Joint Neural Phase Retrieval and Compression for Energy- and Computation-Efficient Holography on the Edge

This repository contains the source codes for the paper Joint Neural Phase Retrieval and Compression for Energy- and Computation-Efficient Holography on the Edge

Prepare the environment

To train/test with the codes, please prepare the conda environment first.

Prepare the Python environment:

1. Install anaconda on your machine. (For example, you can install MiniConda.)

2. After installing anaconda, run the following commands in the terminal:

   conda env create -f environment_dprc.yaml  
   conda activate dprc
   

3. Install the library compressAI with our modifications for entropy decoding (Note that using the library installed from other sources may produce unreasonable results).

   cd external_libraries/compressai
   python setup.py install 
   

Test on your own data:

If you want to run tests on the images collected by yourself, just put the images in ./data/collected, then open a terminal within current folder and run the following command:

cd codes; sh scripts/compress_tests.sh

After running, you'll get the test results in folder ./running/DPRC/results. And the results for 'R,G,B' channels are stored in three different folders.

Training

Please dowload DIV2K dataset and put it under ./data folder, then the final training data path is ./data/DIV2K_train_HR/. Note that the framework needs to be trained on GPUs with more than 15 GB memory when the batch size is set to 1.

Trigger the training by running the scripts:

1. Stage1:

   cd codes; sh scripts/train_stage1.sh
   

2. Stage2: Train the full framework (DPRC):

   cd codes; sh scripts/train_stage2.sh
   

We have provided a few data samples in ./data for a quick start when the full dataset is not ready yet.

Some parameters

In the training scripts, we can adjust the parameters by specifying:

• -g '1': specify the index of gpu on which the codes will be run, e.g, we specify '1' here.
• --dataset: specify which dataset to train/test on, use --dataset DIV2K for train/test on DIV2K dataset and use --dataset collected to test on self-collected images.
• --pixel_pitch 6.4: this will make the model use a pixel pitch of 6.4 um.
• --prop_dist 20: specify a reconstruction distance at 20 cm. You can change propagation distance according to your needs.
• --pretrain_path: since we adopt a two-stage training, it is used to specify the location of the pretrained model file produced by stage 1.
• --vis: if specified, we'll visualize the output images in tensorboard during training.
• --batch_size 1: the batch size used for training or test.
• --lr 1e-4: set the learning rate used for training the framework.
• --channel r: specify the channel of training image and corresponding wavelength, the supported choices include 'r', 'g' and 'b'.
• --compress: this must be specified during training and test in stage2, which will let the network integrate compression related modules.

Bibtex

We are happy to be cited when our codes are useful in your projects.

@article{DPRC,
author = {Wang, Yujie and Chakravarthula, Praneeth and Sun, Qi and Chen, Baoquan},
title = {Joint Neural Phase Retrieval and Compression for Energy- and Computation-Efficient Holography on the Edge},
year = {2022},
volume = {41},
number = {4},
journal = {ACM Trans. Graph.},
month = {jul},
articleno = {110},
numpages = {16}
}

Acknowledgement

Some codes are adapted from neural-holography by @ Suyeon Choi and HiFiC by @ Justin-Tan, we thank the authors for opensourcing their awesome resources.