Note: While this repository is useful for reproducing results from Häusser et al., please consider using the salad domain adaptation library in the future: https://domainadaptation.org
This repository contains an implementation of "Associative Domain Adaptation" [1].
Right now, it features the SVHN -> MNIST transfer as described in the paper.
The results line up the the ones reported in the paper, even slightly better at Accuracy: 98.06 % / Error: 1.94 % on the MNIST Validation set.
This implementation is meant to be minimalistic, for easy adaptation to other projects.
To train a model with standard settings, execute
> python train.py
Notes:
- The hyperparameters where loosely inspired by the ones reported in the original publication, but not too much finetuning was necessary to get to this result.
- Note the use of the InstanceNormalization layer, which is similar, but not exactly similar to the reference implementation provided by the authors.
Original Paper: https://arxiv.org/abs/1708.00938 Official Repo: https://github.com/haeusser/learning_by_association
@inproceedings{haeusser2017associative,
title={Associative domain adaptation},
author={Haeusser, Philip and Frerix, Thomas and Mordvintsev, Alexander and Cremers, Daniel},
booktitle={International Conference on Computer Vision (ICCV)},
volume={2},
number={5},
pages={6},
year={2017}
}
In case of any questions with this repository, either use the issue tracker or contact me directly.