This repository contains all the codes for the Autoencoder question in HW3 of CE-40959: Deep Learning Course, presented by Dr. Soleymani in Sharif University of Technology. In this question, a complete set of Jupyter Notebook and python scripts is prepared for implementing a Conv/Deconv Autoencoder model applied on images.
In the jupyter notebook, we are going to work on farsi OCR dataset. As its name implies, it is like famous MNIST dataset but it consists of images of handwritten digits in farsi.
The problem we define for this dataset is to reconstruct original image after making some random rotations. We want to develop a model which recieves as input a rotated image and outputs its original without rotation. Meanwhile, a latent embedding is learned in the training process and its quality will be examined later by some t-SNE visualizations.
Alongside the notebook, there some python files with TODO sections filled with proper lines of code. For Each TODO section, a comprehensive description of the required code is provided.