owart 0.1.0

Poster-style visualizations of neural network weights.

owart

Generates posters based on neural network weights.

Pick an MLX-compatible model and this script will iterate over those weights and do the following:

• Find the global minimum and maximum weight values.
• Normalize weights to [0, 1].
• Trim the upper and lower percentiles to reduce outlier impact.
• Map normalized weights to colors based on intensity.
• Pack weight matrices into a visually appealing layout.
• Render the packed layout into a high-resolution image with a legend and metadata.

Requirements

• Python 3.13+
• MLX on Apple Silicon (for mlx / mlx-lm)

Dependencies are declared in pyproject.toml.

Usage

Basic:

uvx owart mlx-community/Llama-3.2-1B-Instruct-4bit

Options

• --packing {polyomino,maxrects,guillotine,shelf}: packing algorithm (default: polyomino).
• --polyomino-workers N: parallel search workers for polyomino packing (default: CPU count).
• --palette {tab20,tab20b,tab20c}: color palette (20 colors).
• --color-method {blackwhite,opacity}: how intensity maps to color.
• --shrink N / --scale-method {mean,median,max,sample}: image reduction (shrink auto-computed if omitted).
• --intensity-method {linear,gamma,smoothstep,sigmoid}: intensity curve (default: smoothstep).
• --intensity-clip-percent P: clamp top/bottom P percentile before normalization (default: 0.05).
• --frame-padding N: padding around the visualization.
• --frame-section-gap N: gap between footer sections.
• --frame-bottom-padding N: space below the legend.
• --frame-font-size N: base font size for footer/legend.
• --frame-line-spacing N: line spacing (default: 0.35 * font size).
• --frame-bg #RRGGBB or R,G,B: frame background color.
• --inner-border-width N: border width around the central visual.
• --inner-border-color #RRGGBB or R,G,B: border color.
• --no-frame: output just the visualization.

Example

uvx owart mlx-community/Llama-3.2-1B-Instruct-4bit \
  --packing polyomino --polyomino-workers 8 \
  --palette tab20 --color-method blackwhite \
  --intensity-method smoothstep --intensity-clip-percent 0.05 \
  --frame-padding 80 --frame-font-size 60 --frame-section-gap 80 \
  --frame-bg "#000000" \
  --inner-border-width 5 --inner-border-color "#FFFFFF"