#!/usr/bin/env python # coding: utf-8 # simple_image_data_modified.py - 支持多层级VGG特征提取 import os import torch import numpy as np from PIL import Image import matplotlib.pyplot as plt from tqdm import tqdm import torchvision.transforms as transforms import torchvision.datasets as datasets from torchvision.models import vgg16 import torch.nn.functional as F import random from collections import defaultdict import json import warnings from concurrent.futures import ThreadPoolExecutor, as_completed import threading import multiprocessing as mp import time import hashlib class VGGMultiLevelExtractor: """支持多层级VGG特征提取的模块""" # VGG16各层级配置 VGG_LEVELS = { 3: {'name': 'MaxPool_Block1', 'channels': 64, 'layer_idx': 3}, 8: {'name': 'MaxPool_Block2', 'channels': 128, 'layer_idx': 8}, 15: {'name': 'MaxPool_Block3', 'channels': 256, 'layer_idx': 15}, 22: {'name': 'MaxPool_Block4', 'channels': 512, 'layer_idx': 22}, 29: {'name': 'MaxPool_Block5', 'channels': 512, 'layer_idx': 29} } def __init__(self, device='cuda'): self.device = device self.vgg_model = None self.normalize = None self.vgg_input_size = 224 self._init_vgg_model() def _init_vgg_model(self): """初始化VGG模型""" print(f"Loading VGG16 model on {self.device}...") full_vgg = vgg16(pretrained=True) self.vgg_features = full_vgg.features.to(self.device) self.vgg_features.eval() # ImageNet标准归一化参数 self.normalize = transforms.Normalize( mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) print("VGG16 model loaded successfully!") print(f"Available levels: {list(self.VGG_LEVELS.keys())}") def extract_features_at_level(self, images, level=29): """在指定层级提取VGG特征 Args: images: 图片列表（PIL Images或路径） level: VGG层级（3, 8, 15, 22, 29） Returns: torch.Tensor: 提取的特征，形状为 [N, channels] """ if level not in self.VGG_LEVELS: raise ValueError(f"Unsupported level {level}. Available: {list(self.VGG_LEVELS.keys())}") level_info = self.VGG_LEVELS[level] layer_idx = level_info['layer_idx'] expected_channels = level_info['channels'] print(f"Extracting features at level {level} ({level_info['name']}) - {expected_channels} channels") # 预处理pipeline transform = transforms.Compose([ transforms.Resize((self.vgg_input_size, self.vgg_input_size)), transforms.ToTensor() ]) # 转换图片为张量 img_tensors = [] for img in images: if img is None: continue if isinstance(img, str): # 如果是路径，加载图片 img = Image.open(img).convert('RGB') if isinstance(img, Image.Image): # 确保是RGB模式 if img.mode != 'RGB': img = img.convert('RGB') img_tensor = transform(img) else: # 如果已经是tensor，需要resize到224x224 if img.dim() == 3: img_tensor = F.interpolate(img.unsqueeze(0), size=(self.vgg_input_size, self.vgg_input_size), mode='bilinear', align_corners=False).squeeze(0) else: img_tensor = img # 确保是RGB 3通道 if img_tensor.shape[0] == 1: img_tensor = img_tensor.repeat(3, 1, 1) elif img_tensor.shape[0] != 3: img_tensor = img_tensor[:3] img_tensors.append(img_tensor) if not img_tensors: return torch.zeros(0, expected_channels, device=self.device) batch_tensor = torch.stack(img_tensors).to(self.device) with torch.no_grad(): # ImageNet标准归一化 normalized_batch = torch.stack([self.normalize(img) for img in batch_tensor]) # 分批处理避免内存问题 process_batch_size = 16 features_list = [] for i in range(0, len(normalized_batch), process_batch_size): batch = normalized_batch[i:i+process_batch_size] # 前向传播到指定层 x = batch for idx, layer in enumerate(self.vgg_features): x = layer(x) if idx == layer_idx: break # 全局平均池化 batch_features = F.adaptive_avg_pool2d(x, (1, 1)) # [N, channels, 1, 1] batch_features = batch_features.view(batch_features.size(0), -1) # [N, channels] features_list.append(batch_features) if features_list: features = torch.cat(features_list, dim=0) else: features = torch.zeros(0, expected_channels, device=self.device) print(f"Extracted features shape: {features.shape}") return features class SimpleImageDataLoader: def __init__(self, dataset_type='imagenet100', dataset_path=None, samples_per_block=50): self.dataset_type = dataset_type self.dataset_path = dataset_path self.samples_per_block = samples_per_block self.class_info = {} self.total_blocks_per_class = {} # 🔑 关键修改：支持多层级VGG特征提取 self.vgg_extractor = None # 🔑 关键修改：分别存储train和val数据 self.train_class_info = {} self.val_class_info = {} self.train_blocks_per_class = {} self.val_blocks_per_class = {} # 初始化数据集信息 self._scan_dataset() def _scan_dataset(self): """扫描数据集，获取每个类别的图片路径和块数""" print(f"Scanning {self.dataset_type} dataset...") if self.dataset_type == 'imagenet100': self._scan_imagenet100() elif self.dataset_type == 'imagenet10': self._scan_imagenet10() elif self.dataset_type in ['cifar10', 'cifar100']: self._scan_cifar() elif self.dataset_type == 'folder': self._scan_folder() print(f"Found {len(self.class_info)} classes") for class_idx, info in list(self.class_info.items())[:5]: print(f" Class {class_idx}: {info['total_images']} images, {self.total_blocks_per_class[class_idx]} blocks") if len(self.class_info) > 5: print(f" ... and {len(self.class_info) - 5} more classes") def _scan_imagenet100(self): """修改后的ImageNet100扫描 - 分别存储train和val""" print(f"Scanning {self.dataset_type} dataset...") # 找到所有train文件夹和val文件夹 train_folders = [] val_folder = None print(self.dataset_path) if os.path.exists(self.dataset_path): for item in os.listdir(self.dataset_path): item_path = os.path.join(self.dataset_path, item) if os.path.isdir(item_path): if item.startswith('train.X'): train_folders.append(item_path) elif item == 'val.X': val_folder = item_path train_folders.sort() import pdb # pdb.set_trace() # 收集所有类别 all_classes = set() for train_folder in train_folders: if os.path.exists(train_folder): classes = [f for f in os.listdir(train_folder) if os.path.isdir(os.path.join(train_folder, f)) and f.startswith('n')] all_classes.update(classes) all_classes = sorted(list(all_classes)) # 🔑 关键修改：分别处理train和val数据 for class_idx, class_name in enumerate(all_classes): train_paths = [] val_paths = [] # 收集训练集图片 for train_folder in train_folders: class_path = os.path.join(train_folder, class_name) if os.path.exists(class_path): files = [f for f in os.listdir(class_path) if f.lower().endswith(('.jpg', '.jpeg', '.png'))] files.sort() paths = [os.path.join(class_path, f) for f in files] train_paths.extend(paths) # 收集验证集图片 if val_folder: val_class_path = os.path.join(val_folder, class_name) if os.path.exists(val_class_path): files = [f for f in os.listdir(val_class_path) if f.lower().endswith(('.jpg', '.jpeg', '.png'))] files.sort() paths = [os.path.join(val_class_path, f) for f in files] val_paths.extend(paths) # 🔑 分别存储训练集和验证集信息 self.train_class_info[class_idx] = { 'class_name': class_name, 'image_paths': train_paths, 'total_images': len(train_paths) } self.train_blocks_per_class[class_idx] = max(1, len(train_paths) // self.samples_per_block) self.val_class_info[class_idx] = { 'class_name': class_name, 'image_paths': val_paths, 'total_images': len(val_paths) } self.val_blocks_per_class[class_idx] = max(1, len(val_paths) // self.samples_per_block) # 为了向后兼容，也存储混合信息 all_paths = train_paths + val_paths self.class_info[class_idx] = { 'class_name': class_name, 'image_paths': all_paths, 'total_images': len(all_paths) } self.total_blocks_per_class[class_idx] = max(1, len(all_paths) // self.samples_per_block) print(f"Found {len(all_classes)} classes") print(f"Train images per class (first 3): {[self.train_class_info[i]['total_images'] for i in range(min(3, len(all_classes)))]}") print(f"Val images per class (first 3): {[self.val_class_info[i]['total_images'] for i in range(min(3, len(all_classes)))]}") def _scan_imagenet10(self): """扫描ImageNet10数据集""" # 复用imagenet100的逻辑 self._scan_imagenet100() def _scan_cifar(self): """扫描CIFAR数据集""" try: if self.dataset_type == 'cifar10': dataset = datasets.CIFAR10(self.dataset_path or './data', train=True, download=True) num_classes = 10 else: dataset = datasets.CIFAR100(self.dataset_path or './data', train=True, download=True) num_classes = 100 # 为CIFAR创建虚拟的类别信息 for class_idx in