File size: 12,250 Bytes
c9aee57 | 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 | """
CKA 分析 - Centered Kernel Alignment
用于分析特征相似性:
1. DeCLIP 的 Q 特征 vs V 特征
2. Integrated 输出特征 vs DeCLIP Q 特征
3. Integrated 输出特征 vs DeCLIP V 特征
CKA 是一种测量神经网络表示相似性的方法,值在 [0, 1] 之间,1 表示完全相似
"""
import torch
import torch.nn.functional as F
import numpy as np
from typing import Dict, List, Tuple, Optional
import matplotlib.pyplot as plt
import seaborn as sns
import os
def centering_matrix(n: int, device: str = "cpu") -> torch.Tensor:
"""创建中心化矩阵 H = I - 1/n * ones"""
I = torch.eye(n, device=device)
ones = torch.ones(n, n, device=device) / n
return I - ones
def linear_kernel(X: torch.Tensor) -> torch.Tensor:
"""线性核 K = X @ X.T"""
return X @ X.T
def rbf_kernel(X: torch.Tensor, sigma: float = 1.0) -> torch.Tensor:
"""RBF 核"""
sq_dists = torch.cdist(X, X, p=2) ** 2
return torch.exp(-sq_dists / (2 * sigma ** 2))
def hsic(K: torch.Tensor, L: torch.Tensor, H: torch.Tensor) -> torch.Tensor:
"""
Hilbert-Schmidt Independence Criterion
HSIC(K, L) = trace(KHLH) / (n-1)^2
"""
n = K.shape[0]
KHLH = K @ H @ L @ H
return torch.trace(KHLH) / ((n - 1) ** 2)
def cka(X: torch.Tensor, Y: torch.Tensor, kernel: str = "linear") -> float:
"""
Centered Kernel Alignment
CKA(X, Y) = HSIC(K_X, K_Y) / sqrt(HSIC(K_X, K_X) * HSIC(K_Y, K_Y))
Args:
X: 特征矩阵 (n_samples, n_features_x)
Y: 特征矩阵 (n_samples, n_features_y)
kernel: 核函数类型 ("linear" or "rbf")
Returns:
CKA 值 [0, 1]
"""
assert X.shape[0] == Y.shape[0], "X and Y must have same number of samples"
n = X.shape[0]
device = X.device
# 中心化矩阵
H = centering_matrix(n, device)
# 核矩阵
if kernel == "linear":
K_X = linear_kernel(X)
K_Y = linear_kernel(Y)
elif kernel == "rbf":
K_X = rbf_kernel(X)
K_Y = rbf_kernel(Y)
else:
raise ValueError(f"Unknown kernel: {kernel}")
# HSIC
hsic_xy = hsic(K_X, K_Y, H)
hsic_xx = hsic(K_X, K_X, H)
hsic_yy = hsic(K_Y, K_Y, H)
# CKA
cka_value = hsic_xy / (torch.sqrt(hsic_xx * hsic_yy) + 1e-8)
return cka_value.item()
class CKAAnalyzer:
"""
CKA 特征相似性分析器
"""
def __init__(self, save_dir: str = "cka_analysis_results"):
self.save_dir = save_dir
os.makedirs(save_dir, exist_ok=True)
self.results = {}
def extract_features(
self,
model: torch.nn.Module,
image: torch.Tensor,
mode: str = "vanilla"
) -> Dict[str, torch.Tensor]:
"""
提取特征
Args:
model: CLIP 模型
image: 输入图像 (1, 3, H, W)
mode: 提取模式
Returns:
特征字典
"""
model.eval()
features = {}
with torch.no_grad():
if mode in ["csa_vfm_distill", "qq_vfm_distill"]:
output, context = model.encode_dense(image, normalize=False, keep_shape=True, mode=mode)
features["output"] = output
if isinstance(context, tuple):
if len(context) >= 2:
features["Q"] = context[0]
features["V"] = context[1]
elif len(context) == 1:
features["context"] = context[0]
else:
features["context"] = context
else:
output = model.encode_dense(image, normalize=False, keep_shape=True, mode=mode)
features["output"] = output
return features
def analyze_qv_similarity(
self,
model: torch.nn.Module,
