Ecosystem

Frameworks extending Lux

• Boltz.jl – Prebuilt deep learning models for image classification tasks
• DeepEquilibriumNetworks.jl – Continuous and Discrete Deep Equilibrium Networks
• DiffEqFlux.jl – Neural Differential Equations, Continuous Normalizing Flows, etc.

Extended Julia Ecosystem

As you might have noticed we don't do much apart from Neural Networks. All other parts of the DL training/evaluation pipeline should be offloaded to:

Automatic Differentiation

• Zygote.jl – Currently the default and recommended AD library
• Tracker.jl – Well tested and robust AD library (might fail on edge cases)
• Enzyme.jl – (Very) Experimental Support
• ForwardDiff.jl – For forward mode AD support
• ReverseDiff.jl – Tape based reverse mode AD (might fail on edge cases and doesn't work on GPU)

Data Manipulation and Loading

• Augmentor.jl
• DataLoaders.jl
• Images.jl
• DataAugmentation.jl

Distributed DataParallel Training

• FluxMPI.jl

Neural Network Primitives

• NNlib.jl
• LuxLib.jl

Optimization

• Optimisers.jl
• ParameterSchedulers.jl
• Optimization.jl

Parameter Manipulation

• Functors.jl

Serialization

• Serialization.jl
• JLD2.jl

Testing Utilities

• FiniteDifferences.jl – Finite Differencing. Useful for testing gradient correctness
• JET.jl
• LuxTestUtils.jl

Training Visualization & Logging

• Wandb.jl
• TensorBoardLogger.jl