This is a implementation of Domain-Adversarial Training of Neural Networks
with pytorch. This paper introduced a simple and effective method for accompli-
shing domian adaptation with SGD with a GRL(Gradient Reveral Layer). According
to this paper, domain classifier is used to decrease the H-divergence between
source domain distribution and target domain distribution. For the tensorflow
version, you can see tf-dann.
python3.6.2
pip install -r requirements.txt
In this work, MNIST and MNIST_M datasets are used in experiments. MNIST dataset
can be downloaded withtorchvision.datasets. MINIST_M dataset can be downloa-
ded at Yaroslav Ganin's homepage. Then you can extract the file to your data dire-
ctory and run thepreprocess.pyto make the directory able to be used with
torchvision.datasets.ImageFolder:
python preprocess.py
You can run
main.pyto implements the MNSIT experiments for the paper with the
similar model and same paramenters.The paper's results and this work's results a-
re as follows:
| Method | Target Acc(paper) | Target Acc(this work) |
|---|---|---|
| Source Only | 0.5225 | 0.5189 |
| DANN | 0.7666 | 0.7600 |
Experiment on SVHN->MNIST is added in this project, but some bugs are not fixed.
The accuracies of source and target domains are not good at the same time.
Experiment on SynDig->SVHN is added.