MSDA

A soft version of MSDA

Python 3.6 + Tensorflow 1.12.0

This repo has implemented digits classification experiments in the paper: Multiple Source Domain Adaptation with Adversarial Learning[https://arxiv.org/abs/1705.09684] There a lot of codes that can be written more elegently so if you are interested, feel free to pull requests.

<center><img src="https://github.com/daoyuan98/MSDA/blob/master/images/model.png" width="800"></center>

Acknowledgement

Thanks to the code from https://github.com/pumpikano/tf-dann , This repo has referenced much of the code there and this code is currently a simple extension from his code.

Experiments Results

On digits Classifaction, I have carried out two experiments and have received satisfactory results.

1. Sv+Mm+Sy-->Mt

Domain Accurarcy and Iteration（left） and Model Accurarcy on Target Domain(right)

<div align="center"> <img src="https://github.com/daoyuan98/MSDA/blob/master/images/1_d_acc.png" width="390"/><img src="https://github.com/daoyuan98/MSDA/blob/master/images/1_tar_acc.png" width="390"/> </div>

2. Sv+Mt+Sy-->Mm

Domain Accurarcy and Iteration（left） and Model Accurarcy on Target Domain(right)

<div align="center"> <img src="https://github.com/daoyuan98/MSDA/blob/master/images/2_d_acc.png" width="390"/><img src="https://github.com/daoyuan98/MSDA/blob/master/images/2_tar_acc.png" width="390"/> </div>

Discussion

• I didn't put forward the 3rd experiment in the paper because I found that even after many many iterations(>130k), the model is overfitting but the accurarcy(~.764) is still far from that in the paper(.818).
• The different training epochs may depend on the difficulty of different tasks.

Any advice and comments are welcome!