File size: 13,178 Bytes
3404d44 | 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 | #!/usr/bin/env python3
"""Compute and plot VD/HD/HV entanglement metrics across all layers for all scales.
Definitions (values come from the delta-similarity CSV heatmaps, where each cell is
the cosine similarity between the mean delta vector of row-category and col-category):
VD-entanglement = 1/4 * (mean(above-far, below-close) - mean(above-close, below-far))
HD-entanglement = 1/4 * (mean(left-far, right-close) - mean(left-close, right-far))
HV-entanglement = 1/4 * (mean(left-above, right-below) - mean(left-below, right-above))
Note: 'below' is stored as 'under' in the CSV. The script handles both transparently.
Positive value = the two axes are more entangled in the "expected" direction
(e.g. aboveβfar, leftβabove) than in the "unexpected" direction.
Single directory: color by scale (vanilla=blue, 80k=orange, β¦)
Multiple directories: color by model family, linestyle by scale
Usage (single dir):
python plot_entanglement.py results_short_answer/molmo
python plot_entanglement.py results_short_answer/molmo --subset both_correct
Usage (multiple dirs β compare families):
python plot_entanglement.py results_short_answer/molmo results_short_answer/nvila results_short_answer/qwen
python plot_entanglement.py results_short_answer/molmo results_short_answer/nvila --out-dir /tmp/compare
"""
import argparse
import re
from pathlib import Path
from collections import defaultdict
import numpy as np
import pandas as pd
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
# ββ Scale ordering and colors (kept in sync with swap_analysis.py) ββββββββββββ
SCALE_ORDER = [
'vanilla', '80k', '80k-5pct', '80k-10pct', '80k-20pct', '80k-30pct',
'400k', '400k-5pct', '800k', '800k-5pct', '2m', 'roborefer',
'molmo2', 'qwen3_32b', 'qwen3_235b',
]
# Used in single-dir mode (one color per scale)
SCALE_COLORS = {
'vanilla': '#1f77b4', '80k': '#ff7f0e', '400k': '#2ca02c',
'800k': '#d62728', '2m': '#9467bd', 'roborefer': '#8c564b',
'molmo2': '#17becf', 'qwen3_32b': '#bcbd22', 'qwen3_235b':'#e377c2',
'80k-5pct': '#b2dfdb', '80k-10pct': '#00b894', '80k-20pct': '#00897b',
'80k-30pct': '#004d40', '400k-5pct': '#66bb6a', '800k-5pct': '#ef9a9a',
}
SCALE_DISPLAY_NAMES = {
'80k-5pct': '80k 5%', '80k-10pct': '80k 10%',
'80k-20pct': '80k 20%', '80k-30pct': '80k 30%',
'400k-5pct': '400k 5%', '800k-5pct': '800k 5%',
}
# Used in multi-dir mode (one color per model family, one linestyle per scale)
FAMILY_COLOR_CYCLE = [
'#1f77b4', # blue
'#d62728', # red
'#2ca02c', # green
'#ff7f0e', # orange
'#9467bd', # purple
'#8c564b', # brown
'#e377c2', # pink
'#17becf', # cyan
'#bcbd22', # yellow-green
]
SCALE_LINESTYLE_CYCLE = [
'solid',
'dashed',
'dotted',
'dashdot',
(0, (5, 1)), # long dash
(0, (3, 1, 1, 1)), # dash-dot-dot
(0, (1, 1)), # dotted dense
(0, (5, 5)), # long dash sparse
]
# ββ CSV helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
_CSV_RE = re.compile(r'^delta_similarity_(.+)_L(\d+)_(all_pairs|both_correct)\.csv$')
def _loc(df: pd.DataFrame, row: str, col: str) -> float:
"""Look up (row, col) with 'under' β 'below' aliasing."""
aliases = {'below': 'under', 'under': 'below'}
r = row if row in df.index else aliases.get(row, row)
c = col if col in df.columns else aliases.get(col, col)
if r not in df.index or c not in df.columns:
return float('nan')
return float(df.loc[r, c])
def compute_entanglement(df: pd.DataFrame) -> dict:
"""Compute VD, HD, HV entanglement from a 6Γ6 delta-similarity DataFrame.
Each metric is the difference of two means of cosine similarities (range [-2, 2]),
divided by 4 to normalise to [-0.5, 0.5].
