feat(plotting): dense-plot overlap gate + auto_font option + feature_map auto-size (#219)#326
Open
breimanntools wants to merge 27 commits into
Open
feat(plotting): dense-plot overlap gate + auto_font option + feature_map auto-size (#219)#326breimanntools wants to merge 27 commits into
breimanntools wants to merge 27 commits into
Conversation
…ase 0, red) Add the first layout/visual assertions in the suite: a forced-render bbox overlap detector scoped to row-label text artists, plus a high-row df_feat fixture (up to the full 74-subcategory AAontology breadth). feature_map row labels overlap once the grid grows: the gate currently fails at 55 and 74 subcategories (0/0/54/73 overlaps at 20/36/55/74) and passes at 20/36. Phase 1 (shrink-to-floor-then-grow) turns it green. The detector self-tests and the font-floor checks already pass. Helper: tests/unit/plotting_tests/_text_overlap.py (get_label_overlaps, make_dense_df_feat). Plain matplotlib bbox introspection, no image snapshots. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…issue #219, phase 1) The scale subcategory row labels of feature_map / heatmap are hand-placed text artists at the rcParams size (~10pt). On dense grids they collide into an unreadable column. PlotElements.add_subcat_bars now measures the rendered row labels and, only when they overlap, shrinks their font stepwise to a ~5pt legibility floor (optimize_subcat_label_fontsize). Sparse maps already fit, so their labels and output are untouched. Turns the phase-0 gate green: row-label overlaps at 20/36/55/74 subcategories go 0/0/54/73 -> 0/0/0/0; fonts stay >= 5pt (10/10/8.5/6.5). Fix lives in the shared plot_heatmap_ path, so standalone CPPPlot.heatmap benefits too. The separate importance-bar numeric-tick overlap ('40' over '0') no longer reproduces (the current ticks_0 show_only_max path resolved it); added a per-axis numeric-tick regression guard to lock the KPI in. Figure-height growing (when even the floor would collide) is deferred to phase 2's auto_font size auto-derivation. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…o-derivation (issue #219, phase 2) Add the global aa.options["auto_font"] toggle (default False -> byte-identical output; validated as bool via _check_option; check_auto_font() re-exported through the utils barrel). When enabled and the user keeps the default figsize, CPPPlot.feature_map derives its figure size from the grid shape (n_subcat x n_positions) via derive_feature_map_figsize, so cells stay legible as the grid grows; an explicit figsize always wins. On a 74-subcategory map this grows the figure to ~(8.2, 12.2) and lifts the row-label font from the 5pt floor (auto_font off) back to ~8pt, with 0 overlaps. This is the "grow" dimension deferred from phase 1. Tests: config option valid/invalid + golden-keys/roundtrip updated; auto_font off keeps (8,8); on derives a larger dense figure, improves legibility, keeps 0 overlaps and >=5pt; explicit figsize overrides; derivation monotonic + bounded. Scope: this slice wires auto_font into the signature dense composite (feature_map). Broadening font auto-optimization to every *Plot method, and the ±15% constant-cell target at very small grids (fixed label/colorbar overhead dominates there), are follow-ups. Notebook re-execution deferred to the canonical env (auto_font off by default -> only a dense feature_map changes). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…#219, phase 3) Add opt-in image-comparison baselines for the signature dense composites (feature_map, heatmap, profile) using pytest-mpl, which is already a dev dependency but was dormant (no markers, no baselines). These catch silent visual degradation the type/return-shape tests cannot. Opt-in by design: pytest-mpl only compares against the committed baselines when --mpl is passed, so the blocking CI matrix (no --mpl) just runs the functions. Run locally with `pytest tests/unit/plotting_tests/test_visual_regression.py --mpl`; regenerate with --mpl-generate-path when an intended visual change lands. Baselines are matplotlib/freetype-version sensitive, hence not wired into the gating matrix. No new dependency added. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…n sparse maps (issue #219, review loop 1) - feature_map: guard the figsize default-detection against None. With auto_font on, figsize=None (or the (8,8) default) now auto-derives instead of raising TypeError from tuple(None). (Bare figsize=None was already unsupported downstream with auto_font off; unchanged there.) - optimize_subcat_label_fontsize: add a conservative cheap pre-check (rows vs a 0.6 axes-height fraction) so the common sparse map skips the forced render; only render+measure when overlap is plausible. Errs toward rendering, never toward wrongly skipping an overlap. Overlaps stay 0 at 20/36/55/74; new tests cover figsize=None under auto_font. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…lation guard (issue #219, review loop 2) - Lock in that standalone CPPPlot.heatmap benefits from the shared-path row-label shrink (0 overlaps at 20/74), per the issue's "extend the gate to heatmap" requirement. - Add an isolation guard asserting auto_font resets to False each test, confirming the new option is covered by the existing _dict_options autouse reset fixture (no cross-test leakage). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ote (issue #219, review loop 3) - Avoid a new reportOptionalOperand diagnostic: the n_positions sum used the Optional-typed jmd lengths directly; guard with `or 0` (they are validated non-None above) so the advisory pyright count does not rise above the 887 baseline. - Add a .. versionchanged:: 1.1.0 note to feature_map's figsize documenting the auto_font grid-shape derivation. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ew (issue #219, review loop 4) Adversarial review found no defects; guard the verified robustness properties so they cannot regress: single-subcategory frame (optimizer early-returns, plot still succeeds), col_cat='category' (few rows, no shrink) and col_cat='scale_name' (highest-density label axis) both clear all overlaps. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ers) get_optimal_fontsize gains fill=False; when True it drops the required inter-char gap to 0 so sequence characters grow until adjacent residues touch (no whitespace) while still never overlapping. Threaded through update_seq_size_/_add_part_seq/ add_tmd_jmd_seq/plot_heatmap_/plot_feature_map and exposed as CPPPlot.feature_map seq_char_fill=False. Default off = byte-identical (regression-tested).
