Skip to content

visualizer.md

Latest commit

 

History

History
102 lines (75 loc) · 6.29 KB

visualizer.md

File metadata and controls

102 lines (75 loc) · 6.29 KB
Raw

Visualizer (fuzzy-lab)

fuzzy-lab is a Plotly Dash app that talks to a running RpcService. It renders the model topology, lets you drive crisp inputs with sliders, and hot-swaps the engine's model from any .fcl in your fcl/ directory.

It's an authoring + inspection tool — not part of the Toit package. Lives in python/fuzzy_lab/viz/.

Setup

cd python
uv sync --all-extras

uv sync installs Dash, Plotly, httpx, Lark, and pytest. The first run pulls a few hundred MB of packages and takes about a minute; subsequent runs reuse the cached venv.

Running

# Engine in one shell:
jag toit run examples/device.toit
# fuzzy_logic RpcService listening on :8090 (model=tipper)

# Visualizer in another shell:
cd python
uv run fuzzy-lab --connect http://127.0.0.1:8090
# Dash app on http://127.0.0.1:8050/

Open http://127.0.0.1:8050/ in a browser.

CLI flags

Flag Default Purpose
--connect http://127.0.0.1:8080 Base URL of the RpcService
--port 8050 Dash app port
--poll-ms 500 Interval for GET /state polls
--fcl-dir <repo>/fcl Directory globbed (non-recursively) for *.fcl in the file picker

If the engine is reachable but on a different port — e.g., :8090 like examples/device.toit uses — pass --connect http://127.0.0.1:8090. For a device on the LAN: --connect http://192.168.0.130:8090.

What you see

┌───────────────────────────────────────────────────────────────┐
│ tipper                       [ choose .fcl ▾ ]                │
├───────────────────────────────────────────────────────────────┤
│ Inputs                                                        │
│  ┌─ service ──────────────┐  ┌─ food ──────────────────────┐  │
│  │ poor good excellent    │  │ rancid delicious            │  │
│  │   [membership figure]  │  │   [membership figure]       │  │
│  │   ───●──── slider ──── │  │   ──────●──── slider ────── │  │
│  └────────────────────────┘  └─────────────────────────────┘  │
├───────────────────────────────────────────────────────────────┤
│ Rules                                                         │
│  R1: service is poor OR food is rancid → tip is cheap   [fired] │
│  R2: service is good                   → tip is average        │
│  ...                                                          │
├───────────────────────────────────────────────────────────────┤
│ Outputs                                                       │
│  ┌─ tip ─────────────────────────────────────────────────┐    │
│  │  [composition figure with COG marker]  crisp = 14.8   │    │
│  └───────────────────────────────────────────────────────┘    │
└───────────────────────────────────────────────────────────────┘
  • File picker (top right): lists every *.fcl directly under --fcl-dir. Picking one runs fcl2json in-process, POST /models the result, and re-renders. Errors render inline in red without losing the current model.
  • Input panels: one per input variable. The figure shows each term's membership function color-coded; the slider drives POST /input on every drag.
  • Rules row: lists the rules with synthesized human-readable names; "fired" indicators flip live as inputs change.
  • Output panels: stacked membership functions truncated to current pertinence; the centroid (crisp output value) is marked.

Term names in the input/output panels match the colors of their polygons.

Hot-swap loop

The picker re-parses the chosen .fcl through fcl2json every time you select it. So the iteration loop is:

  1. Edit fcl/your-model.fcl.
  2. In the viz, re-select the file from the dropdown.
  3. New topology is in place; sliders still work, polling resumes.

No need to restart anything. If your edit introduces an unsupported feature, the parser error renders inline and the previous model stays loaded.

Known limitations

  • View + slider only. You can change crisp inputs and hot-swap the whole model, but you cannot drag membership-function vertices to edit shapes in place. To change a shape, edit the .fcl and re-select.
  • HTTP polling, not WebSocket. The page polls GET /state every --poll-ms milliseconds (default 500). Plenty fast for human-driven interaction; not designed for high-frequency telemetry.
  • No persistence. Closing the browser loses the current crisp inputs; on next open the engine state is whatever it was when you reconnected.

Troubleshooting

  • ConnectionRefusedError on startup — the engine isn't running. Start examples/device.toit (or your equivalent) first.
  • Dropdown is empty — --fcl-dir points somewhere with no *.fcl files. Default is the repo's fcl/.
  • All sliders snap to the left — the model on the engine doesn't match the slider ranges. The viz uses min(term.a)/max(term.d) per input from GET /model; if your model has weird negative-bound terms (like container-crane's too_far), the slider may not cover the full meaningful range.

Tests

cd python && uv run pytest -k viz

Covers the figure builders (plots.py) and the HTTP client (rpc.py). The Dash UI itself is exercised manually.