kskdev / DANN

Unsupervised Domain Adaptation by Backpropagation (chainer implementation)

chainer-DANN

Implementation of Unsupervised Domain Adaptation by Backpropagation (Y.Ganin & V.Lempitsky ICML'15)

Environment

Ubuntu 14.04 LTS
Python 3.5.2 with Anaconda3 4.2.0

External Libraries

Library	Version
chainer	2.0.0
cupy	1.0.0
numpy	1.14

Dataset

Source: MNIST
Target: MNIST-M

Dataset link : Mnist2MnistM

MNIST-M original link(Projects Link) : http://yaroslav.ganin.net/

Implementation Result

Validation data : target data (without Train Data).

Train Data(Source/Target)	Accuracy(Paper	Accuracy(Impl
Mnist/Mnist-M	81.49%	80.81%

Accuracy

Loss

train/loss/LP: Label Predictor Loss
train/loss/DC: Domain Classifier Loss

Usage

Open train.py and change data_root path.
Run python train.py.

Class label is acquired from the directory where the image file is located and the directory name must be class ID.
Therefore, image files must be arranged and be renamed directory for each class.