File size: 10,465 Bytes
64cdbf6 6b4e832 64cdbf6 6b4e832 64cdbf6 6b4e832 64cdbf6 6b4e832 64cdbf6 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 | """Build the logbook figures from the reproduction results.
Palette: categorical slots 1 (blue) and 6 (orange) from the dataviz reference
palette -- validated with scripts/validate_palette.js in BOTH light and dark
(all checks pass; worst adjacent CVD dE 24.7 protan). Text stays in ink tokens,
never the series colour. Every chart also emits its raw numbers as CSV so the
figures are auditable and machine-readable.
"""
import json, glob, os, sys
import plotly.graph_objects as go
OUT = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "outputs")
FIG = os.path.join(OUT, "figures")
os.makedirs(FIG, exist_ok=True)
# dataviz reference palette, categorical slots 1 and 6
BLUE_L, ORANGE_L = "#2a78d6", "#eb6834"
INK, INK2, GRID = "#0b0b0b", "#52514e", "rgba(0,0,0,0.10)"
def style(fig, title, ytitle, xtitle=""):
fig.update_layout(
title=dict(text=title, font=dict(size=17, color=INK)),
paper_bgcolor="#fcfcfb", plot_bgcolor="#fcfcfb",
font=dict(family="Inter, system-ui, sans-serif", size=13, color=INK2),
yaxis=dict(title=ytitle, gridcolor=GRID, zerolinecolor=GRID,
linecolor=GRID, title_font=dict(color=INK2)),
xaxis=dict(title=xtitle, gridcolor="rgba(0,0,0,0)", linecolor=GRID,
title_font=dict(color=INK2)),
legend=dict(orientation="h", yanchor="bottom", y=1.02, x=0,
font=dict(color=INK2)),
margin=dict(l=60, r=30, t=80, b=55), height=430,
hovermode="x unified",
)
return fig
def save(fig, name, csv_rows, header):
p = os.path.join(FIG, name + ".html")
fig.write_html(p, include_plotlyjs="cdn", full_html=True)
c = os.path.join(FIG, name + ".csv")
with open(c, "w") as f:
f.write(",".join(header) + "\n")
for r in csv_rows:
f.write(",".join(str(x) for x in r) + "\n")
print("wrote", p, "and", c)
return p, c
def load(name):
p = os.path.join(OUT, name)
return json.load(open(p)) if os.path.exists(p) else None
# ---------------------------------------------------------------- Fig 1: speedup
def fig_speedup():
rows = []
for label, fname, paper_sp in [
("Trip Plan\n(CCD-DS, V=4)", "c3_trip_ccd_ds.json", 3.48),
("HumanEval\n(CCD-DS, V=4)", "c4_he_ccd_ds.json", 3.04),
]:
r = load(fname)
if r:
rows.append((label, paper_sp, r["speedup_vs_uniform"]))
if not rows:
return None
fig = go.Figure()
fig.add_bar(name="Paper (Table 1)", x=[r[0] for r in rows], y=[r[1] for r in rows],
marker_color=BLUE_L, marker_line_width=0,
text=[f"{r[1]:.2f}×" for r in rows], textposition="outside",
textfont=dict(color=INK2))
fig.add_bar(name="This reproduction", x=[r[0] for r in rows], y=[r[2] for r in rows],
marker_color=ORANGE_L, marker_line_width=0,
text=[f"{r[2]:.2f}×" for r in rows], textposition="outside",
textfont=dict(color=INK2))
fig.add_hline(y=1.0, line_dash="dot", line_color=INK2,
annotation_text="no speedup (structural ceiling at V=4, d=3)",
annotation_font=dict(color=INK2, size=11))
style(fig, "CCD-DS decoding speedup: reported vs reproduced (Dream-7B)",
"speedup over uniform b_t=1 (×)")
fig.update_layout(barmode="group", bargap=0.3, bargroupgap=0.08)
return save(fig, "fig_speedup", rows, ["config", "paper_speedup", "repro_speedup"])
# ---------------------------------------------------------------- Fig 2: the k law
def fig_k_law():
rows = []
for V in [1, 2, 4, 8, 16]:
r = load(f"c5_abl_V{V}.json")
if r:
k = 256.0 / r["mean_steps"]
rows.append((V, max(1.0, V / 4.0), k, r["mean_steps"], r["score"]))
if not rows:
return None
if not rows:
return None
fig = go.Figure()
fig.add_scatter(name="predicted k = max(1, V/(d+1))", x=[r[0] for r in rows],
y=[r[1] for r in rows], mode="lines+markers",
line=dict(color=BLUE_L, width=2, dash="dash"),
marker=dict(size=9, color=BLUE_L))
fig.add_scatter(name="measured (Dream-7B, Trip City=3)", x=[r[0] for r in rows],
y=[r[2] for r in rows], mode="lines+markers",
line=dict(color=ORANGE_L, width=2), marker=dict(size=9, color=ORANGE_L))
style(fig, "CCD-DS throughput follows k = max(1, V/(d+1)) (d = 3)",
"tokens decoded per step (k)", "buffer width V")
fig.update_layout(hovermode="x unified")
fig.update_xaxes(type="log", tickvals=[r[0] for r in rows],
ticktext=[str(r[0]) for r in rows])
return save(fig, "fig_k_law", rows,
["V", "predicted_k", "measured_k", "mean_steps", "score"])
# ---------------------------------------------------------------- Fig 3: |I^c_t|
def fig_ic_hist():
r = load("c3_trip_ccd_ds.json")
if not r or "ic_sizes" not in r:
