Jing--Li / SEAL

This is the PyTorch implementation of our AAAI 2024 paper SEAL.

A Separation and Alignment Framework for Black-box Domain Adaptation

This is a PyTorch implementation of our AAAI 2024 paper SEAL.

Start Running SEAL on Office, Office-Home, VisDA, and DomainNet.

Step 1. Data Preparation

Please download the datasets and put them into ./data .

Step 2. Generate information list for dataset

# For Office
python generate_infos.py --ds office31

# For Office-Home
python generate_infos.py --ds office_home

# For VisDA
python generate_infos.py --ds visda17

# For DomainNet
python generate_infos.py --ds DomainNet

Step 3. Training black-box source domain

# For Office on domain A
CUDA_VISIBLE_DEVICES=0 python train_src_v1.py configs/office31/src_a/train_src_a.py

# For Office-Home on domain A
CUDA_VISIBLE_DEVICES=1 python train_src_v1.py configs/office_home/src_A/train_src_A.py

# For VisDA
CUDA_VISIBLE_DEVICES=2 python train_src_v2.py configs/visda17/train_src.py

# For DomainNet
CUDA_VISIBLE_DEVICES=3 python train_src_v1.py configs/DomainNet/src_c/train_src_c.py

Step 4. Adapting to target domain using SEAL

# For Office on domain shift A->D
CUDA_VISIBLE_DEVICES=0 python train_SEAL.py configs/office31/src_a/SEAL_d.py

# For Office-Home on domain shift A->C
CUDA_VISIBLE_DEVICES=1 python train_SEAL.py configs/office_home/src_A/SEAL_C.py

# For VisDA
CUDA_VISIBLE_DEVICES=2 python train_SEAL.py configs/visda17/SEAL.py

# For DomainNet on domain shift C->P
CUDA_VISIBLE_DEVICES=3 python train_SEAL.py configs/DomainNet/src_c/SEAL_p.py