theopfr / cycle-gan-pytorch

This repository contains an implementation of the Cylce-GAN architecture for style transfer along with instructions to train on an own dataset.

🐴🔄🦓 Cycle-GAN implemented in PyTorch

This repository contains an implementation of the Cylce-GAN architecture as proposed in the original paper along with instructions to train on an own dataset.

---

⬇️ setup:

1. install repository:

git clone https://github.com/theopfr/cycle-gan-pytorch.git
cd cycle-gan-pytorch

2. install requirements:

Requirements: Python>=3.7, Pytorch, torchvision, tqdm, numpy

pip install -r requirements.txt

---

🏋️ train:

1. create dataset:

• create a folder inside datasets/ with a descriptive name to store your dataset
• create two sub-folders trainA and trainB
• put all the images of one of the two image categories in one of the folders (e.g put all the images of horses in trainA and all the images of zebras in trainB)

You can find datasets here.

2. run the train script:

• navigate to src/
• run the train.py script by specifying the train arguments and hyperparameters with command line flags (find the train arguments in the table below; the run_name and dataset_name flag have to be set)
• example:

  python .\train.py --run_name "horse-zebra-run" --dataset_name "horse-zebra-dataset" --save_image_intervall 50 --resume False --epochs 100 --image_size 256 --batch_size 1 --num_res_blocks 9 --lr 0.0002 --lr_decay_rate 1 --lr_decay_intervall 200 --gaussian_noise_rate 0.05 --lambda_adversarial 1 --lambda_cycle 10 --lambda_identity 1 
  

---

🚩 train script flags/arguments:

argument	type	default	description
run_name	str	-	Name for the train run (a folder with this name will be created inside runs/ to store train metrics, model checkpoints and generated images).
dataset_name	str	-	Name of the folder which holds the dataset to train on.
resume	str	False	Options: "True", "False"; specifies if the train run should be continued if it was previously interrupted (if set to "False", the run-folder will be reinitialized).
save_image_intervall	int	50	Specifies after how many iterations (not epochs!) generated images should be saved to the run-folder.
epochs	int	100	The amount of epochs to train.
image_size	int	256	The image size to which all images with be resized (images will be quadratic).
batch_size	int	1	The batch size.
num_res_blocks	int	9	Amount of residual blocks in the generator model.
lr	float	0.0002	The learning rate.
lr_decay_rate	float	1.0	Decay rate of the learning rate (will be multiplyed to the learning rate, therefore 1.0 means no decay).
lr_decay_intervall	int	200	Specifies after how many iterations (not epochs!) the learning rate should be decayed (has to be >=1).
gaussian_noise_rate	float	0.05	Specifies how much gaussian noise will be applied to images before being fed into the discriminator model (will be multiplied with random noise and then added to the images).
lambda_adversarial	int	1	Specifies how much to weight the adverarial loss (will be multiplied with the loss).
lambda_cycle	int	10	Specifies how much to weight the cycle loss (will be multiplied with the loss).
lambda_identity	int	1	Specifies how much to weight the identity loss (will be multiplied with the loss).

All the default values are chosen as in the original paper to train on the horse-zebra dataset.

---