CNN Innards Visualizer

This visualisation uses AI Generated code, finetuned for the best visualisation, not code quality

cnn.mov

Interactive C + Raylib visualizer that trains and renders a small CNN live, exposing filters, activations, logits, gradients, and the geometry of information flow through the network.

What This Visualisation Shows

• A real Conv -> ReLU -> GAP -> FC network training in real time
• Filter responses and receptive-field relationships back to the source image
• Training curves, logits, class probabilities, and gradient signals
• Multiple pages for network overview, internals, and node-level graph inspection

Visual Map

flowchart LR
    A["Input Image"]
    B["Conv Filters"]
    C["Feature Maps"]
    D["Pooling / Aggregation"]
    E["Classifier Logits"]
    F["Loss + Gradients"]

    A --> B
    B --> C
    C --> D
    D --> E
    E --> F

Loading

Controls

• q: quit
• 1 / 2 / 3: switch pages
• space: pause or resume training
• r: randomize and restart
• Filter and page-specific inspection controls are shown directly in the app HUD

Run

make run