DiDA：Disentangled Synthesis for Domain Adaptation

Introduction

In this paper, we present a method for boosting Domain Adaptation(DA) performance by leveraging Disentanglement(Di) analysis. The key idea is that by learning to separately extract both the common and the domain-specific features, one can synthesize more target domain data with supervision, thereby boosting the domain adaptation performance.

Usage

DiDA is implemented and tested with Pytorch in python2 and use tensorboard for visualization. Take DSN as backbone and dataset as MNIST-MNISTM as an example：

python train.py --dataroot ./traindata --name DSN_m --model DSN_mv2 --no_dropout --lr 0.01 --lr_de 1 - gpu_ids 0 --lr_policy nopolicy --which_method DSN

run the DA backbone

python train2.py --dataroot ./traindata --name Di_DSN_m --model Di_DSN_m --no_dropout --lr 0.00001 --lr_de 1 - gpu_ids 0 --lr_policy nopolicy --which_method DSN --batchSize 64 --which_epoches_DA [which_epoches_DA] --which_usename_DA DSN_m

run the Di stage

python train3.py --dataroot ./traindata --name DSN_miter1step1 --model DSN_m_iter --no_dropout --lr 0.01 --lr_de 10 - gpu_ids 0 --lr_policy nopolicy --which_method DSN --batchSize 100 --which_epoches_DA [which_epoches_DA] --which_usename_DA DSN_m --which_epoches_Di [which_epoches_Di] --which_usename_Di Di_DSN_m 

iteration of DA

python train2.py --dataroot ./traindata --name Di_DSN_iter1m --model Di_iter_DSN_m --no_dropout --lr 0.00001 --lr_de 1 - gpu_ids 0 --lr_policy nopolicy --which_method DSN --batchSize 64 --which_epoches_DA [which_epoches_DA] --which_usename_DA DSN_miter1step1 --which_epoches_Di [which_epoches_Di] --which_usename_Di Di_DSN_m 

iteration of Di

use tensorboard for visualization:

cd <your path>/DiDA
tensorboard --logdir=runs --port=6006

Citation

If you use this code, please cite the following paper.

@article{cao2018dida,
  title={DiDA: Disentangled Synthesis for Domain Adaptation},
  author={Cao, Jinming and Katzir, Oren and Jiang, Peng and Lischinski, Dani and Cohen-Or, Danny and Tu, Changhe and Li, Yangyan},
  journal={arXiv preprint arXiv:1805.08019},
  year={2018}
}