images: List[torch.Tensor],
mode: str = "csa_vfm_distill"
) -> float:
"""
分析 Q 和 V 特征的相似性
Args:
model: DeCLIP 模型
images: 图像列表
mode: 解耦模式
Returns:
Q 和 V 的 CKA 相似度
"""
all_q = []
all_v = []
for image in images:
features = self.extract_features(model, image, mode)
if "Q" in features and "V" in features:
# 展平特征
q = features["Q"].flatten(start_dim=1) # (B, N*dim)
v = features["V"].flatten(start_dim=1)
all_q.append(q)
all_v.append(v)
if not all_q:
print("No Q/V features extracted")
return 0.0
# 合并所有样本
Q_all = torch.cat(all_q, dim=0) # (total_samples, features)
V_all = torch.cat(all_v, dim=0)
# 计算 CKA
cka_qv = cka(Q_all, V_all, kernel="linear")
self.results["Q_vs_V"] = cka_qv
return cka_qv
def analyze_integrated_vs_decoupled(
self,
decoupled_model: torch.nn.Module,
integrated_model: torch.nn.Module,
images: List[torch.Tensor]
) -> Dict[str, float]:
"""
分析 Integrated 输出与 DeCLIP Q/V 的相似性
Args:
decoupled_model: 解耦蒸馏模型
integrated_model: 集成蒸馏模型
images: 图像列表
Returns:
CKA 相似度字典
"""
all_integrated = []
all_q = []
all_v = []
for image in images:
# 提取 integrated 特征
int_features = self.extract_features(integrated_model, image, mode="vanilla")
int_output = int_features["output"].flatten(start_dim=1)
all_integrated.append(int_output)
# 提取 decoupled 特征
dec_features = self.extract_features(decoupled_model, image, mode="csa_vfm_distill")
if "Q" in dec_features and "V" in dec_features:
q = dec_features["Q"].flatten(start_dim=1)
v = dec_features["V"].flatten(start_dim=1)
all_q.append(q)
all_v.append(v)
if not all_q:
print("No Q/V features extracted from decoupled model")
return {}
# 合并
Int_all = torch.cat(all_integrated, dim=0)
Q_all = torch.cat(all_q, dim=0)
V_all = torch.cat(all_v, dim=0)
# 确保特征维度一致
min_dim = min(Int_all.shape[1], Q_all.shape[1], V_all.shape[1])
Int_all = Int_all[:, :min_dim]
Q_all = Q_all[:, :min_dim]
V_all = V_all[:, :min_dim]
# 计算 CKA
results = {
"Integrated_vs_Q": cka(Int_all, Q_all, kernel="linear"),
"Integrated_vs_V": cka(Int_all, V_all, kernel="linear"),
"Q_vs_V": cka(Q_all, V_all, kernel="linear")
}
self.results.update(results)
return results
def visualize_cka_matrix(
self,
cka_matrix: np.ndarray,
labels: List[str],
save_name: str = "cka_matrix.png",
title: str = "CKA Similarity Matrix"
):
"""
可视化 CKA 相似度矩阵
"""
plt.figure(figsize=(8, 6))
sns.heatmap(
cka_matrix,
xticklabels=labels,
yticklabels=labels,
annot=True,
fmt=".3f",
cmap="YlOrRd",
vmin=0,
vmax=1,
square=True
)
plt.title(title, fontsize=14)
plt.tight_layout()
save_path = os.path.join(self.save_dir, save_name)
plt.savefig(save_path, dpi=150, bbox_inches="tight")
plt.close()
print(f"CKA matrix saved to {save_path}")
def create_comparison_matrix(self) -> Tuple[np.ndarray, List[str]]:
"""
创建对比矩阵
"""
labels = ["DeCLIP Q", "DeCLIP V", "Integrated Output"]
n = len(labels)
matrix = np.eye(n)
if "Q_vs_V" in self.results:
matrix[0, 1] = matrix[1, 0] = self.results["Q_vs_V"]
if "Integrated_vs_Q" in self.results:
