scripts/plot_eval.py

"""Generate publication-style figures from an eval-run's CSVs.

Reads
-----
    {results_dir}/closed_set.csv
    {results_dir}/open_set.csv
    {results_dir}/roc.csv

Writes
------
    {results_dir}/figures/
        ├── 1_roc_curves.png         # open-set ROC, panel per n_refs
        ├── 2_closed_set_metrics.png # bar chart R@1 / R@5 / mAP × method × n_refs
        └── 3_n_refs_effect.png      # R@1 + AUC vs n_refs

If --results-dir is not passed, picks the most recent timestamped folder
under /seed_data/eval_results/.

Usage (one-time install of matplotlib inside the container):
    docker compose exec backend pip install matplotlib

Then run:
    docker compose exec backend python -m scripts.plot_eval
"""
from __future__ import annotations

import argparse
import csv
import logging
from collections import defaultdict
from pathlib import Path

import matplotlib.pyplot as plt
import numpy as np

log = logging.getLogger("plot_eval")

METHOD_ORDER = ("flat", "centroid", "max_sim", "max_sim_bonus")
METHOD_COLORS = {
    "flat":          "#888888",
    "centroid":      "#4c72b0",
    "max_sim":       "#dd8452",
    "max_sim_bonus": "#55a868",
}


def load_csv(path: Path) -> list[dict]:
    with path.open(encoding="utf-8") as f:
        return list(csv.DictReader(f))


def latest_results_dir(root: Path) -> Path | None:
    if not root.exists():
        return None
    candidates = [d for d in root.iterdir() if d.is_dir() and (d / "closed_set.csv").exists()]
    if not candidates:
        return None
    return max(candidates, key=lambda d: d.name)


def trapezoid_auc(points: list[tuple[float, float]]) -> float:
    pts = sorted(set(points))
    pts = [(0.0, 0.0)] + pts + [(1.0, 1.0)]
    pts = sorted(set(pts))
    auc = 0.0
    for (x1, y1), (x2, y2) in zip(pts, pts[1:]):
        auc += (x2 - x1) * (y1 + y2) / 2.0
    return auc


# ---- Figure 1: ROC curves ------------------------------------------------

def plot_roc(roc_rows: list[dict], out_path: Path) -> None:
    n_refs_values = sorted({int(r["n_refs"]) for r in roc_rows})
    methods = [m for m in METHOD_ORDER if any(r["method"] == m for r in roc_rows)]

    fig, axes = plt.subplots(1, len(n_refs_values), figsize=(5.2 * len(n_refs_values), 5),
                             sharey=True)
    if len(n_refs_values) == 1:
        axes = [axes]

    for ax, n_refs in zip(axes, n_refs_values):
        for method in methods:
            pts = [
                (float(r["fpr_mean"]), float(r["tpr_mean"]))
                for r in roc_rows
                if int(r["n_refs"]) == n_refs and r["method"] == method
            ]
            pts.sort()
            xs = [0.0] + [p[0] for p in pts] + [1.0]
            ys = [0.0] + [p[1] for p in pts] + [1.0]
            auc = trapezoid_auc(pts)
            ax.plot(
                xs, ys,
                label=f"{method}  (AUC = {auc:.3f})",
                color=METHOD_COLORS[method],
                marker="o", markersize=4, linewidth=1.8,
            )

        ax.plot([0, 1], [0, 1], color="black", linestyle="--",
                linewidth=0.8, alpha=0.4, label="random")
        ax.set_xlim(-0.02, 1.02)
        ax.set_ylim(-0.02, 1.02)
        ax.set_xlabel("False positive rate  (1 − specificity)")
        ax.set_title(f"n_refs = {n_refs}")
        ax.grid(True, alpha=0.3, linewidth=0.5)
        ax.legend(loc="lower right", fontsize=8.5, frameon=True)
    axes[0].set_ylabel("True positive rate  (sensitivity)")

    fig.suptitle("Open-set ROC — Find My Dog re-identification", fontsize=13.5,
                 fontweight="bold")
    fig.tight_layout()
    fig.savefig(out_path, dpi=140, bbox_inches="tight")
    plt.close(fig)
    log.info("Wrote %s", out_path)


# ---- Figure 2: Closed-set bar chart --------------------------------------

def plot_closed_set(closed_rows: list[dict], out_path: Path) -> None:
    methods = [m for m in METHOD_ORDER if any(r["method"] == m for r in closed_rows)]
    n_refs_values = sorted({int(r["n_refs"]) for r in closed_rows})
    metrics = [("r1", "R@1"), ("r5", "R@5"), ("map", "mAP")]

    agg: dict[tuple[str, int], dict[str, list[float]]] = defaultdict(
        lambda: {m[0]: [] for m in metrics}
    )
    for r in closed_rows:
        key = (r["method"], int(r["n_refs"]))
        for k, _ in metrics:
            agg[key][k].append(float(r[k]))

    fig, axes = plt.subplots(1, 3, figsize=(15, 4.8), sharey=True)
    n_groups = len(methods)
    width = 0.8 / max(len(n_refs_values), 1)
    n_refs_colors = plt.cm.viridis(np.linspace(0.25, 0.85, len(n_refs_values)))

