Adversarial Discriminative Domain Adaptation

A tensorflow implement of ADDA

Environment

• python 3
• tensorflow 1.09
• sklearn
• matplotlib 2.2.2
• numpy 1.14.2

Network

Network achitecture referenced LeNet for mnist digits datasets.

• For encoder:
  • 3 x Conv + 1 x Linear
• For classifier:
  • 1 x Conv + softmax
• For Discriminator:
  • 3 x Linear
• More detial can be seem in "adda.py". You can design your network according to your adaptation task.

Usage

Note: This repository is still semi-finished. Dataset only MNIST and USPS are support.

python main.py --step=1 --epoch=20

• step:
  • Step 1 is training source network.
  • Step 2 is training target encoder and discriminator.
  • Step 3 is evaluation of adda network
• epoch: the training epoch in the step.

Result

	MNIST(Source)	USPS(Target)
Source Encoder + Source Classifier	99.23%	78.52%
Target Encoder + Source Classifier	-	91.38%

• Target accuracy: 78.52%(without adapting) vs 91.38%(Adapted)

Visualization result using t-SNE

• Orign imageand feature map of encoder output after adaptation