tutorials

Small tutorial examples

This directory contains some small code examples showing a series of concepts in a minimal amount of code.

First prepare your environment to run PyTorch Lightning on IPU. Note that --force-reinstall is required to make sure the installed PyTorch version is compatible with both PopTorch and PyTorch Lightning:

python3 -m venv tutorials-venv
source tutorials-venv/bin/activate
pip3 install -U pip
pip3 install <path-to-poptorch-wheel>.whl
pip3 install -r requirements.txt --force-reinstall

Now enable the Poplar SDK:

source <path-to-poplar-sdk>/popart-ubuntu_18_04-2.4.0+2529-969064e2df/enable.sh
source <path-to-poplar-sdk>/poplar-ubuntu_18_04-2.4.0+2529-969064e2df/enable.sh

Each example can be run as follows:

python3 NAME.py

The comments in each one describe what it is doing/showing.

There is a simple model which we use to compose the examples.

• simple_torch_model.py A simple model for us to use in the examples.

The examples show:

• simple_lightning_ipu.py Shows how to run a simple Lightning model on IPU.
• ipu_options_and_dataloading.py Shows how to pass PopTorch options. Includes tutorial on replication and dataloading.
• pipelined_ipu.py Shows how to pipeline a model across IPUs.
• custom_losses.py Shows how to add a custom loss on the IPU in a simple regression model.