A PyTorch implementation for paper Unsupervised Domain Adaptation by Backpropagation
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Fazil Altinel 95d78532ce Logging improvements 4 years ago
core Logging improvements 4 years ago
datasets Fixes for initial training 4 years ago
experiments Looging improvements 4 years ago
models update SYNSIGNS-GTSRB source only expe result, finally succeed. 5 years ago
utils Looging improvements 4 years ago
.gitignore gitignore changes 4 years ago
LICENSE Initial commit 7 years ago
README.md Dataset environment variable fix 4 years ago
requirements.txt add requirements 5 years ago

README.md

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 3.6
  • PyTorch 1.0

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

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 Source Only - - -
This Repo 0.6528 - -

Credit