File size: 19,655 Bytes
d425e71 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 |
"""Instance-based k-NN extension for VLM concept analysis.
This module extends the existing VLM concept analysis with nearest-neighbor
prototype-based classification. It reuses the existing functions and adds
instance-based readout capabilities.
"""
from __future__ import annotations
from collections import defaultdict
from typing import Any, Optional
import numpy as np
# Import from the existing analysis module
from pca import (analyze_concept_trends, cosine_similarity_numpy,
extract_concept_from_filename, group_tensors_by_concept,
load_tensors_by_layer)
def _build_normalized_prototype_bank(
concept_tensors: list[tuple[np.ndarray, Any, int, str]]
) -> tuple[Optional[np.ndarray], list[dict[str, Any]]]:
"""Build an (N,d) bank of L2-normalized prototype vectors and metadata.
Args:
concept_tensors: List of tuples (vec, label, row_id, image_path)
Returns:
Tuple of (X matrix (N,d), meta list of dicts with concept/row_id/image_path)
"""
X_list, meta = [], []
for vec, label, row_id, image_path in concept_tensors:
if vec is None:
continue
norm = np.linalg.norm(vec)
if not np.isfinite(norm) or norm == 0:
continue
X_list.append(vec / norm)
meta.append({
'concept': extract_concept_from_filename(image_path),
'row_id': row_id,
'image_path': image_path,
'label': label
})
if not X_list:
return None, []
X = np.vstack(X_list)
return X, meta
def _nearest_prototypes(
target_vec: np.ndarray,
X_bank: Optional[np.ndarray],
meta: list[dict[str, Any]],
topk: int = 5
) -> list[dict[str, Any]]:
"""Compute cosine similarities target vs all prototypes (already normalized).
Args:
target_vec: Target vector (d,), will be L2-normalized here
X_bank: Prototype bank matrix (N, d), already normalized
meta: List of metadata dicts for each prototype
topk: Number of top neighbors to return
Returns:
Top list of dicts sorted by similarity with keys:
['concept', 'row_id', 'image_path', 'label', 'sim']
"""
if X_bank is None or len(meta) == 0:
return []
# L2-normalize target
t = target_vec
t_norm = np.linalg.norm(t)
if not np.isfinite(t_norm) or t_norm == 0:
return []
t = t / t_norm
sims = X_bank @ t # cosine since both normalized
k = min(topk, sims.shape[0])
# argpartition is O(N); then sort the small top-k slice
idx = np.argpartition(-sims, k - 1)[:k]
idx = idx[np.argsort(-sims[idx])]
out = []
for i in idx:
m = meta[i]
out.append({
'concept': m['concept'],
'row_id': m['row_id'],
'image_path': m['image_path'],
'label': m['label'],
'sim': float(sims[i]),
})
return out
def _knn_weighted_vote(
neighbors: list[dict[str, Any]],
p: float = 1.0
) -> tuple[Optional[str], dict[str, float]]:
"""Weighted majority vote over top-k neighbors.
Args:
neighbors: List of neighbor dicts with 'concept' and 'sim' keys
p: Power for weighting (weight = sim^p, negatives clipped to 0)
Returns:
Tuple of (winner_concept, score_dict)
"""
wsum = defaultdict(float)
for nb in neighbors:
w = max(0.0, nb['sim']) ** p
wsum[nb['concept']] += w
if not wsum:
return None, {}
winner = max(wsum.items(), key=lambda kv: kv[1])[0]
return winner, dict(wsum)
def analyze_target_vs_concepts_with_knn(
target_tensors: list[tuple[np.ndarray, Any, int, str]],
concept_tensors: list[tuple[np.ndarray, Any, int, str]],
layer_name: str,
knn_topk: int = 5,
knn_power: float = 1.0
) -> list[dict[str, Any]]:
"""Analyze similarity between targets and concepts with k-NN instance-based prediction.
Keeps existing per-prototype stats and centroid metrics.
Adds instance-based nearest-neighbor prediction (1-NN + k-NN vote).
