De-Confusing Pseudo-Labels in Source-Free Domain Adaptation [ECCV 2024]

Source-free domain adaptation (SFDA) aims to adapt a source-trained model to an unlabeled target domain without access to the source data. SFDA has attracted growing attention in recent years, where existing approaches focus on self-training that usually includes pseudo-labeling techniques. In this paper, we introduce a novel noise-learning approach tailored to address noise distribution in domain adaptation settings and learn to de-confuse the pseudo-labels. More specifically, we learn a noise transition matrix of the pseudo-labels to capture the label corruption of each class and learn the underlying true label distribution. Estimating the noise transition matrix enables a better true class-posterior estimation, resulting in better prediction accuracy. We demonstrate the effectiveness of our approach when combined with several SFDA methods: SHOT, SHOT++, and AaD. We obtain state-of-the-art results on three domain adaptation datasets: VisDA, DomainNet, and OfficeHome.

[Paper]

---

Prerequisites

• pip install timm==0.4.12
• python==3.8
• pytorch==1.10.0=py3.8_cuda11.3_cudnn8.2.0_0
• cudatoolkit=11.3
• torchvision==0.11.0
• numpy
• scipy
• scikit-learn
• pillow
• tqdm
• natsort``
• six
• opencv
• scikit-image

Datasets

• Please manually download the datasets Office-Home, VisDA-C and DomainNet (126)

Training

Pretrained-source models

It is recommended to use pre-trained source models for consistency.

Pre-trained source models can be downloaded from https://github.com/tim-learn/SHOT

After downloading, place the file under weights folder, directly in the project's root.

Source model training

(please see above for downloading pre-trained source models)

python train_source.py --class_num 126 --output ./output/weights/sketch_vendor_DN_2021 --dset DomainNet --max_epoch 50 --s 3 --layer wn

Target adaptation

DomainNet

sketch->clipart

python train_adapt.py --CM_lr_factor 2.0 --PL_net_to_use SwinBase1K --short_exp_name NADL_DN_126_swin_1K --seed 2021 --dataset_name DomainNet --source sketch --target clipart   

Another example: painting->real

python train_adapt.py --CM_lr_factor 2.0 --PL_net_to_use SwinBase1K --short_exp_name NADL_DN_126_swin_1K --seed 2021 --dataset_name DomainNet --source painting --target real

OfficeHome:

python train_adapt.py  --CM_lr_factor 1.0 --PL_net_to_use SwinBase1K --short_exp_name NADL_OH_swin_1K --load_downloaded_weights 1 --CM_beta 30 --seed 2021 --dataset_name OfficeHome --source Art --target Clipart --lr_value 1e-2 --batch_size 64 --PL_iteration_amount 2 

Another example:

python train_adapt.py  --CM_lr_factor 1.0 --PL_net_to_use SwinBase1K --short_exp_name NADL_OH_swin_1K --load_downloaded_weights 1 --CM_beta 30 --seed 2021 --dataset_name OfficeHome --source Product --target RealWorld --lr_value 1e-2 --batch_size 64  --PL_iteration_amount 2 

VisDA-C:

python train_adapt.py  --cnn_to_use resnet101 --CM_lr_factor 1.0 --CM_prior_loss 1.0 --PL_net_to_use SwinBase1K --short_exp_name VISDA_OH_swin_1K --load_downloaded_weights 1 --CM_beta 5.0 --seed 2021 --dataset_name VisDA-C --target validation --PL_after_epoch 1