt
| ...
office---
| |
| labeled_source_images_amazon.txt
| unlabeled_target_images_amazon_3.txt
| labeled_target_images_amazon_3.txt
| unlabeled_source_images_webcam.txt
| ...
office_home---
|
...
- DomainNet (clipart, painting, real, sketch)
python main.py --dataset multi --source real --target sketch --save_interval 5000 --steps 70000 --net resnet34 --num 3 --save_check
- Office-home (Art, Clipart, Product, Real)
- Office (amazon, dslr, webcam)
- DomainNet (clipart, painting, real, sketch)
python test.py --dataset multi --source real --target sketch --steps 70000
- (DomainNet) Real to Sketch BaseNet / Classifier
We provide 5, 10, 20-shot splits for four domains (clipart, painting, real, sketch) of the DomainNet dataset.
- (DomainNet) splits