Args:
target_tensors: List of target tensor data
concept_tensors: List of concept tensor data
layer_name: Name of the current layer
knn_topk: Number of nearest neighbors to consider
knn_power: Power for weighted voting (weight = sim^p)
Returns:
List of analysis results with added 'instance_knn' section
"""
# Group by concept (existing behavior)
concept_groups = group_tensors_by_concept(concept_tensors)
print(f'Found {len(concept_groups)} concepts: {list(concept_groups.keys())}')
for concept, tensors in concept_groups.items():
print(f' {concept}: {len(tensors)} images')
# Precompute centroids (as before)
concept_centroids = {}
for concept_name, tensor_list in concept_groups.items():
vecs = [t[0] for t in tensor_list]
if len(vecs) > 0:
concept_centroids[concept_name] = np.mean(np.vstack(vecs), axis=0)
else:
concept_centroids[concept_name] = None
# NEW: build prototype bank once for this layer
X_bank, bank_meta = _build_normalized_prototype_bank(concept_tensors)
if X_bank is None:
print('Warning: prototype bank is empty for this layer; skipping instance-NN.')
results = []
for target_data in target_tensors:
target_vec, target_label, target_row_id, target_image_filename = target_data
target_result = {
'layer': layer_name,
'target_row_id': target_row_id,
'target_label': target_label,
'target_image_filename': target_image_filename,
'concept_analysis': {}, # existing per-concept stats live here
'instance_knn': {} # NEW: instance-based readout lives here
}
# --- Existing per-concept stats (unchanged) ---
for concept_name, concept_tensor_list in concept_groups.items():
similarities = []
for concept_data in concept_tensor_list:
concept_vec, concept_label, concept_row_id, concept_image_filename = concept_data
if target_vec.shape != concept_vec.shape:
print(f'Warning: Shape mismatch between target {target_row_id} and concept {concept_row_id}')
continue
sim = cosine_similarity_numpy(target_vec, concept_vec)
similarities.append(sim)
concept_stats = {}
if similarities:
similarities = np.array(similarities)
distances = 1.0 - similarities
concept_stats.update({
'min_similarity': float(np.min(similarities)),
'max_similarity': float(np.max(similarities)),
'mean_similarity': float(np.mean(similarities)),
'min_distance': float(np.min(distances)),
'mean_distance': float(np.mean(distances)),
'num_comparisons': int(len(similarities)),
})
centroid = concept_centroids.get(concept_name, None)
if centroid is not None and centroid.shape == target_vec.shape:
cen_sim = cosine_similarity_numpy(target_vec, centroid)
cen_ang = float(np.degrees(np.arccos(np.clip(cen_sim, -1.0, 1.0))))
concept_stats.update({
'centroid_similarity': float(cen_sim),
'centroid_angular_deg': cen_ang
})
if concept_stats:
target_result['concept_analysis'][concept_name] = concept_stats
# --- NEW: instance-based nearest neighbor prediction ---
if X_bank is not None:
nbs = _nearest_prototypes(target_vec, X_bank, bank_meta, topk=knn_topk)
winner_1nn = nbs[0]['concept'] if nbs else None
voted, vote_scores = _knn_weighted_vote(nbs, p=knn_power) if nbs else (None, {})
target_result['instance_knn'] = {
'top1_concept': winner_1nn,
'top1_similarity': nbs[0]['sim'] if nbs else None,
'topk_neighbors': nbs, # list with concept,row_id,image_path,sim
'topk_voted_concept': voted, # weighted by sim^p over topk (non-negative)
'vote_scores': vote_scores, # dict concept->weight
'topk': knn_topk,
'vote_power': knn_power
}
results.append(target_result)
target_display = target_image_filename if target_image_filename else f'Target_{target_row_id}'
print(f'Analyzed {target_display} against {len(concept_groups)} concepts')
return results
def concept_similarity_analysis_with_knn(
target_db_path: str,
concept_db_path: str,
layer_names: Optional[list[str]] = None,
n_pca_components: Optional[int] = None,
knn_topk: int = 5,
knn_power: float = 1.0,
device: str = 'cpu'
) -> dict[str, dict[str, Any]]:
"""Main function for concept-based similarity analysis with k-NN prediction.
