Spaces:
Running
Running
File size: 21,568 Bytes
9d65593 ba623bd 9d65593 ba623bd 9d65593 e85f452 9d65593 e85f452 9d65593 7ecfc20 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 | # core/plotter.py
"""
Publication-quality figure generation for Wang's Five Laws.
Standards: Nature / PRL / top-conference level.
Canvas: 18Γ20 inches @ 300 DPI, Arial/Helvetica fonts.
Color system:
Q-related β blue (#2166AC)
K-related β red (#D6604D)
V-related β green (#4DAC26)
QK pair β purple (#762A83)
QV pair β cyan (#01665E)
KV pair β orange (#E08214)
Model A (base) β solid line
Model B (RL) β dashed line
Delta β gray fill
"""
import numpy as np
import pandas as pd
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
from matplotlib.lines import Line2D
import io
import os
# ββ Font & style ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
plt.rcParams.update({
"font.family": "DejaVu Sans", # fallback; Arial not always present
"font.size": 9,
"axes.titlesize": 11,
"axes.labelsize": 10,
"xtick.labelsize": 9,
"ytick.labelsize": 9,
"legend.fontsize": 9,
"figure.dpi": 300,
"savefig.dpi": 300,
"axes.linewidth": 0.8,
"grid.linewidth": 0.4,
"lines.linewidth": 1.5,
"legend.framealpha": 0.85,
"legend.edgecolor": "0.7",
"axes.spines.top": False,
"axes.spines.right": False,
})
# ββ Color palette βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
C = {
"Q": "#2166AC", # blue
"K": "#D6604D", # red
"V": "#4DAC26", # green
"QK": "#762A83", # purple
"QV": "#01665E", # cyan/teal
"KV": "#E08214", # orange
"ref": "#555555", # reference line (gray)
"band_alpha": 0.18,
}
BAND_COLORS = {
"Q": "#2166AC",
"K": "#D6604D",
"QK": "#762A83",
"QV": "#01665E",
"KV": "#E08214",
}
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Data helpers
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _aggregate_by_layer(df: pd.DataFrame, col: str):
"""
Pseudo-bulk two-step aggregation per layer (Nature Comms 2021).
Step 1: median across Q heads within each (layer, kv_head) group.
Step 2: median / q25 / q75 across kv_head groups per layer.
Avoids pseudoreplication bias in GQA models (e.g. 4Q:1K).
Excludes kv_shared rows for KV metrics (theoretical-value bias).
"""
kv_cols = {"ssr_KV", "pearson_KV", "cosU_KV", "cosV_KV", "alpha_KV"}
if col in kv_cols:
df = df[df["kv_shared"] == 0] if "kv_shared" in df.columns else df
layers = np.array(sorted(df["layer"].unique()))
med_vals, q25_vals, q75_vals = [], [], []
for layer in layers:
ldf = df[df["layer"] == layer]
# Step 1: median within each kv_head group
if "kv_head" in ldf.columns:
step1 = ldf.groupby("kv_head")[col].median().values
else:
step1 = ldf[col].dropna().values
step1 = step1[~np.isnan(step1)] if len(step1) > 0 else step1
# Step 2: statistics across kv_head medians
med_vals.append(float(np.median(step1)) if len(step1) > 0 else np.nan)
q25_vals.append(float(np.percentile(step1, 25)) if len(step1) > 0 else np.nan)
q75_vals.append(float(np.percentile(step1, 75)) if len(step1) > 0 else np.nan)
return layers, np.array(med_vals), np.array(q25_vals), np.array(q75_vals)
def _global_layers(df: pd.DataFrame):
"""Return list of layer indices where kv_shared==True (Gemma global layers)."""
