M-ADDA: Metric-based Adversarial Discriminative Domain Adaptation [Paper]

Accepted in the ICML 2018 Workshop of Domain Adaptation for Visual Understanding (DAVU).

Description

The idea is to cluster the source dataset using the triplet loss and then cluster the target dataset using adversarial learning and the center-magnet loss. The Figure below shows the resultant clusters after using this method. The different colored stars represent the cluster centers, each corresponding to a different digit category.

Source (MNIST)	Target (USPS)

Requirements

• pytorch 0.4.0.
• torchvision 0.2.0

Download the datasets

cd datasets
bash download.sh

Test results

To obtain the test results, do the following two steps:

1. Download the checkpoints

cd checkpoints
bash download.sh

2. Run the following command,

python main.py -e usps2mnist mnist2usps uspsBig2mnistBig mnistBig2uspsBig -m test_model

The output should be,

mnist2usps          0.955676
mnistBig2uspsBig    0.980541
usps2mnist          0.951500
uspsBig2mnistBig    0.983100

which represent the accuracies obtained on the target test set.

• mnistBig, and uspsBig use the full training set.
• mnist, and usps use 2000 images from MNIST and 1800 images from USPS for training, respectively.

Training the models

0. Download the pretrained models and the usps dataset

cd checkpoints
bash download.sh
cd ..
cd datasets
bash download.sh
cd ..

1. Experiment USPS => MNIST

python main.py -e usps2mnist  -m train -rt 1

2. Experiment MNIST => USPS

python main.py -e mnist2usps  -m train -rt 1

3. Experiment MNIST ALL => USPS ALL

python main.py -e mnistBig2uspsBig  -m train -rt 1

4. Experiment USPS ALL => MNIST ALL

python main.py -e uspsBig2mnistBig  -m train -rt 1

Citation

If you find this useful for your research, please cite:

@Article{laradji2018m,
    title={M-ADDA: Unsupervised Domain Adaptation with Deep Metric Learning},
    author={Laradji, Issam and Babanezhad, Reza},
    journal={arXiv preprint arXiv:1807.02552},
    year = {2018}
}