Autoencoder Implementation with PyTorch

This project details the implementation of a convolutional autoencoder in PyTorch using the Omniglot dataset. The primary goal is to use the encoder-decoder architecture to compress the images into a low-dimensional latent space (latent space) and reconstruct them to be as close as possible to the original images.

Hyperparameters

The following hyperparameters were used during the training process:

Hyperparameter	Value
num_epochs	3
batch_size	64
optimizer	Adam
learning_rate	1e-3
weight_decay	1e-5

Results

Below is a glimpse at the results obtained after training:

Final Epoch Losses

The training and test losses at the final epoch (using the Mean Squared Error) are summarized below:

Metric	Value
Training Loss	$1.933 \times 10^{-5}$
Test Loss	$1.484 \times 10^{-5}$

Useful resources :

• Patrick Loeber's implementation of an autoecnder
• An Introduction to Variational Autoencoders on Arxiv
• Carnegie Mellon University, Pennsylvania lectures