Deprecating matplotlib visualize: what is in it, and why should it leave GPy?

[lawrennd]

Why this thread

GPy.plotting.matplot_dep.visualize is the 2012–2015 interactive latent-space explorer: click or drag in a 2-D latent scatter, reconstruct the observation, and (when Q>2) pick which two latent axes to plot from an ARD bar chart. We are deprecating it and moving that behaviour out of GPy.

This is not a rewrite of static GPy plotting (plot, plot_latent, plot_confidence). Those stay.

Design is in CIP-000B. The replacement lives in a separate package, vidden, which is under construction. The warning slice is on main (#1253): importing visualize, constructing a viewer, or calling data_play emits a DeprecationWarning that points here and at vidden. The file is still importable. This thread remains the place to comment on what is in the module, why it should leave GPy, and what should remain.

Please read the inventory and the two proposed reasons below, then comment — especially if you still use these viewers in teaching or research.

What is in visualize.py

The file is a small GUI toolkit glued to GPLVM predict, not model code.

Name	Role	Used from GPy examples
data_show	Base: modify / close	(base class)
matplotlib_show	Matplotlib axes window	(base class)
vpython_show	VPython scene	(base class; VPython import is already optional)
vector_show	Observation as index vs value	bgplvm_oil, ssgplvm_oil
lvm	2-D latent scatter, click-to-track, reconstruct	stick, brendan_faces, olivetti_faces, cmu_mocap, …
lvm_dimselect	Left/right click ARD bars to choose y/x latent axes	stick_bgplvm, oil dimselect
lvm_subplots	One scatter per pair of latent dims	(essentially unused)
image_show	Reshape a vector and imshow	brendan_faces, olivetti_faces
stick_show	3-D point cloud + connect matrix	stick, stick_play, stick_bgplvm
skeleton_show	Angle skeleton → xyz, then the same 3-D draw	cmu_mocap
mocap_data_show / _vpython	Shared mocap drawing	(base classes)
data_play	Frame loop with time.sleep	stick_play
visual_available	Whether import visual succeeded	bcgplvm_stick / bcgplvm_linear_stick

The behaviour we still want is unchanged: a pointer in latent space, model.predict(X*), a reconstruction pane. That is a pointer stream plus predict, not a matplotlib GUI-backend problem.

Why deprecate it (proposed rationale)

Two reasons. Both are open for comment.

1. Matplotlib callbacks do not sit easily on macOS (or Jupyter)

The old viewers bind button_press_event / motion_notify_event on a matplotlib canvas. That only works when a GUI event loop is integrated with the interpreter. On macOS that is the Cocoa backend (macosx in matplotlib, osx in IPython). That path has broken across non-framework Python builds, conda envs, Jupyter's default inline backend, plt.show(block=False) plus input(), and nested canvas.draw() (RecursionError in transforms).

VPython (import visual) for 3-D skeleton views is effectively dead in the form GPy imports.

ipympl (%matplotlib widget) can recover some of the 2-D callbacks in Jupyter. It is a stopgap: it still uses matplotlib event handlers and does not fix VPython or the desktop demos. We do not want another round of local patches around visualize.py.

2. Interactive visualisation should not ship inside the core library

GPy's job is inference: kernels, models, predict. Interactive viewers pull in GUI backends (matplotlib event loops, historically VPython, and in any rewrite Panel / HoloViews / Bokeh). Those dependencies do not belong on the default pip install gpy path, and they do not belong in the same source tree as the numerical core.

A separate package can evolve on its own release clock, duck-type any model with X and predict, and stay out of GPy's test matrix. That is vidden. GPy remains a client, not a host.

What stays in GPy

• Static GPy.plotting (matplotlib and plotly backends): plot, plot_latent, plot_confidence, scales, and so on.
• GPLVM / BayesianGPLVM / BCGPLVM model code.
• Demo data loaders (via pods: OSU stick, oil, faces, CMU mocap).
• The reconstruction contract: latent point X* of shape (1, Q) in, model.predict(X*)[0] out.

Replacement

vidden is a browser explorer. Pointer events call model.predict in Python. Current demos (from that repository):

pip install -e ".[explore,gpy]"   # in a vidden clone
panel serve examples/stick.py --show
panel serve examples/oil.py --show
panel serve examples/faces.py --show
panel serve examples/doodle.py --show
panel serve examples/tree.py --show

GPy's remaining CIP-000B work is: stop examples importing visualize once those demos cover them, then remove or shrink the file. The warning itself is already on main.

Options

1. Warn now, remove after vidden covers stick / oil / faces (current CIP-000B plan). Warn-now is done on main (#1253). visualize stays importable through one more cycle with a warning; examples still call lvm / image_show / stick_show until those demos move.
2. Keep visualize indefinitely as a deprecated-but-present matplotlib path for anyone whose desktop GUI still works.
3. Optional extra inside GPy (gpy[explore]) instead of a separate package. We already chose against this so GPy does not grow Panel/HoloViews; arguing the other way is still useful.
4. Leave geometry helpers in GPy (connect matrices, skeleton to_xyz) even after the viewers go, if models or other tools still need them.

Questions

• Does anyone still rely on the matplotlib callbacks in a working setup (Linux Qt, teaching lab, …)?
• Is “visualisation lives outside the inference library” the right split, or would you rather keep an extra inside GPy?
• Which helpers, if any, should remain in GPy.util after the viewers leave?
• Is a one-release warning window enough, or do you need a longer deprecation?

Comments welcome — including disagreement with either of the two reasons. If you use lvm / image_show / stick_show in notes or courses, a short “we still need X” is exactly the signal we want.

Asking for eyes

Tagging people who have historically maintained GPy plotting, demos, or teaching use (please ignore if this is no longer your beat — a short “still relevant / not me” is enough):

@mzwiessele @jameshensman @alansaul @MartinBubel @ekalosak @zhenwendai @nfusi @mikecroucher @bobturneruk @letterdaniel @AlexGrig @lionfish0 @ric70x7

[lawrennd]

Update 2026-10-28. The warning slice is on main via #1253.

Importing GPy.plotting.matplot_dep.visualize, constructing a viewer (lvm, image_show, stick_show, …), or calling data_play now emits a DeprecationWarning that points at vidden and this thread. Static plotting (plot, plot_latent, plot_confidence) is unchanged. The module is still importable. GPy examples still call the old classes until vidden covers stick / oil / faces.

That is option 1’s first step. Comments still welcome on the two reasons and on whether a one-release warning window is enough.