TensorHolography Reproduction

Paper info

• Title: Towards real-time photorealistic 3D holography with deep neural networks
• Author: L. Shi, B. Li, C. Kim, P. Kellnhofer and W. Matusik
• DOI: 10.1038/s41586-020-03152-0
• Nature 2021 Vol. 591 Issue 7849 Pages 234-239
• Github repo link (contains all most of the information you need)

Dataset download

• Google drive
• Taobao if can not access to google drive (too slow and easy to crash)

Prepare python environment

• Operating system: Ubuntu
• Package org: Anaconda, pip
• Environment file: environment.yml
• Shell command for prepare tensorholo env:

 # create conda python environment
conda env create -n tensorholo python=3.8.5
conda activate tensorholo

# install the packages manually
pip install opencv-python 
pip install numpy==1.18.5
pip install onnx
pip install tf2onnx

# Tensorflow install
# My tips: In the Github readme.md file: "This code runs with Python 3.8 and Tensorflow 1.15 (NVIDIA-maintained version to support training on the latest NVIDIA GPUS)." which means that it requires the access to nvidia website. My VPN software does not support Ubuntu/Linux, so it will easily fail. I did it in this way: when the install fails, you can download the .wheel file mannually (Ctrl + mouse click ) on Windows and then install it by "pip install your_wheel_file.wheel" for you env.
pip install nvidia-pyindex
pip install nvidia-tensorflow[horovod]

Follow the instructions from README.md for val, eval or re training.