This repository contains a catalogue of application examples that have been optimised to run on Graphcore IPUs for both training and inference. Access reproducible code for a wide range of popular models covering NLP, Computer Vision, Speech, Multimodal, GNNs, AI for Simulation, Recommender Systems, and more. This includes a selection of models that achieve state of the art performance on IPUs, as well as code examples for self-learning.
Run models out-the-box on IPUs integrated with popular ML frameworks and libraries:
To see what's new and easily filter applications by domain and framework, please check out our Model Garden 🌷.
For more detailed benchmark information, please visit our Performance Results page.
The code presented here requires using Poplar SDK 3.0.x
Please install and enable the Poplar SDK following the instructions in the Getting Started guide for your IPU system.
If you require POD128/256 setup and configuration for our applications, please contact our engineering support.
- Computer Vision
- Natural Language Processing
- Speech
- Multimodal
- Graph Neural Network
- AI for Simulation
- Recommender Systems
- Reinforcement Learning
- Sparsity
- Probability
- Miscellaneous
- Archived
| Model | Domain | Type | Links |
|---|---|---|---|
| ResNet | Image Classification | Training & Inference | TensorFlow 1 , TensorFlow 2, PyTorch, PyTorch Lightning |
| ResNeXt | Image Classification | Training & Inference | TensorFlow 1 , PopART (Inference), PyTorch (Inference) |
| EfficientNet | Image Classification | Training & Inference | TensorFlow 1 , PyTorch, PyTorch Lightning |
| MobileNet | Image Classification | Inference | TensorFlow 1 |
| MobileNetv2 | Image Classification | Inference | TensorFlow 1 |
| MobileNetv3 | Image Classification | Training & Inference | PyTorch |
| ViT(Vision Transformer) | Image Classification | Training | PyTorch, Hugging Face Optimum |
| DINO | Image Classification | Training | PyTorch |
| Swin | Image Classification | Training | PyTorch |
| MAE (Masked AutoEncoder) | Image Classification | Training | PyTorch |
| Yolov3 | Object Detection | Training & Inference | TensorFlow 1 |
| Yolov4-P5 | Object Detection | Inference | PyTorch |
| Faster RCNN | Object Detection | Training & Inference | PopART |
| EfficientDet | Object Detection | Inference | TensorFlow 2 |
| SSD | Object Detection | Inference | TensorFlow 1 |
| UNet (Medical) | Image segmentation | Training & Inference | TensorFlow 2 |
| UNet (Industrial) | Image segmentation | Training | TensorFlow 1 |
| Neural Image Fields | Neural Radiance Fields | Training | TensorFlow 2 |
| Model | Domain | Type | Links |
|---|---|---|---|
| BERT | NLP | Training & Inference | TensorFlow 1 , PyTorch , PopART, TensorFlow 2, PopXL, PaddlePaddle, Hugging Face Optimum |
| Group BERT | NLP | Training | TensorFlow 1 |
| Packed BERT | NLP | Training | PyTorch, PopART |
| GPT2 | NLP | Training | PyTorch , Hugging Face Optimum |
| GPTJ | NLP | Training | PopXL |
| GPT3-2.7B | NLP | Training | PopXL |
| RoBERTa | NLP | Training | Hugging Face Optimum |
| DeBERTa | NLP | Training | Hugging Face Optimum |
| HuBERT | NLP | Training | Hugging Face Optimum |
| BART | NLP | Training | Hugging Face Optimum |
| T5 | NLP | Training | Hugging Face Optimum |
| Model | Domain | Type | Links |
|---|---|---|---|
| DeepVoice3 | TTS (TextToSpeech) | Training & Inference | PopART |
| FastSpeech2 | TTS(TextToSpeech) | Training & Inference | TensorFlow 2 |
| Fastpitch | TTS (TextToSpeech) | Training | PyTorch |
| Conformer | STT(SpeechToText) | Training & Inference | PopART, TensorFlow 1, PyTorch |
| Transfomer Transducer | STT(SpeechToText) | Training & Inference | PopART |
| Wav2Vec2 | STT(SpeechToText) | Training | Hugging Face Optimum |
| Model | Domain | Type | Links |
|---|---|---|---|
| miniDALL-E | multimodal | Training | PyTorch |
| CLIP | multimodal | Training | PyTorch |
| LXMERT | multimodal | Training | Hugging Face Optimum |
| Frozen in time | multimodal | Training & Inference | PyTorch |
| ruDalle (Preview) | multimodal | Inference | PopXL |
| Model | Domain | Type | Links |
|---|---|---|---|
| TGN (Temporal Graph Network) | GNN | Training & Inference | TensorFlow 1 |
| MPNN (Message Passing Neural Networks) | GNN | Training & Inference | TensorFlow 2 |
| Spektral GNN library with QM9 | GNN | Training | TensorFlow 2 |
| Cluster GCN | GNN | Training & Inference | TensorFlow 2 |
| Model | Domain | Type | Links |
|---|---|---|---|
| DeepDriveMD | Biology (Protein folding) | Training | TensorFlow 2 |
| CosmoFlow example using 3D Convolutions | Cosmology | Training & Inference | TensorFlow 1 |
| et0 | Evapotransporation | Inference | TensorFlow 1 |
| Approximate Bayesian Computation (ABC) COVID-19 | Medical | Inference | TensorFlow 2 |
| Model | Domain | Type | Links |
|---|---|---|---|
| Deep AutoEncoders for Collaborative Filtering | Recommender Systems | Training & Inference | TensorFlow 1 |
| Click through rate: Deep Interest Network | Recommender Systems | Training & Inference | TensorFlow 1 |
| Model | Domain | Type | Links |
|---|---|---|---|
| RL Policy model | Reinforcement Learning | Training | TensorFlow 1 |
| Model | Domain | Type | Links |
|---|---|---|---|
| MNIST RigL | Dynamic Sparsity | Training | TensorFlow 1 |
| Autoregressive Language Modelling | Dynamic Sparsity | Training | TensorFlow 1 |
| Block-Sparse library | Sparsity | Training & Inference | PopART , TensorFlow 1 |
| Model | Domain | Type | Links |
|---|---|---|---|
| Contrastive Divergence VAE using MCMC methods | Generative Model | Training | TensorFlow 1 |
| mcmc | Statistics | Training & Inference | TensorFlow 1 |
| Model | Domain | Type | Links |
|---|---|---|---|
| Sales forecasting | MLP (Multi-Layer Perceptron) | Training | TensorFlow 1 |
| Monte Carlo Ray Tracing | Graphics | Inference | Poplar |
The following applications have been archived. More information can be provided on request.
