This project implements an autoencoder network that encodes an image to its feature representation. The feature representation of an image can be used to conduct style transfer between a content image and a style image.

• The project is written in Python 3.7 and uses PyTorch 1.1 (also working with PyTorch 1.3).

• requirements.txt lists the python packages needed to run the project.

The architecture consists of an pre-trained VGG-19 encoder network that was trained on an object recognition task. The decoder is initialized randomly and trained with two losses.

Pre-trained models can be found in ./models and loaded with the code.

The Autoencoder is trained with two losses and an optional regularizer. A perceptual loss measures the distance between the feature representation of the original image and the produced image. A per-pixel loss measures the pixel-wise difference between input image and output image.

The configurations-folder specifies three configurations that can be used to train and test the network. The project only gets the exact path to the configuration that is used e.g. python main.py './configurations/train.yaml'.

• train.yaml trains the model from scratch. The default parameters can be found in the file.

• test.yaml tests the model and outputs the input as well as the output image of the network.

• test_multiple.yaml tests several models and displays the results next to each other.

This project is part of a bachelor thesis which was submitted in August 2019. This autoencoder network makes up one chapter of the final thesis. A slightly modified version of the chapter can be found in this repository as a pdf-file. Also, the chapter introduces all related formulas to this work.

The final thesis can be found here in a corrected and modified version.