if "kv_shared" not in df.columns:
return []
return sorted(df[df["kv_shared"] == 1]["layer"].unique().tolist())
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Single-subplot drawing primitives
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _draw_line(ax, layers, med, q25, q75, color, label, linestyle="-",
show_band=True, global_layers=None):
ax.plot(layers, med, color=color, linestyle=linestyle,
linewidth=1.8, label=label, zorder=3)
if show_band:
ax.fill_between(layers, q25, q75, color=color,
alpha=C["band_alpha"], zorder=2)
if global_layers:
for gl in global_layers:
ax.axvline(gl, color="#AAAAAA", linewidth=0.7,
linestyle=":", zorder=1)
def _add_hline(ax, y, label=None, color=None):
color = color or C["ref"]
ax.axhline(y, color=color, linewidth=1.0, linestyle="--",
alpha=0.75, zorder=1, label=label)
def _finalize_ax(ax, title, ylabel, xlabel="Layer index"):
ax.set_title(title, fontweight="bold", pad=4)
ax.set_ylabel(ylabel)
ax.set_xlabel(xlabel)
ax.grid(True, axis="y", alpha=0.35)
ax.legend(loc="best", handlelength=1.5)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# The 12-panel 4Γ3 figure (single model)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def plot_single_model(
df: pd.DataFrame,
model_name: str,
show_band: bool = True,
head_dim: int = 128,
d_model: int = 5120,
) -> plt.Figure:
"""
4Γ3 grid, 12 subplots.
Row 1 β Law 1 & 2 (singular value metrics):
[0,0] pearson_QK [0,1] ssr_QK [0,2] alpha_QK
Row 2 β Law 3 (condition numbers & max singular values):
[1,0] sigma_max_Q [1,1] sigma_max_K [1,2] cond_Q & cond_K (dual line)
Row 3 β Law 4 (output subspace, left singular vectors U):
[2,0] cosU_QK [2,1] cosU_QV [2,2] cosU_KV
+ random baseline 1/βd_head
Row 4 β Law 5 (input subspace, right singular vectors V):
[3,0] cosV_QK [3,1] cosV_QV [3,2] cosV_KV
+ random baseline 1/βd_model
"""
fig, axes = plt.subplots(4, 3, figsize=(18, 20))
fig.suptitle(
f"Wang's Five Laws β {model_name}",
fontsize=14, fontweight="bold", y=0.995
)
gl = _global_layers(df)
baseline_U = 1.0 / np.sqrt(head_dim)
baseline_V = 1.0 / np.sqrt(d_model)
# ββ helper βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def draw(ax, col, color, label, linestyle="-"):
layers, med, q25, q75 = _aggregate_by_layer(df, col)
_draw_line(ax, layers, med, q25, q75, color, label,
linestyle=linestyle, show_band=show_band,
global_layers=gl)
# ββ Row 0: Law 1 & 2 βββββββββββββββββββββββββββββββββββββββββββββββββββββ
ax = axes[0, 0]
draw(ax, "pearson_QK", C["QK"], "Pearson r (QβK)")
_add_hline(ax, 1.0, "Ideal = 1")
_finalize_ax(ax, "Law 1 β Spectral Linear Alignment",
"Pearson r (Q, K spectra)")
ax = axes[0, 1]
draw(ax, "ssr_QK", C["QK"], "SSR (QβK)")
_add_hline(ax, 0.0, "Ideal = 0")
_finalize_ax(ax, "Law 2 β Spectral Shape Fidelity",
"SSR (QβK normalized)")
ax = axes[0, 2]
draw(ax, "alpha_QK", C["QK"], "Ξ± (QβK)")
_add_hline(ax, 1.0, "Ideal = 1")
_finalize_ax(ax, "Law 1+2 β Scale Factor Ξ± (QβK)",
"Scale factor Ξ±")
# ββ Row 1: Law 3 βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
ax = axes[1, 0]