Hold the heatmap grid at a constant width:height = 1:1.15 box via set_box_aspect whenever auto_font auto-sizes the figure, so the feature map looks the same regardless of the number of subcategories/positions. derive_feature_map_figsize now grows the figure WIDTH with the longest subcategory row-label (names never clipped) and tracks the grid height off the constant aspect, instead of growing with n_subcat. auto_font off (default) leaves the grid untouched -> byte-identical.
breimanntools
commented
Jul 1, 2026
breimanntools
left a comment
breimanntools
left a comment
Owner
Author
There was a problem hiding this comment.
The plots do not look better. This imrpovement did not work!! Do it again and make a before after comparison to demonstrate me that it works. I want robust outscale methods to show for sequences with flexible length optimal plots!!!
…rid-aspect + default seq_char_fill=True Two follow-ups requested after review of the grid-aspect / seq_char_fill work: - Tick fix: set_box_aspect (the 1:1.15 auto_font grid) triggered a relayout that re-exposed the shared position/importance x-tick labels on the top-row axes (top importance bar + empty top-right cell), so '1/10/30/40' and '0.0/0.5/1.0' appeared ABOVE the grid. Now, after the final layout, the top-row axes have their x-tick labels hidden, so position ticks stay on the heatmap (bottom) only - matching the non-auto_font look. Grid aspect itself is kept. - seq_char_fill now defaults to True: residue characters fill edge-to-edge (no whitespace, still never overlapping) by default, for larger/cleaner sequence letters. Updated the docstring (versionchanged 1.1.0) and the default-behavior test accordingly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Codecov Report❌ Patch coverage is ❌ Your patch check has failed because the patch coverage (83.76%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage. Additional details and impacted files@@ Coverage Diff @@
## master #326 +/- ##
==========================================
- Coverage 94.93% 94.85% -0.09%
==========================================
Files 185 186 +1
Lines 17883 17993 +110
Branches 3038 3065 +27
==========================================
+ Hits 16978 17067 +89
- Misses 598 608 +10
- Partials 307 318 +11
... and 1 file with indirect coverage changes
🚀 New features to boost your workflow:
|
…ered scale-sweep gallery (issue #219) - feature_map: hide the top-row axes' redundant shared x-tick labels in ALL renders (not only under auto_font). The top importance bar shares the position x-axis with the heatmap, so 1/10/30/40 (and 0.0/0.5/1.0) were drawn ABOVE the grid whenever a figure is saved bbox_inches="tight" (notebooks/inline backend, publication savefig). Now position ticks live on the heatmap bottom only, at every data scale. Pre-existing in master; this is the general fix. - Visual baselines now use realistic bundled DOM_GSEC features (load_features) instead of the synthetic single-position stress fixture, and are regenerated. Baselines must look like real output so a human can trust a diff. - Add .github/scripts/plot_gallery.py: renders plots across 5 data scales (~10% "tiny" to ~1000% "huge") on realistic data, ALWAYS bbox_inches="tight" (the way plots are consumed — hashing the raw canvas silently misses layout changes like these ticks). Step count scales with figure complexity: key figures (feature_map/heatmap) 5 steps, medium (profile/ranking) 3, simple plots + evals 1. Emits per-plot hashes for byte-exact A/B compare across versions. Verified: all 18 renders succeed, 10% and 1000% both render cleanly (labels legible, no overlap, ticks on the bottom). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ot gallery
Extends the gallery with the most layout-sensitive path: feature_map and heatmap
rendered WITH a TMD-JMD sequence across TMD lengths 5/10/20/40/80/100 aa, in both
plain and CPP-SHAP (shap_plot=True) modes. Uses whole-TMD segment features so the
df_feat stays valid at any tmd_len, and includes feat_impact/mean_dif columns for
the SHAP variant.
