asberk/ignite

High-level library to help with training neural networks in PyTorch

TL;DR

Ignite is a high-level library to help with training neural networks in PyTorch:

• ignite helps you write compact but full-featured training loops in a few lines of code
• you get a training loop with metrics, early-stopping, model checkpointing and other features without the boilerplate

Below we show a side-by-side comparison of using pure pytorch and using ignite to create a training loop to train and validate your model with occasional checkpointing:

As you can see, the code is more concise and readable with ignite. Furthermore, adding additional metrics, or things like early stopping is a breeze in ignite, but can start to rapidly increase the complexity of your code when "rolling your own" training loop.

Table of Contents

• Installation
• Why Ignite?
• Documentation
• Structure
• Examples
  • MNIST Example
  • Tutorials
  • Distributed CIFAR10 Example
  • Other Examples
  • Reproducible Training Examples
• Communication
• Contributing
• Projects using Ignite
• About the team

Installation

From pip:

pip install pytorch-ignite

From conda:

conda install ignite -c pytorch

From source:

pip install git+https://github.com/pytorch/ignite

Nightly releases

From pip:

pip install --pre pytorch-ignite

From conda (this suggests to install pytorch nightly release instead of stable version as dependency):

conda install ignite -c pytorch-nightly

Why Ignite?

Ignite's high level of abstraction assumes less about the type of network (or networks) that you are training, and we require the user to define the closure to be run in the training and validation loop. This level of abstraction allows for a great deal more of flexibility, such as co-training multiple models (i.e. GANs) and computing/tracking multiple losses and metrics in your training loop.

Power of Events & Handlers

The cool thing with handlers is that they offer unparalleled flexibility (compared to say, callbacks). Handlers can be any function: e.g. lambda, simple function, class method etc. The first argument can be optionally engine, but not necessary. Thus, we do not require to inherit from an interface and override its abstract methods which could unnecessarily bulk up your code and its complexity.

Execute any number of functions whenever you wish

Examples

trainer.add_event_handler(Events.STARTED, lambda _: print("Start training"))

# attach handler with args, kwargs
mydata = [1, 2, 3, 4]
logger = ...

def on_training_ended(data):
    print("Training is ended. mydata={}".format(data))
    # User can use variables from another scope  
    logger.info("Training is ended")


trainer.add_event_handler(Events.COMPLETED, on_training_ended, mydata)
# call any number of functions on a single event
trainer.add_event_handler(Events.COMPLETED, lambda engine: print("OK"))

@trainer.on(Events.ITERATION_COMPLETED)
def log_something(engine):
    print(engine.state.output)

Built-in events filtering

Examples

# run the validation every 5 epochs
@trainer.on(Events.EPOCH_COMPLETED(every=5))
def run_validation():
    # run validation

# change some training variable once on 20th epoch
@trainer.on(Events.EPOCH_STARTED(once=20))
def change_training_variable():
    # ...

# Trigger handler with customly defined frequency
@trainer.on(Events.ITERATION_COMPLETED(event_filter=first_x_iters))
def log_gradients():
    # ...

Stack events to share some actions

Examples

Events can be stacked together to enable multiple calls:

@trainer.on(Events.COMPLETED | Events.EPOCH_COMPLETED(every=10))
def run_validation():
    # ...

Custom events to go beyond standard events

Examples

Custom events related to backward and optimizer step calls:

class BackpropEvents(EventEnum):
    BACKWARD_STARTED = 'backward_started'
    BACKWARD_COMPLETED = 'backward_completed'
    OPTIM_STEP_COMPLETED = 'optim_step_completed'

def update(engine, batch):
    # ...
    loss = criterion(y_pred, y)
    engine.fire_event(BackpropEvents.BACKWARD_STARTED)
    loss.backward()
    engine.fire_event(BackpropEvents.BACKWARD_COMPLETED)
    optimizer.step()
    engine.fire_event(BackpropEvents.OPTIM_STEP_COMPLETED)
    # ...    

trainer = Engine(update)
trainer.register_events(*BackpropEvents)

@trainer.on(BackpropEvents.BACKWARD_STARTED)
def function_before_backprop(engine):
    # ...

• Complete snippet can be found here.
• Another use-case of custom events: trainer for Truncated Backprop Through Time.

Out-of-the-box metrics

• Metrics for various tasks: Precision, Recall, Accuracy, Confusion Matrix, IoU etc, ~20 regression metrics.

