Visualization

This page describes the visualization components of the package.

Dashboard

The dashboard provides a live, interactive view of training dynamics and updates continuously while training is running.

Currently implemented plots:

• Training loss vs Iteration
• Parameter distance (L2 distance from initial weights)
• Parameter update size (step-to-step distance)
• Gradient norm
• Gradient norm test (measuring signal-to-noise)
• Gradient distribution (1D histogram)

All plots that use Iteration on the X-axis are automatically linked, ensuring a synchronized view of training progress across metrics. Axis limits are automatically adjusted during training to keep all data visible.

Makie integration

The package uses Makie.jl and observable variables to enable live updates.

Dashboard execution (internal)

Visualization is handled by a dedicated render loop that updates plots continuously while training is running. This loop is decoupled from the optimization process to avoid blocking training.

Training and visualization communicate via a Channel. At each iteration, the training loop sends:

• the current step index
• a dictionary containing the values of the tracked quantities

Scalar quantities are appended to time series plots. Vector-valued quantities (such as gradient histograms) replace the current plot data.

The dashboard is rendered using WGLMakie and served in a browser via Bonito.