modulearn 0.1.0

A visual, Unreal-Blueprints-style node-graph editor for building and launching ML training runs.

ModuLearn

A visual, Unreal-Blueprints-style node-graph editor for building and launching machine-learning training runs. Drop Dataset → Model → Train nodes on a canvas, wire hyperparameters, loss and feature transforms into them, hit ▶ Train, and watch the learning curve live on the graph. No forms, no drift between the UI and the code it runs.

ModuLearn is a working name — you assemble training runs from modular nodes and learn from the live curve. Rename the package/repo freely; nothing depends on the name.

Why

Most training UIs are flat forms that slowly rot out of sync with the code behind them. ModuLearn inverts that: your Python is the single source of truth. You declare what you can train once, in a Registry, and the entire editor — palette, typed wiring rules, live validation, launch — is generated from it. Wire something the backend can't accept and it simply won't connect.

Install

pip install -e .          # fastapi + uvicorn + pydantic
modulearn demo            # serve the bundled editor at http://localhost:8000

modulearn demo is a complete, dependency-free editor (its "trainer" is a toy loss curve) so you can click around immediately. Once you've built your own app, serve it the same way:

modulearn run myproject.py        # serves the `app` you built with create_app
modulearn run myproject.py:editor # ...or a differently-named app / factory
modulearn --help

Swap on_train for your real training loop and nothing else changes. (The equivalent source file lives at examples/quickstart.py if you'd rather run it directly with python.)

The whole integration is two things

from modulearn import Registry, Param, create_app

reg = Registry()

reg.add_dataset("iris", title="Iris", kind="tabular",
                features=[...], targets=["species"])
reg.add_model("mlp", title="MLP", requires_kind="tabular",
              params=[Param("hidden", "hidden layers", "int_list", [64, 32])])
reg.add_hyperparameter("lr", label="learning rate", default=1e-3, min=1e-6, max=1)
reg.add_loss([Param("loss", "kind", "enum", "mse", choices=["mse", "cross_entropy"])])

def on_train(compiled, reporter):
    # compiled.dataset, .model, .model_params, .hyperparameters, .loss_params, .transforms
    reporter.state(epochs=100)
    for e in range(100):
        reporter.metric(epoch=e, train=..., val=...)   # drives the live chart
        reporter.state(epoch=e, best_val=...)
    reporter.state(phase="done", test_score=...)

app = create_app(reg, on_train, title="My Project")

1. A Registry — declare datasets, models, hyperparameters, a loss, and optional feature transforms. Each hyperparameter you add automatically grows a matching, type-checked input on the Train sink.
2. on_train(compiled, reporter) — your training loop. compiled is a fully-validated CompiledGraph; reporter publishes progress the editor polls (see the state contract below).

How it fits together

 registry.py   ── you declare nodes ──►  catalog (GET /api/nodes)
      │                                        │
      ▼                                        ▼
 compiler.py   ◄── canvas JSON ────────  static/ (custom canvas engine)
      │  type-checks every wire, runs semantic checks
      ▼
 CompiledGraph ──►  your on_train(compiled, reporter)  ──►  runs/<id>/{state,metrics}.json
                                                                   │
                                                    editor polls ◄─┘  (live panel + chart)

• Typed ports. Links carry a family (dataset, model, loss, run) or a field-specific scalar subtype (scalar/lr), so a learning-rate value fits the lr input and never epochs — enforced both client-side and in the compiler.
• All-or-nothing compile. compile_graph collects every structural, type and semantic error and raises them together, so the UI flags all bad wires at once.
• Composable transforms. Chain Dataset → Transform → … → Model; the compiler walks the chain into compiled.transforms. Mark a transform live=False to show it in the palette but have the compiler refuse it until your backend is ready.
• Stateless live window. Training runs in a background thread and persists to runs/<id>/. Closing the browser never stops a run; reopening shows true state.

The state contract

Your on_train reports through reporter, and the editor reads these keys:

call	keys the UI understands
reporter.state(**kw)	phase (running/done/error), epoch, epochs, best_val, test_score, error
reporter.metric(**row)	epoch, train, val → the live learning curve

Anything else you write is stored and returned by the API, just not charted.

Editor niceties

• Searchable palette with collapsible category dropdowns; ⌘K / / to focus, Enter to drop the first match.
• Wires colored by type, with flowing pulses that speed up on the edges feeding a running Train node; ports glow while you drag a compatible connection.
• Canvas persists to localStorage; reload/fork any past run's blueprint.
• All motion respects prefers-reduced-motion.

Layout

modulearn/
  registry.py    the declarative surface your app fills in
  compiler.py    graph JSON -> validated CompiledGraph (pure, unit-testable)
  server.py      create_app(): FastAPI factory + background job runner
  cli.py         the `modulearn` command (demo / run)
  demo.py        the app `modulearn demo` serves
  static/        graph.html, graph.js (app), graph-engine.js (canvas engine)
examples/
  quickstart.py  a complete, dependency-free editor

License

MIT (see LICENSE). The canvas node-graph engine (static/graph-engine.js) is our own, built from scratch — no third-party graph library.