"""
vd = (_loc(df, 'above', 'far') + _loc(df, 'below', 'close') -
_loc(df, 'above', 'close') - _loc(df, 'below', 'far')) / 4
hd = (_loc(df, 'left', 'far') + _loc(df, 'right', 'close') -
_loc(df, 'left', 'close') - _loc(df, 'right', 'far')) / 4
hv = (_loc(df, 'left', 'above') + _loc(df, 'right', 'below') -
_loc(df, 'left', 'below') - _loc(df, 'right', 'above')) / 4
return {'VD': vd, 'HD': hd, 'HV': hv}
def load_entanglement(csv_dir: Path, subset: str) -> dict:
"""
Returns:
{scale: {layer_int: {'VD': float, 'HD': float, 'HV': float}}}
"""
data = defaultdict(dict)
for fname in sorted(csv_dir.iterdir()):
m = _CSV_RE.match(fname.name)
if not m:
continue
scale, layer_str, file_subset = m.group(1), m.group(2), m.group(3)
if file_subset != subset:
continue
layer = int(layer_str)
try:
df = pd.read_csv(fname, index_col=0)
except Exception as e:
print(f" [warn] Could not read {fname.name}: {e}")
continue
data[scale][layer] = compute_entanglement(df)
return dict(data)
def _scale_sort_key(s):
return SCALE_ORDER.index(s) if s in SCALE_ORDER else 99
# ββ Plotting ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
METRICS = [
('VD', 'Vertical-Distance Entanglement\nmean(above-far, below-close) β mean(above-close, below-far)'),
('HD', 'Horizontal-Distance Entanglement\nmean(left-far, right-close) β mean(left-close, right-far)'),
# ('HV', 'HV-Entanglement\nmean(left-above, right-below) β mean(left-below, right-above)'),
]
def plot_entanglement_single(scale_data: dict, model_type: str,
subset: str, save_path: Path):
"""Single directory: color by scale."""
fig, axes = plt.subplots(1, 2, figsize=(12, 5))
for ax, (metric_key, metric_label) in zip(axes, METRICS):
for scale in SCALE_ORDER:
if scale not in scale_data:
continue
layer_dict = scale_data[scale]
layers = sorted(layer_dict.keys())
vals = [layer_dict[l][metric_key] for l in layers]
if not any(np.isfinite(v) for v in vals):
continue
ax.plot(
layers, vals, '-',
color=SCALE_COLORS.get(scale, 'gray'),
label=SCALE_DISPLAY_NAMES.get(scale, scale),
linewidth=2,
)
_style_ax(ax, metric_label)
tag = 'Both-Correct' if subset == 'both_correct' else 'All Pairs'
fig.suptitle(
f'{model_type.upper()} β Entanglement Metrics Across Layers [{tag}]',
fontsize=13, fontweight='bold',
)
_save(fig, save_path)
def plot_entanglement_multi(family_data: dict, subset: str, save_path: Path):
"""Multiple directories: color by family, linestyle by scale."""
