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)
- Lego House image dataset - this dataset was manually created using the Google Image feature and then processed to remove duplicates.
- 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 is a type of GAN which image-to-image translation without the need to image paid between domains.
- Standard - The first experiment was observed with the standard CycleGAN.
- ResNet - ResNet model as a discriminator was used.
- VGG16 - VGG16 model as a discriminator was used.
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.
- Standard
- Resnet
- VGG16
Links and references used in the project