GT CS7641 SP2020 Project - Team 3 Learning Machines
Domain Adaptation via Semi-supervised Image-to-image Translation
https://ast0414.github.io/semit/
In this project, we propose a (semi-)supervised image-to-image translation framework as a domain adaptation method based on a combination of variational autoencoders and generative adversarial networks upon a shared latent space assumption.
- environment.yml: the conda environment file used in this project.
- 0_data_exploration.ipynb: Data exploration and preprocessing.
- 1_mnist_baseline.ipynb: Train and test a MNIST baseline classification model.
- 2_kannada_baseline.ipynb: Train and test a Kannada-MNIST baseline classification model.
- 3_VAE_examples.ipynb: An example of VAE (from PyTorch Official Example) using our preprocessed MNIST data file.
- 4_train_VAE.py: Practice training of our VAE module separately.
- 5_train_GAN.py: Practice training of our GAN module separately.
- 6_train_UNIT.py: Train an unsupervised translator.
- 7_train_SUIT.py: Train a fully supervised translator.
- 8_train_SEMIT.py: Train a semi-supervised translator.
- 9_test_translator.py: Evaluate the classification accuracy for a trained translator.
- 10_generate_umap.py: Generate visualization of the latent space using UMAP or t-SNE.
- 11_visualize_embedding.ipynb: Plot the visualization from 10.
| Dataset | Input | Reconstruction | Translation |
|---|---|---|---|
| MNIST |  |
 |
 |
| K-MNIST |  |
 |
 |
