Implementation of Unsupervised Domain Adaptation by Backpropagation (Y.Ganin & V.Lempitsky ICML'15)
Ubuntu 14.04 LTS
Python 3.5.2 with Anaconda3 4.2.0
| Library | Version |
|---|---|
| chainer | 2.0.0 |
| cupy | 1.0.0 |
| numpy | 1.14 |
Source: MNIST
Target: MNIST-M
Dataset link : Mnist2MnistM
MNIST-M original link(Projects Link) : https://yaroslav.ganin.net/
Validation data : target data (without Train Data).
| Train Data(Source/Target) | Accuracy(Paper | Accuracy(Impl |
|---|---|---|
| Mnist/Mnist-M | 81.49% | 80.81% |
train/loss/LP: Label Predictor Loss
train/loss/DC: Domain Classifier Loss
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.