# Collect all scales across all families (in canonical order)
all_scales = sorted(
{s for scales in family_data.values() for s in scales},
key=_scale_sort_key,
)
families = list(family_data.keys()) # preserve insertion order
# Assign colors to families and linestyles to scales
family_color = {f: FAMILY_COLOR_CYCLE[i % len(FAMILY_COLOR_CYCLE)]
for i, f in enumerate(families)}
scale_ls = {s: SCALE_LINESTYLE_CYCLE[i % len(SCALE_LINESTYLE_CYCLE)]
for i, s in enumerate(all_scales)}
fig, axes = plt.subplots(1, 2, figsize=(12, 5))
for ax, (metric_key, metric_label) in zip(axes, METRICS):
for family in families:
scale_data = family_data[family]
color = family_color[family]
for scale in all_scales:
if scale not in scale_data:
continue
layer_dict = scale_data[scale]
layers = sorted(layer_dict.keys())
vals = [layer_dict[l][metric_key] for l in layers]
if not any(np.isfinite(v) for v in vals):
continue
scale_disp = SCALE_DISPLAY_NAMES.get(scale, scale)
ax.plot(
layers, vals,
color=color,
linestyle=scale_ls[scale],
label=f'{family} {scale_disp}',
linewidth=2,
)
_style_ax(ax, metric_label)
tag = 'Both-Correct' if subset == 'both_correct' else 'All Pairs'
title_families = ' vs '.join(f.upper() for f in families)
fig.suptitle(
f'{title_families} β Entanglement Metrics Across Layers [{tag}]',
fontsize=13, fontweight='bold',
)
_save(fig, save_path)
def _style_ax(ax, title: str):
ax.axhline(y=0, color='gray', linestyle='--', alpha=0.5, linewidth=1)
ax.set_xlabel('Layer Index', fontsize=11)
ax.set_ylabel('Entanglement', fontsize=11)
ax.set_ylim(-1, 1)
ax.set_title(title, fontsize=10, fontweight='bold')
ax.legend(fontsize=9)
ax.grid(True, alpha=0.3)
def _save(fig, save_path: Path):
plt.tight_layout()
save_path.parent.mkdir(parents=True, exist_ok=True)
plt.savefig(save_path, dpi=300, bbox_inches='tight')
plt.close()
print(f"Saved: {save_path}")
# ββ Main ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main():
parser = argparse.ArgumentParser(
description='Plot VD/HD/HV entanglement metrics from saved delta-similarity CSVs.',
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog=__doc__,
)
parser.add_argument('results_dirs', nargs='+', type=str,
help='One or more results directories '
'(e.g. results_short_answer/molmo results_short_answer/nvila)')
parser.add_argument('--subset', choices=['all_pairs', 'both_correct'], default='all_pairs',
help='Which CSV subset to use (default: all_pairs)')
parser.add_argument('--scales', nargs='+', default=None,
help='Restrict to these scales (default: all found)')
parser.add_argument('--out-dir', type=str, default=None,
help='Output directory. Single dir default: {results_dir}/plots/entanglement/. '
'Multi dir default: {common_parent}/entanglement_compare/')
args = parser.parse_args()
# Resolve and validate all directories
dirs = []
for p in args.results_dirs:
d = Path(p).resolve()
if not d.is_dir():
parser.error(f'Directory not found: {d}')
csv_dir = d / 'csv'
if not csv_dir.is_dir():
parser.error(f'No csv/ subdirectory in: {d}')
dirs.append(d)
multi = len(dirs) > 1
subset = args.subset
tag = 'Both-Correct' if subset == 'both_correct' else 'All Pairs'
# Determine output path
if args.out_dir:
out_dir = Path(args.out_dir)
elif multi:
common = dirs[0].parent
out_dir = common / 'entanglement_compare'
else:
out_dir = dirs[0] / 'plots' / 'entanglement'
print(f"Subset : {subset}")
print(f"Output dir : {out_dir}")
print()
# Load data from all directories
family_data = {} # {model_type: {scale: {layer: entanglement}}}
for d in dirs:
model_type = d.name
scale_data = load_entanglement(d / 'csv', subset)
if not scale_data:
print(f"[warn] No matching CSVs in {d}/csv β skipping")
continue
if args.scales:
scale_data = {s: v for s, v in scale_data.items() if s in args.scales}
found = sorted(scale_data.keys(), key=_scale_sort_key)
print(f" {model_type}: {len(found)} scales β {found}")
for s in found:
layers = sorted(scale_data[s].keys())
deepest = layers[-1]
e = scale_data[s][deepest]
vd = f"{e['VD']:>7.4f}" if np.isfinite(e['VD']) else ' nan'
hd = f"{e['HD']:>7.4f}" if np.isfinite(e['HD']) else ' nan'
hv = f"{e['HV']:>7.4f}" if np.isfinite(e['HV']) else ' nan'
print(f" {s:<15} L{deepest:>2} VD={vd} HD={hd} HV={hv}")
family_data[model_type] = scale_data
if not family_data:
print("[error] No data loaded from any directory.")
return
print()
if multi:
families_tag = '_'.join(family_data.keys())
save_path = out_dir / f'entanglement_{families_tag}_{subset}.png'
plot_entanglement_multi(family_data, subset, save_path)
else:
model_type = list(family_data.keys())[0]
save_path = out_dir / f'entanglement_{subset}.png'
plot_entanglement_single(family_data[model_type], model_type, subset, save_path)
if __name__ == '__main__':
main()
|