draw(ax, "sigma_max_Q", C["Q"], "Ο_max (Q)")
_finalize_ax(ax, "Law 3 β Max Singular Value (Q)",
"Ο_max")
ax = axes[1, 1]
draw(ax, "sigma_max_K", C["K"], "Ο_max (K)")
_finalize_ax(ax, "Law 3 β Max Singular Value (K)",
"Ο_max")
ax = axes[1, 2]
draw(ax, "cond_Q", C["Q"], "ΞΊ(Q)")
draw(ax, "cond_K", C["K"], "ΞΊ(K)")
ax.set_yscale("log")
_finalize_ax(ax, "Law 3 β Condition Number ΞΊ (log scale)",
"Condition number ΞΊ (log)")
# ββ Row 2: Law 4 βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Share y-axis across this row
axU = [axes[2, 0], axes[2, 1], axes[2, 2]]
u_data = {}
for col in ["cosU_QK", "cosU_QV", "cosU_KV"]:
_, med, q25, q75 = _aggregate_by_layer(df, col)
u_data[col] = (med, q25, q75)
all_u = np.concatenate([np.concatenate([v[1], v[2]]) for v in u_data.values()])
all_u = all_u[~np.isnan(all_u)]
if len(all_u) > 0:
u_ymin = max(0, np.nanmin(all_u) * 0.92)
u_ymax = np.nanmax(all_u) * 1.08
else:
u_ymin, u_ymax = 0, 0.15
for (col, color, title_suffix), ax in zip(
[("cosU_QK", C["QK"], "QβK"),
("cosU_QV", C["QV"], "QβV"),
("cosU_KV", C["KV"], "KβV")],
axU
):
draw(ax, col, color, f"cosU ({title_suffix})")
_add_hline(ax, baseline_U,
f"Random = 1/βd_h β {baseline_U:.4f}")
ax.set_ylim(u_ymin, u_ymax)
_finalize_ax(ax, f"Law 4 β Output Subspace cosU ({title_suffix})",
"Mean |cos| (left singular vectors)")
# ββ Row 3: Law 5 βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
axV = [axes[3, 0], axes[3, 1], axes[3, 2]]
v_data = {}
for col in ["cosV_QK", "cosV_QV", "cosV_KV"]:
_, med, q25, q75 = _aggregate_by_layer(df, col)
v_data[col] = (med, q25, q75)
all_v = np.concatenate([np.concatenate([v[1], v[2]]) for v in v_data.values()])
all_v = all_v[~np.isnan(all_v)]
if len(all_v) > 0:
v_ymin = max(0, np.nanmin(all_v) * 0.92)
v_ymax = np.nanmax(all_v) * 1.08
else:
v_ymin, v_ymax = 0, 0.05
for (col, color, title_suffix), ax in zip(
[("cosV_QK", C["QK"], "QβK"),
("cosV_QV", C["QV"], "QβV"),
("cosV_KV", C["KV"], "KβV")],
axV
):
draw(ax, col, color, f"cosV ({title_suffix})")
_add_hline(ax, baseline_V,
f"Random = 1/βD β {baseline_V:.4f}")
ax.set_ylim(v_ymin, v_ymax)
_finalize_ax(ax, f"Law 5 β Input Subspace cosV ({title_suffix})",
"Mean |cos| (right singular vectors)")
# ββ Global layer legend βββββββββββββββββββββββββββββββββββββββββββββββββββ
if gl:
fig.text(
0.5, 0.001,
f"Vertical dotted lines mark global (K=V shared) layers: {gl}",
ha="center", fontsize=8, color="#666666"
)
fig.tight_layout(rect=[0, 0.01, 1, 0.995])
return fig
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Two-model comparison figure (same 4Γ3, dual lines + delta subpanels)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def plot_compare_models(
df_a: pd.DataFrame,
df_b: pd.DataFrame,
name_a: str,
name_b: str,
show_band: bool = True,
show_delta: bool = True,
head_dim: int = 128,
d_model: int = 5120,
) -> plt.Figure:
"""
4Γ3 comparison grid.
Each subplot: Model A (solid) vs Model B (dashed).
Delta (B - A) shown as gray fill when show_delta=True.
