Autoencoder implementation

Example output from the autoencoder with Adam optimiser written in C:

Original input (28x28):

Decoded output (16x1) ⟶ (28x28):

Description

This project implements an autoencoder (AE) trained on the MNIST dataset. The AE uses no convolutional layers and the encoder and decoder parts are comprised of 2 dense layers each. A variant of the Adam optimiser is used in each implementation.

The source code is written in Python using Pytorch/Numpy, Cython and C and execution times during training will soon be compared.

References:

• C autoencoder fork from C based MNIST classifier here
• Python autoencoder referenced from here
• Basic Neural Network in C
• Normal distribution generator for C
• PyTorch AE
• DEMON Adam optimiser

Current progress:

• AE implemented using Adam optimiser in C
• Implemented batch training and composite matrix operations in C
• AE implemented in PyTorch and Cython with DEMON Adam optimiser

Future aims:

• Add DEMON to the Adam optimiser to C source code.
• Implement a disentangled VAE in all languages.
• Test and compare execution times.