matrix[0, 2] = matrix[2, 0] = self.results["Integrated_vs_Q"]
if "Integrated_vs_V" in self.results:
matrix[1, 2] = matrix[2, 1] = self.results["Integrated_vs_V"]
return matrix, labels
def print_summary(self):
"""打印分析摘要"""
print("\n" + "="*50)
print("CKA Analysis Summary")
print("="*50)
for key, value in self.results.items():
print(f" {key}: {value:.4f}")
print("\nInterpretation:")
if "Q_vs_V" in self.results:
qv = self.results["Q_vs_V"]
if qv < 0.5:
print(f" Q and V are fairly different (CKA={qv:.3f}), indicating successful decoupling")
else:
print(f" Q and V are similar (CKA={qv:.3f}), suggesting features are mixed")
print("="*50 + "\n")
def save_results(self, filename: str = "cka_results.json"):
"""保存分析结果"""
import json
filepath = os.path.join(self.save_dir, filename)
with open(filepath, "w") as f:
json.dump(self.results, f, indent=2)
print(f"CKA results saved to {filepath}")
def run_cka_analysis(
decoupled_checkpoint: str,
integrated_checkpoint: str,
image_paths: List[str],
model_name: str = "EVA02-CLIP-B-16",
save_dir: str = "cka_analysis_results"
):
"""
运行完整的 CKA 分析
Args:
decoupled_checkpoint: 解耦蒸馏模型的 checkpoint 路径
integrated_checkpoint: 集成蒸馏模型的 checkpoint 路径
image_paths: 测试图像路径列表
model_name: 模型名称
save_dir: 保存目录
"""
from open_clip import create_model
from torchvision import transforms
from PIL import Image
print("Loading models...")
device = "cuda" if torch.cuda.is_available() else "cpu"
# 加载模型
decoupled_model = create_model(model_name, pretrained="eva", device=device)
decoupled_model.load_state_dict(torch.load(decoupled_checkpoint, map_location=device)["state_dict"])
decoupled_model.eval()
integrated_model = create_model(model_name, pretrained="eva", device=device)
integrated_model.load_state_dict(torch.load(integrated_checkpoint, map_location=device)["state_dict"])
integrated_model.eval()
# 预处理
transform = transforms.Compose([
transforms.Resize((560, 560)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
# 加载图像
print(f"Loading {len(image_paths)} images...")
images = []
for path in image_paths:
image = Image.open(path).convert("RGB")
image_tensor = transform(image).unsqueeze(0).to(device)
images.append(image_tensor)
# CKA 分析
analyzer = CKAAnalyzer(save_dir)
print("Analyzing Q vs V similarity...")
analyzer.analyze_qv_similarity(decoupled_model, images)
print("Analyzing Integrated vs Decoupled similarity...")
analyzer.analyze_integrated_vs_decoupled(decoupled_model, integrated_model, images)
# 可视化
matrix, labels = analyzer.create_comparison_matrix()
analyzer.visualize_cka_matrix(matrix, labels, save_name="cka_comparison.png")
# 保存结果
analyzer.print_summary()
analyzer.save_results()
print("Done!")
if __name__ == "__main__":
print("CKA Analysis Tool")
print("Usage:")
print(" from decoupling_analysis.cka_analysis import CKAAnalyzer, cka")
print(" analyzer = CKAAnalyzer('save_dir')")
print(" analyzer.analyze_qv_similarity(model, images)")
print(" analyzer.analyze_integrated_vs_decoupled(dec_model, int_model, images)")
|