return None
from collections import Counter
flat = [v for seq in r["ic_sizes"] for v in seq]
c = Counter(flat)
xs = sorted(c)
rows = [(x, c[x]) for x in xs]
fig = go.Figure()
fig.add_bar(x=[str(x) for x in xs], y=[c[x] for x in xs], marker_color=BLUE_L,
marker_line_width=0, text=[c[x] for x in xs], textposition="outside",
textfont=dict(color=INK2), name="steps")
style(fig, "Intersection size |I<sup>c</sup><sub>t</sub>| at the paper's V=4, d=3 "
"(Dream-7B, Trip Plan)",
"number of decoding steps", "|I^c_t| (0 = fallback to baseline)")
fig.update_layout(showlegend=False, bargap=0.35)
return save(fig, "fig_ic_hist", rows, ["ic_size", "steps"])
# ---------------------------------------------------------------- Fig 4: main table
def fig_scores():
specs = [
("Trip Plan", [("baseline", "c3_trip_baseline.json", 15.10),
("CCD", "c3_trip_ccd.json", 16.93),
("CCD-DS", "c3_trip_ccd_ds.json", 19.01)]),
# prefer the merged n=64 arms (problems 0..63) where both halves exist,
# so the figure matches the n reported in the analysis table
("HumanEval", [("baseline", "c4merged_he_baseline.json", 52.66),
("CCD", "c4merged_he_ccd.json", 57.31),
("CCD-DS", "c4_he_ccd_ds.json", 56.71)]),
]
rows, labels, paper, repro = [], [], [], []
for task, items in specs:
for meth, fname, pv in items:
r = load(fname)
if r:
labels.append(f"{task}<br>{meth}")
paper.append(pv); repro.append(r["score"])
rows.append((task, meth, pv, r["score"], r["n_examples"]))
if not rows:
return None
fig = go.Figure()
fig.add_bar(name="Paper (Table 1)", x=labels, y=paper, marker_color=BLUE_L,
marker_line_width=0, text=[f"{v:.1f}" for v in paper],
textposition="outside", textfont=dict(color=INK2))
fig.add_bar(name="This reproduction", x=labels, y=repro, marker_color=ORANGE_L,
marker_line_width=0, text=[f"{v:.1f}" for v in repro],
textposition="outside", textfont=dict(color=INK2))
style(fig, "Benchmark scores: reported vs reproduced (Dream-7B)", "score")
fig.update_layout(barmode="group", bargap=0.3, bargroupgap=0.08)
return save(fig, "fig_scores", rows, ["task", "method", "paper", "repro", "n"])
def fig_ablation_score():
rows = []
base = load("c5_abl_baseline.json")
for V in [1, 2, 4, 8, 16]:
r = load(f"c5_abl_V{V}.json")
if r:
rows.append((V, r["score"], r["mean_steps"]))
if not rows or not base:
return None
fig = go.Figure()
fig.add_scatter(name="CCD-DS (measured)", x=[r[0] for r in rows], y=[r[1] for r in rows],
mode="lines+markers", line=dict(color=ORANGE_L, width=2),
marker=dict(size=9, color=ORANGE_L))
fig.add_hline(y=base["score"], line_dash="dot", line_color=INK2,
annotation_text=f"baseline {base['score']:.1f}",
annotation_font=dict(color=INK2, size=11))
fig.add_scatter(name="paper's reported peak (buffer 4 → 70%)", x=[4], y=[70.0],
mode="markers", marker=dict(size=15, color=BLUE_L, symbol="star"))
style(fig, "Claim 5: accuracy vs buffer width (Trip City=3, n=40)",
"exact-match score", "buffer width V (d = 3)")
fig.update_xaxes(type="log", tickvals=[r[0] for r in rows],
ticktext=[str(r[0]) for r in rows])
return save(fig, "fig_ablation_score", rows, ["V", "score", "steps"])
def fig_temperature():
import csv as _csv
p = os.path.join(OUT, "temperature.csv")
if not os.path.exists(p):
return None
rows = []
with open(p) as f:
for d in _csv.DictReader(f):
rows.append((float(d["temperature"]), float(d["baseline"]), float(d["ccd_ds"])))
fig = go.Figure()
fig.add_scatter(name="baseline", x=[r[0] for r in rows], y=[r[1] for r in rows],
mode="lines+markers", line=dict(color=BLUE_L, width=2),
marker=dict(size=9, color=BLUE_L))
fig.add_scatter(name="CCD-DS", x=[r[0] for r in rows], y=[r[2] for r in rows],
mode="lines+markers", line=dict(color=ORANGE_L, width=2),
marker=dict(size=9, color=ORANGE_L))
fig.add_vrect(x0=0.05, x1=0.75, fillcolor="rgba(0,0,0,0.05)", line_width=0,
annotation_text="Dream's confidence ranking collapses here (top_p=0.9)",
annotation_position="top left",
annotation_font=dict(color=INK2, size=10))
style(fig, "Claim 6: score vs temperature (HumanEval, n=16, 256 steps, top_p=0.9)",
"pass@1", "sampling temperature")
return save(fig, "fig_temperature", rows, ["temperature", "baseline", "ccd_ds"])
if __name__ == "__main__":
made = []
for fn in (fig_speedup, fig_k_law, fig_ic_hist, fig_scores,
fig_ablation_score, fig_temperature):
try:
r = fn()
if r:
made.append(r[0])
else:
print(f"skip {fn.__name__}: inputs missing")
except Exception as e:
print(f"skip {fn.__name__}: {e}")
print(f"\n{len(made)} figures written to {FIG}")
|