    for ax, (mkey, mlabel) in zip(axes, metrics):
        x = np.arange(n_groups)
        for j, (n_refs, color) in enumerate(zip(n_refs_values, n_refs_colors)):
            offset = (j - (len(n_refs_values) - 1) / 2) * width
            heights = [
                np.mean(agg[(m, n_refs)][mkey]) * 100 if agg[(m, n_refs)][mkey] else 0
                for m in methods
            ]
            errs = [
                np.std(agg[(m, n_refs)][mkey]) * 100 if agg[(m, n_refs)][mkey] else 0
                for m in methods
            ]
            bars = ax.bar(
                x + offset, heights, width, yerr=errs,
                label=f"n_refs = {n_refs}",
                color=color, edgecolor="white", linewidth=0.5, capsize=3,
            )
            # Bar value labels.
            for xi, h in zip(x + offset, heights):
                ax.text(xi, h + 1.5, f"{h:.0f}", ha="center", fontsize=7.5, color="#333")

        ax.set_xticks(x)
        ax.set_xticklabels(methods, rotation=18, ha="right", fontsize=9)
        ax.set_title(mlabel)
        ax.grid(axis="y", alpha=0.3, linewidth=0.5)
        ax.set_ylim(0, 110)
        if mkey == "r1":
            ax.legend(loc="lower right", fontsize=8.5)
    axes[0].set_ylabel("metric value (%)")

    fig.suptitle("Closed-set metrics — assumes correct dog is in gallery",
                 fontsize=13.5, fontweight="bold")
    fig.tight_layout()
    fig.savefig(out_path, dpi=140, bbox_inches="tight")
    plt.close(fig)
    log.info("Wrote %s", out_path)


# ---- Figure 3: n_refs effect --------------------------------------------

def plot_n_refs_effect(closed_rows: list[dict], roc_rows: list[dict],
                       out_path: Path) -> None:
    methods = [m for m in METHOD_ORDER if any(r["method"] == m for r in closed_rows)]
    n_refs_values = sorted({int(r["n_refs"]) for r in closed_rows})

    fig, axes = plt.subplots(1, 2, figsize=(12, 4.8))

    # Left panel: closed-set R@1
    ax = axes[0]
    for method in methods:
        means, stds = [], []
        for n_refs in n_refs_values:
            vals = [
                float(r["r1"]) for r in closed_rows
                if r["method"] == method and int(r["n_refs"]) == n_refs
            ]
            means.append(np.mean(vals) * 100 if vals else np.nan)
            stds.append(np.std(vals) * 100 if vals else 0)
        ax.errorbar(
            n_refs_values, means, yerr=stds,
            label=method, color=METHOD_COLORS[method],
            marker="o", capsize=4, linewidth=2,
        )
    ax.set_xlabel("number of reference photos per dog")
    ax.set_ylabel("Closed-set R@1 (%)")
    ax.set_title("Closed-set R@1 vs. cluster size")
    ax.set_xticks(n_refs_values)
    ax.set_ylim(50, 105)
    ax.grid(alpha=0.3, linewidth=0.5)
    ax.legend(loc="lower right", fontsize=9)

    # Right panel: open-set AUC
    ax = axes[1]
    by_mn = defaultdict(list)
    for r in roc_rows:
        by_mn[(r["method"], int(r["n_refs"]))].append(
            (float(r["fpr_mean"]), float(r["tpr_mean"]))
        )
    for method in methods:
        aucs = []
        for n_refs in n_refs_values:
            pts = by_mn.get((method, n_refs), [])
            aucs.append(trapezoid_auc(pts) if pts else np.nan)
        ax.plot(
            n_refs_values, aucs,
            label=method, color=METHOD_COLORS[method],
            marker="o", linewidth=2,
        )
    ax.axhline(y=0.5, color="black", linestyle="--",
               linewidth=0.8, alpha=0.4, label="random")
    ax.set_xlabel("number of reference photos per dog")
    ax.set_ylabel("Open-set ROC AUC")
    ax.set_title("Open-set AUC vs. cluster size")
    ax.set_xticks(n_refs_values)
    ax.set_ylim(0.4, 1.02)
    ax.grid(alpha=0.3, linewidth=0.5)
    ax.legend(loc="lower right", fontsize=9)

    fig.suptitle("Effect of reference cluster size",
                 fontsize=13.5, fontweight="bold")
    fig.tight_layout()
    fig.savefig(out_path, dpi=140, bbox_inches="tight")
    plt.close(fig)
    log.info("Wrote %s", out_path)


# ---- Main ----------------------------------------------------------------

def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--results-dir", type=Path, default=None,
        help="Path to a specific eval result folder. Defaults to latest under /seed_data/eval_results/.",
    )
    parser.add_argument(
        "--root", type=Path, default=Path("/seed_data/eval_results"),
        help="Root used to find latest result if --results-dir not given.",
    )
    args = parser.parse_args()

    logging.basicConfig(level=logging.INFO, format="%(levelname)s %(name)s: %(message)s")

    results_dir = args.results_dir or latest_results_dir(args.root)
    if not results_dir:
        raise SystemExit(f"No eval results found under {args.root}.")

    log.info("Using results dir: %s", results_dir)

    closed = load_csv(results_dir / "closed_set.csv")
    roc = load_csv(results_dir / "roc.csv")

    figdir = results_dir / "figures"
    figdir.mkdir(parents=True, exist_ok=True)

    plot_roc(roc, figdir / "1_roc_curves.png")
    plot_closed_set(closed, figdir / "2_closed_set_metrics.png")
    plot_n_refs_effect(closed, roc, figdir / "3_n_refs_effect.png")

    log.info("All figures in: %s", figdir)


if __name__ == "__main__":
    main()