My implementation for a simple CNN-based autoencoder for 2D image data.
Requirements
Python 3.7 or greater, along the following libraries:
Numpy Matplotlib PyTorch
Example
The files driver.py and example_auto2D.py illustrate using autoencoder to learn a 2-D representation of a 128x128 images of Gaussian distributions. Elements in the dataset differ only by their mean position, which is chosen at random from a circle. Thus, the optimal representation of this data in 2-D latent space is a circle (representing the position). It can be seen in the results that the autoencoder successfully learns a topologically close approximation of this.
Example application of autoencoder to 128x128 pixelated images of Gaussian data
Average mean-squared error on training set w.r.t. epoch