PyTorch-DANN

This repo is mostly based on https://github.com/wogong/pytorch-dann.

A PyTorch implementation for paper Unsupervised Domain Adaptation by Backpropagation

InProceedings (icml2015-ganin15)
Ganin, Y. & Lempitsky, V.
Unsupervised Domain Adaptation by Backpropagation
Proceedings of the 32nd International Conference on Machine Learning, 2015

Environment

• Python 3.8.5
• PyTorch 1.6.0

Note

Before running the training code, make sure that DATASETDIR environment variable is set to dataset directory.

• MNISTmodel()
  • basically the same network structure as proposed in the paper, expect for adding dropout layer in feature extractor
  • large gap exsits between with and w/o dropout layer
  • better result than paper
• SVHNmodel()
  • network structure proposed in the paper may be wrong for both 32x32 and 28x28 inputs
  • change last conv layer's filter to 4x4, get similar(actually higher) result
• GTSRBmodel()
• AlexModel
  • not successful, mainly due to the pretrained model difference
• ResNet50
  • Better and more stable results than AlexNet.

Run

For training on Office dataset using ResNet-50, first set configs in experiments/office.py, then run

$ python experiments/office.py

Result

	MNIST-MNISTM	SVHN-MNIST	SYNDIGITS-SVHN	SYNSIGNS-GTSRB
Source Only	0.5225	0.5490	0.8674	0.7900
DANN (paper)	0.7666	0.7385	0.9109	0.8865
This Repo Source Only	-	-	-	0.9100
This Repo	0.8400	0.7339	0.8200	-

	AMAZON-WEBVCAM	DSLR-WEBCAM	WEBCAM-DSLR
Source Only	0.6420	0.9610	0.9780
DANN (paper)	0.7300	0.9640	0.9920
This Repo (ResNet50)	0.8151	-	-

Credit

• https://github.com/wogong/pytorch-dann
• https://github.com/fungtion/DANN
• https://github.com/hehai131/DANN
• https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
• https://github.com/corenel/torchsharp
• https://github.com/corenel/pytorch-starter-kit