GANPOP

Code, dataset, and trained models for "GANPOP: Generative Adversarial Network Prediction of Optical Properties from Single Snapshot Wide-field Images"

If you use this code, please cite:

Chen, Mason T., et al. "GANPOP: Generative Adversarial Network Prediction of Optical Properties from Single Snapshot Wide-field Images." arXiv preprint arXiv:1906.05360 (2019).

Setup

Prerequisites

Linux (Tested on Ubuntu 16.04)
NVIDIA GPU (Tested on Nvidia P100 using Google Cloud)
CUDA CuDNN (CPU mode and CUDA without CuDNN may work with minimal modification, but untested)
Pytorch>=0.4.0
torchvision>=0.2.1
dominate>=2.3.1
visdom>=0.1.8.3
scipy

Dataset Organization

All image pairs must be 256x256 and paired together in 256x512 images. '.png' and '.jpg' files are acceptable. Data needs to be arranged in the following order:

GANPOP_Pytorch # Path to all the code
└── Datasets # Datasets folder
      └── XYZ_Dataset # Name of your dataset
            ├── test
            └── train

Training

To train a model:

python train.py --dataroot <datapath> --name <experiment_name>  --gpu_ids 0 --display_id 0 
--lambda_L1 60 --niter 100 --niter_decay 100 --pool_size 64 --loadSize 256 --fineSize 256 --gan_mode lsgan --lr 0.0002 --which_model_netG fusion

To view epoch-wise intermediate training results, ./checkpoints/<experiment_name>/web/index.html
--lambda_L1 weight of L1 loss in the cost function
--niter number of epochs with constant learning rate
--niter_decay number of epochs with linearly decaying learning rate
--pool_size number of past results to sample from by the discriminator
--lr learning rate
--gan_mode type of GAN used, either lsgan or vanilla
--which_model_netG generator type; fusion, unet_256, or resnet_9blocks

Pre-trained Models

Example pre-trained models for each experiment can be downloaded here.

"AC" and "DC" specify the type of input images, and "corr" stands for profilometry-corrected experiment.
These models are all trained on human esophagus samples 1-6, human hands and feet 1-6, and 6 phantoms.
Test patches are available under dataset folder, including human esophagus 7-8, hands and feet 7-8, 4 ex-vivo pigs, 1 live pig, and 12 phantoms. To validate the models, please save the downloaded subfolders with models under checkpoints and follow the directions in the next section ("Testing").

Testing

To test the model:

python test.py --dataroot <datapath> --name <experiment_name> --gpu_ids 0 --display_id 0 
--loadSize 256 --fineSize 256 --model pix2pix --which_model_netG fusion

The test results will be saved to a html file here: ./results/<experiment_name>/test_latest/index.html.

Dataset

The full-image dataset can be downloaded here. Folders are structured in the same way as pre-trained models (AC and DC, with "corr" being profilometry-corrected). Please refer to README.txt for more details.

Issues

Please open new threads or report issues to mason@jhmi.edu

License

Acknowledgments

This code is inspired by pytorch-CycleGAN-and-pix2pix and FusionNet_Pytorch

Subsidized computing resources were provided by Google Cloud.

Reference

If you find our work useful in your research please consider citing our paper:

@article{chen2019ganpop,
  title={GANPOP: Generative Adversarial Network Prediction of Optical Properties from Single Snapshot Wide-field Images},
  author={Chen, Mason T and Mahmood, Faisal and Sweer, Jordan A and Durr, Nicholas J},
  journal={arXiv preprint arXiv:1906.05360},
  year={2019}
}

faisalml/GANPOP_Pytorch