""" PCA 可视化 - Feature Space Visualization 对比 DeCLIP 解耦蒸馏和 Integrated 集成蒸馏的特征质量 使用方法： 1. 加载训练好的模型 2. 提取特征 3. PCA 降维到 3 维 4. 映射为 RGB 图像 """ import torch import torch.nn.functional as F import numpy as np from PIL import Image import matplotlib.pyplot as plt from sklearn.decomposition import PCA from typing import Tuple, Optional, Dict import os class PCAFeatureVisualizer: """ PCA 特征可视化器 用于对比不同方法的特征质量 """ def __init__(self, save_dir: str = "pca_visualization_results"): self.save_dir = save_dir os.makedirs(save_dir, exist_ok=True) def extract_and_visualize( self, model: torch.nn.Module, image: torch.Tensor, mode: str = "vanilla", title: str = "Feature PCA", save_name: Optional[str] = None ) -> np.ndarray: """ 提取特征并进行 PCA 可视化 Args: model: CLIP 模型 image: 输入图像 (1, 3, H, W) mode: 特征提取模式 ("vanilla" for integrated, "csa_vfm_distill" for decoupled) title: 图像标题 save_name: 保存文件名 Returns: PCA RGB 图像 (H, W, 3) """ model.eval() with torch.no_grad(): if mode == "vanilla": # Integrated: 使用完整 forward 的最终输出 features = model.encode_dense(image, normalize=False, keep_shape=True, mode="vanilla") elif mode in ["csa_vfm_distill", "qq_vfm_distill", "vv_vfm_distill"]: # Decoupled: 使用解耦模式的输出 features, context = model.encode_dense(image, normalize=False, keep_shape=True, mode=mode) else: features = model.encode_dense(image, normalize=False, keep_shape=True, mode=mode) # features: (B, C, H, W) B, C, H, W = features.shape # 转换为 (H*W, C) 用于 PCA features_flat = features[0].permute(1, 2, 0).reshape(-1, C).cpu().numpy() # PCA 降维到 3 维 pca = PCA(n_components=3) features_pca = pca.fit_transform(features_flat) # 归一化到 [0, 1] features_pca = (features_pca - features_pca.min()) / (features_pca.max() - features_pca.min() + 1e-8) # 重塑为图像 pca_image = features_pca.reshape(H, W, 3) # 可视化 if save_name: self._save_visualization(pca_image, title, save_name) return pca_image def compare_decoupled_vs_integrated( self, decoupled_model: torch.nn.Module, integrated_model: torch.nn.Module, image: torch.Tensor, original_image: Optional[np.ndarray] = None, save_name: str = "comparison.png" ): """ 对比 DeCLIP 解耦蒸馏和 Integrated 集成蒸馏的特征 Args: decoupled_model: 解耦蒸馏模型 integrated_model: 集成蒸馏模型 image: 输入图像 (1, 3, H, W) original_image: 原始图像 (H, W, 3) 用于对比显示 save_name: 保存文件名 """ # 提取解耦特征 decoupled_pca = self.extract_and_visualize( decoupled_model, image, mode="csa_vfm_distill", title="DeCLIP (Decoupled)" ) # 提取集成特征 integrated_pca = self.extract_and_visualize( integrated_model, image, mode="vanilla", title="Integrated" ) # 创建对比图 fig, axes = plt.subplots(1, 3 if original_image is not None else 2, figsize=(15, 5)) idx = 0 if original_image is not None: axes[idx].imshow(original_image) axes[idx].set_title("Original Image", fontsize=14) axes[idx].axis("off") idx += 1 axes[idx].imshow(decoupled_pca) axes[idx].set_title("DeCLIP (Decoupled)\nExpected: Clear boundaries, semantic coherence", fontsize=12) axes[idx].axis("off") idx += 1 axes[idx].imshow(integrated_pca) axes[idx].set_title("Integrated\nExpected: Blurry, mixed features", fontsize=12) axes[idx].axis("off") 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"Comparison saved to {save_path}") def extract_qv_features( self, model: torch.nn.Module, image: torch.Tensor, save_name: str = "qv_features.png" ): """ 提取并可视化 DeCLIP 的 Q 和 V 特征 用于展示解耦效果 """ model.eval() with torch.no_grad(): # 使用 csa_vfm_distill 模式提取 Q 和 K（实际上是 Q 和 V 的 self-attention） features, context = model.encode_dense(image, normalize=False, keep_shape=True, mode="csa_vfm_distill") # context 应该是 (q_feature, k_feature) 或类似结构 if isinstance(context, tuple) and len(context) >= 2: q_feature, v_feature = context[0], context[1] # 可视化 Q 特征 B, N, nHead, dim = q_feature.shape H = W = int(np.sqrt(N - 1)) # 减去 cls token q_flat = q_feature[0, 1:].reshape(H, W, -1).cpu().numpy() v_flat = v_feature[0, 1:].reshape(H, W, -1).cpu().numpy() # PCA pca_q = PCA(n_components=3) pca_v = PCA(n_components=3) q_pca = pca_q.fit_transform(q_flat.reshape(-1, q_flat.shape[-1])) v_pca = pca_v.fit_transform(v_flat.reshape(-1, v_flat.shape[-1])) q_pca = (q_pca - q_pca.min()) / (q_pca.max() - q_pca.min() + 1e-8) v_pca = (v_pca - v_pca.min()) / (v_pca.max() - v_pca.min() + 1e-8) q_image = q_pca.reshape(H, W, 3) v_image = v_pca.reshape(H, W, 3) # 可视化 fig, axes = plt.subplots(1, 2, figsize=(12, 5)) axes[0].imshow(q_image) axes[0].set_title("Q Features (Content)", fontsize=14) axes[0].axis("off") axes[1].imshow(v_image) axes[1].set_title("V Features (Context)", fontsize=14) axes[1].axis("off") 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"Q/V features saved to {save_path}") return q_image, v_image else: print("Could not extract Q and V features. Context format unexpected.") return None, None def _save_visualization(self, pca_image: np.ndarray, title: str, save_name: str): """保存可视化结果""" plt.figure(figsize=(8, 8)) plt.imshow(pca_image) plt.title(title, fontsize=14) plt.axis("off") save_path = os.path.join(self.save_dir, save_name) plt.savefig(save_path, dpi=150, bbox_inches="tight") plt.close() print(f"Visualization saved to {save_path}") def visualize_feature_comparison( decoupled_checkpoint: str, integrated_checkpoint: str, image_path: str, model_name: str = "EVA02-CLIP-B-16", save_dir: str = "pca_visualization_results" ): """ 便捷函数：加载模型并进行特征对比可视化 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_path: 测试图像路径 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 # 加载模型 print("Loading models...") device = "cuda" if torch.cuda.is_available() else "cpu" 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() # 加载图像 print(f"Loading image: {image_path}") image = Image.open(image_path).convert("RGB") original_image = np.array(image) # 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]) ]) image_tensor = transform(image).unsqueeze(0).to(device) # 可视化 visualizer = PCAFeatureVisualizer(save_dir) visualizer.compare_decoupled_vs_integrated( decoupled_model, integrated_model, image_tensor, original_image=original_image, save_name="decoupled_vs_integrated.png" ) print("Done!") if __name__ == "__main__": print("PCA Feature Visualizer") print("Usage:") print(" from decoupling_analysis.pca_visualization import PCAFeatureVisualizer") print(" visualizer = PCAFeatureVisualizer('save_dir')") print(" visualizer.compare_decoupled_vs_integrated(decoupled_model, integrated_model, image)")