File size: 20,848 Bytes
7f9ee11 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 | # ============================================================================
# INTERNAL ANALYZER: CaptionBERT-8192
#
# Sees inside the model, not just the output. Five diagnostic lenses:
# 1. Spectral trajectories β eigenvalue evolution per layer
# 2. Effective dimensionality β how deeply each input is understood
# 3. Cross-layer divergence β where computation actually happens
# 4. Token influence β which input tokens drive the output
# 5. Neighborhood structure β local geometry at each layer
#
# Usage:
# analyzer = InternalAnalyzer(model, tokenizer)
# report = analyzer.analyze(["girl", "woman", "subtraction", "multiplication"])
# analyzer.print_report(report)
# analyzer.compare(report, "girl", "subtraction")
# ============================================================================
import torch
import torch.nn.functional as F
import numpy as np
from collections import defaultdict
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
class InternalAnalyzer:
def __init__(self, model, tokenizer, max_len=512):
self.model = model.to(DEVICE).eval()
self.tokenizer = tokenizer
self.max_len = max_len
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# CORE: Extract all layer representations
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@torch.no_grad()
def extract_layers(self, texts):
"""Get per-layer mean-pooled representations for each input."""
if isinstance(texts, str):
texts = [texts]
inputs = self.tokenizer(
texts, max_length=self.max_len, padding="max_length",
truncation=True, return_tensors="pt").to(DEVICE)
outputs = self.model(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
output_hidden_states=True)
mask = inputs["attention_mask"].unsqueeze(-1).float()
n_tokens = inputs["attention_mask"].sum(-1)
# Mean-pool each layer
layer_pooled = []
for h in outputs.hidden_states:
pooled = (h * mask).sum(1) / mask.sum(1).clamp(min=1)
layer_pooled.append(pooled.cpu())
return {
"texts": texts,
"layer_pooled": layer_pooled, # list of (B, 384) per layer
"layer_raw": outputs.hidden_states, # tuple of (B, L, 384) per layer
"final_embedding": outputs.last_hidden_state.cpu(), # (B, 768)
"attention_mask": inputs["attention_mask"].cpu(),
"n_tokens": n_tokens.cpu(),
}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 1. SPECTRAL TRAJECTORIES
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def spectral_trajectory(self, data):
"""
Eigenvalue spectrum at each layer for each input.
Shows how the representation's internal structure evolves.
"""
results = []
n_layers = len(data["layer_pooled"])
B = data["layer_pooled"][0].shape[0]
for b in range(B):
trajectory = []
for layer_idx in range(n_layers):
# For single vector: compute singular values of the
# raw token-level representation (before pooling)
h = data["layer_raw"][layer_idx][b].cpu().float() # (L, 384)
mask = data["attention_mask"][b]
n_real = mask.sum().int().item()
h = h[:n_real] # only real tokens
if n_real < 2:
trajectory.append({"spectrum": [], "eff_dim": 0, "entropy": 0})
continue
# SVD of token representations
h_centered = h - h.mean(0, keepdim=True)
try:
S = torch.linalg.svdvals(h_centered)
except Exception:
trajectory.append({"spectrum": [], "eff_dim": 0, "entropy": 0})
continue
# Normalized spectrum
S_norm = S / (S.sum() + 1e-12)
# Effective dimensionality (participation ratio)
eff_dim = (S.sum() ** 2) / (S.pow(2).sum() + 1e-12)
# Spectral entropy
S_pos = S_norm[S_norm > 1e-12]
entropy = -(S_pos * S_pos.log()).sum()
trajectory.append({
"spectrum": S[:20].tolist(), # top 20 singular values
"eff_dim": eff_dim.item(),
"entropy": entropy.item(),
"top1_ratio": (S[0] / (S.sum() + 1e-12)).item(),
})
results.append({
"text": data["texts"][b],
"trajectory": trajectory,
})
return results
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 2. EFFECTIVE DIMENSIONALITY (output space)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def effective_dimensionality(self, data, k_neighbors=50):
"""
Local effective dimensionality around each embedding.