Args:
target_db_path: Path to target images database
concept_db_path: Path to concept images database
layer_names: List of layer names to analyze (None for all common layers)
n_pca_components: Number of PCA components (None to skip PCA)
knn_topk: Number of nearest neighbors for k-NN prediction
knn_power: Power for weighted voting in k-NN
device: PyTorch device
Returns:
Dictionary of analysis results by layer with k-NN predictions
"""
print('Starting concept-based similarity analysis with k-NN...')
print(f'Target DB: {target_db_path}')
print(f'Concept DB: {concept_db_path}')
print(f'PCA components: {n_pca_components}')
print(f'k-NN parameters: topk={knn_topk}, power={knn_power}')
# Load tensors from both databases (reuse existing function)
print(f'\nLoading tensors from {target_db_path}...')
target_tensors = load_tensors_by_layer(target_db_path, device)
print(f'Loading tensors from {concept_db_path}...')
concept_tensors = load_tensors_by_layer(concept_db_path, device)
# Find common layers
common_layers = set(target_tensors.keys()) & set(concept_tensors.keys())
print(f'\nFound {len(common_layers)} common layers: {sorted(common_layers)}')
if not common_layers:
print('No common layers found between databases!')
return {}
# Determine which layers to analyze
if layer_names is None:
layers_to_analyze = sorted(common_layers)
print('Analyzing all common layers')
else:
if isinstance(layer_names, str):
layer_names = [layer_names]
layers_to_analyze = [layer for layer in layer_names if layer in common_layers]
print(f'Analyzing specified layers: {layers_to_analyze}')
# Warn about missing layers
missing_layers = set(layer_names) - common_layers
if missing_layers:
print(f'Warning: Requested layers not found: {missing_layers}')
if not layers_to_analyze:
print('No valid layers to analyze!')
return {}
all_results = {}
# Process each layer
for layer in layers_to_analyze:
print(f'\n{"=" * 50}')
print(f'Processing Layer: {layer}')
print(f'{"=" * 50}')
target_layer_tensors = target_tensors[layer]
concept_layer_tensors = concept_tensors[layer]
print(f'Target tensors: {len(target_layer_tensors)}')
print(f'Concept tensors: {len(concept_layer_tensors)}')
# Apply PCA if requested (reuse existing function)
if n_pca_components is not None:
# Import the PCA function
from pca import apply_pca_to_layer
target_layer_tensors, concept_layer_tensors, pca_model = apply_pca_to_layer(
target_layer_tensors, concept_layer_tensors, n_pca_components
)
else:
pca_model = None
# Analyze similarities with k-NN
layer_results = analyze_target_vs_concepts_with_knn(
target_layer_tensors, concept_layer_tensors, layer,
knn_topk=knn_topk, knn_power=knn_power
)
all_results[layer] = {
'results': layer_results,
'pca_model': pca_model,
'n_pca_components': n_pca_components,
'knn_topk': knn_topk,
'knn_power': knn_power
}
# Print layer summary
if layer_results:
print(f"\nLayer \'{layer}\' Summary:")
print(f' Analyzed {len(layer_results)} target images')
# Get all concept names from first result
if layer_results[0]['concept_analysis']:
concept_names = list(layer_results[0]['concept_analysis'].keys())
print(f' Against {len(concept_names)} concepts: {concept_names}')
# Print k-NN summary
knn_predictions = []
for result in layer_results:
ik = result.get('instance_knn', {})
if ik.get('top1_concept'):
knn_predictions.append(ik['top1_concept'])
if knn_predictions:
from collections import Counter
pred_counts = Counter(knn_predictions)
print(f' k-NN Predictions: {dict(pred_counts)}')
return all_results
def save_knn_analysis_results(
results: dict[str, dict[str, Any]],
output_file: str = 'output/knn_similarity_analysis.txt'
) -> None:
"""Save k-NN analysis results to a text file.
