Project Title

Generative Models: Convolutional Autoencoders and Variational Autoencoders

Task and Model

The task is to generate as similar images (fake images) as the given dataset (real images).
The Convolutional AutoEncoder(CAE) and Variational AutoEncoder(VAE) are created here, to reconstruct images for the UT Zappos50K Dataset.

Prerequisites

I uploaded the zipped dataset(625 Mb) to Google drive, and used Google Colab to load & unzip the data, the dataset can be found here https://drive.google.com/file/d/1nYEgytPOkFyUjDQfBGzwCQbszf6OE143/view?usp=sharing. You can also directly upload the unzipped dataset to your Colab, just remember to change the path in Step1 -'DATASETS/UTZappos50K'.

Introduction

The UT Zappos50K Dataset used here contains 4 types of images, i.e. Boots, Sandals, Shoes, Slippers.
The Adam optimizer was utilized with learning rate of 0.001.
As for the loss function, for CAE, I used mean squared loss; For VAE, I used binary cross entropy (reconstruction loss) plus KL-divergence (regularization loss).

Data Visualization

For CAE, the training loss looks like:

CAE's outcome display:

original images(left) v.s generated images(right)

For VAE, the Reconstruction Loss, KL-Divergence Loss, and total loss look like:

VAE's outcome display:

original images

generated images

In addition, the video obtained from linear interpolation:

image transforming gif

Acknowledge

Special thanks to CIS522 course's TA and professor, for providing the data set and guidance of the training process