High = rich understanding. Low = surface-level placement.
"""
embeddings = data["final_embedding"].float() # (B, 768)
B = embeddings.shape[0]
if B < k_neighbors + 1:
k_neighbors = max(B - 1, 2)
# Pairwise distances
sim = embeddings @ embeddings.T
results = []
for b in range(B):
# Get k nearest neighbors
sims = sim[b].clone()
sims[b] = -1 # exclude self
_, topk_idx = sims.topk(k_neighbors)
neighbors = embeddings[topk_idx] # (k, 768)
# Local PCA
centered = neighbors - neighbors.mean(0, keepdim=True)
try:
S = torch.linalg.svdvals(centered)
except Exception:
results.append({"eff_dim": 0, "local_variance": 0})
continue
# Participation ratio
eff_dim = (S.sum() ** 2) / (S.pow(2).sum() + 1e-12)
# How fast do eigenvalues decay?
S_norm = S / (S.sum() + 1e-12)
decay_rate = (S_norm[:5].sum() / S_norm.sum()).item()
results.append({
"text": data["texts"][b],
"eff_dim": eff_dim.item(),
"decay_rate": decay_rate, # high = concentrated, low = spread
"local_spread": centered.norm(dim=-1).mean().item(),
})
return results
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 3. CROSS-LAYER DIVERGENCE
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def cross_layer_divergence(self, data):
"""
How much does the representation change between layers?
High change = computation happening. Low change = pass-through.
"""
results = []
n_layers = len(data["layer_pooled"])
B = data["layer_pooled"][0].shape[0]
for b in range(B):
profile = []
for i in range(n_layers - 1):
h_curr = data["layer_pooled"][i][b].float()
h_next = data["layer_pooled"][i + 1][b].float()
# Cosine between consecutive layers
cos = F.cosine_similarity(h_curr.unsqueeze(0),
h_next.unsqueeze(0)).item()
# L2 distance
l2 = (h_next - h_curr).norm().item()
# Direction change (how much the direction rotates)
h_curr_n = F.normalize(h_curr, dim=0)
h_next_n = F.normalize(h_next, dim=0)
angle = torch.acos(torch.clamp(
(h_curr_n * h_next_n).sum(), -1, 1)).item()
profile.append({
"layer": f"{i}β{i+1}",
"cosine": cos,
"l2_shift": l2,
"angle_rad": angle,
})
# Total path length through representation space
total_path = sum(p["l2_shift"] for p in profile)
# Where did most change happen?
max_shift_layer = max(range(len(profile)),
key=lambda i: profile[i]["l2_shift"])
results.append({
"text": data["texts"][b],
"profile": profile,
"total_path": total_path,
"max_shift_layer": max_shift_layer,
"input_output_cos": F.cosine_similarity(
data["layer_pooled"][0][b].unsqueeze(0).float(),
data["layer_pooled"][-1][b].unsqueeze(0).float()
).item(),
})
return results
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 4. TOKEN INFLUENCE (gradient-based)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def token_influence(self, texts):
"""
Which tokens influence the output most?
Uses gradient of output norm w.r.t. input embeddings.
