File size: 12,114 Bytes
a831c4c | 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 | """
CKA 对比分析:EVA-CLIP 预训练 vs DeCLIP+ 训练后
目标:验证解耦蒸馏是否使 Q 和 K 特征变得更加不同
- 如果 CKA(Q,K) 在 DeCLIP+ 训练后降低,说明特征成功解耦
- 如果 CKA(Q,K) 保持不变或升高,说明特征没有解耦
使用方法:
cd DeCLIP_private
python decoupling_analysis/run_cka_comparison.py
"""
import sys
import os
import subprocess
# 添加 src 到路径
sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'src'))
import torch
import torch.nn.functional as F
from PIL import Image
from torchvision import transforms
import numpy as np
import json
from typing import List, Tuple, Dict
from collections import defaultdict
# 从 cka_analysis.py 导入核心函数
from cka_analysis import cka, CKAAnalyzer
def download_checkpoint_if_needed(target_path: str, repo_id: str = "xiaomoguhzz/xiaomogu_pami", filename: str = "declip_plus_seg/epoch_6.pt"):
"""
如果权重文件不存在,自动从 HuggingFace 下载
"""
if os.path.exists(target_path):
print(f"Checkpoint already exists: {target_path}")
return True
print(f"Downloading checkpoint to {target_path}...")
target_dir = os.path.dirname(target_path)
os.makedirs(target_dir, exist_ok=True)
try:
# 使用 huggingface-cli 下载
cmd = f"huggingface-cli download {repo_id} {filename} --local-dir {target_dir}"
result = subprocess.run(cmd, shell=True, capture_output=True, text=True)
if result.returncode == 0:
# 下载后文件会在 target_dir/declip_plus_seg/epoch_6.pt
downloaded_path = os.path.join(target_dir, filename)
if os.path.exists(downloaded_path) and downloaded_path != target_path:
# 如果下载路径和目标路径不同,移动文件
os.makedirs(os.path.dirname(target_path), exist_ok=True)
if not os.path.exists(target_path):
os.rename(downloaded_path, target_path)
print(f"Download complete: {target_path}")
return True
else:
print(f"Download failed: {result.stderr}")
return False
except Exception as e:
print(f"Download error: {e}")
return False
def load_eva_clip_model(checkpoint_path: str = None, device: str = "cuda"):
"""
加载 EVA-CLIP 模型
Args:
checkpoint_path: 如果为 None,使用预训练权重;否则加载指定 checkpoint
device: 计算设备
"""
from open_clip import create_model
# 创建模型并加载预训练权重
model = create_model("EVA02-CLIP-B-16", pretrained="eva", device=device)
if checkpoint_path is not None:
print(f"Loading checkpoint from {checkpoint_path}")
checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=False)
# 处理不同格式的 checkpoint
if "state_dict" in checkpoint:
state_dict = checkpoint["state_dict"]
elif "model" in checkpoint:
state_dict = checkpoint["model"]
else:
state_dict = checkpoint
# 移除 "module." 前缀(如果有)
state_dict = {k.replace("module.", ""): v for k, v in state_dict.items()}
# 只加载 visual encoder 相关的权重
visual_state_dict = {}
for k, v in state_dict.items():
if k.startswith("visual."):
visual_state_dict[k.replace("visual.", "")] = v
if visual_state_dict:
# 加载 visual encoder 权重
missing, unexpected = model.visual.load_state_dict(visual_state_dict, strict=False)
print(f"Loaded visual encoder weights. Missing: {len(missing)}, Unexpected: {len(unexpected)}")
if missing:
print(f" Missing keys (first 5): {missing[:5]}")
if unexpected:
print(f" Unexpected keys (first 5): {unexpected[:5]}")
else:
# 尝试直接加载整个模型
missing, unexpected = model.load_state_dict(state_dict, strict=False)
print(f"Loaded full model weights. Missing: {len(missing)}, Unexpected: {len(unexpected)}")
model.eval()
return model
def extract_qv_features(
model: torch.nn.Module,
images: List[torch.Tensor],
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
提取 Q 和 V 特征
DeCLIP 设计:
- Q 特征 → content loss(细粒度特征)
- V 特征 → context loss(语义特征)
Args:
model: CLIP 模型
images: 图像 tensor 列表,每个为 (1, 3, H, W)
Returns:
Q_all, V_all: 聚合后的 Q 和 V 特征
"""
model.eval()
all_q = []
all_v = []
with torch.no_grad():
for image in images:
# 分别使用 qq_vfm_distill 和 vv_vfm_distill 提取 Q 和 V
# 提取 Q 特征
output_q = model.visual.encode_dense(image, keep_shape=True, mode="qq_vfm_distill")
if isinstance(output_q, tuple) and len(output_q) == 2:
_, q = output_q
q_flat = q.flatten()
all_q.append(q_flat)
# 提取 V 特征
output_v = model.visual.encode_dense(image, keep_shape=True, mode="vv_vfm_distill")
if isinstance(output_v, tuple) and len(output_v) == 2:
_, v = output_v
v_flat = v.flatten()
all_v.append(v_flat)
if not all_q or not all_v:
print("Warning: No Q/V features extracted!")
return None, None
# 对于 CKA,我们需要 (n_samples, n_features) 格式
# 每张图的特征作为一个 sample
Q_all = torch.stack(all_q, dim=0) # (n_images, features)
V_all = torch.stack(all_v, dim=0) # (n_images, features)
print(f"Extracted features: Q shape = {Q_all.shape}, V shape = {V_all.shape}")
return Q_all, V_all
def run_comparison(
pretrained_path: str = None, # None 表示使用默认预训练权重
declip_checkpoint: str = None,
image_dir: str = None,
num_images: int = 50,
save_dir: str = "cka_analysis_results",
device: str = "cuda"
):
"""
运行 CKA 对比分析
Args:
pretrained_path: 预训练模型路径(None 使用默认)
declip_checkpoint: DeCLIP+ 训练后的 checkpoint 路径
image_dir: 测试图像目录
num_images: 使用的图像数量
save_dir: 结果保存目录
device: 计算设备
"""
os.makedirs(save_dir, exist_ok=True)
# 准备图像预处理
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])
])
# 收集图像
print(f"\n{'='*60}")
print("Loading test images...")