"""
fig, axes = plt.subplots(4, 3, figsize=(18, 20))
fig.suptitle(
f"Wang's Five Laws β {name_a} vs {name_b}",
fontsize=14, fontweight="bold", y=0.995
)
gl_a = _global_layers(df_a)
gl_b = _global_layers(df_b)
gl = sorted(set(gl_a) | set(gl_b))
baseline_U = 1.0 / np.sqrt(head_dim)
baseline_V = 1.0 / np.sqrt(d_model)
def draw_pair(ax, col, color, label_a, label_b, hline=None, hline_label=None):
"""Draw Model A (solid) and Model B (dashed) on the same axes."""
lay_a, med_a, q25_a, q75_a = _aggregate_by_layer(df_a, col)
lay_b, med_b, q25_b, q75_b = _aggregate_by_layer(df_b, col)
_draw_line(ax, lay_a, med_a, q25_a, q75_a, color, label_a,
linestyle="-", show_band=show_band, global_layers=gl)
_draw_line(ax, lay_b, med_b, q25_b, q75_b, color, label_b,
linestyle="--", show_band=show_band, global_layers=None)
# Delta fill
if show_delta:
common = np.intersect1d(lay_a, lay_b)
if len(common) > 1:
idx_a = np.isin(lay_a, common)
idx_b = np.isin(lay_b, common)
delta = med_b[idx_b] - med_a[idx_a]
pos = np.maximum(delta, 0)
neg = np.minimum(delta, 0)
ax.fill_between(common, 0, pos,
color="#AAAAAA", alpha=0.25, zorder=0)
ax.fill_between(common, 0, neg,
color="#AAAAAA", alpha=0.25, zorder=0)
if hline is not None:
_add_hline(ax, hline, hline_label)
# ββ Row 0 ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
ax = axes[0, 0]
draw_pair(ax, "pearson_QK", C["QK"],
f"{name_a} Pearson r", f"{name_b} Pearson r", hline=1.0, hline_label="Ideal=1")
_finalize_ax(ax, "Law 1 β Spectral Linear Alignment", "Pearson r (Q, K)")
ax = axes[0, 1]
draw_pair(ax, "ssr_QK", C["QK"],
f"{name_a} SSR", f"{name_b} SSR", hline=0.0, hline_label="Ideal=0")
_finalize_ax(ax, "Law 2 β Spectral Shape Fidelity", "SSR (QβK)")
ax = axes[0, 2]
draw_pair(ax, "alpha_QK", C["QK"],
f"{name_a} Ξ±", f"{name_b} Ξ±", hline=1.0, hline_label="Ideal=1")
_finalize_ax(ax, "Law 1+2 β Scale Factor Ξ± (QβK)", "Scale factor Ξ±")
# ββ Row 1 ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
ax = axes[1, 0]
draw_pair(ax, "sigma_max_Q", C["Q"],
f"{name_a} Ο_max(Q)", f"{name_b} Ο_max(Q)")
_finalize_ax(ax, "Law 3 β Max Singular Value (Q)", "Ο_max")
ax = axes[1, 1]
draw_pair(ax, "sigma_max_K", C["K"],
f"{name_a} Ο_max(K)", f"{name_b} Ο_max(K)")
_finalize_ax(ax, "Law 3 β Max Singular Value (K)", "Ο_max")
ax = axes[1, 2]
# cond: draw both Q and K for both models β 4 lines
lay_a, med_a, q25_a, q75_a = _aggregate_by_layer(df_a, "cond_Q")
lay_b, med_b, q25_b, q75_b = _aggregate_by_layer(df_b, "cond_Q")
_draw_line(ax, lay_a, med_a, q25_a, q75_a, C["Q"],
f"{name_a} ΞΊ(Q)", "-", show_band, gl)
_draw_line(ax, lay_b, med_b, q25_b, q75_b, C["Q"],
f"{name_b} ΞΊ(Q)", "--", show_band, None)
lay_a, med_a, q25_a, q75_a = _aggregate_by_layer(df_a, "cond_K")
lay_b, med_b, q25_b, q75_b = _aggregate_by_layer(df_b, "cond_K")
_draw_line(ax, lay_a, med_a, q25_a, q75_a, C["K"],
f"{name_a} ΞΊ(K)", "-", show_band, None)
_draw_line(ax, lay_b, med_b, q25_b, q75_b, C["K"],
f"{name_b} ΞΊ(K)", "--", show_band, None)
ax.set_yscale("log")