"""
if isinstance(texts, str):
texts = [texts]
results = []
for text in texts:
inputs = self.tokenizer(
[text], max_length=self.max_len, padding="max_length",
truncation=True, return_tensors="pt").to(DEVICE)
# Get embedding layer output with gradients
input_ids = inputs["input_ids"]
attention_mask = inputs["attention_mask"]
n_real = attention_mask.sum().item()
# Hook into embedding
emb = self.model.token_emb(input_ids) + \
self.model.pos_emb(torch.arange(input_ids.shape[1],
device=DEVICE).unsqueeze(0))
emb = self.model.emb_drop(self.model.emb_norm(emb))
emb.retain_grad()
# Forward through encoder
kpm = ~attention_mask.bool()
x = emb
for layer in self.model.encoder.layers:
x = layer(x, src_key_padding_mask=kpm)
# Pool and project
mask = attention_mask.unsqueeze(-1).float()
pooled = (x * mask).sum(1) / mask.sum(1).clamp(min=1)
output = F.normalize(self.model.output_proj(pooled), dim=-1)
# Gradient of output norm w.r.t embeddings
output.sum().backward()
grad = emb.grad[0].cpu()
# Per-token influence = gradient norm
influence = grad.norm(dim=-1)[:int(n_real)] # only real tokens
influence = influence / (influence.sum() + 1e-12) # normalize
# Decode tokens
token_ids = input_ids[0][:int(n_real)].cpu().tolist()
tokens = self.tokenizer.convert_ids_to_tokens(token_ids)
results.append({
"text": text,
"tokens": tokens,
"influence": influence.tolist(),
"top_tokens": sorted(zip(tokens, influence.tolist()),
key=lambda x: -x[1])[:10],
"concentration": (influence.max() / influence.mean()).item(),
})
self.model.zero_grad()
return results
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 5. FULL ANALYSIS
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def analyze(self, texts):
"""Run all analyses on a set of texts."""
if isinstance(texts, str):
texts = [texts]
print(f" Analyzing {len(texts)} inputs...")
data = self.extract_layers(texts)
spectral = self.spectral_trajectory(data)
eff_dim = self.effective_dimensionality(data)
divergence = self.cross_layer_divergence(data)
influence = self.token_influence(texts)
report = {}
for i, text in enumerate(texts):
report[text] = {
"embedding": data["final_embedding"][i],
"n_tokens": data["n_tokens"][i].item(),
"spectral": spectral[i],
"eff_dim": eff_dim[i] if i < len(eff_dim) else {},
"divergence": divergence[i],
"influence": influence[i],
}
return report
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# PRINTING
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def print_report(self, report):
"""Print full analysis report."""
print(f"\n{'='*70}")
print("INTERNAL ANALYSIS REPORT")
print(f"{'='*70}")
# Summary table
print(f"\n {'Text':<25} {'Tokens':>6} {'EffDim':>7} {'Path':>7} "
f"{'MaxShift':>9} {'InOutCos':>8} {'Concentrate':>11}")
print(f" {'-'*75}")
for text, r in report.items():
label = text[:24]
ed = r["eff_dim"].get("eff_dim", 0)
tp = r["divergence"]["total_path"]
ms = r["divergence"]["max_shift_layer"]
ioc = r["divergence"]["input_output_cos"]
conc = r["influence"]["concentration"]
print(f" {label:<25} {r['n_tokens']:>6} {ed:>7.1f} {tp:>7.2f} "
f" layer {ms:>2} {ioc:>7.3f} {conc:>10.1f}")
# Spectral evolution
print(f"\n SPECTRAL TRAJECTORY (effective dim per layer):")
print(f" {'Text':<25}", end="")
n_layers = len(next(iter(report.values()))["spectral"]["trajectory"])
for i in range(n_layers):
print(f" L{i:>2}", end="")
print()
print(f" {'-'*75}")
for text, r in report.items():
label = text[:24]
print(f" {label:<25}", end="")
for step in r["spectral"]["trajectory"]:
ed = step.get("eff_dim", 0)
print(f" {ed:>4.0f}", end="")
print()
# Spectral entropy per layer
print(f"\n SPECTRAL ENTROPY (information content per layer):")
print(f" {'Text':<25}", end="")
for i in range(n_layers):
print(f" L{i:>2}", end="")
print()
print(f" {'-'*75}")
for text, r in report.items():
label = text[:24]
print(f" {label:<25}", end="")
for step in r["spectral"]["trajectory"]:
ent = step.get("entropy", 0)
print(f" {ent:>4.1f}", end="")
print()
# Cross-layer divergence profiles
print(f"\n COMPUTATION PROFILE (L2 shift between layers):")
print(f" {'Text':<25}", end="")
for i in range(n_layers - 1):
print(f" {i}β{i+1:>2}", end="")
print()
print(f" {'-'*75}")
for text, r in report.items():
label = text[:24]
print(f" {label:<25}", end="")
for step in r["divergence"]["profile"]:
print(f" {step['l2_shift']:>4.1f}", end="")
print()
# Token influence for each input
print(f"\n TOKEN INFLUENCE (top contributing tokens):")
for text, r in report.items():
top = r["influence"]["top_tokens"][:5]
tok_str = " ".join(f"{t}={v:.3f}" for t, v in top)
print(f" {text[:40]:<42} {tok_str}")
def compare(self, report, text_a, text_b):
"""Compare internal representations of two specific inputs."""
