EXoN: EXplainable encoder Network

This repository is the official implementation of EXoN: EXplainable encoder Network with Tensorflow 2.0.

Package Dependencies

python==3.7
numpy==1.19.5
tensorflow==2.4.0

Additional package requirements for this repository are described in requirements.txt.

Training & Evaluation

0. directory and codes

.
+-- logs (folder which contains results of all experiments: tensorboard)
|       +-- mnist_100 (with 100 labeled datasets)
|       +-- mnist_pre_design (Appendix A.2)
|       +-- cifar10_4000 (with 4000 labeled datasets)

+-- mnist (folder which contains source codes for MNIST dataset experiments)
|   +-- configs (folder which contains configuration informations of MNIST dataset experiments)
|       +-- mnist_100.yaml
|       +-- mnist_pre_design.yaml
|   +-- preprocess.py
|   +-- model.py
|   +-- criterion.py
|   +-- mixup.py
|   +-- main.py
|   +-- result.py
|   +-- pre_design.py (Appendix A.2)
|   +-- pre_design_result.py (Appendix A.2)

(source codes for CIFAR-10 dataset experiments)
+-- configs (folder which contains configuration informations of CIFAR-10 dataset experiments)
|       +-- cifar10_4000.yaml
+-- preprocess.py
+-- model.py
+-- criterion.py
+-- mixup.py
+-- main.py
+-- result.py
+-- requirements.txt
+-- accuracy.py
+-- accuracy.txt
+-- LICENSE
+-- README.md

1. How to Training & Evaluation

labeled_examples is the number of labeled datsets for running and we provide configuration .yaml files for 100 labeled datsets of MNIST and 4000 labeled datasets of CIFAR-10. And we add required tests and evaluations at the end of code.

1. MNIST datset experiment

python mnist/main.py --config_path "configs/mnist_{labeled_examples}.yaml"

• for Appedix A.2,

python mnist/pre_design.py --config_path "configs/mnist_{labeled_examples}.yaml"

To get visualization results of MNIST dataset experiment and Appendix A.2, use mnist/result.py and mnist/pre_design_result.py (following codes line by line).

2. CIFAR-10 dataset experiment

python main.py --config_path "configs/cifar10_{labeled_examples}.yaml"

To get visualization results of CIFAR-10 dataset experiment, use result.py (following the code line by line). To get summarized results of repeated CIFAR-10 dataset experiments, use accuracy.py (following the code line by line).

Results

1. MNIST

Effect of Regularizations

2. CIFAR-10

The activated latent subspace determines features of generated image

interpolation

• between same classes

• between different classes

Citation

@article{an2023custom,
	author = {SeungHwan An and Jong-June Jeon},
	doi = {https://doi.org/10.1016/j.patrec.2023.11.018},
	issn = {0167-8655},
	journal = {Pattern Recognition Letters},
	title = {Customization of latent space in semi-supervised variational autoencoder},
	url = {https://www.sciencedirect.com/science/article/pii/S0167865523003288},
	year = {2023},