photonviz

GPU-accelerated (WebGL2) charts for Jupyter, JupyterLab and Google Colab.

· 📖 Docs · ▶ Live demo

The Python bridge to Photon. NumPy arrays and torch tensors cross to the browser as binary buffers, so a million points still pan and zoom at 60fps inside a notebook cell — no image round-trip, no JSON blow-up.

pip install photonviz

Nothing else to install: the widget ships one self-contained ESM bundle, so there is no CDN fetch, no jupyter labextension install, and no separate JS package.

_{Real capture from JupyterLab — 200k points, interactive: wheel to zoom, drag to pan, hover for a tooltip.}

Quick start

import numpy as np
import photonviz as pv

x = np.linspace(0, 10, 200_000)
pv.line(x, np.sin(x) + 0.3 * np.sin(x * 7), name="signal", plot={"theme": "dark", "legend": True})

Charts chain, and the last expression in a cell renders itself:

(pv.Plot(theme="dark", title="Two series", legend=True)
   .line(x, np.sin(x), name="sin", color="#60a5fa")
   .line(x, np.cos(x), name="cos", color="#f472b6", dash=[6, 4])
   .hline(0, color="#64748b"))

Every keyword maps 1:1 onto the TypeScript options, so the JS reference applies verbatim — color, width, step, colorBy, yAxis, renderType, and so on.

Google Colab

Run once per session, then use photonviz normally:

from google.colab import output
output.enable_custom_widget_manager()

What you can draw

pv.scatter(x, y, sizes=area, colors=hex_list)     # bubble chart
pv.histogram(samples, bins=40)
pv.heatmap(z, cols, rows, extent={"x": [0, 1], "y": [0, 1]}, colormap="magma")
pv.candlestick(t, o, h, l, c)                     # + heikin_ashi, bollinger, drawdown…
pv.regression(x, y, band=2)                       # OLS + confidence band (or method="loess")
pv.corr_matrix([a, b, c], names=["a", "b", "c"])  # diverging, locked to ±1
pv.psd(signal, sampleRate=1000)                   # Welch spectrum
pv.confusion_matrix(y_true, y_pred)               # + roc_curve, pr_curve, calibration, embedding…
pv.surface(z, cols, rows)                         # 3D — orbit with the mouse
pv.polar_line(theta, r)

Plot, Plot3D and Polar are the full objects; the module-level names are one-line shortcuts. Pass plot={...} to a shortcut to configure the plot itself:

pv.scatter(x, y, size=4, plot={"theme": "dark", "pick": "xy", "colorbar": False})

Model architecture

Hand a PyTorch, Keras, scikit-learn or ONNX model straight to model_graph — the export happens in Python, the layout in the browser.

import torch, torchvision, photonviz as pv

model = torchvision.models.resnet18()

# 2D — a Netron-style DAG, with residual connections routed around the trunk.
pv.model_graph(model, example_input=torch.randn(1, 3, 224, 224), direction="horizontal")

# 3D — one cuboid per layer, sized from its output tensor: feature maps shrink
# while channel depth grows.
pv.model_graph_3d(
    model,
    example_input=torch.randn(1, 3, 224, 224),
    labels="full",
    plot={"aspectMode": "data", "projection": "orthographic", "showAxes": False},
)

_{Each cuboid is one layer: the visible face is its feature map, the thickness its channel count. Drag to orbit.}

The exporters are usable on their own — pv.from_torch(model), pv.from_keras, pv.from_sklearn, pv.from_onnx — and a hand-written {"nodes": [...], "edges": [...]} graph works too.

Example notebooks

Runnable, in the repo — examples/notebooks/:

How it works

photonviz is an anywidget, which is why the same object renders in Jupyter Notebook 7, JupyterLab 4, VS Code and Colab with no per-frontend code.

Chart calls build a plain dict. On sync, every array-like is swapped for a {"$buffer": i} marker and its raw bytes are appended to a buffer list; ipywidgets ships those as binary. The browser rebuilds typed-array views over them and hands them to @photonviz/core, which uploads straight to GPU buffers. Colours, names, extents and other structural values stay JSON.

All plots on a page share one WebGL2 context, so a notebook with dozens of charts will not exhaust the browser's context limit.

Development

From a checkout of the monorepo:

pnpm install
pnpm build:python          # bundles the widget into src/photonviz/static/widget.js
cd python
pip install -e ".[dev]"
pytest

License

MIT