I explore a few variations of autoencoders (AE) and variational autoencoders (VAE) for tasks such as image generation and denoising.
Environment setup- I use few basic python libraries such as numpy, torch, torchvision, scipy, PIL, matplotlib, torchsummary and other builtin packages such as math for this work. These can installed individually using pip/conda in your local system.
Details of experiments:
-
Vanilla autoencoder that can generate new images of handwritten digits that are similar to the images in the MNIST dataset.
-
Denoising Convolutional Autoencoder that is capable of generating clear images from noisy inputs.
-
Fully Connected Variational Autoencoder (FC-VAE) for generating new images of handwritten digits.
References: [1] https://avandekleut.github.io/vae/ [2] https://github.com/seloufian/Deep-Learning-Computer-Vision/blob/master/eecs498-007/A6/variational_autoencoders.ipynb