custom_gradient

Graphcore

---

Creating a TensorFlow Custom Operator with Gradient

Creates a custom operator (a batched dot product) defining both the forward operator and its gradient in Poplar code. Uses the custom operator in a simple logistic regression optimisation program which checks the results with the custom operator match those from the built-in operator.

File structure

• Makefile Simple Makefile that builds the Poplar shared object.
• product.cpp Poplar code that describes the forward and grad operators.
• regression.py TensorFlow program that uses the custom operator to do logistic regression.
• requirements.txt Required packages.
• test_regression.py Script for testing this example.
• README.md This file.

How to use this demo

1. Prepare the environment.

   Install the Poplar SDK following the instructions in the Getting Started guide for your IPU system. Make sure to source the enable.sh script for Poplar.

2. Build the custom operator and then run the Python code:

make
python3 regression.py