Adversarial Domain Adaptation Paper Implementations
This repository is a collection of some unsupervised adversarial domain adaptation models and their pre-trained checkpoints.
This repository is inspired by & heavily borrows code from - pytorch-domain-adaptation.
Currently, 3 popular Adversarial Domain Adaptation model implementations have been provided which were trained on MNIST and MNIST-M datasets as their source and target datasets. Visualization of activations and test results from target dataset for the 3 models are also provided using Jupyter Notebook. Ongoing work for adding more implementations of latest research papers in domain-adaptation to the repository.
Please follow the installation instructions from the parent repository.
Note
This work is purely for experimental purposes. If you are using this work, please ensure you cite the original authors of the papers and the base repository from which this code was borrowed and modified.
Implemented papers
Paper: Unsupervised Domain Adaptation by Backpropagation, Ganin & Lemptsky (2014)
Link: https://arxiv.org/abs/1409.7495
Description: Negates the gradient of the discriminator for the feature extractor to train both networks simultaneously.
Implementation: revgrad.py
Paper: Adversarial Discriminative Domain Adaptation, Tzeng et al. (2017)
Link: https://arxiv.org/abs/1702.05464
Description: Adapts the weights of a classifier pretrained on source data to produce similar features on the target data.
Implementation: adda.py
Paper: Wasserstein Distance Guided Representation Learning, Shen et al. (2017)
Link: https://arxiv.org/abs/1707.01217
Description: Uses a domain critic to minimize the Wasserstein Distance (with Gradient Penalty) between domains.
Implementation: wdgrl.py
Instructions
Note: There is no need to download any dataset for training / testing. The required datasets are provided in datasets directory. Pre-trained model checkpoints are present in trained_models directory.
- In a Python 3.6 environment, run:
$ conda install pytorch torchvision numpy -c pytorch
$ pip install tqdm opencv-python
- Train a model on the source dataset using the notebook
$ train_source.ipynb
- Choose an algorithm and pass it the pretrained network, for example, the notebook:
$ python revgrad.ipynb
- To test a model, use the below notebook Note: Uncomment the MODEL_FILE line corresponding to the path of the saved pre-trained checkpoint for the model you're testing:
$ python test_model.ipynb