Brief Description of MaTEx TensorFlow

Vinay Amatya edited this page Jul 13, 2018 · 9 revisions

MaTEx-TensorFlow provides a user-transparent distributed memory implementation of TensorFlow using MPI.

For more details, please read our arXiv paper User-transparent Distributed TensorFlow.

Features:

  1. Automatic parallelization of TensorFlow using MPI for GPU and CPU Clusters.
  2. Support for data parallelism -- which is sufficient for the majority of DL implementations
  3. Support for parallel data readers including formats such as MNIST, CIFAR-10/100, CSV, HDF5, and Parallel NETCDF (PNETCDF).

This git folder has all the requirements to install and run of MPI enabled TensorFlow. MaTEx-TensorFlow is based on the 1.7.0 release of TensorFlow.