Spaces:
Sleeping
Sleeping
| """ | |
| training/plot_curves.py — Generate reward curve HTML from training results | |
| ========================================================================== | |
| Reads checkpoints/cpu_run/reward_curve.csv and produces a self-contained | |
| HTML file with interactive Plotly charts — no matplotlib, no GPU. | |
| Run: | |
| python -m training.plot_curves --input checkpoints/cpu_run/reward_curve.csv | |
| # Opens reward_curves.html in browser | |
| """ | |
| from __future__ import annotations | |
| import argparse | |
| import csv | |
| import json | |
| import os | |
| import webbrowser | |
| def load_csv(path: str) -> dict[str, list]: | |
| data: dict[str, list] = {} | |
| with open(path) as f: | |
| reader = csv.DictReader(f) | |
| for row in reader: | |
| for k, v in row.items(): | |
| data.setdefault(k, []).append(float(v)) | |
| return data | |
| def generate_html(data: dict[str, list], output_path: str) -> str: | |
| steps = data.get("step", []) | |
| f1 = data.get("overseer_f1", []) | |
| det = data.get("detection_rate", []) | |
| fp = data.get("false_positive_rate", []) | |
| # Compute improvement annotations | |
| f1_before = f1[0] if f1 else 0 | |
| f1_after = f1[-1] if f1 else 0 | |
| gap = f1_after - f1_before | |
| html = f"""<!DOCTYPE html> | |
| <html> | |
| <head> | |
| <meta charset="UTF-8"> | |
| <title>NegotiArena Training Curves</title> | |
| <script src="https://cdn.plot.ly/plotly-2.26.0.min.js"></script> | |
| <style> | |
| body {{ background:#0f0f1a; color:#e0e0e0; font-family:sans-serif; | |
| max-width:1100px; margin:40px auto; padding:20px; }} | |
| h1 {{ color:#4a9eff; }} | |
| h3 {{ color:#aaa; font-weight:400; }} | |
| .metric-row {{ display:flex; gap:20px; margin:24px 0; }} | |
| .card {{ background:#1a1a2e; border:1px solid #333; border-radius:8px; | |
| padding:20px; flex:1; text-align:center; }} | |
| .card .val {{ font-size:2.2em; font-weight:bold; margin:8px 0; }} | |
| .card .label {{ color:#aaa; font-size:0.9em; }} | |
| .green {{ color:#00cc66; }} | |
| .red {{ color:#ff4a4a; }} | |
| .blue {{ color:#4a9eff; }} | |
| footer {{ color:#555; font-size:0.8em; margin-top:40px; text-align:center; }} | |
| </style> | |
| </head> | |
| <body> | |
| <h1>🏛️ NegotiArena — Training Results</h1> | |
| <h3>REINFORCE on CPU → same reward signal used for GRPO on GPU</h3> | |
| <div class="metric-row"> | |
| <div class="card"> | |
| <div class="label">Overseer F1 (Before)</div> | |
| <div class="val red">{f1_before:.3f}</div> | |
| <div class="label">Random baseline</div> | |
| </div> | |
| <div class="card"> | |
| <div class="label">Overseer F1 (After)</div> | |
| <div class="val green">{f1_after:.3f}</div> | |
| <div class="label">After {int(steps[-1]) if steps else 0} training steps</div> | |
| </div> | |
| <div class="card"> | |
| <div class="label">Improvement</div> | |
| <div class="val blue">+{gap:.3f}</div> | |
| <div class="label">{gap/f1_before*100:.0f}% relative gain</div> | |
| </div> | |
| <div class="card"> | |
| <div class="label">FP Rate (After)</div> | |
| <div class="val green">{fp[-1]:.1%}</div> | |
| <div class="label">Was {fp[0]:.1%}</div> | |
| </div> | |
| </div> | |
| <div id="chart_f1"></div> | |
