| """Generate the reproduction's figures (Plotly HTML + CSV raw data).""" |
| from __future__ import annotations |
|
|
| import argparse |
| import glob |
| import json |
| import os |
| from collections import defaultdict |
|
|
| import numpy as np |
| import plotly.graph_objects as go |
| from plotly.subplots import make_subplots |
|
|
| |
| C_OURS = "#D55E00" |
| C_BASE = ["#0072B2", "#009E73", "#CC79A7", "#56B4E9", "#E69F00", |
| "#7570B3", "#666666", "#8C6D31", "#1B9E77", "#A6761D"] |
| LABEL = { |
| "stgfn": "ST-GFN (Ours)", "tb": "TB", "fm": "FM", "subtb": "SubTB", "db": "DB", |
| "eflownet": "EFlowNet", "stochastic_gfn": "Stochastic-GFN", "tb_rnd": "TB+RND", |
| "tb_novelty": "TB+Novelty", "tb_icm": "TB+ICM", "tb_cv": "TB+ControlVar", |
| "stgfn_no_spectral": "ST-GFN w/o spectral", "stgfn_no_intrinsic": "ST-GFN w/o intrinsic", |
| } |
| ORDER = ["stgfn", "tb", "fm", "subtb", "db", "eflownet", "stochastic_gfn", |
| "tb_rnd", "tb_novelty", "tb_icm", "tb_cv"] |
|
|
| LAYOUT = dict( |
| template="plotly_white", font=dict(family="Inter, system-ui, sans-serif", size=13), |
| margin=dict(l=60, r=30, t=60, b=55), hovermode="x unified", |
| ) |
|
|
|
|
| def load(out_dir): |
| data = defaultdict(lambda: defaultdict(list)) |
| for path in glob.glob(os.path.join(out_dir, "*.json")): |
| with open(path) as f: |
| r = json.load(f) |
| data[r["env"]][r["method"]].append(r) |
| return data |
|
|
|
|
| def color(m, i): |
| return C_OURS if m.startswith("stgfn") else C_BASE[i % len(C_BASE)] |
|
|
|
|
| def curve_figure(data, env, metric, ylabel, title, out_html, out_csv): |
| fig = go.Figure() |
| rows = [("method", "iter", "mean", "lo", "hi")] |
| for i, m in enumerate([m for m in ORDER if m in data[env]]): |
| runs = data[env][m] |
| iters = [c["iter"] for c in runs[0]["curve"]] |
| vals = [] |
| for r in runs: |
| v = [c.get(metric) for c in r["curve"]] |
| if all(x is not None for x in v) and len(v) == len(iters): |
| vals.append(v) |
| if not vals: |
| continue |
| arr = np.array(vals, float) |
| mu = arr.mean(0) |
| ci = 1.96 * arr.std(0, ddof=1) / np.sqrt(len(arr)) if len(arr) > 1 else np.zeros_like(mu) |
| col = color(m, i) |
| wide = m == "stgfn" |
| fig.add_trace(go.Scatter( |
| x=iters + iters[::-1], y=list(mu + ci) + list(mu - ci)[::-1], |
| fill="toself", fillcolor=col, opacity=0.15, line=dict(width=0), |
| hoverinfo="skip", showlegend=False)) |
| fig.add_trace(go.Scatter(x=iters, y=mu, name=LABEL.get(m, m), |
| line=dict(color=col, width=3.5 if wide else 1.9))) |
| for k, it in enumerate(iters): |
| rows.append((m, it, mu[k], mu[k] - ci[k], mu[k] + ci[k])) |
| fig.update_layout(title=title, xaxis_title="training iteration", |
| yaxis_title=ylabel, **LAYOUT) |
| fig.write_html(out_html, include_plotlyjs="cdn", full_html=True) |
| with open(out_csv, "w") as f: |
| for r in rows: |
| f.write(",".join(str(x) for x in r) + "\n") |
| print("wrote", out_html) |
|
|
|
|
| def bar_figure(data, env, metric, ylabel, title, out_html, out_csv, higher_better=True): |
| means, cis, labels, cols = [], [], [], [] |
| rows = [("method", "mean", "ci95", "n_seeds")] |
| for i, m in enumerate([m for m in ORDER if m in data[env]]): |
| vals = [r["final"].get(metric) for r in data[env][m]] |
| vals = [v for v in vals if v is not None] |
| if not vals: |
| continue |
| mu = float(np.mean(vals)) |
| ci = float(1.96 * np.std(vals, ddof=1) / np.sqrt(len(vals))) if len(vals) > 1 else 0.0 |
| means.append(mu); cis.append(ci); labels.append(LABEL.get(m, m)); cols.append(color(m, i)) |
| rows.append((m, mu, ci, len(vals))) |
| order = np.argsort(means)[::-1] if higher_better else np.argsort(means) |
