code/model_vis_tools/vis_sd_featsv5.py

from diffusion_model.stable_diffusion import diffusion
import torch
import numpy as np
import os
from PIL import Image
from pycocotools.coco import COCO
from torchvision.transforms import Normalize, Compose, RandomResizedCrop, InterpolationMode, ToTensor, Resize, \
    CenterCrop
from open_clip.transform import ResizeLongest, _convert_to_rgb
from torchvision import transforms
import matplotlib.pyplot as plt
import cv2
from functools import reduce
import torch.nn.functional as F


class SDNormalize(object):
    def __call__(self, img):
        return 2.0 * img - 1.0


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

def visualize_self_att_raw(
    ori_img, self_att_raw, token_choosen, output_dir, attn_map_hw=(35, 35), vis_hw=(560, 560)
):
    """
    可视化self_att_raw中所有注意力图的token choosen位置的结果。
    每幅图单独保存。
    :param ori_img: 输入图像
    :param self_att_raw: 原始注意力图，形状[10, 1225, 1225]
    :param token_choosen: 选择的token坐标 (row, col)
    :param output_dir: 输出文件名前缀
    :param attn_map_hw: 注意力图的分辨率 (H, W)
    :param vis_hw: 可视化图像的分辨率 (H, W)
    """
    # 1. 原图
    if isinstance(ori_img, torch.Tensor):
        img = ori_img
        if img.ndim == 4:  # [1,3,H,W] -> [3,H,W]
            img = img[0]
        img = img.cpu().numpy()
        img = np.transpose(img, (1, 2, 0))  # [H, W, 3]
        img = (img * 255).clip(0, 255).astype(np.uint8)
    elif isinstance(ori_img, Image.Image):
        img = np.array(ori_img)
        if img.dtype != np.uint8:
            img = (img * 255).clip(0, 255).astype(np.uint8)
    elif isinstance(ori_img, np.ndarray):
        img = ori_img
        if img.dtype != np.uint8:
            img = (img * 255).clip(0, 255).astype(np.uint8)
    else:
        raise ValueError("ori_img should be torch.Tensor, PIL.Image, or np.ndarray")
    if img.shape[2] == 4:  # RGBA to RGB
        img = img[:, :, :3]
    img_resized = cv2.resize(img, vis_hw, interpolation=cv2.INTER_LINEAR)

    # 2. token坐标
    h_attn, w_attn = attn_map_hw
    row, col = token_choosen
    y_vis = int((row + 0.5) * vis_hw[0] / h_attn)
    x_vis = int((col + 0.5) * vis_hw[1] / w_attn)

    img_with_dot = img_resized.copy()
    cv2.circle(img_with_dot, (x_vis, y_vis), radius=13, color=(0, 0, 0), thickness=-1)  # 黑色圆
    cv2.circle(img_with_dot, (x_vis, y_vis), radius=11, color=(255, 0, 0), thickness=-1)  # 红色圆

    # 3. 遍历 self_att_raw
    num_layers = self_att_raw.shape[0]
    for i in range(num_layers):
        layer_att_map = self_att_raw[i, row * w_attn + col].to(torch.float32).detach().cpu().numpy().reshape(h_attn, w_attn)
        layer_att_map = (layer_att_map - layer_att_map.min()) / (layer_att_map.max() - layer_att_map.min() + 1e-8)
        layer_att_map_up = cv2.resize(layer_att_map, vis_hw, interpolation=cv2.INTER_LINEAR)

        # 绘制图像
        fig, axs = plt.subplots(1, 2, figsize=(12, 6))
        axs[0].imshow(img_with_dot)
        axs[0].set_title(f"Original (with token) - Layer {i}")
        axs[0].axis('off')

        axs[1].imshow(layer_att_map_up, cmap='jet')
        axs[1].set_title(f"Self-Attention (Layer {i})")
        axs[1].axis('off')

        plt.tight_layout()

