Getting started

We have created a package to provide a modular monitoring and debugging pipeline for training neural networks, implementing a modular training loop using Flux.jl. This package allows users to specify observable quantities (e.g. gradient norm, curvature). Users can see the live dashboard that we implemented with Makie.jl.

The JL_Cockpit module provides plots for the live visualisation for neural network training. It uses the GLMakie module and their Observable to provide the live functionality.

Example Workflow

To add the JL_Cockpit module to your Julia environment run

]
add https://github.com/Themightyfirefly/JL_Cockpit

Once the module has been added, activate the module

using JL_Cockpit

To run the visualised training, execute

training_loop()

During the visualisation, users can press the r key to reset the plots (in case they zoomed into one) and the s key to save a screenshot of the current figure.