orange3-nnviewer

A minimal Orange Data Mining add-on with one widget: Neural Network Viewer — takes the Model output of Orange's Neural Network widget (scikit-learn MLP) and draws the network on a native Qt graphics scene:

• one column of nodes per layer (inputs, hidden layers, output),
• feature names on the input neurons, target/class names on the output,
• edge width proportional to |weight|, blue = positive, red = negative,
• the numeric weight printed beside each edge (for small networks),
• biases printed under the neurons they feed,
• hover: an edge thickens, fills its two end nodes with its colour and tints the associated labels; its tooltip shows "source → target, weight = …", node tooltips show name and bias.

The diagram always scales to fit the window — there is no manual zoom or panning.

Build & install

Build a wheel (a single self-contained plugin file):

cd orange3-nnviewer
uv build            # -> dist/orange3_nnviewer-<ver>-py3-none-any.whl

Install it into the same Python environment that runs Orange (--no-deps: Orange's own dependencies are already present, and re-resolving them can trigger source builds against the bundled framework Python):

uv pip install --python <python-that-runs-orange> --no-deps dist/orange3_nnviewer-*.whl

(For the macOS Orange.app bundle, the interpreter is /Applications/Orange.app/Contents/MacOS/python.)

During development, an editable install is more convenient (edits are live on Orange restart):

uv pip install --python <python-that-runs-orange> -e .

Then start (or restart) Orange Canvas — the widget appears in the toolbox under the "Neural Networks" category, named "NN Viewer":

python -m Orange.canvas

Publish to PyPI

Once published, the widget is installable from Orange's own UI (Options → Add-ons… → Add more… → orange3-nnviewer):

uv build
uv publish          # needs a PyPI token (UV_PUBLISH_TOKEN or --token)

Test outside Canvas

Trains an MLP on the bundled housing dataset and opens the widget alone:

python -m orangecontrib.nnviewer.widgets.ow_nn_viewer

Tests

Run with the Python that runs Orange (Qt renders offscreen, no display needed):

python -m pytest -q

Use in a workflow

File  →  Neural Network  →  Neural Network Viewer
                      Model ─┘

Knobs to tweak (in ow_nn_viewer.py)

• MAX_LABELED_EDGES — above this edge count, numbers are suppressed (tooltips still show every value).
• weight label format — f"{weight:.2f}" in _draw.
• colours — POS_COLOR / NEG_COLOR.
• edge width scale — 1.0 + 4.0 * abs(weight) / max_abs.
• node size — radii = [min(0.05, 0.3 * g) for g in y_gaps].
• label placement — t = 0.60/0.40 along the line in _draw.