| Model | Domain | Type | Framework |
|---|---|---|---|
| Minigo | Reinforcement Learning | Training | TensorFlow 1 |
- Documentation: Explore our software documentation, user guides, and technical notes
- Tutorials: Hands-on code tutorials, simple application and feature examples
- How-to Videos: Watch practical how-to videos and demos by Graphcore engineers
- Research Papers: Read publications from Graphcore's Research team and IPU innovators
To easily run the examples with tested and optimised configurations and to reproduce the performance shown on our performance results page, you can use the examples-utils benchmarking module, which comes with every example when you install its requirements. To use this simple, shared interface for almost any of the examples provided here, locate and look through the example's benchmarks.yml file and run:
python3 -m examples_utils benchmark --spec <path to benchmarks.yml file> --benchmark <name of benchmark>For more information on using the examples-utils benchmarking module, please refer to the README.
Visualise your code's inner workings with a user-friendly, graphical interface to optimise your machine learning models.
Download PopVision to analyse IPU performance and utilisation.
If you encounter a problem or want to suggest an improvement to our examples please raise a Github issue or contact us at support@graphcore.ai.
The utils/ folder contains utilities libraries and scripts that are used across the other code examples. This includes:
- utils/examples_tests - Common Python helper functions for the repository's unit tests
- utils/benchmarks - Common Python helper functions for running benchmarks on the IPU in different frameworks
Unless otherwise specified by a LICENSE file in a subdirectory, the LICENSE referenced at the top level applies to the files in this repository.
Sep 2022
- Added those models below to reference models
- Vision : MAE (PyTorch), G16 EfficientNet (PyTorch)
- NLP : GPTJ (PopXL), GPT3-2.7B (PopXL)
- Multimodal : Frozen in time (PyTorchs), ruDalle(Preview) (PopXL)
- Deprecating all TensorFlow 1 applications. Support will be removed in the next release.
Aug 2022
- Change the folder name of models
- NLP : from gpt to gpt2
- Speech : from wenet-conformer to conformer
July 2022
- Major reorganisation of all the apps so that they are arranged as: problem domain / model / framework.
- Problem domains: Vision, NLP, Speech, GNN, Sparsity, AI for Simultation, Recomender systems, Reinforcement learning, Probability, Multimodal, and Miscellaneous.
- Added those models below to reference models
- Vision : Swin (PyTorch) , ViT (Hugging Face Optimum)
- NLP : GPT2 Small/Medium/Large (PyTorch), BERT-Base/Large (PopXL), BERT-Base(PaddlePaddle), BERT-Base/Large(Hugging Face Optimum), GPT2 Small/Medium (Hugging Face Optimum), RoBERTa Base/Large(Hugging Face Optimum), DeBERTa(Hugging Face Optimum), HuBERT(Hugging Face Optimum), BART(Hugging Face Optimum), T5 small(Hugging Face Optimum)
- Speech : Fastpitch (PyTorch), WeNet-Conformer-Medium(PyTorch) ,Wav2Vec2(Hugging Face Optimum)
- Multimodal : CLIP (PyTorch), LXMERT(Hugging Face Optimum)
- AI for Simulation : et0(TensorFlow 1)
- Removed Conformer-small/large (PyTorch)
- Archived Minigo (TensorFlow 1)
May 2022
- Added those models below to reference models
- Vision : ViT-pretraining(PyTorch), DINO(PyTorch), EfficientDet-inference(TensorFlow 2), Neural Image Fields (TensorFlow 2)
- NLP : PackedBERT(PyTorch, PopART), BERT-Large(TensorFlow 2)
- Speech : FastSpeech2-inference(TensorFlow 2), Conformer-Large(PyTorch)
- GNN : Cluster GCN(TensorFlow 2)
- AI for Simulation : DeepDriveMD(TensorFlow 2)
December 2021
- Added those models below to reference models
- Vision : miniDALL-E(PyTorch), Faster RCNN(PopART), UNet(TensorFlow 2), ResNet50(TensorFlow 2)
- NLP : BERT(TensorFlow 2)
- Speech : FastSpeech2(TensorFlow 2), Transfomer Transducer(PopART), Conformer-Small(PyTorch)
- GNN : TGN(TensorFlow 1), MPNN(TensorFlow 2)