| <div id="chart_det"></div> | |
| <script> | |
| const steps = {json.dumps(steps)}; | |
| const f1 = {json.dumps(f1)}; | |
| const det = {json.dumps(det)}; | |
| const fp = {json.dumps(fp)}; | |
| const layout = {{ | |
| paper_bgcolor: '#0f0f1a', | |
| plot_bgcolor: '#1a1a2e', | |
| font: {{ color: '#cccccc' }}, | |
| xaxis: {{ title: 'Training Step', gridcolor: '#333' }}, | |
| legend: {{ bgcolor: '#1a1a2e' }}, | |
| height: 360, | |
| margin: {{ t: 50, b: 50 }}, | |
| }}; | |
| // Chart 1: F1 curve with before/after annotation | |
| Plotly.newPlot('chart_f1', [ | |
| {{ | |
| x: steps, y: f1, mode: 'lines+markers', | |
| name: 'Overseer F1', line: {{ color: '#00cc66', width: 3 }}, | |
| marker: {{ size: 6 }} | |
| }}, | |
| {{ | |
| x: [steps[0]], y: [f1[0]], mode: 'markers', | |
| name: 'Before', marker: {{ color: '#ff4a4a', size: 14, symbol: 'circle' }} | |
| }}, | |
| {{ | |
| x: [steps[steps.length-1]], y: [f1[f1.length-1]], mode: 'markers', | |
| name: 'After', marker: {{ color: '#00cc66', size: 14, symbol: 'star' }} | |
| }} | |
| ], {{ | |
| ...layout, | |
| title: {{ text: 'Overseer F1 — Coalition Detection Improvement', font: {{ color: '#4a9eff', size: 16 }} }}, | |
| yaxis: {{ title: 'F1 Score', range: [0, 1], gridcolor: '#333' }}, | |
| annotations: [{{ | |
| x: steps[steps.length-1], y: f1[f1.length-1], | |
| text: ` After: ${{f1[f1.length-1].toFixed(3)}}`, | |
| showarrow: true, arrowhead: 2, | |
| font: {{ color: '#00cc66', size: 13 }}, | |
| bgcolor: '#1a2e1a', bordercolor: '#00cc66', | |
| }}] | |
| }}); | |
| // Chart 2: Detection rate vs FP rate | |
| Plotly.newPlot('chart_det', [ | |
| {{ | |
| x: steps, y: det, mode: 'lines', name: 'Detection Rate', | |
| line: {{ color: '#4a9eff', width: 2 }} | |
| }}, | |
| {{ | |
| x: steps, y: fp, mode: 'lines', name: 'False Positive Rate', | |
| line: {{ color: '#ff9a4a', width: 2, dash: 'dot' }} | |
| }} | |
| ], {{ | |
| ...layout, | |
| title: {{ text: 'Detection Rate vs False Positive Rate', font: {{ color: '#4a9eff', size: 16 }} }}, | |
| yaxis: {{ title: 'Rate', range: [0, 1], gridcolor: '#333' }}, | |
| }}); | |
| </script> | |
| <footer> | |
| NegotiArena | Meta × Scaler OpenEnv Hackathon | Theme 1: Multi-Agent Interactions<br> | |
| CPU training uses REINFORCE with the same reward signal as GPU GRPO training. | |
| </footer> | |
| </body> | |
| </html>""" | |
| with open(output_path, "w", encoding="utf-8") as f: | |
| f.write(html) | |
| return output_path | |
| def main(): | |
| parser = argparse.ArgumentParser() | |
| parser.add_argument("--input", default="checkpoints/cpu_run/reward_curve.csv") | |
| parser.add_argument("--output", default="reward_curves.html") | |
| parser.add_argument("--open", action="store_true", default=True, | |
| help="Open in browser after generating") | |
| args = parser.parse_args() | |
| if not os.path.exists(args.input): | |
| print(f"❌ File not found: {args.input}") | |
| print(" Run training first: python -m training.train_cpu --steps 200") | |
| return | |
| data = load_csv(args.input) | |
| out = generate_html(data, args.output) | |
| print(f"✅ Chart saved to {out}") | |
| if args.open: | |
| webbrowser.open(f"file://{os.path.abspath(out)}") | |
| if __name__ == "__main__": | |
| main() |