range(num_classes): class_indices = [i for i, (_, label) in enumerate(dataset) if label == class_idx] self.class_info[class_idx] = { 'class_name': f'cifar_class_{class_idx}', 'image_paths': class_indices, # 存储索引而不是路径 'total_images': len(class_indices) } self.total_blocks_per_class[class_idx] = max(1, len(class_indices) // self.samples_per_block) except Exception as e: print(f"Error scanning CIFAR dataset: {e}") def _scan_folder(self): """扫描文件夹结构的数据集 - 手动分割train/val""" if not os.path.exists(self.dataset_path): print(f"Dataset path does not exist: {self.dataset_path}") return # 假设文件夹结构为: dataset_path/class_name/images class_folders = [f for f in os.listdir(self.dataset_path) if os.path.isdir(os.path.join(self.dataset_path, f))] class_folders.sort() # 设置分割比例 (可以作为参数传入) train_ratio = 0.8 # 80% 用于训练，20% 用于验证 for class_idx, class_name in enumerate(class_folders): class_path = os.path.join(self.dataset_path, class_name) files = [f for f in os.listdir(class_path) if f.lower().endswith(('.jpg', '.jpeg', '.png'))] files.sort() image_paths = [os.path.join(class_path, f) for f in files] # 🔑 手动分割数据 total_images = len(image_paths) train_split_idx = int(total_images * train_ratio) # 分割图片路径 train_paths = image_paths[:train_split_idx] val_paths = image_paths[train_split_idx:] # 🔑 分别存储训练集和验证集信息 self.train_class_info[class_idx] = { 'class_name': class_name, 'image_paths': train_paths, 'total_images': len(train_paths) } self.train_blocks_per_class[class_idx] = max(1, len(train_paths) // self.samples_per_block) self.val_class_info[class_idx] = { 'class_name': class_name, 'image_paths': val_paths, 'total_images': len(val_paths) } self.val_blocks_per_class[class_idx] = max(1, len(val_paths) // self.samples_per_block) # 为了向后兼容，也存储混合信息 self.class_info[class_idx] = { 'class_name': class_name, 'image_paths': image_paths, 'total_images': len(image_paths) } self.total_blocks_per_class[class_idx] = max(1, len(image_paths) // self.samples_per_block) print(f"Found {len(class_folders)} classes") print(f"Train/Val split ratio: {train_ratio:.1f}/{1-train_ratio:.1f}") if class_folders: print(f"Train images per class (first 3): {[self.train_class_info[i]['total_images'] for i in range(min(3, len(class_folders)))]}") print(f"Val images per class (first 3): {[self.val_class_info[i]['total_images'] for i in range(min(3, len(class_folders)))]}") def _get_current_class_info(self, is_test_mode=False): """根据是否为测试模式返回相应的class_info""" if is_test_mode: return self.val_class_info, self.val_blocks_per_class else: return self.train_class_info, self.train_blocks_per_class def _get_vgg_cache_dir(self, args, level=29): """获取VGG特征缓存目录 - 支持多层级""" if not args or not hasattr(args, 'dataset_path'): return None base_path = getattr(args, 'dataset_path', '') if not base_path: return None cache_components = [f'vgg_level{level}'] if hasattr(args, 'embedding_noise_level') and args.embedding_noise_level > 0: cache_components.append(f"noise{args.embedding_noise_level}") if hasattr(args, 'normalize_features') and args.normalize_features: cache_components.append("norm") cache_suffix = '_'.join(cache_components) # print(cache_dir) cache_dir = f"{base_path}_{cache_suffix}" cache_dir = cache_dir.replace('/hpc/group/chenglab/zl310/spring2025_projects/InterpStableDiffusion/results/interp_res_slerp','/work/jf381/data/icl_jay/interp_res_slerp') return cache_dir def get_vgg_features(self, images, device='cuda', args=None, level=29): """获取VGG特征 - 支持多层级和缓存机制""" if self.vgg_extractor is None: self.vgg_extractor = VGGMultiLevelExtractor(device=device) if not images: expected_channels = VGGMultiLevelExtractor.VGG_LEVELS[level]['channels'] return torch.zeros(0, expected_channels, device=device) # 检查是否有缓存目录配置 cache_dir = self._get_vgg_cache_dir(args, level) if cache_dir: return self._get_vgg_features_with_cache(images, device, args, cache_dir, level) else: return self._extract_vgg_features(images, device, args, level) def _get_image_cache_path(self, image_path, cache_dir, level=29): """获取单张图片的缓存路径 - 包含层级信息""" # 使用图片路径和层级的hash作为缓存文件名 content = f"{image_path}_level{level}" path_hash = hashlib.md5(content.encode()).hexdigest() return os.path.join(cache_dir, f"{path_hash}.pt") def _get_vgg_features_with_cache(self, images, device, args, cache_dir, level=29): """带缓存的VGG特征提取 - 支持多层级""" os.makedirs(cache_dir, exist_ok=True) # 分离图片路径和PIL对象 image_paths = [] pil_images = [] for img in images: if isinstance(img, str): # 如果是路径字符串 image_paths.append(img) try: pil_img = Image.open(img).convert('RGB') pil_images.append(pil_img) except: pil_images.append(None) elif hasattr(img, 'filename'): # PIL Image对象可能有filename属性 image_paths.append(getattr(img, 'filename', '')) pil_images.append(img) else: # 没有路径信息，直接处理 image_paths.append('') pil_images.append(img) # 检查缓存 cached_features = [] uncached_indices = [] uncached_images = [] for i, (img_path, pil_img) in enumerate(zip(image_paths, pil_images)): if img_path and os.path.exists(img_path): cache_path = self._get_image_cache_path(img_path, cache_dir, level) if os.path.exists(cache_path): try: # 读取缓存的特征 cached_feature = torch.load(cache_path, map_location=device) cached_features.append((i, cached_feature)) continue except: pass # 需要重新计算 uncached_indices.append(i) uncached_images.append(pil_img) # 计算未缓存的特征 if uncached_images: level_name = VGGMultiLevelExtractor.VGG_LEVELS[level]['name'] print(f"Computing VGG level {level} ({level_name}) features for {len(uncached_images)} uncached images...") new_features = self._extract_vgg_features(uncached_images, device, args, level) # 保存新计算的特征到缓存 for j, (original_idx, feature) in enumerate(zip(uncached_indices, new_features)): img_path = image_paths[original_idx] if img_path and os.path.exists(img_path): cache_path = self._get_image_cache_path(img_path, cache_dir, level) try: torch.save(feature.cpu(), cache_path) except Exception as e: print(f"Failed to cache feature for {img_path}: {e}") else: expected_channels = VGGMultiLevelExtractor.VGG_LEVELS[level]['channels'] new_features = torch.zeros(0, expected_channels, device=device) # 合并缓存的和新计算的特征 expected_channels = VGGMultiLevelExtractor.VGG_LEVELS[level]['channels'] all_features = torch.zeros(len(images), expected_channels, device=device) # 填入缓存的特征 for original_idx, cached_feature in cached_features: all_features[original_idx] = cached_feature.to(device) # 填入新计算的特征 new_feature_idx = 0 for original_idx in uncached_indices: if new_feature_idx < len(new_features): all_features[original_idx] = new_features[new_feature_idx] new_feature_idx += 1 return all_features def _extract_vgg_features(self, images, device, args, level=29): """实际的VGG特征提取逻辑 - 支持多层级""" # 使用多层级提取器 features = self.vgg_extractor.extract_features_at_level(images, level) # 应用嵌入噪声 if args and hasattr(args, 'embedding_noise_level') and args.embedding_noise_level > 0: noise = torch.randn_like(features) * args.embedding_noise_level features = features + noise # 特征归一化 if args and hasattr(args, 'normalize_features') and args.normalize_features: features = F.normalize(features, p=2, dim=1) # Feature dropout（训练时使用） if args and hasattr(args, 'feature_dropout') and args.feature_dropout > 0: if hasattr(self.vgg_extractor, 'training') and self.vgg_extractor.training: features = F.dropout(features, p=args.feature_dropout, training=True) return features def load_images_for_class_block(self, class_idx, block_idx, args=None, is_test_mode=None): """加载指定类别和块的图片 - 支持train/val切换""" # 🔑 如果没有传入 is_test_mode，从 args 中推断 if is_test_mode is None: # 如果args中有epoch信息，可以从中推断 if args and hasattr(args, 'current_epoch'): is_test_mode = args.current_epoch >= 99999 else: is_test_mode = False # 默认使用训练模式 # 根据测试模式选择数据源 if hasattr(self, 'train_class_info') and hasattr(self, 'val_class_info'): # 新版本：支持train/val分离 class_info = self.val_class_info if is_test_mode else self.train_class_info else: # 原版本：使用混合数据 class_info = self.class_info import pdb # pdb.set_trace() if class_idx not in class_info: return [] class_data = class_info[class_idx] if self.dataset_type in ['cifar10', 'cifar100']: return self._load_cifar_block(class_idx, block_idx, is_test_mode) else: # ImageNet类型数据集或folder image_paths = class_data['image_paths'] start_idx = block_idx * self.samples_per_block end_idx = min(start_idx + self.samples_per_block, len(image_paths)) if start_idx >= len(image_paths): # 如果块索引超出范围，使用最后一块 start_idx = max(0, len(image_paths) - self.samples_per_block) end_idx = len(image_paths) block_paths = image_paths[start_idx:end_idx] # 检查是否使用VGG特征且启用缓存 use_vgg = getattr(args, 'use_vgg_features', False) if args else False cache_dir = self._get_vgg_cache_dir(args, getattr(args, 'vgg_level', 29)) if args else None if use_vgg and cache_dir: # 返回图片路径而不是PIL对象，让VGG特征提取器处理缓存 return block_paths else: # 传统模式：加载PIL图片 images = [] for img_path in block_paths: try: img = Image.open(img_path).convert('RGB') # 应用图像增强 if args and hasattr(args, 'image_noise_level') and args.image_noise_level > 0: aug_type = getattr(args, 'image_aug_type', 'pixel') img_tensor = self.apply_image_augmentation(img, args.image_noise_level, aug_type) # 转回PIL Image img = transforms.ToPILImage()(img_tensor) images.append(img) except Exception as e: print(f"Error loading {img_path}: {e}") continue return images def _load_cifar_block(self, class_idx, block_idx, is_test_mode=False): """加载CIFAR数据集的指定块""" try: if self.dataset_type == 'cifar10': dataset = datasets.CIFAR10(self.dataset_path or './data', train=not is_test_mode, download=True) else: dataset = datasets.CIFAR100(self.dataset_path or './data', train=not is_test_mode, download=True) # 收集该类别的所有样本索引 class_indices = [i for i, (_, label) in enumerate(dataset) if label == class_idx] # 选择指定块 start_idx = block_idx * self.samples_per_block end_idx = min(start_idx + self.samples_per_block, len(class_indices)) if start_idx >= len(class_indices): start_idx = max(0, len(class_indices) - self.samples_per_block) end_idx = len(class_indices) selected_indices = class_indices[start_idx:end_idx] # 加载图片 images = [] for idx in selected_indices: img, _ = dataset[idx] images.append(img) return images except Exception as e: print(f"Error loading CIFAR block: {e}") return [] def apply_image_augmentation(self, image, noise_level=0.0, aug_type='pixel'): """图像增强函数""" if noise_level <= 0: return image if isinstance(image, Image.Image): image = transforms.ToTensor()(image) if aug_type == 'pixel': # 像素级高斯噪声 noise = torch.randn_like(image) * noise_level image = torch.clamp(image + noise, 0, 1) elif aug_type == 'color': # 颜色域增强 rand_val = random.random() if rand_val < 0.33: # Brightness brightness_factor = 1 + (random.random() - 0.5) * noise_level image = torch.clamp(image * brightness_factor, 0, 1) elif rand_val < 0.66: # Contrast mean_val = image.mean() contrast_factor = 1 + (random.random() - 0.5) * noise_level image = torch.clamp((image - mean_val) * contrast_factor + mean_val, 0, 1) else: # Saturation if image.shape[0] == 3: gray = 0.299 * image[0] + 0.587 * image[1] + 0.114 * image[2] saturation_factor = 1 + (random.random() - 0.5) * noise_level image = torch.clamp(gray.unsqueeze(0) + (image - gray.unsqueeze(0)) * saturation_factor, 0, 1) return image def get_epoch_mapping(self, epoch, batch_size, class_combination_seed=42, is_test_mode=False): """修改后的epoch映射 - 支持train/val切换""" # import pdb # pdb.set_trace() # 🔑 根据测试模式选择相应的数据信息 class_info, blocks_per_class = self._get_current_class_info(is_test_mode) num_classes = len(class_info) # 设置随机种子确保可重复性 random.seed(class_combination_seed + epoch) np.random.seed(class_combination_seed + epoch) # 计算总的类别组合数 total_class_combinations = num_classes * (num_classes - 1) # 计算最大块数 max_blocks = max(blocks_per_class.values()) if blocks_per_class else 1 batch_mappings = [] # print(total_class_combinations,global_batch_id) for batch_idx in range(batch_size): global_batch_id = epoch * batch_size + batch_idx print(total_class_combinations,global_batch_id) # 确定当前是第几轮遍历（第几块） block_round = global_batch_id // total_class_combinations # 确定在当前轮中是第几个类别组合 combination_idx = global_batch_id % total_class_combinations # 将组合索引转换为具体的类别对 class1 = combination_idx // (num_classes - 1) class2_offset = combination_idx % (num_classes - 1) class2 = class2_offset if class2_offset < class1 else class2_offset + 1 # 确定每个类别使用第几块 block1 = block_round % blocks_per_class.get(class1, 1) block2 = block_round % blocks_per_class.get(class2, 1) batch_mappings.append({ 'batch_idx': batch_idx, 'global_batch_id': global_batch_id, 'class1': class1, 'class2': class2, 'block1': block1, 'block2': block2, 'block_round': block_round, 'is_test_mode': is_test_mode }) return batch_mappings def generate_batch_data(self, epoch, batch_size, device='cuda', args=None): use_vgg_features = getattr(args, 'use_vgg_features', False) vgg_level = getattr(args, 'vgg_level', 29) # 🔑 新增：支持VGG层级选择 downsample_size = getattr(args, 'downsample_size', 32) scale_rbf = getattr(args, 'scale_rbf', 1.0) k_nn = getattr(args, 'k_nn', 10) class_combination_seed = getattr(args, 'class_combination_seed', 42) n_samples_per_class = getattr(args, 'n_samples_per_class', 50) # 🔑 判断是否为测试模式 is_test_mode = epoch >= 99999 # 获取epoch的映射 batch_mappings = self.get_epoch_mapping(epoch, batch_size, class_combination_seed, is_test_mode) data_source = "VALIDATION" if is_test_mode else "TRAINING" level_info = VGGMultiLevelExtractor.VGG_LEVELS[vgg_level] print(f"Loading epoch {epoch} from {data_source} set...") if use_vgg_features: print(f"Using VGG level {vgg_level} ({level_info['name']}) with {level_info['channels']} channels") all_raw_data = [] all_labels = [] all_laplacians = [] all_adjacencies = [] for mapping in tqdm(batch_mappings, desc=f"Loading {data_source} epoch {epoch}"): class1, class2 = mapping['class1'], mapping['class2'] block1, block2 = mapping['block1'], mapping['block2'] # 🔑 传递测试模式标志 images1 = self.load_images_for_class_block(class1, block1, args, is_test_mode) images2 = self.load_images_for_class_block(class2, block2, args, is_test_mode) # 限制每个类别的样本数 if len(images1) > n_samples_per_class: images1 = images1[:n_samples_per_class] if len(images2) > n_samples_per_class: images2 = images2[:n_samples_per_class] # 合并图片和标签 all_images = images1 + images2 all_batch_labels = [0] * len(images1) + [1] * len(images2) import pdb # pdb.set_trace() if len(all_images) == 0: # 创建虚拟数据 if use_vgg_features: img_dim = level_info['channels'] else: img_dim = 32 * 32 * 3 raw_data = torch.zeros(100, img_dim, device=device) labels = torch.randint(0, 2, (100,), dtype=torch.long, device=device) laplacian = torch.eye(100, device=device) adjacency = torch.eye(100, device=device) * 1e-6 else: # 特征提取 if use_vgg_features: # 🔑 使用指定层级的VGG特征提取（带缓存） print(f"Extracting VGG level {vgg_level} features for batch (class {class1} & {class2})...") raw_data = self.get_vgg_features(all_images, device, args, level=vgg_level) else: # 传统像素特征 processed_images = [] for img in all_images: if isinstance(img, str): # 如果是路径，加载图片 img = Image.open(img).convert('RGB') if isinstance(img, Image.Image): if downsample_size and downsample_size != img.size[0]: img = img.resize((downsample_size, downsample_size), Image.LANCZOS) img_tensor = transforms.ToTensor()(img) else: img_tensor = img if img_tensor.shape[0] == 3: img_tensor = transforms.functional.rgb_to_grayscale(img_tensor) processed_images.append(img_tensor) if processed_images: img_data_tensor = torch.stack(processed_images) raw_data = img_data_tensor.view(img_data_tensor.shape[0], -1).to(device) else: raw_data = torch.zeros(0, 32*32, device=device) # 填充到100个样本 current_size = raw_data.shape[0] if current_size < 100: padding_size = 100 - current_size feature_dim = raw_data.shape[1] padding_data = torch.zeros(padding_size, feature_dim, device=device) raw_data = torch.cat([raw_data, padding_data], dim=0) padding_labels = torch.randint(0, 2, (padding_size,), dtype=torch.long, device=device) all_batch_labels.extend(padding_labels.tolist()) elif current_size > 100: raw_data = raw_data[:100] all_batch_labels = all_batch_labels[:100] labels = torch.tensor(all_batch_labels, dtype=torch.long, device=device) # 计算邻接矩阵和拉普拉斯矩阵 distances = torch.cdist(raw_data, raw_data, p=2) adjacency = torch.exp(-scale_rbf * distances ** 2) # k近邻 adjacency_copy = adjacency.clone() adjacency_copy.fill_diagonal_(0.0) _, nn_indices = torch.topk(adjacency_copy, k_nn, dim=1) # 构建稀疏邻接矩阵 adj_matrix = torch.zeros_like(adjacency, device=device) batch_indices = torch.arange(100, device=device).unsqueeze(1).expand(-1, k_nn) adj_matrix[batch_indices, nn_indices] = adjacency[batch_indices, nn_indices] adj_matrix[nn_indices, batch_indices] = adjacency[nn_indices, batch_indices] adj_matrix.fill_diagonal_(1e-6) adjacency = adj_matrix # 拉普拉斯矩阵 degree = adjacency.sum(dim=1) degree = torch.clamp(degree, min=1e-10) D_inv_sqrt = torch.diag(degree.pow(-0.5)).to(device) laplacian = torch.eye(100, device=device) - D_inv_sqrt @ adjacency @ D_inv_sqrt all_raw_data.append(raw_data) all_labels.append(labels) all_laplacians.append(laplacian) all_adjacencies.append(adjacency) # 堆叠所有batch数据 final_raw_data = torch.stack(all_raw_data, dim=0) final_labels = torch.stack(all_labels, dim=0) final_laplacians = torch.stack(all_laplacians, dim=0) final_adjacencies = torch.stack(all_adjacencies, dim=0) return final_raw_data, final_laplacians, final_labels, final_adjacencies # 缓存管理工具函数 def clean_vgg_cache(dataset_path, level=29, noise_level=None): """清理VGG特征缓存""" cache_components = [f'vgg_level{level}'] if noise_level is not None: cache_components.append(f"noise{noise_level}") cache_dir = f"{dataset_path}_{'_'.join(cache_components)}" if os.path.exists(cache_dir): import shutil shutil.rmtree(cache_dir) print(f"Cleaned VGG level {level} cache: {cache_dir}") else: print(f"Cache directory does not exist: {cache_dir}") def get_vgg_cache_info(dataset_path, level=29, noise_level=None): """获取VGG缓存信息""" cache_components = [f'vgg_level{level}'] if noise_level is not None: cache_components.append(f"noise{noise_level}") cache_dir = f"{dataset_path}_{'_'.join(cache_components)}" if not os.path.exists(cache_dir): return {"exists": False, "path": cache_dir, "level": level} # 统计缓存文件 cache_files = [f for f in os.listdir(cache_dir) if f.endswith('.pt')] total_size = sum(os.path.getsize(os.path.join(cache_dir, f)) for f in cache_files) level_info = VGGMultiLevelExtractor.VGG_LEVELS[level] return { "exists": True, "path": cache_dir, "level": level, "level_name": level_info['name'], "channels": level_info['channels'], "num_files": len(cache_files), "total_size_mb": total_size / (1024 * 1024), "files": cache_files[:10] # 显示前10个文件名 } # 主要接口函数，保持与原版本兼容 def get_or_generate_data_image( epoch, batch_size, n_samples, scale_rbf, k_nn, device, label_percent, context_size, k_feat, data_dir="./cached_data", image_dir="./data", force=False, downsample_size=32, scale_factor=1.0, pixel_scale_factor=None, manifold_list=None, prod_threshold=None, args=None ): """ 主数据生成函数 - 保持原有接口兼容性，新增VGG层级支持 """ # os.makedirs(data_dir, exist_ok=True) # 获取参数 dataset_type = getattr(args, 'dataset_type', 'imagenet100') use_vgg_features = getattr(args, 'use_vgg_features', False) vgg_level = getattr(args, 'vgg_level', 29) # 🔑 新增：VGG层级参数 class_combination_seed = getattr(args, 'class_combination_seed', 42) test_class_combination_seed = getattr(args, 'test_class_combination_seed', 12345) # 判断是否为测试阶段 is_test = epoch >= 99999 seed_to_use = test_class_combination_seed if is_test else class_combination_seed # 生成缓存文件名 cache_components = [ f"epoch_{epoch}", f"mode_{'test' if is_test else 'train'}", f"bs{batch_size}", f"seed{seed_to_use}", f"dataset{dataset_type}", f"vgg_level{vgg_level}" if use_vgg_features else "pixel", # 🔑 层级信息 f"ds{downsample_size}" if downsample_size else "nods", f"scale{scale_factor}" if scale_factor != 1.0 else "noscale", f"scale_rbf{scale_rbf}" ] # 添加噪声相关的缓存标识 if hasattr(args, 'image_noise_level') and args.image_noise_level > 0: cache_components.append(f"imgnoise{args.image_noise_level}_{getattr(args, 