This repo contains the code for digits experiments SVHN->MNIST of the Implicit Class-Conditioned Domain Alignment for Unsupervised Domain Adaptation paper.
For installation instructions, please refer to implicit_alignment.
The digits data are defined in digit.py. By default, this repo runs experiments for RS-UT where both source and target domains are unbalanced under mild unbalance. This configuration can be modified in digit.py.
Running this repo with makefile should yield the following results (with variations of random seeds and pytorch versions):
- source-only: 61%
- DANN: 67%
- DANN+implicit: 88%
DANN is implemented based on modifications of the MDD code.
Note, this repo only supports SVHN->MNIST, for other datasets please refer to the main repo implicit_alignment.
Please create an issue for question/bug reports. Thanks!