/complex_tf

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Complex Tensorflow

This library extends functionality to several core Tensorflow ops. Many Tensorflow ops do not support computing complex-valued dtypes on the GPU. In a few cases the ops support complex-valued variables but they have not been registered and/or tested for complex dtypes.

In other cases a kernel needs to be explicitly written for complex-valued computation. The goal of this project is build up the capabilities of TensorFlow for complex dtypes. When the repository has been sufficiently developed and tested, I'll likely port it and create a a TF pull request.

Building

This repository provides Makefiles that should compile a shared library named complextf.so. You may find that some of the settings in complex_tf\core\kernels\Makefile need to be adjusted. This repository was wiritten as a private repository for my particular environment:

  • Ubuntu 16.04
  • Cuda 8.0
  • gcc 5.4
  • bazel 0.3.2
  • Python 2
  • bleeding-edge tensorflow

In order to make this library it is currently required to have the TensorFlow source files and the python tensorflow python package installed. The Makefile looks for TF's python package include directory by calling tf.sysconfig.get_include(). It relies on many of the headers in the source ditribution. It also requires the environment variable TF_SRC to be set with the path to the tensorflow source files.

Contributions

Contributions are welcome! If you decide to add an op, please follow these steps:

  • Make sure that a complex-valued version of yhe op is not already registered in the TF master branch
  • Investigate whether the op can just be "turned on", i.e. whether simply registering the op works
  • If not, write a kernel for the op.
  • Please add a few test cases for the op (and its gradient)

In adition, please feel free to make pull requests that will help configure the library more generally for other peoples' environments.

Using

After building the library pip install or pip develop the package. Then in python: python import tensorflow as tf import complex_tf as ctf The new shared ops will be loaded as a plugin and should work silently when using Tensorflow. If you see an error about an op having multiple registrations, then it is because tensorflow has added complex functionality to an op where it didn't exist before. At this time the only workaround is to use an older version of Tensorflow or edit the Makefile in complex_tf/core/kernels so that the offending op is not compiled.