SFDA - Domain Adaptation without Source Data

Prerequisites

• Ubuntu 18.04
• Python 3.6+
• PyTorch 1.5+ (recent version is recommended)
• NVIDIA GPU (>= 12GB)
• CUDA 10.0 (optional)
• CUDNN 7.5 (optional)

Getting Started

Installation

• Configure virtual (anaconda) environment

conda create -n env_name python=3.6
source activate env_name
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch

• Install python libraries

conda install -c conda-forge matplotlib
conda install -c anaconda yaml
conda install -c anaconda pyyaml 
conda install -c anaconda scipy
conda install -c anaconda scikit-learn 
conda install -c conda-forge easydict
pip install easydl

Download this repository

• We provide two versions of the repository (with dataset / without dataset) for a flexible experiment

• Full SFDA repository (with dataset): download link

  • In this case, go to training and testing step directly

• SFDA repository (without dataset): download link

• Visualization of repository structure (Full SFDA repository)

|-- APM_update.py
|-- SFDA_test.py
|-- SFDA_train.py
|-- config.py
|-- data.py
|-- lib.py
|-- net.py
|-- office-train-config.yaml
|-- data
|   `-- office
|       |-- domain_adaptation_images
|       |   |-- amazon
|       |   |   `-- images
|       |   |-- dslr
|       |   |   `-- images
|       |   `--- webcam
|       |       `-- images         
|       |-- amazon_31_list.txt
|       |-- dslr_31_list.txt
|       `-- webcam_31_list.txt
|-- pretrained_weights
|   `-- 02
|       `-- domain02accBEST_model_checkpoint.pth.tar
`-- source_pretrained_weights
    `-- 02
        `-- model_checkpoint.pth.tar

Download dataset

• Download the Office31 dataset (link) and unzip in ./data/office

• Download the text file (link), e.g., amazon_31_list.txt, dslr_31_list.txt, webcam_31_list.txt, in ./data/office

Download source-pretrained parameters (Fs and Cs of Figure 2 in our main paper)

• Download source-pretrained parameters (link) in ./source_pretrained_weights/[scenario_number]

ex) Source-pretrained parameters of A[0] -> W[2] senario should be located in ./source_pretrained_weights/02

Training and testing

• Arguments required for training and testing are contained in office-train-config.yaml
• Here is an example of running an experiment on Office31 (default: A -> W)
• Scenario can be changed by editing source: 0, target: 2 in office-train-config.yaml
• We will update the full version of our framework including settings for OfficeHome and Visda-C

Training

• Run the following command

python SFDA_train.py --config office-train-config.yaml

Testing (on pretrained model)

• As a first step, download SFDA pretrained parameters (link) in ./pretrained_weights/[scenario_number]

  ex) SFDA pretrained parameters of A[0] -> W[2] senario should be located in ./pretrained_weights/02

• or run the training code to obtain pretrained weights

• Run the following command

python SFDA_test.py --config office-train-config.yaml

Experimental results on Office31

• Results using the provided code

	A→W	D→W	W→D	A→D	D→A	W→A	Avg
Accuracy (%)	91.06	97.35	98.99	91.96	71.60	68.62	86.60