Deep learning phase recovery (DLPR)

Kaiqiang Wang and Edmund Y. Lam, "Deep learning phase recovery: data-driven, physics-driven, or a combination of both?", [Advanced Photonics Nexus].

@article{wangDeepLearningPhase2024b,
  title = {Deep Learning Phase Recovery: Data-Driven, Physics-Driven, or a Combination of Both?},
  author = {Wang, Kaiqiang and Lam, Edmund Y.},
  year = {2024},
  journal = {Advanced Photonics Nexus},
  volume = {3},
  number = {05},
  doi = {10.1117/1.APN.3.5.056006}
}

Environment requirements

• python == 3.8.18
• pytorch == 2.0.0
• matplotlib == 3.7.2
• numpy == 1.24.3
• opencv_python == 4.8.1.78
• scipy == 1.12.0
• scikit-image == 0.21.0
• tqdm == 4.65.0

The following software is recommended:

• CUDA == 11.7
• cuDNN == 8.5.0

File structure

DLPR
├── dataset_generation.py
├── train.py
├── one_infer.py
├── multiple_infer.py
├── functions
│   ├── data_read.py
│   ├── network.py
│   ├── propagation.py
│   └── train_func.py
├── datasets
│   ├── raw_images
│   │   ├── ...
│   ├── train_in
│   │   ├── ...
│   ├── train_gt
│   │   ├── ...
│   ├── test_in
│   │   ├── ...
│   └── test_gt
│       ├── ...
├── models_and_results
│   ├── model_weights
│   │   ├── ...
│   ├── results_DD_tPD_CD
│   │   ├── ...
│   └── results_uPD_tPDr
└──     ├── ...

Folders or files	Use
dataset_generation.py	generate paired hologram-phase datasets
train.py	train the neural network in the strategies of DD, tPD, and CD
one_infer.py	infer the neural network in the strategies of DD, tPD, or CD
multiple_infer.py	infer the neural network in the strategies of uPD or tPDr
/functions/	contains functions to be called
/datasets/raw_images/	contains raw images used for dataset generation
/datasets/train_in/	contains the generated holograms for network training
/datasets/train_gt/	contains the generated phases for network training
/datasets/test_in/	contains the generated holograms for network inference
/datasets/test_gt/	contains the generated phases for network inference
/models_and_results/model_weights/	contains the trained network weights and training process
/models_and_results/results_DD_tPD_CD/	contains the inference results of DD, tPD, and CD.
/models_and_results/results_uPD_tPDr/	contains the inference results of uPD and tPDr

Step 1: dataset generation

• Download image datasets such as ImageNet, CIFAR, COCO, LFW, and MNIST. Alternatively, randimage can be used to generate random unnatural images.
• Put the downloaded or generated images in the folder /datasets/raw_images/.
• Run dataset_generation.py to generate paired hologram-phase datasets.

python dataset_generation.py

• Then, hologram-phase datasets are generated in the folders /datasets/train_in/, /datasets/train_gt/, /datasets/test_in/, and /datasets/test_gt/.

Step 2: network training

• Run train.py to train the neural network in the strategy of dataset-driven(DD), trained physics-driven (tPD), or co-driven (CD).

python train.py -s 'DD'

python train.py -s 'tPD'

python train.py -s 'CD'

• Then, .pth, .csv, and .png are saved in the folder /models_and_results/model_weights/.

Step 3: network inference

Step 3.1: inference for DD, tPD, and CD strategies.

• Run one_infer.py.

python one_infer.py -s 'DD'

python one_infer.py -s 'tPD'

python one_infer.py -s 'CD'

• Then, inference results are saved in the folder /models_and_results/results_DD_tPD_CD/.

Step 3.2: inference for the strategies of untrained physics-driven (uPD) and trained physics-driven with refinement (tPDr).

• Run multiple_infer.py.

python multiple_infer.py -s 'uPD'

Note that python multiple_infer.py -s 'uPD' can be run before Step 2.

python multiple_infer.py -s 'tPDr'

Note that the default setting for the "cycles" is 1000, which can be adjusted as needed.

• Then, inference results are saved in the folder /models_and_results/results_uPD_tPDr/.