Generating Monet Paintings via CycleGAN

This repository contains code to train a CycleGAN and analyze the results.

---

Data

• The dataset used was downloaded from Kaggle and contains two directories: Monet and Photo.
• The Monet directory contains 300 Monet paintings sized 256x256 in JPEG format.
• The Photo directory contains 7028 photos sized 256x256 in JPEG format.

---

Use pre-trained model

• The notebook "GenerateMonetFromExcistingModel.ipynb" uses google drive to upload the data. In order for that notebook to run smoothly, uploading a zip of the generate.zip folder to your google drive is necessary. Here is a link to the folder: https://drive.google.com/file/d/1NZMds1WLYqjvS8Qrhb3fIaOAhrWGK7XJ/view?usp=sharing
• The runbook generates Monets from 4 different pictures using a pre-existing model (the best model) for two different epoch numbers (epoch 2 and epoch 100). These results are then combined into one .png image called "GeneratedMonets.png".

---

To train model:

• Running main.py will train a model
• To tweak parameters you can select a subset of hyperparamters, by chaning the index in the training step.
• To monitor performance you can add your own Weight and Biasis api-key.
• 6 types of images will be generated during training:
  • Original photos
  • Generated Monet
  • Reconstructed
  • Original Monet
  • Generated photo
  • Reconstructed Monet

---

Analysis

• Data for the best model can be downloaded: https://drive.google.com/file/d/115vhzf6M-uzOwk8RGOGQh5i3a4YOnCUZ/view?usp=share_link
  • The data has already been pre-processed (so this step can be skipped in the analysis)
• The analysis conists of:
  • A RGB comparison of images from a previous training with 3 different metrics (MSE,PSNR,SSIM)
  • A grey-scale comparison of images from a previous training
  • A comparison of original photos and reconstructed photos