Adversarial Teacher-Student Representation Learning for Domain Generalization

This is a PyTorch implementation of the paper Adversarial Teacher-Student Representation Learning for Domain Generalization

Paper | Supp

Usage

Setup

See the requirements.txt for environment configuration.

pip install -r requirements.txt

Datasets

• PACS: I prepared a version of the PACS dataset (ACS-P) here, where A, C, and S domains are used as the source domains and P is set as the target domain. You also can change to your own dataset, but keep the following structure:

ACS-P/
├── ACS
│   ├── 1
│   │   ├── 0
│   │   ├── 1
│   │   ├── 2
│   │   ├── 3
│   │   ├── 4
│   │   ├── 5
│   │   └── 6
│   ├── 2
│   │   ├── 0
│   │   ├── 1
│   │   ├── 2
│   │   ├── 3
│   │   ├── 4
│   │   ├── 5
│   │   └── 6
│   └── 3
│       ├── 0
│       ├── 1
│       ├── 2
│       ├── 3
│       ├── 4
│       ├── 5
│       └── 6
└── P
    └── 0
        ├── 0
        ├── 1
        ├── 2
        ├── 3
        ├── 4
        ├── 5
        └── 6

Training

Warmup Phase

cd warmup/main/
python train.py

Adversarial Teacher-Student Representation Learning

cd main/main/
python train.py

Results

The reproduced result is a little bit lower than the reported result in the paper. Hope that someone could help me fill the gap.

Target	ResNet-18		ResNet-50
	Reproduced Accuracy	Reported Accuracy	Reproduced Accuracy	Reported Accuracy
Photo	94.4	97.3	97.7	98.9

You can use the file samples.ipynb to view images generated from the Augmentor. For examples: