plot

Plot

Drop-in matplotlib templates for publication-quality figures. Each template is one .py file producing one subplot, with no save logic — copy, swap data, add your savefig.

The shared style.py provides:

• Font. Palatino body + STIX math, matching the LaTeX mathpazo package used in arxiv-style templates. DejaVu Sans tail-fallback handles unicode glyphs (❄, ⚡) that Palatino lacks.
• Palette. Google brand colors (Blue/Red/Yellow/Green/Grey + extended Purple) softened to a paper-friendly tier by default.
• Tier system. Same color, five softness levels: brand → medium → paper (default) → soft → mute. One knob switches the global feel.
• Helpers. family_4(base) for ordered categorical gradients; paper(base) / lighten / darken for one-off color tweaks; arrow(label, 'down'|'up') to append ↓ / ↑ to titles.

Templates

File	Type	Use when
00_bar_vertical.py	Vertical bar with reference baseline	Comparing a metric across discrete methods, optionally vs. a baseline value
01_bar_horizontal.py	Horizontal bar with value labels + dashed group separators	Component ablation rows where each row adds/removes a piece, value-labeled
10_box_horizontal.py	Horizontal boxplot with 4-step family gradient	One categorical factor with ordered levels (e.g. progressively more compute)
20_line_multi.py	Multi-line plot with markers (linear xy)	Multiple model variants tracked across a hyperparameter sweep
21_line_broken_y.py	Multi-line with broken y-axis	Two groups of curves on disjoint y-ranges, both must stay visible
22_line_logx.py	Single line on log-x scale	Saturation as x sweeps orders of magnitude (data fraction, token count)
23_line_loglog_compare.py	Log-log multi-line with reference dashed line	Comparing scaling exponents across settings against a known reference
30_scatter_powerlaw.py	Log-log scatter + linear fit line	Clean power-law y = a · C^b; closed-form line + sample points
31_scatter_isoflops.py	Multi-curve parabola scatter with fits	IsoFLOPs-style — each compute budget yields a U-shape, marker size scales with parameter count

Quick start

1. Copy style.py and the chosen template into your figures directory.
2. Edit the data block at the top of the template (data is illustrative — replace with yours).
3. Replace placeholder labels (Model A, Metric A, Task A, Modality A, Component 1, Setup A, Baseline, Reference) with the real names.
4. Add a save call before the script ends:

   fig.savefig('out.pdf')
   fig.savefig('out.png', dpi=200)
   

5. Run python <template>.py.

Style knobs

from style import (
    apply_style,                           # one-shot rcParams setup
    G_BLUE, G_RED, G_YELLOW, G_GREEN,
    G_GREY, G_PURPLE,                      # Google brand constants
    apply_tier, paper,                     # softness control
    family_4,                              # 4-step gradient (idx 0=lightest, 3=darkest)
    lighten, darken,                       # one-off color adjustments
    arrow,                                 # `Metric A ↓` title helper
)

apply_style()                              # call once at the top of every figure script

Switch the global softness tier: edit DEFAULT_TIER = 'paper' in style.py to one of brand / medium / paper / soft / mute. All paper(base) and family_4(base) calls follow.

Per-color tier: apply_tier(G_BLUE, 'soft') overrides the default for one usage.

Regenerate the per-tier hex table in style.py's docstring: python style.py.

Naming convention (pre-genericized)

Every label in the templates is a placeholder, picked so the figure parses without context:

Placeholder	Replace with
Model A, Model B, ...	Method / model names
Metric A, Metric B, ...	Metric names (PPL, accuracy, FID, ...)
Task A, Task B, ...	Task / dataset categories
Method A, Method B, ...	Approach categories
Modality A, Modality B	Domain / modality (vision, language, ...)
Component 1..4	Ablation increments
Setup A, Setup A (2×)	Configuration / scale variants
Baseline	The reference value (e.g. text-only PPL)
Reference	A canonical comparison curve (e.g. Chinchilla scaling)

Replace before saving — never ship a figure with placeholder names in a paper.

Conventions

• One subplot per file. If a figure needs two panels of the same type, call the template twice on different axes; if it needs two different types, copy two templates and lay them out with gridspec.
• Templates do not call plt.show() or fig.savefig(...) — add yours.
• All templates assume style.py is importable from the same directory.
• Output sizes target a single-column or half-page paper figure (figsize ranges roughly 4×3 to 5.6×3.4 inches). Adjust as needed.

Dependencies

• matplotlib >= 3.6 (per-glyph font fallback)
• numpy >= 1.20
• macOS: Palatino ships with the system; on Linux install tex-gyre (TeX Gyre Pagella) or any Palatino clone listed in font.serif of style.py.