This repository containes different SOTA unsupervised domain adaptation implementations in pytorch lightning.
Paper: Deep CORAL: Correlation Alignment for Deep Domain Adaptation (DEEP_CORAL)
link: https://arxiv.org/abs/1607.01719
Paper: Unsupervised Domain Adaptation by Backpropagation, Ganin & Lemptsky (2014) (DANN)
link: https://arxiv.org/abs/1409.7495
Paper: Adversarial Discriminative Domain Adaptation, Tzeng et al. (2017) (ADDA)
link: https://arxiv.org/abs/1702.05464
Paper: Conditional Adversarial Domain Adaptation (CDAN)
link: https://arxiv.org/abs/1705.10667
All models implemented work at features level, in other words, they all use a backbone to extract features from the input data and a classifier network to correctly classy the source domain.
Next, depending on the unsupervised domain adaptation method the alignment network, the adversarial network as well as the gradient reversal layer are used.
Depending on the chosen block different models can be imple-mented.
During training, the downstream task, as well as unsupervised domain adap-tation, are performed.
As a result, the feature extractor (backbone) and the Classifier Network can be used at testing time to make inference on the unlabelled target domain
To completely install the requirements run:
pip install -U -r requirements.txt
After setting the right configuration in ./conf/config.yaml run:
python main.py
During training one can inspect the model behaviour with tensorboard using the following command:
tensorboard --logdir ./logs