_finalize_ax(ax, "Law 3 β Condition Number ΞΊ (log scale)", "Condition number ΞΊ (log)")
# ββ Row 2: Law 4 βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
u_cols = [("cosU_QK", C["QK"], "QβK"),
("cosU_QV", C["QV"], "QβV"),
("cosU_KV", C["KV"], "KβV")]
# Compute shared y range
u_vals = []
for col, _, _ in u_cols:
for df_ in [df_a, df_b]:
_, med, q25, q75 = _aggregate_by_layer(df_, col)
u_vals.extend(q25[~np.isnan(q25)].tolist())
u_vals.extend(q75[~np.isnan(q75)].tolist())
u_ymin = max(0, min(u_vals) * 0.92) if u_vals else 0
u_ymax = (max(u_vals) * 1.08) if u_vals else 0.15
for (col, color, suffix), ax in zip(u_cols, axes[2]):
draw_pair(ax, col, color,
f"{name_a}", f"{name_b}",
hline=baseline_U,
hline_label=f"Random 1/βd_h β {baseline_U:.4f}")
ax.set_ylim(u_ymin, u_ymax)
_finalize_ax(ax, f"Law 4 β cosU ({suffix})",
"Mean |cos| (U)")
# ββ Row 3: Law 5 βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
v_cols = [("cosV_QK", C["QK"], "QβK"),
("cosV_QV", C["QV"], "QβV"),
("cosV_KV", C["KV"], "KβV")]
v_vals = []
for col, _, _ in v_cols:
for df_ in [df_a, df_b]:
_, med, q25, q75 = _aggregate_by_layer(df_, col)
v_vals.extend(q25[~np.isnan(q25)].tolist())
v_vals.extend(q75[~np.isnan(q75)].tolist())
v_ymin = max(0, min(v_vals) * 0.92) if v_vals else 0
v_ymax = (max(v_vals) * 1.08) if v_vals else 0.05
for (col, color, suffix), ax in zip(v_cols, axes[3]):
draw_pair(ax, col, color,
f"{name_a}", f"{name_b}",
hline=baseline_V,
hline_label=f"Random 1/βD β {baseline_V:.4f}")
ax.set_ylim(v_ymin, v_ymax)
_finalize_ax(ax, f"Law 5 β cosV ({suffix})",
"Mean |cos| (V)")
# ββ Legend for line styles ββββββββββββββββββββββββββββββββββββββββββββββββ
solid_patch = Line2D([0], [0], color="#333333", linewidth=1.8,
linestyle="-", label=f"Solid = {name_a}")
dashed_patch = Line2D([0], [0], color="#333333", linewidth=1.8,
linestyle="--", label=f"Dashed = {name_b}")
fig.legend(handles=[solid_patch, dashed_patch],
loc="lower center", ncol=2, fontsize=9,
bbox_to_anchor=(0.5, 0.001))
if gl:
fig.text(
0.5, 0.0045,
f"Vertical dotted lines mark global (K=V shared) layers: {gl}",
ha="center", fontsize=8, color="#666666"
)
fig.tight_layout(rect=[0, 0.015, 1, 0.995])
return fig
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Export helpers
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def save_figure(fig: plt.Figure, base_path: str):
"""
Save figure to PNG (300 dpi), PDF (vector), and SVG (vector).
base_path: path without extension, e.g. "/tmp/wang_laws_gemma"
Returns list of saved file paths.
"""
paths = []
for fmt, kwargs in [
("png", {"dpi": 300, "bbox_inches": "tight"}),
("pdf", {"bbox_inches": "tight"}),
("svg", {"bbox_inches": "tight"}),
]:
p = f"{base_path}.{fmt}"
fig.savefig(p, format=fmt, **kwargs)
paths.append(p)
return paths
def fig_to_png_bytes(fig: plt.Figure) -> bytes:
"""Return PNG bytes for Gradio Image component."""
buf = io.BytesIO()
fig.savefig(buf, format="png", dpi=150, bbox_inches="tight")
buf.seek(0)
return buf.read()
# fig_to_plotly removed β use core/plotter_plotly.py for native Plotly figures. |