This repository provides a Pytorch implementation of DANN (Domain Adversarial Neural Networks) introduced by Ganin et al. [1]. The code includes the supporting scripts for reproducing the results obtained in [1] for the domain adaptation task, explicitly using the datasets mnist and mnist modified.
1- Python 3.7.4
2- Pytorch 1.13
Aiming at simplifying Python environment issues, we provide the docker container used to conduct the experiments' results obtained with this code.
This code reproduces the experiments carried out among the mnist and mnist-modified datasets. The difference between such datasets is represented in the figure from [1].
Although other network architectures can be considered in this repository, here is expressly provided the network architectures used in [1] for mnist vs. mnist modified experiments. The following figure, also taken from [1], specifies the architectural details:
Additionally, during the experiments, the same set of hyper-parameters were used.
The results were computed regarding Accuracy and F1-Score; the following table shows them.
| Source | Mnist | |
|---|---|---|
| Target | Mnist-M | |
| Metrics | Accuracy | F1-Score |
| No Domain Adaptation | 45 | 49.8 |
| DANN | 70.4 | 70.8 |
[1] Ganin and V. Lempitsky, “Unsupervised domain adaptation by backpropagation,”arXiv preprint arXiv:1409.7495, 2014.