Skip to content

Latest commit

 

History

History

noiseGAN

README.md

Synthesizing Camera Noise using Generative Adversarial Networks

noiseGAN

We provide a PyTorch implementations for paper entitled "Synthesizing Camera Noise using Generative Adversarial Networks".

This implementation is based on the original PyTorch implementation of the CycleGAN. We used an old branch of this repository. For more information, check below.

If you wish to run our code with a newer PyTorch version (>0.3.1), please check this implementation.

Prerequisites

  • Python 3
  • NVIDIA GPU + CUDA CuDNN

Getting Started

Installation

We used PyTorch 0.3.1, please check https://pytorch.org/get-started/previous-versions/ for installation instructions. You can try install with the following pip command:

pip install torch==0.3.1  torchvision==0.2.1

If the above command fails, you'll need to find the correct whl for installing the pytorch 0.3.1.

Also, you must install the requirements inside requirements.txt:

pip install -r requirements.txt

Training a model

  • Run python -m visdom.server in a console and enter http://localhost:8097 in order to view training results and loss plots.
  • Run the train script:
python train.py --dataroot_A <path-to-lower-ISO> --dataroot_B <path-to-higher-ISO> --name <name-of-model> --model cycle_gan --no_dropout --lambda_std 10 --lambda_low_freq 10 --niter 40 --niter_decay 0

For checking the parameters, please check the files in options folder.

Downloading and testing our models

  • Download our trained models (~500MB) and extract them inside checkpoints. You can run the following bash:
bash ./download_trained_models.sh
  • Run the test script
python generate_noise.py --dataroot_A <input-imgs-dir> --dataroot_B None --name <name-of-model> --model test --dataset_mode single --no_dropout
  • By using the following command, you can generate noise for 3 sample images
python generate_noise.py --dataroot_A ./sample_imgs/SIDD_N_S6_clean --dataroot_B None --name sidd_cleanTo3200_S6 --model test --dataset_mode single --no_dropout

Running on Docker

If you are having problems on installing the PyTorch 0.3.1, you can try our Docker. First, you will need to install the following requirements:

Them, you can build the Docker image (this may be take a while) by using:

docker build -t noise-gan -f Dockerfile .

You can run the script for generating noise by:

docker run --gpus all --rm noise-gan python3 generate_noise.py --dataroot_A ./sample_imgs/SIDD_N_S6_clean --dataroot_B None --name sidd_cleanTo3200_S6 --model test --dataset_mode single --no_dropout

PS: Some warnings might come due to visdom visualization, but the test code will run just fine.

PS2: for any changes in the directories and/or code, you should rebuild the Docker image (which will be much faster due to cached files).