print(f"{'='*60}")
image_paths = []
for root, dirs, files in os.walk(image_dir):
for f in files:
if f.endswith(('.png', '.jpg', '.jpeg')):
image_paths.append(os.path.join(root, f))
if len(image_paths) >= num_images:
break
if len(image_paths) >= num_images:
break
print(f"Found {len(image_paths)} images")
# 加载图像
images = []
for path in image_paths:
try:
img = Image.open(path).convert("RGB")
img_tensor = transform(img).unsqueeze(0).to(device)
images.append(img_tensor)
except Exception as e:
print(f"Failed to load {path}: {e}")
print(f"Successfully loaded {len(images)} images")
results = {}
# ==================== 分析 1: 原生 EVA-CLIP ====================
print(f"\n{'='*60}")
print("Analysis 1: Original EVA-CLIP (Pretrained)")
print(f"{'='*60}")
model_pretrained = load_eva_clip_model(checkpoint_path=None, device=device)
Q_pre, V_pre = extract_qv_features(model_pretrained, images)
if Q_pre is not None and V_pre is not None:
cka_pretrained = cka(Q_pre, V_pre, kernel="linear")
print(f"\n>>> CKA(Q, V) for Pretrained EVA-CLIP: {cka_pretrained:.4f}")
results["pretrained_qv_cka"] = cka_pretrained
del model_pretrained
torch.cuda.empty_cache()
# ==================== 分析 2: DeCLIP+ 训练后 ====================
if declip_checkpoint and os.path.exists(declip_checkpoint):
print(f"\n{'='*60}")
print("Analysis 2: DeCLIP+ (After Training)")
print(f"{'='*60}")
model_declip = load_eva_clip_model(checkpoint_path=declip_checkpoint, device=device)
Q_dec, V_dec = extract_qv_features(model_declip, images)
if Q_dec is not None and V_dec is not None:
cka_declip = cka(Q_dec, V_dec, kernel="linear")
print(f"\n>>> CKA(Q, V) for DeCLIP+: {cka_declip:.4f}")
results["declip_qv_cka"] = cka_declip
del model_declip
torch.cuda.empty_cache()
else:
print(f"\nWarning: DeCLIP checkpoint not found at {declip_checkpoint}")
# ==================== 结果总结 ====================
print(f"\n{'='*60}")
print("SUMMARY: CKA(Q, V) Comparison Results")
print(f"{'='*60}")
print("\nDeCLIP 设计:Q 用于 content loss(细节),V 用于 context loss(语义)")
print("如果解耦成功,训练后 Q 和 V 应该更加不同(CKA 降低)\n")
if "pretrained_qv_cka" in results:
print(f" Pretrained EVA-CLIP CKA(Q,V): {results['pretrained_qv_cka']:.4f}")
if "declip_qv_cka" in results:
print(f" DeCLIP+ (trained) CKA(Q,V): {results['declip_qv_cka']:.4f}")
if "pretrained_qv_cka" in results and "declip_qv_cka" in results:
diff = results["declip_qv_cka"] - results["pretrained_qv_cka"]
print(f"\n Δ CKA = {diff:+.4f}")
if diff < -0.05:
print("\n ✓ 结论:DeCLIP+ 训练后 Q 和 V 特征更加不同(CKA 降低)")
print(" 这说明解耦蒸馏成功:Q 学习细节特征,V 学习语义特征。")
elif diff > 0.05:
print("\n ✗ 结论:DeCLIP+ 训练后 Q 和 V 特征更加相似(CKA 升高)")
print(" 这与预期不符,可能需要检查训练配置。")
else:
print("\n ~ 结论:DeCLIP+ 训练后 Q 和 V 的相似度变化不大")
print(" 可能需要更多样本或不同的分析方法来验证。")
# 保存结果
results_path = os.path.join(save_dir, "cka_comparison_results.json")
with open(results_path, "w") as f:
json.dump(results, f, indent=2)
print(f"\nResults saved to: {results_path}")
return results
if __name__ == "__main__":
# 配置路径 - 使用服务器上的标准路径
BASE_DIR = "/opt/tiger/xiaomoguhzz"
# DeCLIP+ 权重路径
DECLIP_CHECKPOINT = os.path.join(BASE_DIR, "declip_plus_seg/epoch_6.pt")
# 测试图像目录 - 使用 COCO val2017
IMAGE_DIR = os.path.join(BASE_DIR, "standard_coco/val2017")
# 如果 COCO 目录不存在,尝试使用 ReflectionBench 的图像
if not os.path.exists(IMAGE_DIR):
IMAGE_DIR = os.path.join(
os.path.dirname(__file__),
"..",
"..",
"ReflectionBenchv2_3/images"
)
SAVE_DIR = os.path.join(
os.path.dirname(__file__),
"cka_analysis_results"
)
# 自动下载权重(如果不存在)
download_checkpoint_if_needed(DECLIP_CHECKPOINT)
# 运行对比分析
run_comparison(
pretrained_path=None, # 使用默认预训练权重
declip_checkpoint=DECLIP_CHECKPOINT,
image_dir=IMAGE_DIR,
num_images=50, # 使用 50 张图像
save_dir=SAVE_DIR,
device="cuda" if torch.cuda.is_available() else "cpu"
)
|