zchary1106/figurecraft/figurecraft-neural-networks

FigureCraft Neural Networks

Create an architecture plate, not a generic sequence of equal rectangles.

Required information layers

1. Macro path: input, stem, stages, head, prediction.
2. Stage facts: spatial size, channels, repeat count, transition, and stride.
3. Representative block inset: expand only the key block once.
4. Visual grammar or stage schedule: explain glyphs and exact configuration.

Workflow

1. Establish the factual topology from model code, configuration, or explicit

user input. Never infer hidden layer sizes.

1. Use diagram.cnn-architecture for professional CNN/residual plates.

Use explicit diagram.neural-network nodes/edges for U-Net, Transformer, or arbitrary branching. The CNN plate does not reconstruct an executable model.

1. Encode spatial size in bounded tensor face area and channels in bounded prism

depth, while always printing exact H×W×C.

1. Use explicit glyphs for add, concat, downsample, projection, residual,

repeated blocks, and training-only paths.

1. Reserve separate lanes before routing skip connections.
2. Locate sibling figurecraft/scripts/figure.py, validate, render, and lint.
3. Inspect the image at final paper size and deliver SVG, source, manifest, and

editable Draw.io when requested.

Publication rules

• Residual paths terminate at a visible add node.
• Repeated blocks use ×N; do not draw all copies.
• Downsampling shows stride or ↓2, not only a smaller tensor.
• Use one inset for block detail and a separate table/legend.
• Color encodes operation family, not arbitrary stage decoration.
• Distinguish conceptual scale from exact executable structure.
• Declare visible provenance and final output medium. Use core render-set

for a compact stage overview and separate block details when the complete plate fails final-size readability. Do not solve it by shrinking annotations.

Blocking failures

Invalid topology, inconsistent dimension order, overlapping block/shortcut cards, skip lines through nodes, missing merge semantics, or unreadable paper scale.