An interactive 2D visualizer for CNN and RNN architectures, written in Odin on top of raylib. Native, lightweight, real-time. Designed as a faster, more focused alternative to web-only tools like TensorFlow Playground.

• Layer types: Input, Conv2D, MaxPool, AveragePool, Flatten, Dense, SimpleRNN, LSTM, GRU, Output.
• Live animated forward pass: colored pulses propagate along connections; speed adjustable at runtime.
• Recurrent self-loops: SimpleRNN/LSTM/GRU layers render with a t-1 self-loop arc.
• Time-unrolled RNN view: press U to expand recurrent layers into N copies labelled t-1, t, t+1, t+2 and connected by explicit time-step arrows. Downstream layers shift right automatically; render-only, no data-model change.
• Multi-channel feature maps: layers with channels > 1 render with offset "stacked card" shadows scaled by channel count.
• Parameter count and FLOPs: live totals in the status bar, plus per-layer numbers in the property panel. Standard formulas (Conv: (k²·in_ch+1)·filters; LSTM: 4·(in+units+1)·units; MAC counted as 2 FLOPs). Bump a filters value and watch the totals jump.
• Full graph editor:
  • Right-click any layer to open a property panel with -/+ and </> controls.
  • N (or the + Add Layer toolbar button) opens a layer-type picker.
  • Drag from a layer's right-edge handle to another layer to create a connection.
  • Click a connection to select it; Delete removes it.
  • Move any layer with left-drag; pan with middle-drag; zoom around the cursor with the mouse wheel.
• Live shape propagation: editing any param (filters, kernel, units, etc.) re-derives output_shape and propagates input_shape through the rest of the graph via a topological forward pass.
• Undo / Redo with Ctrl+Z / Ctrl+Y (depth 64), plus toolbar buttons that show stack depth.
• Copy / Paste any layer with Ctrl+C / Ctrl+V. Pasted copies land near the camera target (consecutive pastes step diagonally so they don't fully overlap) and inherit the source's parameters; connections are not copied.
• JSON persistence: S saves to architecture.json, L loads. Pretty-printed and hand-editable.
• Screenshot export: P writes visualizer_NNN.png.
• Four built-in demos: CNN, SimpleRNN, LSTM, GRU. Tab cycles between them.

CNN forward pass with animated pulses, multi-channel stacked feature maps.	LSTM with recurrent self-loop arcs and output handles.
Property panel with live in/out shapes, param count and FLOPs.	Layer picker modal — pick any of 9 layer types.
Time-unrolled LSTM: each recurrent layer expanded into 4 copies (t-1…t+2) with explicit time-step arrows.

Requires the Odin compiler with the vendored raylib bindings (default in current Odin builds).

odin build . -out:nn-odinray
./nn-odinray

Tested on Odin dev-2026-05 with raylib 5.5. The code uses only core: and vendor:raylib, so no external dependencies beyond the toolchain.

The JSON format is intentionally flat and human-editable:

{
  "version": 1,
  "anim_speed": 0.6,
  "layers": [
    {
      "id": 0,
      "type": "input",
      "pos": [80.0, 308.0],
      "input_shape": [3, 32, 32],
      "output_shape": [3, 32, 32],
      "params": { "channels": 3, "height": 32, "width": 32 }
    },
    {
      "id": 1,
      "type": "conv2d",
      "pos": [260.0, 270.0],
      "input_shape": [3, 32, 32],
      "output_shape": [16, 32, 32],
      "params": { "filters": 16, "kernel": 3, "stride": 1, "padding": 1 }
    }
  ],
  "connections": [
    { "from": 0, "to": 1, "weight": 1.0 }
  ]
}

Layer type values: input, conv2d, maxpool, avgpool, flatten, dense, rnn, lstm, gru, output.

Param fields not relevant to a given layer type are written as zero/empty but ignored on load.

main.odin         window/loop, input handling, CLI flags, undo/redo wiring
model.odin        layer/connection types, params, shape propagation, param/FLOPs, mutations, hit testing
demos.odin        built-in CNN/RNN/LSTM/GRU architectures
visualizer.odin   layout, drawing (cards, connections, self-loops, unrolled view), pulse animation
io.odin           JSON persistence + undo/redo history snapshots
ui.odin           toolbar, status bar, help panel, property panel, layer picker

~2600 lines of Odin across six files. No external libraries beyond Odin's core: and vendor:raylib.

This project is visualization only. It deliberately does not:

• Train models or run backpropagation.
• Render activation values per-pixel (deferred — would require a sampling backend).
• Render in 3D (raylib's 2D pipeline is enough for this scope).
• Implement a full deep-learning framework.

MIT © phiat