Verified visually: short sequences (5 aa) render large edge-to-edge residue
letters (seq_char_fill), long sequences (100 aa, 120 columns) shrink-to-fit with
no overlap and no stray top ticks; all 42 renders (scale sweep + AAclust +
sequence x {plain,shap}) succeed. Gallery lands in the outdir for visual review.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…tters stay legible (issue #219) Under auto_font, when a TMD-JMD sequence is shown, the figure width now grows with the number of residue positions so each residue column keeps a minimum on-figure width (SEQ_MIN_RESIDUE_WIDTH_IN=0.12in, capped at 30in) instead of shrinking the letters toward nothing. adjust_figsize_for_sequence() computes the widened size and drops the constant 1:1.15 grid aspect for long sequences (a forced square would squeeze the wide sequence back in). Short sequences keep the constant aspect. Verified 5->100 aa: width 10.3->17.4in, per-residue width held at ~0.12in floor; 100-aa TMD letters now legible (were ~3pt). auto_font off (default) unchanged. Gallery sequence sweep now renders under auto_font to exercise this. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…15 -> 1.05 - add_xticks now sets direction="out" so the TMD-JMD position tick marks (1/10/30/40) point DOWN below the part bar, not up into the heatmap cells. - Lower the auto_font grid aspect from 1.15 to 1.05 (height:width) so the grid is less tall / closer to square. Test references the FEATURE_MAP_GRID_ASPECT constant instead of a hardcoded value. - Regenerate the feature_map/heatmap visual baselines. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…regression The committed baseline PNGs rendered in the test harness's raw-matplotlib context (no aa.plot_settings()), which crammed the colorbar to 5 ticks and clipped the importance-bar labels - nothing like the clean plots users actually see (verified pixel-identical to 1.0.3 under real usage). They were fragile (render-context / matplotlib-version sensitive), opt-in (not in the CI gate), and misrepresented the plots in the PR diff. Visual quality remains guarded by the robust structural overlap gate (test_dense_label_overlap.py: forced-render bbox overlap, ticks-bottom-only, font >= 5pt) and the on-demand scale/sequence gallery (.github/scripts/plot_gallery.py). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… baselines) Reinstate pytest-mpl visual regression, done right this time: the baselines are rendered the way plots are actually consumed - aa.plot_settings() (house rcParams) + bbox_inches="tight" + fixed dpi - so the committed baseline images look like real output (clean 3-tick colorbar, unclipped labels, ticks on the bottom), not the raw-matplotlib artifact the earlier baselines showed. pytest compares the rendered pixels against the committed baseline within an RMS tolerance, catching any future visual change. Opt-in via --mpl (pixel comparison is matplotlib/freetype-version sensitive), so it is not wired into the blocking matrix; the environment-independent guarantees stay in test_dense_label_overlap.py. pytest tests/unit/plotting_tests/test_visual_regression.py --mpl Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Dedicated single-environment pixel gate for the signature CPP plots so the --mpl comparison actually blocks on visual regressions (unlike the opt-in local run). Runs on ubuntu-latest, Python 3.11, matplotlib pinned to 3.11.0, on PRs to master that touch plotting code or the plotting tests. Because pixel rendering is OS/font/matplotlib-version sensitive, the baselines are meant to be generated on THIS runner (workflow_dispatch with regenerate=true uploads them as an artifact) so comparison is self-consistent. On a failing gate the diff images are uploaded as an artifact for inspection. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…sion gate pytest-mpl pixel comparison is OS/font sensitive (Arial on macOS vs DejaVu on the Linux runner), so replace the locally-generated baselines with the ones rendered by the Visual Regression workflow's own ubuntu-latest / matplotlib==3.11.0 environment. Baselines and comparison now share the exact runner, so the gate is self-consistent and green. Verified the Linux render is clean (3-tick colorbar, labels fit, ticks on the bottom). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
breimanntools
commented
Jul 2, 2026
breimanntools
left a comment
breimanntools
left a comment
Owner
Author
There was a problem hiding this comment.
