File size: 8,133 Bytes
7956103 | 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 | """Publication figures from completed exact/certified phase-three outputs."""
import argparse
from fractions import Fraction
import json
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
BLUE, RED, TEAL, VIOLET, GRAY = "#0F4D92", "#B64342", "#42949E", "#9A4D8E", "#555555"
def val(x):
return float(Fraction(x["rational"] if isinstance(x, dict) else x))
def style():
plt.rcParams.update({"font.family": "DejaVu Sans", "font.size": 19,
"axes.labelsize": 22, "axes.titlesize": 20, "xtick.labelsize": 18,
"ytick.labelsize": 18, "legend.fontsize": 18, "legend.frameon": False,
"axes.spines.top": False, "axes.spines.right": False, "axes.linewidth": 1.5,
"lines.linewidth": 2.8, "lines.markersize": 8, "axes.titlepad": 14,
"pdf.fonttype": 42, "ps.fonttype": 42, "savefig.facecolor": "white"})
def save(fig, folder, name):
folder.mkdir(parents=True, exist_ok=True)
for ext in ("pdf", "png"):
fig.savefig(folder / f"{name}.{ext}", dpi=300, bbox_inches="tight", pad_inches=.16)
plt.close(fig)
def schedules(data, folder):
assert data["status"] == "complete" and data["self_test"]["passed"]
rows = data["rows"]
n = max(r["n"] for r in rows)
specs = [("geometric", "tv_exact", "Geometric, fixed labels", BLUE, "o", "-"),
("equal", "tv_exact", "Equal, fixed labels", RED, "s", "-"),
("best_fixed", "tv_exact", "Best fixed-label batches", TEAL, "^", "-"),
("best_symmetrized", "symmetrized_tv_exact", "Best batches + random labels", VIOLET, "x", "--")]
fig, axes = plt.subplots(1, 2, figsize=(14.2, 7.5))
fig.subplots_adjust(left=.09, right=.985, bottom=.37, top=.87, wspace=.30)
for ax, K in zip(axes, (3, 4)):
group = sorted((r for r in rows if r["n"] == n and r["K"] == K),
key=lambda r: Fraction(r["budget"]))
B = [val(r["budget"]) for r in group]
for key, field, label, color, marker, line in specs:
ax.plot(B, [val(r[key][field]) for r in group], color=color,
marker=marker, ls=line, label=label)
lo = np.array([val(r["all_policy"]["lower"]) for r in group])
hi = np.array([val(r["all_policy"]["upper"]) for r in group])
mid = (lo+hi)/2
ax.errorbar(B, mid, yerr=np.array([mid-lo, hi-mid]), color=GRAY,
fmt="d:", capsize=5, markerfacecolor="white", zorder=6,
label="Certified all-policy optimum")
ax.set(xlabel="Joint-KL budget $B$", ylabel="Full-output TV",
title=rf"$n={n}$ pairs, $K={K}$ rounds", xticks=B)
ax.set_ylim(bottom=0)
h, l = axes[0].get_legend_handles_labels()
fig.legend(h, l, loc="lower center", bbox_to_anchor=(.53,.055), ncol=2,
columnspacing=1.5, handlelength=2.6, fontsize=18)
fig.text(.53, .015, "One frozen model per budget; identical model across round counts.",
ha="center", fontsize=17, color=GRAY)
save(fig, folder, "fixed_schedule_benchmark")
fig, axes = plt.subplots(1, 2, figsize=(13.8, 6.4))
fig.subplots_adjust(left=.10, right=.985, bottom=.27, top=.86, wspace=.30)
for ax, K in zip(axes, (3, 4)):
for n, color, marker in zip(sorted({r["n"] for r in rows}),
(BLUE, RED, TEAL), ("o", "s", "^")):
group = sorted((r for r in rows if r["n"] == n and r["K"] == K),
key=lambda r: Fraction(r["budget"]))
lo = np.array([val(r["geometric_excess_over_optimum_lower"]) for r in group])
hi = np.array([val(r["geometric_excess_over_optimum_upper"]) for r in group])
mid = (lo+hi)/2
ax.errorbar([val(r["budget"]) for r in group], mid,
yerr=np.array([mid-lo,hi-mid]), fmt=marker+"-", color=color,
capsize=5, label=rf"$n={n}$")
ax.axhline(0, color=GRAY, lw=1, ls=":")
ax.set(xlabel="Joint-KL budget $B$", ylabel="TV above all-policy optimum",
