transfer/code/prompt_selection/visualize_per_metric.py

#!/usr/bin/env python3
"""Per-metric bar charts: PS vs BL across all perturbations, with % difference."""
from __future__ import annotations

import sys
from pathlib import Path

_THIS_DIR = Path(__file__).resolve().parent
if str(_THIS_DIR.parent) not in sys.path:
    sys.path.insert(0, str(_THIS_DIR.parent))

import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

from prompt_selection import config as cfg

CSV_PATH = cfg.EVAL_DIR / "all_comparison.csv"
OUTPUT_DIR = cfg.EVAL_DIR / "per_metric_charts"

# Metrics to visualize
SELECTED_METRICS = [
    "pearson_delta",
    "de_direction_match",
    "de_sig_genes_recall",
    "roc_auc",
    "mse",
    "mae",
]

DISPLAY_NAMES = {
    "pearson_delta": "Pearson Delta",
    "de_direction_match": "DE Direction Match",
    "de_sig_genes_recall": "DE Sig Genes Recall",
    "roc_auc": "ROC AUC",
    "mse": "MSE",
    "mae": "MAE",
}

LOWER_IS_BETTER = {"mse", "mae"}


def plot_one_metric(df_metric: pd.DataFrame, metric_name: str, output_dir: Path):
    """Generate a grouped bar chart for one metric across all perturbations."""
    display_name = DISPLAY_NAMES.get(metric_name, metric_name)
    lower_better = metric_name in LOWER_IS_BETTER

    # Sort by PS value (descending for quality, ascending for error)
    df_metric = df_metric.sort_values("prompt_selection", ascending=lower_better).reset_index(drop=True)

    # Shorten long perturbation names for display
    short_names = {
        "O-Demethylated Adapalene": "O-Demeth. Adapalene",
        "Porcn Inhibitor III": "Porcn Inhib. III",
        "Dimethyl Sulfoxide": "DMSO",
    }
    df_metric["display_pert"] = df_metric["perturbation"].map(short_names).fillna(df_metric["perturbation"])

    n = len(df_metric)
    y = np.arange(n)
    bar_h = 0.35

    fig, ax = plt.subplots(figsize=(12, max(6, n * 0.55)))

    bars_ps = ax.barh(y - bar_h / 2, df_metric["prompt_selection"], bar_h,
                      label="Prompt Selection", color="#4C72B0", edgecolor="white", linewidth=0.5)
    bars_bl = ax.barh(y + bar_h / 2, df_metric["random_baseline"], bar_h,
                      label="Random Baseline", color="#DD8452", edgecolor="white", linewidth=0.5)

    ax.set_yticks(y)
    ax.set_yticklabels(df_metric["display_pert"], fontsize=11)
    ax.invert_yaxis()
    ax.set_xlabel(display_name, fontsize=12)
    ax.legend(loc="lower right", fontsize=10, framealpha=0.9)
    ax.grid(axis="x", alpha=0.3, linestyle="--")
    ax.set_axisbelow(True)

    if lower_better:
        subtitle = "(lower is better)"
    else:
        subtitle = "(higher is better)"
    ax.set_title(f"{display_name} — Prompt Selection vs Random Baseline\n{subtitle}",
                 fontsize=14, fontweight="bold", pad=12)

    # Annotate percentage difference
    for idx, row in df_metric.iterrows():
        ps_val = row["prompt_selection"]
        bl_val = row["random_baseline"]
        max_val = max(abs(ps_val), abs(bl_val))

        if abs(bl_val) > 1e-12:
            pct = (ps_val - bl_val) / abs(bl_val) * 100
        else:
            pct = 0.0

        if abs(pct) < 0.01:
            label = "0%"
            color = "gray"
        else:
            sign = "+" if pct > 0 else ""
            label = f"{sign}{pct:.1f}%"
            if lower_better:
                color = "#388E3C" if pct < 0 else "#D32F2F"  # green if lower (better)
            else:
                color = "#388E3C" if pct > 0 else "#D32F2F"  # green if higher (better)

        # Position label to the right of the longer bar
        text_x = max(ps_val, bl_val)
        if text_x < 0:
            text_x = min(ps_val, bl_val)
            ax.text(text_x * 1.02, idx, label,
                    va="center", ha="right", fontsize=10, fontweight="bold", color=color)
        else:
            ax.text(text_x * 1.02 + max_val * 0.01, idx, label,
                    va="center", ha="left", fontsize=10, fontweight="bold", color=color)

    # Add margin for labels
    x_vals = pd.concat([df_metric["prompt_selection"], df_metric["random_baseline"]])
    x_min, x_max = x_vals.min(), x_vals.max()
    margin = (x_max - x_min) * 0.2 if x_max > x_min else abs(x_max) * 0.3
    if x_min < 0:
        ax.set_xlim(left=x_min - margin * 0.5)
    ax.set_xlim(right=x_max + margin)

    plt.tight_layout()
    out_path = output_dir / f"{metric_name}.png"
    fig.savefig(out_path, dpi=150, bbox_inches="tight", facecolor="white")
    plt.close(fig)
    print(f"Saved: {out_path}")


def main():
    df = pd.read_csv(CSV_PATH)
    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

    for metric in SELECTED_METRICS:
        df_metric = df[df["metric"] == metric].copy()
        df_metric = df_metric.dropna(subset=["prompt_selection", "random_baseline"])

        if df_metric.empty:
            print(f"No data for {metric}, skipping.")
            continue

        plot_one_metric(df_metric, metric, OUTPUT_DIR)

    print(f"\nAll charts saved to: {OUTPUT_DIR}")


if __name__ == "__main__":
    main()