a = report[text_a]
b = report[text_b]
cos = F.cosine_similarity(
a["embedding"].unsqueeze(0),
b["embedding"].unsqueeze(0)).item()
print(f"\n{'='*70}")
print(f"COMPARISON: '{text_a}' vs '{text_b}'")
print(f"{'='*70}")
print(f" Output cosine: {cos:.4f}")
print(f" Tokens: {a['n_tokens']} vs {b['n_tokens']}")
# Effective dim comparison
ed_a = a["eff_dim"].get("eff_dim", 0)
ed_b = b["eff_dim"].get("eff_dim", 0)
print(f" Effective dim: {ed_a:.1f} vs {ed_b:.1f} (Ξ={abs(ed_a-ed_b):.1f})")
# Path comparison
pa = a["divergence"]["total_path"]
pb = b["divergence"]["total_path"]
print(f" Total path: {pa:.2f} vs {pb:.2f} (Ξ={abs(pa-pb):.2f})")
# Layer-by-layer spectral comparison
print(f"\n Effective dim trajectory:")
print(f" {'Layer':<8} {'A':>8} {'B':>8} {'Ξ':>8}")
traj_a = a["spectral"]["trajectory"]
traj_b = b["spectral"]["trajectory"]
for i in range(len(traj_a)):
ea = traj_a[i].get("eff_dim", 0)
eb = traj_b[i].get("eff_dim", 0)
print(f" L{i:<6} {ea:>8.1f} {eb:>8.1f} {abs(ea-eb):>8.1f}")
# Divergence profile comparison
print(f"\n Computation profile (L2 shift):")
print(f" {'Transition':<10} {'A':>8} {'B':>8} {'Ξ':>8}")
for i in range(len(a["divergence"]["profile"])):
sa = a["divergence"]["profile"][i]["l2_shift"]
sb = b["divergence"]["profile"][i]["l2_shift"]
label = a["divergence"]["profile"][i]["layer"]
print(f" {label:<10} {sa:>8.2f} {sb:>8.2f} {abs(sa-sb):>8.2f}")
# Token influence comparison
print(f"\n Top tokens:")
print(f" A: {' '.join(f'{t}={v:.3f}' for t,v in a['influence']['top_tokens'][:5])}")
print(f" B: {' '.join(f'{t}={v:.3f}' for t,v in b['influence']['top_tokens'][:5])}")
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# RUN
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
if __name__ == "__main__":
from transformers import AutoModel, AutoTokenizer
REPO_ID = "AbstractPhil/geolip-captionbert-8192"
print("Loading model...")
model = AutoModel.from_pretrained(REPO_ID, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(REPO_ID)
analyzer = InternalAnalyzer(model, tokenizer)
# Test words spanning known-domain and unknown-domain
test_words = [
# Known domain (captions)
"girl",
"woman",
"dog",
"sunset",
"painting",
# Unknown domain (abstract)
"subtraction",
"multiplication",
"prophetic",
"differential",
"adjacency",
# Phrases
"a girl sitting near a window",
"a dog playing on the beach",
"the differential equation of motion",
]
report = analyzer.analyze(test_words)
analyzer.print_report(report)
# Direct comparisons
analyzer.compare(report, "girl", "woman")
analyzer.compare(report, "girl", "subtraction")
analyzer.compare(report, "a girl sitting near a window",
"the differential equation of motion")
print(f"\n{'='*70}")
print("DONE")
print(f"{'='*70}") |