plot_eval

plot_eval(df_eval, metric=None, score_col=None, figsize=None)

Decompose a find_features() sweep into a set of publication-ready evaluation figures.

The high-dimensional structural sweep (Part x Split x Scale) is decomposed rather than crammed into one plot: the two most-informative axes (largest marginal-mean impact on the score) become each heatmap’s rows and columns, and the least-informative axis becomes the slice — one clean 2D viridis heatmap per slice level, all sharing a single color scale so panels are directly comparable, every cell annotated with its score and the selected configuration boxed. Alongside the heatmaps it returns a marginal-impact bar panel (how much each lever moves the score) and an ``n_filter`` refinement panel. Every figure is returned separately so each can be saved and placed individually in a publication.

Parameters:

• df_eval (pd.DataFrame) – Per-configuration sweep table — the third return value of find_features() (with a stage column and one <metric>_mean column per metric), or any compatible eval table. An is_selected column (if present) marks the winner; otherwise the maximum score is used.

• metric (str, optional) – Metric whose <metric>_mean column is colored. If None, the first *_mean column.

• score_col (str, optional) – Explicit score column to color by (overrides metric).

• figsize (tuple of float, optional) – Per-figure size in inches. If None, a size is derived from each panel’s grid.

Returns:

figs – The evaluation figures (heatmap slices, then the marginal-impact panel, then the n_filter panel). Empty when no axis varies (a single configuration, e.g. search="fast").

Return type:

list of matplotlib.figure.Figure

See also

• find_features() : the CPP AutoML pipeline whose df_eval this visualizes; it attaches these figures to the returned ax as ax.eval when plot=True.

Examples

aap.plot_eval turns a :func:aap.find_features sweep table (df_eval) into a set of publication-ready evaluation figures. The high-dimensional Part × Split × Scale grid is decomposed into clean 2D viridis heatmaps — the two most-informative axes on each panel, the least-informative as the slice — sharing one color scale with the selected configuration starred, plus a marginal-impact panel and an n_filter panel. It returns the list of figures, so each can be saved and placed individually in a paper. find_features(plot=True) calls it for you and attaches the figures as ax.eval.

import warnings
import matplotlib.pyplot as plt
import aaanalysis as aa
import aaanalysis.pipe as aap

aa.options["verbose"] = False
aa.plot_settings()
# Silence the small-demo-data 'n_filter' shortfall advisory (moot on real datasets).
warnings.filterwarnings("ignore", message=r"'n_filter'", category=RuntimeWarning)

df_seq = aa.load_dataset(name="DOM_GSEC", n=20)
labels = df_seq["label"].to_list()

# A balanced search sweeping the Split pattern modes and the Scale sets (two structural axes).
df_feat, ax, df_eval = aap.find_features(labels=labels, df_seq=df_seq, search="balanced",
                                         kws={"n_split_max": 15}, plot=False,
                                         random_state=42, n_jobs=1)
aa.display_df(df_eval, n_rows=10, show_shape=True)

DataFrame shape: (61, 14)

	stage	list_parts	split_types	pattern_mode	n_split_max	scale	n_jmd	n_filter	n_features	balanced_accuracy_mean	balanced_accuracy_std	is_pareto	is_selected	rank
1	sensitivity	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment,Pattern	p1	15	explain:50	10	150	150	1.000000	0.000000	True	False	1
2	sensitivity	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment,Pattern	p1	15	explain:all	10	150	150	1.000000	0.000000	True	False	2
3	sensitivity	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment,PeriodicPattern	p2	15	explain:40	10	150	150	1.000000	0.000000	True	False	3
4	sensitivity	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment,Pattern...PeriodicPattern	p1+p2	15	explain:50	10	150	150	1.000000	0.000000	True	False	4
5	sensitivity	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment,Pattern...PeriodicPattern	p1+p2	15	explain:all	10	150	150	1.000000	0.000000	True	False	5
6	n_filter	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment,Pattern	p1	15	explain:50	10	125	125	1.000000	0.000000	True	False	6
7	n_filter	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment,Pattern	p1	15	explain:50	10	150	150	1.000000	0.000000	True	False	7
8	refine	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment,Pattern	p1	15	explain:50	10	125	82	1.000000	0.000000	True	True	8
9	sensitivity	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment	none	15	explain:40	10	150	150	0.975000	0.050000	False	False	9
10	sensitivity	tmd,jmd_n_tmd_n,tmd_c_jmd_c	Segment	none	15	explain:50	10	150	150	0.975000	0.050000	False	False	10

Calling plot_eval on the table returns the list of figures (heatmap slice(s), then the marginal-impact panel, then the n_filter panel). plt.show() renders them all:

figs = aap.plot_eval(df_eval)
print(f"{len(figs)} publication eval figures")
plt.show()

6 publication eval figures

Color by a specific metric (when several were optimized), pin the score column with score_col, or set the per-figure figsize. Each figure is a standalone matplotlib.figure.Figure you can save for a publication:

figs = aap.plot_eval(df_eval, metric="balanced_accuracy", figsize=(5, 4))
# figs[0].savefig("sweep_heatmap.png", dpi=300, bbox_inches="tight")
print(f"saved-ready figures: {len(figs)}")
plt.show()

saved-ready figures: 6