Visualization & Diagnostics¶

from torchregress.viz import set_style, save_figure, plot_residuals

set_style()                                    # consistent look
fig = plot_residuals(y_pred, y_true, return_figure=True)
save_figure(fig, "residuals", formats=["png", "pdf"])

from torchregress.viz import plot_residuals, plot_residual_histogram, plot_qq_plot

plot_residuals(y_pred, y_true, y_pred_std=y_pred_std, clip_outliers=True)
plot_residual_histogram(y_pred, y_true, show_kde=True, return_figure=True)
plot_qq_plot(y_pred, y_true, return_figure=True)

from torchregress.viz import plot_prediction_intervals, plot_distribution_comparison

# For 95% Gaussian intervals
plot_prediction_intervals(y_pred, y_pred - 1.96 * std, y_pred + 1.96 * std, y_true)

# For ensemble / Monte Carlo samples — shape [n_samples, batch_size]
plot_distribution_comparison(predicted_samples, y_true, n_samples_to_show=4, credible_interval=0.95)

from torchregress.viz import (
    plot_reliability_diagram, plot_gaussian_reliability_diagram,
    plot_pit_histogram, plot_calibration_curve,
)

# Quantile reliability
plot_reliability_diagram({0.1: q10, 0.5: q50, 0.9: q90}, y_true)

# Gaussian reliability
plot_gaussian_reliability_diagram(y_pred, y_pred_std, y_true)

# PIT calibration — uniform = perfectly calibrated
plot_pit_histogram(y_pred, y_pred_std, y_true, n_bins=20)

# Binary probability calibration
plot_calibration_curve(y_pred_probs, y_true_binary)

from torchregress.viz import (
    plot_uncertainty_vs_error, plot_binned_metrics,
    plot_target_density_error_overlap, plot_conditional_density_slices,
    plot_censored_survival_curves,
)

# Are confident predictions actually more accurate?
plot_uncertainty_vs_error(y_pred, y_pred_std, y_true)

# RMSE binned by predicted value
plot_binned_metrics(y_pred, y_pred_std, y_true, metric="rmse")

# Overlap of target density and error distribution
plot_target_density_error_overlap(y_true, y_pred)

# Predictive density at selected x-values (density_fn maps (x, y_grid) -> pdf)
plot_conditional_density_slices(density_fn, x_slices, y_grid, y_true_slices=y_true_at_slices)

# Survival curves for censored data
plot_censored_survival_curves(
    predicted_survival, time_grid, observed_times, censoring_indicators
)

from torchregress.viz import (
    plot_learning_curves, plot_validation_metrics,
    plot_early_stopping, plot_lr_find_results,
)

train_hist = {"loss": train_losses, "rmse": train_rmse}
val_hist = {"loss": val_losses, "rmse": val_rmse}

# Combined train/val curves
plot_learning_curves(train_hist, val_hist, log_scale=["loss"], smoothing=0.2)

# Validation-only metrics
plot_validation_metrics(epochs, val_hist, return_figure=True)

# Early stopping — patience=5 finds the minimum before overfitting
plot_early_stopping(train_losses, val_losses, patience=5)

# LR finder — 'valley' picks the steepest descent point
plot_lr_find_results(lr_values, loss_values, suggestion_method="valley")

from torchregress.viz import (
    plot_performance_comparison, plot_parameter_sensitivity,
    plot_feature_importance, plot_model_ensemble_contributions,
    plot_risk_coverage_curve, plot_causal_uplift_qini,
    plot_simex_extrapolation,
)

# Bar / radar / heatmap comparison
metrics = {"Model A": {"RMSE": 0.25, "MAE": 0.18}, "Model B": {"RMSE": 0.22, "MAE": 0.16}}
plot_performance_comparison(metrics, plot_type="bar", highlight_best=True)
plot_performance_comparison(metrics, plot_type="radar")

# Sensitivity of RMSE to hidden dimension
param_vals = {"hidden_dim": [32, 64, 128, 256]}
sens_metrics = {"RMSE": [0.55, 0.48, 0.45, 0.46]}
plot_parameter_sensitivity(param_vals, sens_metrics)

# Feature importance (automatically sorts)
plot_feature_importance(feature_names, importance_scores, top_n=10, horizontal=True)

# Ensemble member contributions
plot_model_ensemble_contributions(member_predictions, ensemble_mean)

# Risk-coverage — trade off coverage for lower risk
plot_risk_coverage_curve(y_true, y_pred, rejection_scores)

# Uplift Qini curve
plot_causal_uplift_qini(uplift_scores, treatment, outcome)

# SIMEX extrapolation
plot_simex_extrapolation(lambdas, coefficient_estimates, extrapolated_coef)

from torchregress.viz import (
    set_style, create_grid_figure, create_color_palette,
    add_identity_line, add_annotations, save_figure,
    enable_latex_rendering, format_metric_label,
)

# Global style
set_style()

# Multi-panel diagnostic report
fig, axes = create_grid_figure(n_plots=6, ncols=3, figsize=(18, 12))
plot_residuals(y_pred, y_true, ax=axes[0], title="Residuals")
plot_qq_plot(y_pred, y_true, ax=axes[1], title="Q-Q")
# ... fill remaining axes ...
save_figure(fig, "diagnostic_report", formats=["png", "pdf"])

# Color palette
palette = create_color_palette(5, palette_name="viridis")

# Reference lines
add_identity_line(ax)              # y = x

# Annotations
add_annotations(ax, {"RMSE": 0.45, "PICP": 0.91})

# Format metric names
label = format_metric_label("expected_calibration_error")
print(label)  # "Expected Calibration Error"