| fig = go.Figure(go.Bar( |
| x=[labels[i] for i in order], y=[means[i] for i in order], |
| error_y=dict(type="data", array=[cis[i] for i in order], thickness=1.4), |
| marker_color=[cols[i] for i in order], |
| )) |
| fig.update_layout(title=title, yaxis_title=ylabel, xaxis_tickangle=-35, **LAYOUT) |
| fig.write_html(out_html, include_plotlyjs="cdn", full_html=True) |
| with open(out_csv, "w") as f: |
| for r in rows: |
| f.write(",".join(str(x) for x in r) + "\n") |
| print("wrote", out_html) |
|
|
|
|
| def spectral_diag_figure(diag_path, out_html, out_csv): |
| with open(diag_path) as f: |
| d = json.load(f) |
| names = list(d.keys()) |
| fig = make_subplots(rows=1, cols=2, subplot_titles=( |
| "RFF feature norm ||z(s)||² (theory: exactly 1)", |
| "Spectral regularizer energy ||V̂(s,a)||²")) |
| fig.add_trace(go.Bar(x=names, y=[d[n]["mean_z_norm_sq"] for n in names], |
| error_y=dict(type="data", array=[d[n]["std_z_norm_sq"] for n in names]), |
| marker_color="#0072B2", name="||z||²"), row=1, col=1) |
| fig.add_trace(go.Bar(x=names, y=[d[n]["mean_energy"] for n in names], |
| error_y=dict(type="data", array=[d[n]["std_energy"] for n in names]), |
| marker_color=C_OURS, name="||V̂||²"), row=1, col=2) |
| fig.add_hline(y=1.0, line_dash="dash", line_color="#666", row=1, col=1) |
| fig.add_hline(y=1.0, line_dash="dash", line_color="#666", row=1, col=2) |
| fig.update_layout(title="Why the spectral regularizer is inert in deterministic environments", |
| showlegend=False, xaxis_tickangle=-25, xaxis2_tickangle=-25, **LAYOUT) |
| fig.write_html(out_html, include_plotlyjs="cdn", full_html=True) |
| with open(out_csv, "w") as f: |
| f.write("env,mean_z_norm_sq,std_z_norm_sq,mean_energy,std_energy,mean_action_spread\n") |
| for n in names: |
| r = d[n] |
| f.write(f"{n},{r['mean_z_norm_sq']},{r['std_z_norm_sq']},{r['mean_energy']}," |
| f"{r['std_energy']},{r['mean_across_action_spread']}\n") |
| print("wrote", out_html) |
|
|
|
|
| if __name__ == "__main__": |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--out", default="../outputs/main") |
| ap.add_argument("--fig", default="../figures") |
| args = ap.parse_args() |
| os.makedirs(args.fig, exist_ok=True) |
| data = load(args.out) |
| F = args.fig |
|
|
| specs = [ |
| ("hypergrid", "modes_found", "modes discovered (of 64)", "HyperGrid: mode discovery", True), |
| ("bitsequence", "top100_reward", "normalised top-100 reward", "BitSequence: reward under 90% action failure", True), |
| ("bitsequence", "l1_to_target", "L1 distance to P* ∝ R", "BitSequence: distributional error", False), |
| ("tictactoe", "win_pct", "win rate (%)", "TicTacToe vs 90%-optimal minimax", True), |
| ("tictactoe", "optimal_pct", "optimal moves (%)", "TicTacToe: move quality", True), |
| ("singlecell_proxy", "target_corr", "correlation with true reward", "SingleCell proxy: combinatorial generalisation", True), |
| ] |
| CURVE_OK = {"modes_found", "top100_reward", "mean_reward", "win_pct", |
| "l1_to_target", "coverage_pct", "diversity"} |
| for env, metric, ylab, title, hib in specs: |
| if env not in data: |
| continue |
| bar_figure(data, env, metric, ylab, title, |
| f"{F}/{env}_{metric}_bar.html", f"{F}/{env}_{metric}_bar.csv", hib) |
| if metric in CURVE_OK: |
| curve_figure(data, env, metric, ylab, title + " (training curves)", |
| f"{F}/{env}_{metric}_curve.html", f"{F}/{env}_{metric}_curve.csv") |
|
|
| diag = "../outputs/spectral_diagnostics.json" |
| if os.path.exists(diag): |
| spectral_diag_figure(diag, f"{F}/spectral_diagnostics.html", f"{F}/spectral_diagnostics.csv") |
|
|