        # 保存每一层的可视化结果
        output_path = os.path.join(output_dir,f"layer_{i}.png")
        plt.savefig(output_path, dpi=200, bbox_inches='tight')
        plt.close()

def visualize_sd_dino_att(
    ori_img, self_att, sim_dino_soft, sim_dino_refined, 
    token_choosen, filename, attn_map_hw=(64, 64), vis_hw=(512, 512)
):
    """
    可视化SD传播对DINO自相关的细化效果, 4列分别为:
    1. 原图带token
    2. SD自注意力传播
    3. DINO自相关
    4. 细化后的DINO自相关
    """
    # 1. 原图
    # 支持PIL.Image、np.ndarray、tensor三种输入
    if isinstance(ori_img, torch.Tensor):
        img = ori_img
        if img.ndim == 4:  # [1,3,H,W] -> [3,H,W]
            img = img[0]
        img = img.cpu().numpy()
        img = np.transpose(img, (1, 2, 0))  # [H, W, 3]
        img = (img * 255).clip(0, 255).astype(np.uint8)
    elif isinstance(ori_img, Image.Image):
        img = np.array(ori_img)
        if img.dtype != np.uint8:
            img = (img * 255).clip(0, 255).astype(np.uint8)
    elif isinstance(ori_img, np.ndarray):
        img = ori_img
        if img.dtype != np.uint8:
            img = (img * 255).clip(0, 255).astype(np.uint8)
    else:
        raise ValueError("ori_img should be torch.Tensor, PIL.Image, or np.ndarray")
    if img.shape[2] == 4:  # RGBA to RGB
        img = img[:, :, :3]
    img_resized = cv2.resize(img, vis_hw, interpolation=cv2.INTER_LINEAR)

    # 2. token坐标
    h_attn, w_attn = attn_map_hw
    row, col = token_choosen
    y_vis = int((row + 0.5) * vis_hw[0] / h_attn)
    x_vis = int((col + 0.5) * vis_hw[1] / w_attn)
    img_with_dot = img_resized.copy()
    # 绘制黑色边缘的圆
    cv2.circle(img_with_dot, (x_vis, y_vis), radius=13, color=(0, 0, 0), thickness=-1)  # 黑色圆

    # 绘制红点
    cv2.circle(img_with_dot, (x_vis, y_vis), radius=11, color=(255, 0, 0), thickness=-1)  # 红色圆

    # 3. SD自注意力传播
    att_map_sd = self_att[row * w_attn + col].to(torch.float32).detach().cpu().numpy().reshape(h_attn, w_attn)
    att_map_sd = (att_map_sd - att_map_sd.min()) / (att_map_sd.max() - att_map_sd.min() + 1e-8)
    att_map_sd_up = cv2.resize(att_map_sd, vis_hw, interpolation=cv2.INTER_LINEAR)

    # 4. DINO自相关（传播前）
    att_map_dino = sim_dino_soft[row * w_attn + col].detach().cpu().numpy().reshape(h_attn, w_attn)
    att_map_dino = (att_map_dino - att_map_dino.min()) / (att_map_dino.max() - att_map_dino.min() + 1e-8)
    att_map_dino = att_map_dino.astype(np.float32) 
    att_map_dino_up = cv2.resize(att_map_dino, vis_hw, interpolation=cv2.INTER_LINEAR)

    # 5. DINO传播后
    att_map_dino_ref = sim_dino_refined[row * w_attn + col].detach().cpu().numpy().reshape(h_attn, w_attn)
    att_map_dino_ref = (att_map_dino_ref - att_map_dino_ref.min()) / (att_map_dino_ref.max() - att_map_dino_ref.min() + 1e-8)
    att_map_dino_ref = att_map_dino_ref.astype(np.float32) 
    att_map_dino_ref_up = cv2.resize(att_map_dino_ref, vis_hw, interpolation=cv2.INTER_LINEAR)