title=rf"Geometric fixed labels: $K={K}$", xticks=[.02,.05,.10])
ax.ticklabel_format(axis="y", style="plain", useOffset=False)
h,l = axes[0].get_legend_handles_labels()
fig.legend(h,l,loc="lower center",bbox_to_anchor=(.54,.07),ncol=3,fontsize=20)
fig.text(.54,.015,"Arithmetic enclosures; connecting lines are guides, not fitted laws.",
ha="center",fontsize=17,color=GRAY)
save(fig,folder,"geometric_schedule_overhead")
def biased(data, folder):
cases = [c for c in data["cases"] if c.get("status") == "certified"]
if not cases:
raise ValueError("No certified biased cases available for a result figure")
ns = sorted({c["n"] for c in cases})
fig, axes = plt.subplots(1, 2, figsize=(13.8, 7.2))
fig.subplots_adjust(left=.10,right=.985,bottom=.34,top=.86,wspace=.31)
round_specs = [(2,BLUE,"o"),(3,RED,"s"),(4,TEAL,"^")]
for ax,n in zip(axes,ns[:2]):
group = [c for c in cases if c["n"] == n]
model = next(m for m in data["models"] if m["n"] == n)
biases = sorted({Fraction(c["a"]) for c in data["cases"] if c["n"] == n})
for K,color,marker in round_specs:
if not any(c["K"] == K for c in data["cases"] if c["n"] == n):
continue
found = {Fraction(c["a"]):c for c in group if c["K"] == K}
lo = np.array([val(found[a]["certificate"]["tv_lower"]) if a in found else np.nan for a in biases])
hi = np.array([val(found[a]["certificate"]["tv_upper"]) if a in found else np.nan for a in biases])
mid=(lo+hi)/2
ax.errorbar(list(map(float,biases)),mid,yerr=np.array([mid-lo,hi-mid]),
fmt=marker+"-",color=color,capsize=5,label=rf"$K={K}$")
ax.axhline(val(model["diagnostics"]["model_tv"]),color=GRAY,ls="--",label="Serial model TV")
ax.set(xlabel="Representative bias $a$",ylabel="Certified optimum TV",
title=rf"$n={n}$ pairs",xticks=list(map(float,biases)))
endpoints = [val(model["diagnostics"]["model_tv"])]
endpoints += [val(c["certificate"][key]) for c in group for key in ("tv_lower","tv_upper")]
span = max(endpoints)-min(endpoints)
padding = max(span*.16, .0005)
ax.set_ylim(min(endpoints)-padding,max(endpoints)+padding)
missing = sum(c.get("status") != "certified" for c in data["cases"] if c["n"] == n)
if missing:
ax.text(.04,.04,f"{missing} missing cases (see report)",transform=ax.transAxes,fontsize=17)
if len(ns)==1:
n=ns[0]
group=sorted((m for m in data["models"] if m["n"]==n),key=lambda m:Fraction(m["a"]))
axes[1].plot([float(Fraction(m["a"])) for m in group],
[val(m["diagnostics"]["all_context_head_error"]) for m in group],
color=VIOLET,marker="d")
axes[1].axhline(.25,color=GRAY,ls=":")
axes[1].set(xlabel="Representative bias $a$",ylabel="Worst conditional-head TV",
title="All-context accuracy check",xticks=[.25,.4,.5])
axes[1].set_ylim(bottom=0)
h,l=axes[0].get_legend_handles_labels()
fig.legend(h,l,loc="lower center",bbox_to_anchor=(.54,.14),ncol=4,fontsize=18)
fig.text(.54,.015,"Vertical axes zoomed; n=6 evaluated at K=3,4 only.\nFixed mismatch mixture and KL at each n.\nUniform head errors are measured, not assumed equal across biases.",
ha="center",va="bottom",fontsize=17,color=GRAY,linespacing=1.2)
save(fig,folder,"biased_depth_sensitivity")
if __name__ == "__main__":
parser=argparse.ArgumentParser()
parser.add_argument("--run-dir",type=Path,required=True)
parser.add_argument("--figure-dir",type=Path)
args=parser.parse_args()
style()
output=args.figure_dir or args.run_dir/"figures"
schedules(json.loads((args.run_dir/"schedules.json").read_text()),output)
biased(json.loads((args.run_dir/"biased_results.json").read_text()),output)
print("Exported three PDF/PNG figure pairs.",flush=True)
|