Deep Phase Retrieval (DPR)

DPR is a deep neural network for direct phase retrievals of single-particle diffraction patterns from single-pulse coherent diffraction imaging experiments using X-ray free electron lasers.

Requirements

Python = 3.11.5

NumPy = 1.26.0

SciPy = 1.11.3

scikit-image = 0.20.0

PIL = 10.0.1

h5py = 3.9.0

PyTorch = 2.1.0

Torchvision = 0.16.0

CUDA = 11.8

cuDNN = 8.7.0

Notes

1. The pretrained network is for the diffraction patterns with oversampling ratios in the range of 10 to 20 along each axis and total diffracted intensities in the range of 10⁶ to 10⁷.

2. Coefficients for the loss function might require to be adjusted for training datasets with different conditions.

3. We used NVIDIA Irregular Mask Dataset from https://research.nvidia.com/labs/adlr/publication/partialconv-inpainting. Please check the file paths in 'module.dataset.IrregularMaskDataset' when using 'generate_dataset.ipynb'. Other datasets, EMNIST and CIFAR-100, are from torchvision library.

4. We imported following codes.

   Partial Convolution from https://github.com/NVIDIA/partialconv

   Fast Fourier Convolution (FFC) from https://github.com/pkumivision/FFC

   FFC ResNet Block from https://github.com/advimman/lama

   Partial Convolution from https://github.com/NVIDIA/partialconv

   Adaptive Sharpness-Aware Minimization (ASAM) from https://github.com/SamsungLabs/ASAM

5. When using DPR or weighted partial convolution, please cite our paper with proper references.

   https://doi.org/10.1038/s41524-025-01569-7

6. Contact: Sung Yun Lee, sungyun98@postech.ac.kr