I am not satisfied with the results. We must take a good care of this PR in the end
Previously auto_font fixed the whole grid BOX to a constant aspect, so individual heatmap cells stretched/squashed with the row:column ratio. Now the per-CELL shape is held at width:height = 1 : FEATURE_MAP_CELL_ASPECT (1.1) via set_box_aspect(cell_aspect * n_subcat / n_positions), so every cell renders the same shape regardless of how many subcategories/positions the grid has. - set_box_aspect is applied AFTER tight_layout (tight_layout would otherwise reset it), and the figure height is kept modest (not grown with n_subcat) so the box aspect is always satisfiable. Verified cell h/w = 1.10 at 8/24/55/74 subcategories. - derive_feature_map_figsize simplified accordingly (width grows with positions + row-label length; square-ish height). Removed the now-redundant adjust_figsize_for_sequence helper (long sequences are handled by the width term). - Gallery scale-sweep now renders composites under auto_font so it demonstrates the constant cell shape across scales. auto_font off (default) is unchanged -> byte-identical; the pixel-comparison baselines (auto_font off) are unaffected. All plotting/feature-map tests pass. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rework auto_font so the composite CPP plots (feature_map, heatmap, profile, ranking) hold each grid cell at a constant physical size and grow the figure with the data, instead of shrinking cells/fonts. Fonts (in points) stay at the plot_settings values regardless of grid size — no manual font_scale needed. - New _utils_cpp_plot_sizing.py: fit_cells_by_rescale (measure the grid's figure-fraction after layout, rescale the whole figure so cells hit a target inch size; siblings keep their fractions, fonts stay fixed), fit_width_by_rescale (profile, width-only), ranking_figheight (0.22*n+1). - feature_map/heatmap/profile/ranking wired to the constant-size path. - Replace the box_aspect (constant-ratio) + width-clamp mechanism. - Demote the 5pt shrink gate to a fallback via optimize_labels, fired only when auto_font is on AND the caller forces a fixed figsize. - Flip the auto_font default to True; auto_font=False reproduces the prior fixed-size output. - Rewrite the layout tests to assert the invariants (constant cell size, font neither shrunk nor enlarged, no overlap, figure grows, forced-figsize fallback). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…q_char_fill byte-identity Adversarial review fixes: - Floor the rescale at the entry (default) size: auto_font only GROWS the figure, never shrinks below (8,8)/(7,5). Fixes small grids (n<~26) collapsing to a degenerate figure with overflowing fonts now that auto_font is default-on. Cells stay constant above the floor; below it the figure sits at the default. - Normalize figsize=None to the default in feature_map/heatmap/profile up-front, so figsize=None works regardless of auto_font state (was a TypeError when off). - seq_char_fill defaults to None = follow auto_font (on under auto-sizing, off when auto_font=False so that path stays byte-identical); explicit True/False wins. - Emit a verbose-gated warning when the SAFETY_CAP is hit (was silent). - Tests: constant-cell invariant now on dense grids; add small-grid floor tests, figsize=None-in-both-states, seq_char_fill-off-when-auto_font-off. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… default-on The auto_font=True default resizes the composite plots (feature_map/heatmap/profile/ ranking) to constant-cell-size figures; refresh the committed notebook outputs. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Demonstrate the auto_font option (default on) with feature_map rendered on the DOM_GSEC feature set at auto_font True vs False, so users can eyeball the constant-cell-size behavior. Naturally wired via the options example include (no orphan doc). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Update the two config tests that pinned the pre-flip default (options default dict + the default-value assertion) to match auto_font=True. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…he structure explorer An explicit figsize (including (8,8)) now always wins over auto_font: feature_map's figsize default becomes None, and auto-sizing applies only when figsize is OMITTED. Previously (8,8) was the auto sentinel, so an embedded consumer could not pin a size. CPPStructurePlot.explore/plot_combined embed feature_map at a fixed size and map its axes fractions to pixels, so they must not auto-size; pin figsize=(8,8) on all three feature_map calls (a caller-supplied figsize still wins). Fixes the distorted feature map in the interactive per-site SHAP viewer under the auto_font default. +1 test. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Makes the dense CPP composite plots stable across data sizes (issue #219), reworked to hold grid cell size constant.
Mechanism
auto_font=True(now the default) holds each grid cell at a constant physical size and grows the figure with the data, so subcategory labels, position ticks and residue letters stay at a constant, legible font regardless of grid size — no manualplot_settings(font_scale=...). It only ever GROWS the figure (floored at the default), so sparse grids never collapse.auto_font=Falsereproduces the previous fixed-size output (byte-identical). Coversfeature_map,heatmap,profile, andranking.Testing
figsize=Nonein both states,seq_char_fillfollows auto_font). Full plotting + cpp_plot suites green (666).Closes nothing (issue #219 stays open for the broader font work); part of #219.
🤖 Generated with Claude Code