Direct Domain Adaptation Through Reciprocal Linear Transformations

This is the PyTorch implementation of workflow and experiments from "Direct Domain Adaptation through reciprocal linear transformations" by Tariq Alkhalifah and Oleg Ovcharenko, 2021.

Method

We propose the direct domain adaptation (DDA) approach to reduce the domain gap between two datasets. The process includes a series of linear operations on both datasets :

1. Cross-correlation of the input data with the a sample pixel (or the mean of sample pixels) from images of the same domain.
2. Convolution of the previous result with the mean of the auto-correlated input data from the other domain.

In the training stage, the input data are from the source domain and the mean of auto-correlated images is from the target domain. In the inference/application stage, the input data are from the target domain and the mean of auto-correlated images is from the source domain. The proposed method does not explicitly interfere with the training workflow and network architecture so might be incorporated into existing routines.

---

The only difference between data manipulations at the training and inference stages is that during training a random pixel from the same dataset is used for cross-correlation, rather than the mean of random pixels at the inference stage.

Classification accuracy

The notebooks in this repository explore performance of the DDA method against several setups when transferring knowledge from MNIST to MNIST-M datasets.

• The experiment 1 simply reproduces the DANN training, where a generative adversarial networks reduces the domain gap in the layers of the network.
• In the experiment 2, we extract the generator part of the previous and use it as a benchmark CNN architecture when no any domain adaptation methods are applied.
• The experiment 3 is similar to the previous but processing the data on the fly according to the DDA workflow.

	DANN	CNN	CNN+DDA
MNIST	0.988	0.996	0.990
MNIST-M	0.911	0.344	0.734

Download MNIST-M dataset

Follow instructions in this repo to download the target MNIST-M dataset.

Reference

@misc{alkhalifah2021direct,
      title={Direct domain adaptation through reciprocal linear transformations}, 
      author={Tariq Alkhalifah and Oleg Ovcharenko},
      year={2021},
      eprint={2108.07600},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Acknowledgements

The code is built over Unsupervised Domain Adaptation by Backpropagation.