import os import h5py import torch import os import numpy as np from PIL import Image,ImageDraw import torch.nn.functional as F from open_clip.transform import ResizeMaxSize,_convert_to_rgb,det_image_transform,ResizeLongest from torchvision.transforms import ToTensor,Normalize import matplotlib.pyplot as plt from torchvision import transforms from pycocotools.coco import COCO from sklearn.decomposition import PCA from math import sqrt from third_party.utils.utils_correspondence import pca_reduce_features import math from sklearn.cluster import DBSCAN os.environ["CUDA_VISIBLE_DEVICES"] = "0" from matplotlib import colors import matplotlib as mpl def dbscan_cluster_from_self_similarity(self_corr_matrix, eps=1.1, min_samples=5): if isinstance(self_corr_matrix, torch.Tensor): mat = self_corr_matrix.squeeze(0).cpu().numpy() else: mat = np.squeeze(self_corr_matrix) assert mat.ndim == 2 and mat.shape[0] == mat.shape[1], "输入应为(N,N)自相关矩阵" X = mat clustering = DBSCAN(eps=eps, min_samples=min_samples) labels = clustering.fit_predict(X) return labels # shape (N,) def gaussian_window(dim1, dim2, std=1.): constant = 1 / (std * math.sqrt(2)) ks = list() for dim in [dim1, dim2]: start = -(dim - 1) / 2.0 k = torch.linspace(start=start * constant, end=(start + (dim - 1)) * constant, steps=dim, dtype=torch.float) ks.append(k) dist_square_to_mu = (torch.stack(torch.meshgrid(*ks, indexing='ij')) ** 2).sum(0) return torch.exp(-dist_square_to_mu) def get_attention_addition(dim1, dim2, window, adjust_for_cls=True): m = torch.einsum('ij,kl->ijkl', torch.eye(dim1), torch.eye(dim2)) m = m.permute((0, 3, 1, 2)).contiguous() # m[ijkl] = 1 iff (i, j) == (k, l) out = F.conv2d(m.view(-1, dim1, dim2).unsqueeze(1), window.unsqueeze(0).unsqueeze(1), padding='same').squeeze(1) out = out.view(dim1 * dim2, dim1 * dim2) if adjust_for_cls: v_adjusted = torch.vstack([torch.zeros((1, dim1 * dim2)), out]) out = torch.hstack([torch.zeros((dim1 * dim2 + 1, 1)), v_adjusted]) return out class UnNormalize(object): def __init__(self, mean, std): self.mean = mean self.std = std def __call__(self, image): image2 = torch.clone(image) if len(image2.shape) == 4: # batched image2 = image2.permute(1, 0, 2, 3).contiguous() for t, m, s in zip(image2, self.mean, self.std): t.mul_(s).add_(m) return image2.permute(1, 0, 2, 3).contiguous() elif len(image2.shape) == 3: # single image for t, m, s in zip(image2, self.mean, self.std): t.mul_(s).add_(m) return image2 else: raise ValueError(f"Unsupported image shape: {image2.shape}") def plot_pca(f, path,target_size): # f: numpy array, shape [N, D] pca = PCA(n_components=3) pca.fit(f) pca_img = pca.transform(f) # n x 3 h = w = int(sqrt(pca_img.shape[0])) pca_img = pca_img.reshape(h, w, 3) pca_img_min = pca_img.min(axis=(0, 1)) pca_img_max = pca_img.max(axis=(0, 1)) pca_img = (pca_img - pca_img_min) / (pca_img_max - pca_img_min + 1e-8) # 防止除零 pca_img = Image.fromarray((pca_img * 255).astype(np.uint8)) pca_img = transforms.Resize(target_size, interpolation=transforms.InterpolationMode.NEAREST)(pca_img) pca_img.save(path) def load_data(coco): image_ids=[] img_ids = coco.getImgIds() cat_ids = coco.getCatIds() for img_id in img_ids: img = coco.loadImgs(img_id)[0] ann_ids = coco.getAnnIds(imgIds=img['id'], catIds=cat_ids, iscrowd=None) anns = coco.loadAnns(ann_ids) anns = [ann for ann in anns if ann['iscrowd'] == 0] if len(anns) == 0: continue image_ids.append(img_id) torch.manual_seed(42) image_ids = [image_ids[i] for i in torch.randperm(len(image_ids))] return image_ids def build_DINOv2(): model_name='dinov2_vitb14_reg' hub_path = '/mnt/SSD8T/home/wjj/.cache/torch/hub/facebookresearch_dinov2_main' try: vfm = torch.hub.load(hub_path, model_name, source='local').half() except Exception as e: raise RuntimeError(f"Failed to load DINOv2 model '{model_name}': {e}") return vfm mean=[0.485, 0.456, 0.406] std=[0.229, 0.224, 0.225] normalize = Normalize(mean=mean, std=std) DINO_transform=transforms.Compose([ ResizeLongest(35*14, fill=0), _convert_to_rgb, ToTensor(), normalize ]) _transform=transforms.Compose([ ResizeLongest(560, fill=0), _convert_to_rgb,]) with torch.no_grad(): device="cuda" coco_path='/mnt/SSD8T/home/wjj/dataset/standard_coco/annotations/instances_train2017.json' img_path='/mnt/SSD8T/home/wjj/dataset/standard_coco/train2017' cache_path = "/mnt/SSD8T/home/wjj/code/distilldift/train/cache/Dift_COCO_dift_merged_fp16.h5" cache=h5py.File(cache_path, 'r')[str(0)] weights = torch.load("/mnt/SSD8T/home/wjj/code/DeCLIP/EVAB_COCO_117K_topk10.pth", map_location="cpu") coco=COCO(coco_path) image_ids=load_data(coco) dino=build_DINOv2().to(device) image_select=200 img_name = coco.loadImgs(image_ids[image_select])[0]['file_name'] weight = weights[image_select] match_id = weight[np.random.choice(10)] matching_sample = image_ids[match_id] matching_sample_info = coco.imgs[matching_sample] knn_image_name=matching_sample_info['file_name'] knn_image_path = os.path.join(img_path, knn_image_name) knn_image= Image.open(knn_image_path) knn_image_tensor = DINO_transform(knn_image).unsqueeze(0).to(torch.float16).to(device) knn_dino_feats_raw = dino.get_intermediate_layers(knn_image_tensor, reshape=True)[0].flatten(start_dim=-2).transpose(-2,-1) # 未归一化 knn_sd_feats_raw = torch.from_numpy(cache[knn_image_name][()]).unsqueeze(0).flatten(start_dim=-2).transpose(-2,-1).to(torch.float32).to(device) # 未归一化 knn_dino_feats = F.normalize(knn_dino_feats_raw, dim=2) knn_sd_feats = F.normalize(knn_sd_feats_raw, dim=2) image_path = os.path.join(img_path, img_name) image = Image.open(image_path) sd_feats_raw=torch.from_numpy(cache[img_name][()]).unsqueeze(0) # sd_feats_raw=pca_reduce_features(sd_feats_raw, 256) sd_feats_raw =sd_feats_raw.flatten(start_dim=-2).transpose(-2,-1).to(torch.float32).to(device) # 未归一化 image_tensor = DINO_transform(image).unsqueeze(0).to(torch.float16).to(device) dino_feats_raw = dino.get_intermediate_layers(image_tensor, reshape=True)[0].flatten(start_dim=-2).transpose(-2,-1) # 未归一化 # 归一化用于 similarity sd_feats = F.normalize(sd_feats_raw, dim=2) dino_feats = F.normalize(dino_feats_raw, dim=2) w_dino=0.5 w_sd=0.5 dino_weighted = (w_dino ** 0.5) * dino_feats sd_weighted = (w_sd ** 0.5) * sd_feats # 拼接 (未归一化) —— 用于 PCA sd_dino_feats_raw = torch.cat([sd_weighted, dino_weighted], dim=2) sd_dino_feats=F.normalize(sd_dino_feats_raw, dim=2) # 计算 similarity sim_sd = torch.einsum('bic,bjc->bij', sd_feats, sd_feats) sim_dino = torch.einsum('bic,bjc->bij', dino_feats, dino_feats) sim_sd_dino = torch.einsum('bic,bjc->bij', sd_dino_feats, sd_dino_feats) dino_dbscan_labels=dbscan_cluster_from_self_similarity(sim_dino) target_size = (560, 560) low_res_size = (35, 35) low_res_token_choosen = (14, 12) token_chosen = int( low_res_token_choosen[0] * low_res_size[1] + low_res_token_choosen[1] ) token_x_low_res = token_chosen % low_res_size[0] token_y_low_res = token_chosen // low_res_size[0] token_x_img = int(((token_x_low_res+ 0.5) / low_res_size[0]) * target_size[0]) token_y_img = int(((token_y_low_res+ 0.5) / low_res_size[1]) * target_size[1]) output_dir = "sd_vis" if not os.path.exists(output_dir): os.mkdir(output_dir) unnormalize = UnNormalize(mean, std) sim_sd = sim_sd[:, token_chosen, :] # 1, h*w sim_dino = sim_dino[:, token_chosen, :] # 1, h*w sim_sd_dino = sim_sd_dino[:, token_chosen, :] # 1, h*w vis_img1=_transform(image) window_size = [side * 2 - 1 for side in low_res_size] window = gaussian_window(*window_size, std=15) addition = get_attention_addition(*low_res_size, window,adjust_for_cls=False).unsqueeze(0) # 1, low_res_size*low_res_size, low_res_size*low_res_size addition=addition[:,token_chosen].to(sim_dino.device) # 1, h*w # sim_dino_gaussian=sim_dino*addition addition = addition.to(sim_dino.dtype) mask = sim_dino > 0 # [bs, h*w]，布尔掩码 sim_dino_new = sim_dino.clone() sim_dino_new[mask] = sim_dino[mask] * addition.expand_as(sim_dino)[mask] sim_dino_gaussian=sim_dino_new # 2. sim_sd, sim_dino, sim_sd_dino热力图 sim_maps = [sim_sd, sim_dino, sim_dino_gaussian, sim_sd_dino] sim_np_maps = [] # 1. 处理聚类标签 cluster_map = dino_dbscan_labels.reshape(low_res_size) # (35, 35) cluster_map_tensor = torch.from_numpy(cluster_map).unsqueeze(0).unsqueeze(0).float() # [1,1,35,35] cluster_map_up = F.interpolate(cluster_map_tensor, size=target_size, mode="nearest") # [1,1,560,560] cluster_map_np = cluster_map_up.squeeze().cpu().numpy() # 2. 计算聚类数量，包含-1噪声 all_labels = np.unique(dino_dbscan_labels) print(all_labels) n_clusters = len(all_labels) # 3. 构建颜色和norm cluster_colors = ['black'] + [mpl.colormaps['tab20'](i) for i in range(n_clusters-1)] cmap = mpl.colors.ListedColormap(cluster_colors) bounds = np.arange(-1, n_clusters) # -1,0,1,...,n_clusters-2 norm = mpl.colors.BoundaryNorm(bounds, cmap.N) for sim in sim_maps: sim_map = sim.view(1, 1, low_res_size[0], low_res_size[1]) sim_map_up = F.interpolate(sim_map, size=target_size, mode="bilinear", align_corners=False) sim_map_np = sim_map_up.squeeze().cpu().numpy() sim_np_maps.append(sim_map_np) # ================== 第一张图：Similarity ================== fig, axes = plt.subplots(1, 6, figsize=(30, 5.5)) # 第一列：原图 + 选点 axes[0].imshow(vis_img1) axes[0].scatter([token_x_img], [token_y_img], c='red', s=200, marker="o", edgecolors='black', linewidths=2) axes[0].set_title('Image') axes[0].axis('off') # 后四列：三个 similarity 热力图 + 聚类结果 sim_titles = [ 'Token Similarity (SD)', 'Token Similarity (DINOv2)', 'Token Similarity (DINOv2_gaussian)', 'Token Similarity (SD+DINOv2)' ] for i in range(len(sim_np_maps)): axes[i+1].imshow(sim_np_maps[i], cmap='jet') axes[i+1].set_title(sim_titles[i]) axes[i+1].axis('off') # 新增：DBSCAN聚类可视化 im = axes[-1].imshow(cluster_map_np, cmap=cmap, norm=norm, interpolation='nearest') axes[-1].set_title('DBSCAN Clusters') axes[-1].axis('off') plt.tight_layout() plt.savefig(os.path.join(output_dir, "token_similarity_vis_plus_cluster.png")) plt.close(fig) # ===== PCA 部分 ===== # feats_to_pca 用未归一化特征 pca_paths = [] feats_to_pca = [dino_feats_raw, sd_feats_raw, sd_dino_feats_raw] pca_titles = ["DINOv2 PCA", "SD PCA", "SD+DINOv2 PCA"] for i, feats in enumerate(feats_to_pca): feats_np = feats[0].cpu().numpy() pca_path = os.path.join(output_dir, f"pca_vis_{i}.png") plot_pca(feats_np, pca_path, target_size) pca_paths.append(pca_path) # 2. 读取PCA图片 pca_imgs = [ transforms.Resize(target_size, interpolation=transforms.InterpolationMode.NEAREST)(Image.open(p)) for p in pca_paths ] # ================== 第二张图：PCA ================== fig, axes = plt.subplots(1, 4, figsize=(20, 6)) # 第一列：原图 axes[0].imshow(vis_img1) axes[0].set_title('Image') axes[0].axis('off') # 后三列：PCA 图片 pca_titles = ["DINOv2 PCA", "SD PCA", "SD+DINOv2 PCA"] for i in range(3): axes[i+1].imshow(pca_imgs[i]) axes[i+1].set_title(pca_titles[i]) axes[i+1].axis('off') plt.tight_layout() plt.savefig(os.path.join(output_dir, "pca_vis.png")) plt.close(fig)