File size: 5,717 Bytes
ce00c50 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 | """
Plot training curves from a real metrics.json produced by training/train.py.
Usage:
python3 scripts/plot_from_metrics.py [metrics.json] [output.png]
Defaults: ./metrics.json -> ./training_curves.png
The metrics.json schema (see training/metrics.py) is:
{
"step": [...],
"arbitration_accuracy": [...],
"merge_success_rate": [...],
"avg_reward": [...],
"curriculum_phase": [...],
"conflict_detection_rate": [...],
"false_alarm_rate": [...],
"wrong_agent_rate": [...]
}
"""
import json
import sys
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
def rolling_mean(arr, window):
arr = np.asarray(arr, dtype=float)
if len(arr) == 0:
return arr
if len(arr) < window:
return arr
out = np.empty_like(arr)
cumsum = np.cumsum(np.insert(arr, 0, 0))
for i in range(len(arr)):
lo = max(0, i - window + 1)
out[i] = (cumsum[i + 1] - cumsum[lo]) / (i - lo + 1)
return out
def main(metrics_path: str = "metrics.json", out_path: str = "training_curves.png"):
if not Path(metrics_path).exists():
sys.exit(f"metrics.json not found at {metrics_path}")
with open(metrics_path) as f:
h = json.load(f)
steps = np.asarray(h.get("step", []))
if len(steps) == 0:
sys.exit("metrics.json is empty (no step entries yet)")
accs = np.asarray(h.get("arbitration_accuracy", [])) * 100
rewards = np.asarray(h.get("avg_reward", []))
merge = np.asarray(h.get("merge_success_rate", [])) * 100
phases = np.asarray(h.get("curriculum_phase", []))
plt.style.use("dark_background")
fig, axes = plt.subplots(2, 2, figsize=(14, 9))
fig.suptitle(
f"Conflict Arbitration Agent - Training Progress ({len(steps)} steps logged)",
fontsize=14, fontweight="bold", color="#e6e6f0",
)
# 1. Reward
ax = axes[0, 0]
ax.scatter(steps, rewards, alpha=0.3, c="#8be9d6", s=15, label="per-step reward")
if len(rewards) >= 20:
ax.plot(steps, rolling_mean(rewards, 20), color="#ff79c6", linewidth=2.5,
label="rolling avg (window=20)")
ax.axhline(0, color="#444", linestyle="--", linewidth=1, alpha=0.6)
ax.set_title("Average reward over time", color="#e6e6f0")
ax.set_xlabel("Training step")
ax.set_ylabel("Reward")
ax.legend(loc="lower right", framealpha=0.3)
ax.grid(True, alpha=0.15)
# 2. Accuracy
ax = axes[0, 1]
ax.scatter(steps, accs, alpha=0.3, c="#50fa7b", s=15, label="per-step accuracy")
if len(accs) >= 20:
ax.plot(steps, rolling_mean(accs, 20), color="#f1fa8c", linewidth=2.5,
label="rolling avg (window=20)")
ax.axhline(33.3, color="#ff5555", linestyle="--", linewidth=1.5, alpha=0.7,
label="random baseline (33.3%)")
ax.set_title("Arbitration accuracy over time", color="#e6e6f0")
ax.set_xlabel("Training step")
ax.set_ylabel("Accuracy (%)")
ax.set_ylim(-5, 105)
ax.legend(loc="lower right", framealpha=0.3)
ax.grid(True, alpha=0.15)
# 3. Merge success rate
ax = axes[1, 0]
if len(merge) > 0:
ax.scatter(steps, merge, alpha=0.3, c="#bd93f9", s=15, label="per-step")
if len(merge) >= 20:
ax.plot(steps, rolling_mean(merge, 20), color="#ffb86c", linewidth=2.5,
label="rolling avg (window=20)")
ax.set_title("Merge success rate", color="#e6e6f0")
ax.set_xlabel("Training step")
ax.set_ylabel("Success (%)")
ax.set_ylim(-5, 105)
ax.legend(loc="lower right", framealpha=0.3)
ax.grid(True, alpha=0.15)
else:
ax.axis("off")
# 4. Summary stats
ax = axes[1, 1]
ax.axis("off")
n = len(steps)
head = max(1, min(100, n // 4))
tail = max(1, min(100, n // 4))
head_r = float(np.mean(rewards[:head]))
tail_r = float(np.mean(rewards[-tail:]))
head_a = float(np.mean(accs[:head]))
tail_a = float(np.mean(accs[-tail:]))
pos = int((rewards > 0).sum())
above = int((accs > 33.3).sum())
phase_summary = ""
if len(phases) > 0:
unique, counts = np.unique(phases, return_counts=True)
phase_summary = "\nPHASE TIME\n" + "\n".join(
f" Phase {int(p)}: {int(c)} steps ({100*c/n:.0f}%)"
for p, c in zip(unique, counts)
)
text = f"""TRAINING SUMMARY
{'='*40}
Steps logged: {n}
First step / Last: {int(steps[0])} / {int(steps[-1])}
REWARD
First {head} mean: {head_r:+.2f}
Last {tail} mean: {tail_r:+.2f}
Improvement: {tail_r - head_r:+.2f}
Best: {float(rewards.max()):+.2f} (step {int(steps[int(np.argmax(rewards))])})
Positive steps: {pos} / {n} ({100*pos/n:.0f}%)
ACCURACY
First {head} mean: {head_a:.1f}%
Last {tail} mean: {tail_a:.1f}%
Best: {float(accs.max()):.1f}%
Above-chance: {above} / {n} ({100*above/n:.0f}%)
Random baseline: 33.3%
{phase_summary}
"""
ax.text(0.02, 0.98, text, transform=ax.transAxes, fontsize=10,
verticalalignment="top", fontfamily="monospace", color="#c8c8e8")
plt.tight_layout()
plt.savefig(out_path, dpi=150, bbox_inches="tight", facecolor="#0a0a14")
print(f"Saved {out_path}")
print(f"\nFirst-{head} reward: {head_r:+.3f} Last-{tail} reward: {tail_r:+.3f} delta: {tail_r-head_r:+.3f}")
print(f"First-{head} acc: {head_a:.2f}% Last-{tail} acc: {tail_a:.2f}% delta: {tail_a-head_a:+.2f}pp")
if __name__ == "__main__":
metrics = sys.argv[1] if len(sys.argv) > 1 else "metrics.json"
out = sys.argv[2] if len(sys.argv) > 2 else "training_curves.png"
main(metrics, out)
|