TensorFlow Graph Optimizations

Tatiana Shpeisman


This talk gives an overview of TensorFlow graph optimization technologies. The talk starts from an introduction to TensorFlow. It then goes into an in-depth dive into Grappler - a graph optimizer for TensorFlow graphs. The second part of the talk focuses on MLIR - a new compiler infrastructure designed to support optimizations not only on TensorFlow graphs but also on multiple other representations used by components of the TensorFlow software stack (TensorFlow Lite, XLA etc.).