This project was made specifically for the Kaggle Competition "I'm something of a Painter Myself", you can download the Monet dataset from Kaggle to get started.
CycleGAN is a powerful technique for unsupervised image-to-image translation, which learns to map images from one domain to another without requiring paired examples. This repository implements the CycleGAN model as proposed in 2017 in the paper "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks" by Jun-Yan Zhu et al.
Install required packages with simple
pip install -r requirements.txt
Note that PyTorch will be installed with cu118 by default as specified in the requirements.txt
You can download the dataset directly from the following Kaggle's website. This implementation uses the .jpg files rather than tfrec by default. The structure of your dataset folder should look as follows
- data/
- monet_jpg/ # All monet images
- photo_jpg/ # All photos
Once you've got your dataset and virtualenv all in place, navigate to the project dir and run main.py with relevant arguments. For example, the following will train the model on a batch_size=64 and for num_epochs=10 whilst loading data from data_root=/data
cd CycleGAN/
python main.py --data_root=/data --batch_size=64 --num_epochs=10
Running with Weights and Biases, also setting wandb.watch(model) to True
python main.py --wandb=True --wandb_watch=True
If you're not using main.py you can use Trainer class as standalone manager for training the model. The standard way of fitting with Trainer is by calling Trainer.fit(**kwargs), however, if you want to run fit with wandb then instead use Trainer.fit_wandb(**kwargs) which just wraps fit(**kwargs) into the wandb context manager that initialises wandb project with a default name of Monet_CycleGAN which you can either modify or make the project in your wandb account.
Checkout the results directory to see some sample images from various experiments and see how the model evolved over time.
- Discriminators History
- Correct Loss functions
- Identity Loss function
- Better Architecture, acutally based on the paper
I thought it might be useful to be able to visualise the output of CycleGAN in sort of a live mode. Therefore, I decided to write a simple and interactive tool to explore how the input of the Generator gets transformed from domain A to B and back to A (cycle) and also see how the Discriminator judges its input. In a case of a CycleGAN the output of a Discriminator is not a single value (True/False or Real/Fake), but rather a grid. That grid can be visualised to see which areas of an input are more likely to be judged as Real/Fake.
Below is a small demonstration of the tool. To use it, simply run the following notebook notebooks/qualitative_eval.ipynb
