File size: 5,601 Bytes
448d6a5 | 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 | """Generate logbook figures (plotly HTML + raw CSV) from claim result JSONs."""
import os, sys, json
import numpy as np
import plotly.graph_objects as go
BASE = os.path.join(os.path.dirname(__file__), "..")
OUT = os.path.join(BASE, "outputs", "figures")
os.makedirs(OUT, exist_ok=True)
def _save(fig, name):
# embeddable fragment (div + inline script + one cdn plotly load) for logbook cells
fig.write_html(os.path.join(OUT, name + ".html"), include_plotlyjs="cdn", full_html=False,
default_width="100%", default_height="430px")
print("wrote", name + ".html", flush=True)
def load(claim):
p = os.path.join(BASE, "outputs", claim, f"{claim}_results.json")
return json.load(open(p)) if os.path.exists(p) else None
# ---- Claim 1: supernet operator heatmap ----
def fig_claim1():
d = load("claim1")
if not d: return
ops = d["ctrl_ops"]; heat = d["heatmap"]
domains = list(heat.keys())
Z = [[heat[dom][o] for o in ops] for dom in domains]
fig = go.Figure(go.Heatmap(z=Z, x=ops, y=domains, colorscale="Viridis", zmin=0, zmax=1,
colorbar=dict(title="P(include)")))
fig.update_layout(title="Claim 1 — Mind supernet operator inclusion probability by topic domain",
xaxis_title="Thinking operator", yaxis_title="Domain", height=430)
_save(fig, "claim1_supernet_heatmap")
# CSV
with open(os.path.join(OUT, "claim1_supernet_heatmap.csv"), "w") as f:
f.write("domain," + ",".join(ops) + "\n")
for dom in domains:
f.write(dom + "," + ",".join(f"{heat[dom][o]:.4f}" for o in ops) + "\n")
# ---- Claim 2: per-method Overall + MOScores bar ----
def fig_claim2():
d = load("claim2")
if not d: return
methods = [k for k in d if not k.startswith("_")]
rows = [(m, d[m]["agg"]["Overall"], d[m]["agg"]["MOScore_PF"], d[m]["agg"]["MOScore_PS"], d[m]["agg"].get("novelty_mean", float("nan"))) for m in methods]
rows.sort(key=lambda r: r[1])
names = [r[0] for r in rows]
fig = go.Figure()
fig.add_bar(y=names, x=[r[2] for r in rows], name="MOScore PF", orientation="h")
fig.add_bar(y=names, x=[r[3] for r in rows], name="MOScore PS", orientation="h")
fig.add_trace(go.Scatter(y=names, x=[r[1] for r in rows], name="Overall", mode="markers",
marker=dict(size=12, color="black", symbol="diamond")))
fig.update_layout(barmode="group", title="Claim 2 — Win-rate MOScore by method (MindFlow best aggregate)",
xaxis_title="Win-rate score vs expert reference", height=430)
_save(fig, "claim2_methods_bar")
with open(os.path.join(OUT, "claim2_methods.csv"), "w") as f:
f.write("method,Overall,MOScore_PF,MOScore_PS,novelty\n")
for r in sorted(rows, key=lambda x: -x[1]):
f.write(f"{r[0]},{r[1]:.4f},{r[2]:.4f},{r[3]:.4f},{r[4]:.4f}\n")
# ---- Claim 3: learning curves + reward discrimination + distribution shift ----
def fig_claim3():
d = load("claim3")
if not d: return
fig = go.Figure()
colors = {"tournament": "#2ca02c", "pointwise": "#d62728"}
for mode in ["tournament", "pointwise"]:
if mode not in d: continue
h = d[mode]["hist"]
fig.add_trace(go.Scatter(x=h["eval_iter"], y=h["eval_overall"], mode="lines+markers",
name=f"{mode}", line=dict(color=colors.get(mode))))
fig.update_layout(title="Claim 3 — Held-out win-rate MOScore vs controller-optimization iteration",
xaxis_title="Optimization iteration", yaxis_title="Eval Overall MOScore", height=400)
_save(fig, "claim3_learning_curve")
# reward discrimination (mean reward_std per mode)
fig2 = go.Figure()
for mode in ["tournament", "pointwise"]:
if mode not in d: continue
h = d[mode]["hist"]
fig2.add_trace(go.Scatter(x=h["iter"], y=h["reward_std"], mode="lines+markers",
name=f"{mode}", line=dict(color=colors.get(mode))))
fig2.update_layout(title="Claim 3 — Reward signal discrimination (std of intra-group reward)",
xaxis_title="Iteration", yaxis_title="Reward std (higher = less judgment collapse)", height=400)
_save(fig2, "claim3_reward_discrimination")
# distribution shift for tournament
if "tournament" in d:
t = d["tournament"]; ops = t["ctrl_ops"]
init = np.array(t["init_probs"]).mean(axis=0) # mean over layers
fin = np.array(t["final_probs"]).mean(axis=0)
fig3 = go.Figure()
fig3.add_bar(x=ops, y=init, name="initial controller")
fig3.add_bar(x=ops, y=fin, name="after tournament optimization")
fig3.update_layout(barmode="group", title="Claim 3 — Supernet operator-inclusion shift after optimization",
xaxis_title="Operator", yaxis_title="Mean P(include)", height=400)
_save(fig3, "claim3_distribution_shift")
with open(os.path.join(OUT, "claim3_curves.csv"), "w") as f:
f.write("mode,iter,eval_overall\n")
for mode in ["tournament", "pointwise"]:
if mode in d:
h = d[mode]["hist"]
for it, v in zip(h["eval_iter"], h["eval_overall"]):
f.write(f"{mode},{it},{v:.4f}\n")
if __name__ == "__main__":
which = sys.argv[1] if len(sys.argv) > 1 else "all"
if which in ("all", "claim1"): fig_claim1()
if which in ("all", "claim2"): fig_claim2()
if which in ("all", "claim3"): fig_claim3()
print("figures ->", OUT)
|