FluxTraining.jl

Docs (master)

A Julia package for using and writing powerful, extensible training loops for deep learning models.

What does it do?

• Implements a training loop to take the boilerplate out of training deep learning models
• Lets you add features to training loops through reusable callbacks
• Comes with callbacks for many common use cases like hyperparameter scheduling, metrics tracking and logging, checkpointing, early stopping, and more...
• Is extensible by creating custom, reusable callbacks or even custom training loops

When should you use FluxTraining.jl?

• You don't want to implement your own metrics tracking and hyperparameter scheduling or insert common training feature here for the 10th time
• You want to use composable and reusable components that enhance your training loop
• You want a simple training loop with reasonable defaults that can grow to the needs of your project

How do you use it?

Install like any other Julia package using the package manager:

]add FluxTraining

After installation, import it, create a Learner from a Flux.jl model, data iterators, an optimizer, and a loss function. Finally train with fit!.

using FluxTraining

learner = Learner(model, lossfn)
fit!(learner, 10, (trainiter, validiter))

Next, you may want to read

• Getting started
• A full example training an image classifier on the MNIST dataset
• The documentation of FastAI.jl which features many end-to-end examples

Acknowledgements

The design of FluxTraining.jl's two-way callbacks is adapted from fastai's training loop.