This GitHub repository contains the Keras implementation on the MNIST Dataset and the Supplementary Material of the Supervised Dimensionality Reduction and Image Classification Utilizing Convolutional Autoencoders.
- Clone the code to local machine.
git clone https://github.com/JohnNellas/CSAE
- Create the environment from the
environment.ymlfile.
conda env create -f environment.yml
- Activate the environment.
conda activate csae_env
- Run the experiment on MNIST dataset and produce a dimensions_latent_space dimensional latent space.
python3 main.py dimensions_latent_space
Run the experiment for a CSAE with a 10D Latent Space.
python3 main.py 10
The script execution produces a checkpoint folder along with a folder that contains the classification and visualization results, which are entitled as checkpoints_classification_ae and classification_and_visualization_results_CSAE, correspondingly.
The script execution involves the evaluation of the utilized classification methodologies and the creation of visualizations that emerge from the Convolutional Supervised Autoencoder.
The Classification methodologies are evaluated during the script execution and the results are saved in the classification_results sub folder of the directory classification_and_visualization_results_CSAE.
Additionally, the script execution creates a set of visualizations. Initially, a Latent Space visualization is conducted. In more detail, if the derived Latent Space has a dimensionality equal to two (2) or three (3), then a 2D or 3D scatter plot of the raw embeddings is constructed respectively, while for a dimensionality greater than 3, the Latent Space is visualized by creating a 2D scatter plot of the embeddings produced by applying the PCA or t-SNE algorithms onto the resulting Latent Representations. Moreover, concerning the case where the Latent Space of CSAE is two dimensional, a 2D scatterplot of the raw embeddings along with the Classifier's Decision boundary is constructed, where some points are replaced by the corresponding original images, while for the same case, a figure of the images derived from decoding a grid of latent representations is also generated. Finally, a plot of the accuracy, reconstruction and classification error per epoch during the training and validation procedure is constructed. The resulting visualizations are saved in the visualization_results sub folder of the directory classification_and_visualization_results_CSAE. Some examples of the afforemetioned figures, for the two dimensional latent space are presented below.
Finally, the supplementary material is contained in the pdf file entitled as supplementary_material.pdf.