Visualization¶

All plot functions live in h2ml.plots.plots. Each accepts an optional save_path; omit it to call plt.show().

Model scores¶

from h2ml.plots.plots import model_scores, pipeline_scores

# Step 1 scores only
model_scores(result, save_path="plots/step1_scores.png")

# All stages (default, reduced, optimized) overlaid
pipeline_scores(result, save_path="plots/all_stages.png")

CV diagnostics¶

Classification panel (ROC, PR curve, calibration, confusion matrix) or regression panel (residuals, actual vs predicted, error distribution):

from h2ml.plots.plots import cv_diagnostics

cv_diagnostics(result, save_path="plots/diagnostics.png")

SHAP importance¶

from h2ml.plots.plots import shap_importance, shap_summary_plot, shap_dependence

# Horizontal bar chart — mean absolute SHAP per feature
shap_importance(result.selector, save_path="plots/shap_bar.png")

# Beeswarm — direction and magnitude per sample for the final best model
shap_summary_plot(result, save_path="plots/shap_beeswarm.png")

# Scatter + LOWESS with bootstrap CI band for the top-N most important features
shap_dependence(result, n_features=6, save_path="plots/shap_dependence.png")

Spatial fold assignment¶

Visualise how samples are distributed across spatial CV folds:

from h2ml.plots.plots import spatial_folds

spatial_folds(result, store, save_path="plots/folds.png")

Only available when result.cv_type == "spatial".

Spatial blocks¶

A two-panel scatter showing each sample's block ID (left) and the fold it is held out in (right). Call .plot() directly on a fitted splitter, or use the exported plot_spatial_blocks function:

from h2ml.plots import plot_spatial_blocks

# Shortcut method on the splitter (result.splitter is the fitted instance)
result.splitter.plot(save_path="plots/blocks.png")

# Equivalent explicit call
plot_spatial_blocks(result.splitter, save_path="plots/blocks.png")

lon_col / lat_col (default 1 / 0) select which columns of the splitter's coords are longitude and latitude. result.splitter is None after PipelineResult.load() — rebuild via build_final_model() first.