DeepLearning2

This repository contains the code used to produce the Practical 2 project report titled 'From Lego Bricks to Actual Bricks: StyleTransfer with CycleGAN'.

Our aim with this project was to inspect and develop method to perform image-to-image translation using CycleGAN and two data-sets (Lego House and Real House images)

Data Sets

1. Lego House image dataset - this dataset was manually created using the Google Image feature and then processed to remove duplicates.
2. Real House image dataset - House price estimation from visual and textual features was used to gather real house images and then processed to have only the images which contain front images.

CycleGAN

CycleGAN is a type of GAN which image-to-image translation without the need to image paid between domains.

Networks Inspected

1. Standard - The first experiment was observed with the standard CycleGAN.
2. ResNet - ResNet model as a discriminator was used.
3. VGG16 - VGG16 model as a discriminator was used.

How to run the code

The repository is divided in three major parts.

• Directory Run_files_v1_house cantains the standard CycleGAN implementation.
• Directory Run_files_v2_house contains the ResNet based discriptor implementation for CycleGAN.
• Directory Run_files_v3_house contains the VGG16 based discriptor implementation for CycleGAN.
• Other parts are the results and minor development efforts.

To run the files, open the relevant file and use 'python <filename>' from terminal.

Results

1. Standard

2. Resnet

3. VGG16

Links

Links and references used in the project

• https://machinelearningmastery.com/cyclegan-tutorial-with-keras
• https://machinelearningmastery.com/how-to-develop-cyclegan-models-from-scratch-with-keras/