ADDA

Adversarial domain discrimative adaptation

a tensorflow version implement of CVPR 2017 paper ADDA: https://arxiv.org/pdf/1702.05464

SVHN to MNIST

Requirements

python 3.6
tensorflow 1.4.0
numpy
scipy

Encoder network achitecture

lenet-5:
input (32,32,3)
conv1 filter (5,5,20) output: (28,28,20)
maxpool output:(14,14,20)
conv2 filter (5,5,50) output: (10,10,50)
maxpool output:(5,5,50)
flat1 output: (1250,120)
flat2 output: (120,84)
classifier output: (84,10)

More details can be seen in adda.py if you want to design another CNN network

Usage

This repositority only implement SVHN to MNIST, you can change another dataset such as USPS to MNIST or MNIST to USPS if you are interested.
**

Run python main.py

step1: training the source network
step2: training the target and discriminator network.
step3: test target dataset.

target accuracy is 63% (only source) and 77% (after adda).