lfcnn - A TensorFlow framework for light field CNNs

Installation

It is recommended to use Conda to setup a new environment with tensorflow and GPU support. To install with GPU support, run

conda create -n lfcnn python=3.8 tensorflow-gpu=2.2 tensorflow numpy scipy imageio h5py cudnn cudatoolkit
conda activate lfcnn

Then, install the provided package using pip:

pip install lfcnn

Optional dependencies

Optionally, for some of lfcnn's features, install the following:

• matplotlib (via conda or pip)
• sacred (via pip)
• pymongo (via conda or pip)
• mdbh (via pip)

Installation on Windows

LFCNN is mostly compatible with all TF versions TensorFlow >= 2.0, however there is a bug in tf.keras that causes OOMs with data generators (which LFCNN uses) and multithreading and -processing. Therefore, we specify tensorflow >= 2.2 as a dependency, for which this bug has been resolved.

However, as of June 2020, TF 2.2 is not released on Anaconda for Windows. So for Windows, it is necessary to install TF via pip. However, installation of the compatible cuDNN and CUDA should still be performed via conda for simplicity. To setup the new environment with the correct CUDA and cuDNN versions, run

conda create -n lfcnn python=3.8 numpy scipy imageio h5py cudnn=7.6.5 cudatoolkit=10.1
conda activate lfcnn
pip install tensorflow==2.2 tensorflow-gpu==2.2

Furthermore, the Visual C++ redistributable has to be installed on Windows.

Finally, install LFCNN via pip as usual:

pip install lfcnn