'image_aug_type', 'pixel')}") if hasattr(args, 'embedding_noise_level') and args.embedding_noise_level > 0: cache_components.append(f"embnoise{args.embedding_noise_level}") cache_file = os.path.join(data_dir, f"data_{'_'.join(cache_components)}.pt") print(f"Cache file: {cache_file}") # 检查缓存 if not force and os.path.exists(cache_file): try: print(f"Loading cached data from {cache_file}") cached_data = torch.load(cache_file, map_location=device) raw_data = cached_data['raw_data'].to(device) real_lap = cached_data['real_lap'].to(device) labels_tensor = cached_data['labels_tensor'].to(device) real_adj = cached_data['real_adj'].to(device) real_ev = cached_data['real_ev'].to(device) # 生成新的索引 n_labeled = int(100 * label_percent / 100) labeled_indices = torch.stack([torch.randperm(100, device=device)[:n_labeled] for _ in range(batch_size)]) context_indices = torch.stack([torch.randperm(n_labeled, device=device)[:context_size] for _ in range(batch_size)]) all_indices = torch.arange(n_labeled, device=device).expand(batch_size, n_labeled) mask = torch.zeros(batch_size, n_labeled, dtype=torch.bool, device=device) for i in range(batch_size): mask[i].scatter_(0, context_indices[i], True) query_indices = all_indices[~mask].view(batch_size, n_labeled - context_size) return raw_data, real_lap, labels_tensor, real_adj, labeled_indices, context_indices, query_indices, real_ev except Exception as e: print(f"Failed to load cached data: {e}") # 创建数据加载器 if args: args_copy = type(args)() for attr in dir(args): if not attr.startswith('_'): setattr(args_copy, attr, getattr(args, attr)) args_copy.class_combination_seed = seed_to_use args = args_copy loader = SimpleImageDataLoader( dataset_type=dataset_type, dataset_path=image_dir, samples_per_block=getattr(args, 'n_samples_per_class', 50) if args else 50 ) # 生成数据 start_time = time.time() if use_vgg_features: level_info = VGGMultiLevelExtractor.VGG_LEVELS[vgg_level] print(f"Generating data for epoch {epoch} using VGG level {vgg_level} ({level_info['name']})...") else: print(f"Generating data for epoch {epoch}...") raw_data, real_lap, labels_tensor, real_adj = loader.generate_batch_data( epoch=epoch, batch_size=batch_size, device=device, args=args ) import pdb # pdb.set_trace() # 计算特征向量 real_eigs = [] for b in range(batch_size): try: _, vecs = torch.linalg.eigh(real_lap[b]) real_eigs.append(vecs[:, :k_feat]) except Exception: real_eigs.append(torch.eye(100, k_feat, device=device)) real_ev = torch.stack(real_eigs, dim=0) # 生成索引 n_labeled = int(100 * label_percent / 100) labeled_indices = torch.stack([torch.randperm(100, device=device)[:n_labeled] for _ in range(batch_size)]) context_indices = torch.stack([torch.randperm(n_labeled, device=device)[:context_size] for _ in range(batch_size)]) all_indices = torch.arange(n_labeled, device=device).expand(batch_size, n_labeled) mask = torch.zeros(batch_size, n_labeled, dtype=torch.bool, device=device) for i in range(batch_size): mask[i].scatter_(0, context_indices[i], True) query_indices = all_indices[~mask].view(batch_size, n_labeled - context_size) generation_time = time.time() - start_time print(f"Data generation completed in {generation_time:.2f} seconds") # 缓存数据 cached_data = { 'raw_data': raw_data.cpu(), 'real_lap': real_lap.cpu(), 'labels_tensor': labels_tensor.cpu(), 'real_adj': real_adj.cpu(), 'real_ev': real_ev.cpu(), 'generation_time': generation_time, 'dataset_type': dataset_type, 'epoch': epoch, 'seed_used': seed_to_use, 'vgg_level': vgg_level if use_vgg_features else None # 🔑 保存层级信息 } # rsync -acz /hpc/group/chenglab/zl310/spring2025_projects/InterpStableDiffusion/results/interp_res_swissroll /work/jf381/data/icl_jay/interp_res_swissroll try: torch.save(cached_data, cache_file) print(f"Data cached to {cache_file}") except Exception as e: print(f"Failed to cache data: {e}") return raw_data, real_lap, labels_tensor, real_adj, labeled_indices, context_indices, query_indices, real_ev