Graphcore

PopART MNIST Training Demo

This example trains a network on the MNIST dataset using PopART.

File structure

popart_mnist.py The main PopART program that uses a linear network;
get_data.sh Script to fetch the images and labels;
README.md This file;
requirements.txt Specifies the required modules.

How to use this demo

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.

The PopART Python API only supports Python 3. It is recommended to use a virtualenv.

Install the required modules:
```
pip3 install -r requirements.txt
```
Download the data.
```
./get_data.sh
```
This will create and populate a data directory.
Run the program.
```
python3 popart_mnist.py
```

Options

The popart_mnist.py script has a few command line options.

-h                    Show usage information.

--batch-size          Sets the batch size. This must be a multiple of the
                      replication factor.

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

--epochs              Number of epoch to train for.

--num-ipus            Number of IPUs to use.

--pipeline            Pipeline the model over IPUs. Only valid for this model
                      if the number of IPUs is double the replication factor.

--replication-factor  Number of times to replicate the graph to perform data
                      parallel training. This must be a factor of the number of
                      IPUs. Defaults to 1.

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

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

Examples:

python popart_mnist.py --num-ipus 2 --pipeline
python popart_mnist.py --num-ipus 4 --replication-factor 2 --pipeline