Args:
results: Dictionary of analysis results by layer
output_file: Output filename
"""
import os
os.makedirs(os.path.dirname(output_file), exist_ok=True)
with open(output_file, 'w') as f:
f.write('VLM Concept Analysis with Instance-based k-NN Prediction\n')
f.write('=' * 60 + '\n\n')
for layer, layer_data in results.items():
layer_results = layer_data['results']
n_pca_components = layer_data['n_pca_components']
knn_topk = layer_data.get('knn_topk', 5)
knn_power = layer_data.get('knn_power', 1.0)
f.write(f'Layer: {layer}\n')
if n_pca_components:
f.write(f'PCA Components: {n_pca_components}\n')
f.write(f'k-NN Parameters: topk={knn_topk}, power={knn_power}\n')
f.write('-' * 40 + '\n\n')
for result in layer_results:
target_display = result['target_image_filename'] or f'Target_{result["target_row_id"]}'
f.write(f'Target: {target_display}\n')
# k-NN predictions
ik = result.get('instance_knn', {})
if ik:
f.write(f' 1-NN Concept: {ik.get("top1_concept")} (sim={ik.get("top1_similarity", 0):.4f})\n')
if ik.get('topk_voted_concept') is not None and ik.get('topk', 1) > 1:
f.write(f' k-NN Vote (k={ik["topk"]}, p={ik["vote_power"]}): {ik["topk_voted_concept"]}\n')
# Show top neighbors
neighbors = ik.get('topk_neighbors', [])
if neighbors:
f.write(' Top Neighbors:\n')
for i, nb in enumerate(neighbors[:3], 1): # Show top 3
f.write(f' {i}. {nb["concept"]} (sim={nb["sim"]:.4f})\n')
# Original concept analysis
for concept_name, stats in result['concept_analysis'].items():
f.write(f' Concept {concept_name}:\n')
if 'centroid_similarity' in stats:
f.write(f' Centroid Similarity: {stats["centroid_similarity"]:.4f}\n')
if 'mean_similarity' in stats:
f.write(f' Mean Similarity: {stats["mean_similarity"]:.4f}\n')
f.write('\n')
f.write('\n')
print(f'k-NN results saved to {output_file}')
def analyze_knn_accuracy(
results: dict[str, dict[str, Any]],
ground_truth_concept_extractor: Optional[callable] = None
) -> None:
"""Analyze k-NN prediction accuracy if ground truth is available.
Args:
results: Dictionary of analysis results by layer
ground_truth_concept_extractor: Function to extract true concept from target filename
"""
if ground_truth_concept_extractor is None:
ground_truth_concept_extractor = extract_concept_from_filename
print(f'\n{"=" * 50}')
print('k-NN PREDICTION ACCURACY ANALYSIS')
print(f'{"=" * 50}')
for layer, layer_data in results.items():
layer_results = layer_data['results']
knn_topk = layer_data.get('knn_topk', 5)
print(f'\nLayer: {layer}')
print('-' * 30)
if not layer_results:
print('No results for this layer')
continue
correct_1nn = 0
correct_knn = 0
total = 0
for result in layer_results:
# Extract ground truth
true_concept = ground_truth_concept_extractor(result['target_image_filename'])
if not true_concept:
continue
ik = result.get('instance_knn', {})
if not ik:
continue
total += 1
# Check 1-NN accuracy
pred_1nn = ik.get('top1_concept')
if pred_1nn == true_concept:
correct_1nn += 1
# Check k-NN vote accuracy
pred_knn = ik.get('topk_voted_concept')
if pred_knn == true_concept:
correct_knn += 1
if total > 0:
acc_1nn = correct_1nn / total
acc_knn = correct_knn / total
print(f' 1-NN Accuracy: {correct_1nn}/{total} = {acc_1nn:.3f}')
print(f' k-NN Accuracy (k={knn_topk}): {correct_knn}/{total} = {acc_knn:.3f}')
else:
print(' No valid predictions to evaluate')
if __name__ == '__main__':
# Configuration
target_db_path = 'output/llava.db'
concept_db_path = 'output/llava-concepts-colors.db'
# Analysis parameters
layer_names = None # None for all layers
n_pca_components = 5 # None for raw embeddings
knn_topk = 5
knn_power = 1.0
print('=' * 60)
print('VLM CONCEPT ANALYSIS WITH INSTANCE-BASED k-NN')
print('=' * 60)
try:
# Run k-NN analysis
results = concept_similarity_analysis_with_knn(
target_db_path=target_db_path,
concept_db_path=concept_db_path,
layer_names=layer_names,
n_pca_components=n_pca_components,
knn_topk=knn_topk,
knn_power=knn_power,
device='cpu'
)
if results:
# Save detailed results
output_file = 'output/knn_similarity_analysis.txt'
save_knn_analysis_results(results, output_file)
# Analyze k-NN accuracy
analyze_knn_accuracy(results)
# Show aggregate trends (reuse existing function)
analyze_concept_trends(results)
print(f'\n{"=" * 50}')
print('k-NN ANALYSIS COMPLETE')
print(f'{"=" * 50}')
print(f'Processed {len(results)} layers')
print(f'Results saved to: {output_file}')
else:
print('No results generated. Check database compatibility and parameters.')
except Exception as e:
print(f'Error during analysis: {e}')
import traceback
traceback.print_exc()
|