"""Plot robustness curves from a sweep JSON. Reads JSON produced by run_robustness_sweep.sh (or single evaluate_robustness.py runs appended to the same file), draws AUROC / AP / Acc / Acc@EER as a function of perturbation level, one line per perturbation kind. Output: outputs/analysis/robustness/figs_/ ├── robustness_auroc.{png,pdf} ├── robustness_ap.{png,pdf} ├── robustness_acc.{png,pdf} ├── robustness_acceer.{png,pdf} ├── robustness_grid.png (4-in-1 paper figure) └── robustness_table.csv Usage: python3 scripts/analysis/plot_robustness.py \\ --json outputs/analysis/robustness/cta_runs_20260615_205515.json """ from __future__ import annotations import argparse import json import os from collections import defaultdict from pathlib import Path import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt # Order chosen so visually similar perturbations are adjacent on the legend PERT_ORDER = [ "gaussian_noise", "block_wise", "jpeg_quality", "color_saturation", "color_contrast", "gaussian_blur", "pixelate", ] # Color palette: noise/block/jpeg use warm colors (the dangerous ones), # color/blur/pixelate use cool (the harmless ones) COLORS = { "gaussian_noise": "#C0392B", # red "block_wise": "#E67E22", # orange "jpeg_quality": "#F1C40F", # yellow "color_saturation": "#16A085", # teal "color_contrast": "#2980B9", # blue "gaussian_blur": "#8E44AD", # purple "pixelate": "#7F8C8D", # grey } PRETTY = { "gaussian_noise": "Gaussian noise", "block_wise": "Block occlusion", "jpeg_quality": "JPEG compression", "color_saturation": "Color saturation", "color_contrast": "Color contrast", "gaussian_blur": "Gaussian blur", "pixelate": "Pixelation", } def collect(json_path: str): """Returns dict[perturbation] -> {level: {AUROC, AP, Accuracy, Acc@EER}}.""" with open(json_path) as f: blob = json.load(f) runs = blob.get("runs", []) out = defaultdict(dict) for r in runs: p = r.get("perturbation") L = r.get("level") if not p or not L: continue o = r.get("overall", {}) out[p][L] = { "AUROC": o.get("AUROC"), "AP": o.get("AP"), "Accuracy": o.get("Accuracy"), "Acc@EER": o.get("Acc@EER"), "param": r.get("param"), } return out def _plot_one(ax, data, metric_key, title, ylabel, ylim=None): for p in PERT_ORDER: if p not in data: continue levels = sorted(data[p].keys()) ys = [data[p][L].get(metric_key) for L in levels] if all(y is None for y in ys): continue ax.plot(levels, ys, marker="o", linewidth=2.0, markersize=6, color=COLORS[p], label=PRETTY[p]) ax.set_xticks([1, 2, 3, 4, 5]) ax.set_xlabel("Perturbation level (1 = clean, 5 = strongest)") ax.set_ylabel(ylabel) ax.set_title(title) if ylim is not None: ax.set_ylim(ylim) ax.grid(True, alpha=0.3, linestyle=":") ax.spines["top"].set_visible(False) ax.spines["right"].set_visible(False) def save_table(data, out_csv): """Wide CSV: rows = perturbation × level, cols = AUROC, AP, Acc, Acc@EER, param.""" with open(out_csv, "w") as f: f.write("perturbation,level,param,AUROC,AP,Accuracy,Acc@EER,delta_AUROC_vs_L1\n") for p in PERT_ORDER: if p not in data: continue base = data[p].get(1, {}).get("AUROC") for L in sorted(data[p].keys()): row = data[p][L] d = (row.get("AUROC") - base) if (base is not None and row.get("AUROC") is not None) else float("nan") f.write( f"{p},{L},{row.get('param')}," f"{row.get('AUROC'):.4f},{row.get('AP'):.4f}," f"{row.get('Accuracy'):.4f},{row.get('Acc@EER'):.4f}," f"{d:+.4f}\n" ) print(f"[plot] wrote {out_csv}") def plot_grid(data, out_path): """4-in-1 figure for paper.""" fig, axes = plt.subplots(2, 2, figsize=(12, 8)) _plot_one(axes[0, 0], data, "AUROC", "AUROC vs perturbation level", "AUROC") _plot_one(axes[0, 1], data, "AP", "AP vs perturbation level", "Average Precision") _plot_one(axes[1, 0], data, "Accuracy","Accuracy vs perturbation level","Accuracy @ 0.5") _plot_one(axes[1, 1], data, "Acc@EER", "Acc@EER vs perturbation level", "Acc @ EER threshold") # one shared legend at the top handles, labels = axes[0, 0].get_legend_handles_labels() fig.legend(handles, labels, loc="upper center", ncol=6, bbox_to_anchor=(0.5, 1.005), frameon=False, fontsize=9) fig.tight_layout(rect=(0, 0, 1, 0.97)) fig.savefig(out_path, dpi=200, bbox_inches="tight") fig.savefig(out_path.replace(".png", ".pdf"), bbox_inches="tight") plt.close(fig) print(f"[plot] wrote {out_path}") def plot_single(data, metric_key, title, ylabel, out_path, ylim=None): fig, ax = plt.subplots(figsize=(7.5, 5)) _plot_one(ax, data, metric_key, title, ylabel, ylim=ylim) ax.legend(frameon=False, loc="best", fontsize=9) fig.tight_layout() fig.savefig(out_path, dpi=200, bbox_inches="tight") fig.savefig(out_path.replace(".png", ".pdf"), bbox_inches="tight") plt.close(fig) print(f"[plot] wrote {out_path}") def main(): p = argparse.ArgumentParser() p.add_argument("--json", required=True, help="Path to robustness JSON.") p.add_argument("--out_dir", default=None, help="Output dir. Default: alongside the JSON, named figs_") p.add_argument("--exclude", nargs="*", default=[], help="Perturbation names to skip in the plots " "(e.g. --exclude gaussian_noise). Useful for cleaner figures " "when one perturbation is an outlier.") p.add_argument("--include", nargs="*", default=None, help="If given, only these perturbations are plotted. " "Mutually exclusive with --exclude.") args = p.parse_args() json_path = Path(args.json).resolve() suffix_bits = [] if args.exclude: suffix_bits.append("noex_" + "_".join(args.exclude)) if args.include: suffix_bits.append("only_" + "_".join(args.include)) suffix = ("_" + "_".join(suffix_bits)) if suffix_bits else "" if args.out_dir is None: out_dir = json_path.parent / f"figs_{json_path.stem}{suffix}" else: out_dir = Path(args.out_dir).resolve() out_dir.mkdir(parents=True, exist_ok=True) print(f"[plot] reading {json_path}") data = collect(str(json_path)) if args.include: data = {k: v for k, v in data.items() if k in set(args.include)} print(f"[plot] include filter: keeping {sorted(data.keys())}") if args.exclude: excluded = set(args.exclude) data = {k: v for k, v in data.items() if k not in excluded} print(f"[plot] exclude filter: dropping {sorted(excluded)}") print(f"[plot] perturbations to plot: {sorted(data.keys())}") print(f"[plot] writing to {out_dir}") plot_single(data, "AUROC", "Robustness — AUROC", "AUROC", str(out_dir / "robustness_auroc.png")) plot_single(data, "AP", "Robustness — AP", "Average Precision", str(out_dir / "robustness_ap.png")) plot_single(data, "Accuracy","Robustness — Accuracy","Accuracy @ 0.5", str(out_dir / "robustness_acc.png")) plot_single(data, "Acc@EER", "Robustness — Acc@EER", "Acc @ EER threshold", str(out_dir / "robustness_acceer.png")) plot_grid(data, str(out_dir / "robustness_grid.png")) save_table(data, str(out_dir / "robustness_table.csv")) if __name__ == "__main__": main()