Skip to main content

PYRO-NN is the high level Python API to the PYRO-NN-Layers known operators.

Project description


.. image:: :target: :alt: PyPI version

The python framework for the PYRO-NN layers implemented in (


PYRO-NN brings state-of-the-art reconstruction algorithm to neural networks integrated into Tensorflow.

To use pyronn you need to build the operators from sources or install the provided binaries from

The publication can be found under (https://frameworkpaper)


Install via pip :

.. code-block:: bash

pip install pyronn

or if you downloaded this repository ( using:

.. code-block:: bash

pip install -e .

If you encounter a problem during the installation have a look at our wiki:


Can be found <>_.


You can start with PYRO-NN

Potential Challenges

Memory consumption on the graphics card can be a problem with CT datasets. For the reconstruction operators the input data is passed via a Tensorflow tensor, which is already allocated on the graphicscard by Tensorflow itself. In fact without any manual configuration Tensorflow will allocate most of the graphics card memory and handle the memory management internally. This leads to the problem that CUDA malloc calls in the operators itself will allocate memory outside of the Tensorflow context, which can easily lead to out of memory errors, although the memory is not full.

There exist two ways of dealing with this problem:

  1. A convenient way is to reduce the initially allocated memory by Tensorflow itself and allow a memory growth. We suggest to always use this mechanism to minimize the occurrence of out of memory errors:

.. code-block:: python

config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
# ------------------ Call Layers ------------------
with tf.Session(config=config) as sess:
  1. The memory consuming operators like 3D cone-beam projection and back-projection have a so called hardware_interp flag. This means that the interpolation for both operators are either done by the CUDA texture or based on software interpolation. To use the CUDA texture, and thus have a fast hardware_interpolation, the input data need to be copied into a new CUDA array, thus consuming the double amount of memory. In the case of large data or deeper networks it could be favorable to switch to the software interpolation mode. In this case the actual Tensorflow pointer can directly be used in the kernel without any duplication of the data. The downside is that the interpolation takes nearly 10 times longer.

Note that the hardware interpolation is the default setup for all operators.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distribution

pyronn-0.0.1-py3-none-any.whl (41.6 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page