draw-neural-networks-typst

name	draw-neural-networks-typst
description	Empower the agent to generate high-quality Neural Network architecture diagrams in Typst using the neural-netz package.

Skill: Drawing Neural Networks with Typst (neural-netz)

🎯 Objective

Empower the agent to generate high-quality Neural Network architecture diagrams in Typst using the neural-netz package.

📦 Setup and Import

Whenever the user requests drawing or visualizing a neural network in Typst, always start by importing the main function from the neural-netz package:

#import "@preview/neural-netz:0.3.0": draw-network

🛠️ Basic Usage (draw-network)

The core function is draw-network. It accepts the following parameters:

#draw-network(
  layers,                 // (Array) Required. List of dictionaries defining the layers
  connections: (),        // (Array) Optional. List of non-linear connections (skip connections)
  palette: "warm",        // (String) Optional. Color palette: "warm" (default) or "cold"
  show-legend: false,     // (Boolean) Optional. If true, displays the chart legend
  legend-title: "Layers", // (String) Optional. Title for the legend
  scale: 100%,            // (Relative) Optional. Overall scale of the drawing (e.g., 50% to fit the page)
  stroke-thickness: 1,    // (Float) Optional. Line thickness
  depth-multiplier: 0.3,  // (Float) Optional. Multiplier for 3D visual depth
  show-relu: false        // (Boolean) Optional. If true, draws a darker color on convolutions to indicate ReLU
)

🧱 Defining Layers (layers)

The main parameter of draw-network is an array of dictionaries. The order in the array defines the sequential order (from left to right), and they are connected automatically by default.

Predefined Layer Types (type)

• "input": Input layer (can take the key image: "path.jpg" or image: none).
• "conv": Convolution.
• "pool": Pooling. By default, it sticks to the previous layer.
• "convres": Residual Convolution.
• "fc": Fully Connected (Usually 2D, set depth: 0).
• "custom": Fully customizable layer (using fill, bandfill, opacity).

Main Layer Attributes

• type: The type of the layer (described above).
• offset: Space/distance from the previous layer (e.g., offset: 2).
• widths: Array of widths for layer subdivisions (e.g., (1, 1) or (1, 2)).
• height and depth: Height and depth (e.g., height: 6, depth: 8). For pure 2D, set depth: 0.
• channels: Array with labels for the channels that appear on the axes (e.g., (32, 64, 128) or ("10",)).
• label: Name/Text displayed below the block.
• connection-label: Name displayed on the arrow connecting to the NEXT layer.
• name: Unique identifier used to create manual connections (e.g., name: "conv1").
• show-connection: (Boolean) Whether the automatic arrow/connection after this layer should be drawn.

🔗 Manual Connections (connections)

Used for skip-connections (like in ResNets or U-Nets). Each connection is a dictionary in the connections array:

• from: Name of the origin layer (defined in the name attribute of the layer).
• to: Name of the destination layer.
• type: Usually "skip".
• mode: Connection style. Options: "depth", "flat", or "air".
• label: Text label over the connection.
• pos: Vertical/horizontal offset of the connection arc (e.g., pos: 5).
• touch-layer: (Boolean) If true and mode: "air", the arrow touches the layer instead of ending with a loose arrowhead.

📝 Complete Practical Example

Use this example as a baseline to understand the assembly:

#import "@preview/neural-netz:0.3.0": draw-network

#draw-network((
    // Input
    (type: "input", label: "Image", name: "in", show-connection: true),
    
    // Convolutional Block
    (type: "conv", widths: (1.5, 1.5), height: 4, depth: 4, label: "Conv1", name: "c1", offset: 2),
    (type: "pool", channels: ("64",), height: 2, depth: 2, connection-label: "Max Pool"),
    
    // Block that does not propagate arrow automatically
    (type: "convres", channels: (128, 256), height: 5, depth: 5, label: "ResBlock", name: "c2", offset: 3, show-connection: false),
    
    // Linear / Fully Connected 2D Layer (depth: 0)
    (type: "fc", channels: ("10",), height: 6, depth: 0, label: "Linear", name: "out", offset: 2)
  ),
  
  // Extra connections (skip connections)
  connections: (
    (from: "in", to: "c2", type: "skip", mode: "air", label: "Skip Connection", pos: 6, touch-layer: true),
  ),
  
  // Global customizations
  palette: "warm",
  show-legend: true,
  legend-title: "Architecture",
  show-relu: true
)

🧠 Behavioral Instructions for the Agent

1. When receiving instructions about architecture design (e.g., "Create a VGG16" or "Make a CNN with a 28x28 input and a 10 output"), map the requested layers using a coherent visual structure (increasing/decreasing height/depth proportions according to Spatial Resolution or Channels).
2. If the resulting image is too wide, suggest or apply the use of the scale parameter (e.g., scale: 75%).
3. Support custom colors only if explicitly requested (via type: "custom" with fill: rgb("#...")), otherwise, encourage the use of native palettes (warm or cold).
4. You MUST use depth: 0 for purely Dense/Fully Connected layers to highlight the difference between 3D tensors and 1D vectors.