    # 6. 绘图
    fig, axs = plt.subplots(1, 4, figsize=(18, 5))

    axs[0].imshow(img_with_dot)
    axs[0].set_title("Original (with token)")
    axs[0].axis('off')

    axs[1].imshow(att_map_sd_up, cmap='jet')
    axs[1].set_title(f'SD propagation (token {token_choosen})')
    axs[1].axis('off')

    axs[2].imshow(att_map_dino_up, cmap='jet')
    axs[2].set_title(f'DINO sim (pre-propagate)')
    axs[2].axis('off')

    axs[3].imshow(att_map_dino_ref_up, cmap='jet')
    axs[3].set_title(f'DINO sim (post-propagate)')
    axs[3].axis('off')

    plt.tight_layout()
    plt.savefig(filename, dpi=200, bbox_inches='tight')
    plt.close()

with torch.no_grad():
    attention_layers_to_use= [-4, -6]
    sd_version='v2.1'
    time_step=45
    device="cuda:6"
    # coco_path='/mnt/SSD8T/home/wjj/dataset/standard_coco/annotations/instances_train2017.json'
    # img_path='/mnt/SSD8T/home/wjj/dataset/standard_coco/train2017'

    # coco=COCO(coco_path)
    # image_ids=load_data(coco)
    dino=build_DINOv2().to(device)
    sd=diffusion(attention_layers_to_use=attention_layers_to_use,model=sd_version, time_step=time_step, device=device,dtype=torch.float16)
    # image_select=5
    # img_name = coco.loadImgs(image_ids[image_select])[0]['file_name']
    # image_path = os.path.join(img_path, img_name)
    image_root='/mnt/SSD8T/home/wjj/dataset/standard_coco/train2017'
    image_file=os.listdir(image_root)[999]
    image_name=os.path.join(image_root, image_file)
    # image = Image.open('demo_images/horses.jpg')
    image = Image.open("demo_images/bird.jpg")
    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(490, fill=0),
                                    _convert_to_rgb,
                                    ToTensor(),
                                    normalize])
    sd_transform=transforms.Compose([ResizeLongest(560, fill=0),  _convert_to_rgb,ToTensor(), SDNormalize()])
    img_transform=transforms.Compose([ResizeLongest(560, fill=0)])
    dino_img=DINO_transform(image).unsqueeze(0).to(torch.float16).to(device)
    sd_img = sd_transform(image).unsqueeze(0).to(torch.float16).to(device)

    # dino_feats_raw=dino.get_intermediate_layers(dino_img, reshape=True)[0].flatten(start_dim=-2).transpose(-2,-1)
    dino_feats_raw=dino.get_intermediate_layers(dino_img, reshape=True)[0].flatten(start_dim=-2).transpose(-2,-1)
    dino_feats = F.normalize(dino_feats_raw, dim=2)
    sim_dino = torch.einsum('bic,bjc->bij', dino_feats, dino_feats).squeeze(0)
    # sd preprocess
    # 1. 
    sd.forward_wo_preprocess(sd_img, "")
    vis_img=img_transform(image)
    self_att_raw = torch.cat([sd.attention_maps[idx] for idx in attention_layers_to_use]).float()

    self_att = self_att_raw / torch.amax(self_att_raw, dim=-2, keepdim=True) + 1e-5
    self_att = torch.where(self_att < 0.2, 0, self_att)
    self_att /= self_att.sum(dim=-1, keepdim=True) + 1e-5
    self_att = reduce(torch.matmul, self_att, torch.eye(self_att.shape[-1], device=self_att.device)).to(sim_dino.dtype)
    refined_sim_dino = self_att @ sim_dino @ self_att.transpose(0, 1)
    alpha = 0.8 
    refined_sim_dino = (1 - alpha) * sim_dino + alpha * refined_sim_dino

    output_dir = "sd_vis"
    if not os.path.exists(output_dir):
        os.mkdir(output_dir)

    token_choosen=(12, 20)

    visualize_sd_dino_att(
        ori_img=vis_img,                # [1,3,H,W]，归一化 0-1 float
        self_att=self_att,             # (4096,4096)
        sim_dino_soft=sim_dino,   # (4096,4096)
        sim_dino_refined=refined_sim_dino, # (4096,4096)
        token_choosen=token_choosen,
        filename=os.path.join(output_dir, f"vis.png"),
        attn_map_hw=(35, 35),
        vis_hw=(560, 560)
    )
    visualize_self_att_raw(
        ori_img=vis_img,
        self_att_raw=self_att_raw,
        token_choosen=token_choosen,
        output_dir=output_dir,
        attn_map_hw=(35, 35),
        vis_hw=(560, 560)
    )