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

• 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

Credit

• https://github.com/fungtion/DANN
• https://github.com/corenel/torchsharp
• https://github.com/corenel/pytorch-starter-kit