• Users can also compose their own metrics with ease from existing ones using arithmetic operations or torch methods:

precision = Precision(average=False)
recall = Recall(average=False)
F1_per_class = (precision * recall * 2 / (precision + recall))
F1_mean = F1_per_class.mean()  # torch mean method
F1_mean.attach(engine, "F1")

Documentation

• Stable API documentation and an overview of the library: https://pytorch.org/ignite/
• Development version API documentation: https://pytorch.org/ignite/master/
• FAQ and "Questions on Github".
• Project's Roadmap

Additional Materials

• 8 Creators and Core Contributors Talk About Their Model Training Libraries From PyTorch Ecosystem
• Ignite Posters from Pytorch Developer Conferences:
  • 2019
  • 2018

Structure

• ignite: Core of the library, contains an engine for training and evaluating, all of the classic machine learning metrics and a variety of handlers to ease the pain of training and validation of neural networks!
• ignite.contrib: The Contrib directory contains additional modules that can require extra dependencies. Modules vary from TBPTT engine, various optimisation parameter schedulers, logging handlers and a metrics module containing many regression metrics (ignite.contrib.metrics.regression)!

The code in ignite.contrib is not as fully maintained as the core part of the library.

Examples

We provide several examples ported from pytorch/examples using ignite to display how it helps to write compact and full-featured training loops in a few lines of code:

MNIST Example

Basic neural network training on MNIST dataset with/without ignite.contrib module:

• MNIST with ignite.contrib TQDM/Tensorboard/Visdom loggers
• MNIST with native TQDM/Tensorboard/Visdom logging

Tutorials

• Text Classification using Convolutional Neural Networks
• Variational Auto Encoders
• Convolutional Neural Networks for Classifying Fashion-MNIST Dataset
• Training Cycle-GAN on Horses to Zebras with Nvidia/Apex - logs on W&B
• Another training Cycle-GAN on Horses to Zebras with Native Torch CUDA AMP - logs on W&B
• Finetuning EfficientNet-B0 on CIFAR100
• Hyperparameters tuning with Ax
• Basic example of LR finder on MNIST
• Benchmark mixed precision training on Cifar100: torch.cuda.amp vs nvidia/apex
• MNIST training on a single TPU

Distributed CIFAR10 Example

Training a small variant of ResNet on CIFAR10 in various configurations: 1) single gpu, 2) single node multiple gpus, 3) multiple nodes and multilple gpus.

• CIFAR10

Other Examples

• DCGAN
• Reinforcement Learning
• Fast Neural Style

Reproducible Training Examples

Inspired by torchvision/references, we provide several reproducible baselines for vision tasks:

• ImageNet
• Pascal VOC2012

Features:

• Distributed training with mixed precision by nvidia/apex
• Experiments tracking with MLflow or Polyaxon

Communication

• GitHub issues: questions, bug reports, feature requests, etc.

• Discuss.PyTorch, category "Ignite".

• PyTorch Slack at #pytorch-ignite channel. Request access.

User feedback

We have created a form for "user feedback". We appreciate any type of feedback and this is how we would like to see our community:

• If you like the project and want to say thanks, this the right place.
• If you do not like something, please, share it with us and we can see how to improve it.

Thank you !

Contributing

Please see the contribution guidelines for more information.

As always, PRs are welcome :)

Projects using Ignite

Research papers

• BatchBALD: Efficient and Diverse Batch Acquisition for Deep Bayesian Active Learning
• A Model to Search for Synthesizable Molecules
• Localised Generative Flows
• Extracting T Cell Function and Differentiation Characteristics from the Biomedical Literature
• Variational Information Distillation for Knowledge Transfer
• XPersona: Evaluating Multilingual Personalized Chatbot
• CNN-CASS: CNN for Classification of Coronary Artery Stenosis Score in MPR Images
• Bridging Text and Video: A Universal Multimodal Transformer for Video-Audio Scene-Aware Dialog
• Adversarial Decomposition of Text Representation

Blog articles, tutorials, books

• State-of-the-Art Conversational AI with Transfer Learning
• Tutorial on Transfer Learning in NLP held at NAACL 2019
• Deep-Reinforcement-Learning-Hands-On-Second-Edition, published by Packt
• Once Upon a Repository: How to Write Readable, Maintainable Code with PyTorch

Toolkits

• Project MONAI - AI Toolkit for Healthcare Imaging
• DeepSeismic - Deep Learning for Seismic Imaging and Interpretation
• Nussl - a flexible, object oriented Python audio source separation library

Others

• Implementation of "Attention is All You Need" paper
• Implementation of DropBlock: A regularization method for convolutional networks in PyTorch
• Kaggle Kuzushiji Recognition: 2nd place solution
• Unsupervised Data Augmentation experiments in PyTorch
• Hyperparameters tuning with Optuna
• Logging with ChainerUI

See other projects at "Used by"

If your project implements a paper, represents other use-cases not covered in our official tutorials, Kaggle competition's code or just your code presents interesting results and uses Ignite. We would like to add your project in this list, so please send a PR with brief description of the project.

About the team

Project is currently maintained by a team of volunteers. See the "About us" page for a list of core contributors.