mnist

Graphcore

---

PyTorch PopART MNIST Training Demo

This example demonstrates how to train a network defined in PyTorch on the MNIST dataset using a popart backend which is a port of ONNX that targets Graphcore's Poplar Libraries.

File structure

• pytorch_popart_mnist.py The main file.
• 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 scripts for Poplar and PopART.

2. Run the program. Note that the PopART Python API only supports Python 3. Data will be automatically downloaded using torch vision utils.

   python3 pytorch_popart_mnist.py
   

Options

The program has a few command-line options:

-h Show usage information.

--batch-size Sets the batch size.

--batches-per-step Number on mini-batches to perform on the device before returning to the host.

--epochs Number of epoch to train for.

--simulation Run with the IPU_MODEL device instead of hardware.

--log-graph-trace Turn on IR logging to display the graph's ops.