# utils/data_preprocessing.py """ 数据预处理和诊断工具 用于检查和修复foundation model特征 """ import torch import numpy as np import pandas as pd from torch.utils.data import DataLoader from tqdm import tqdm def diagnose_dataset(dataset, dataloader, device='cuda'): """ 诊断数据集的各种统计信息 """ print("=" * 80) print("数据集诊断报告") print("=" * 80) # 1. 基本统计 print(f"\n[1] 基本信息") print(f" 样本数量: {len(dataset)}") # 2. 收集所有数据 all_pre = [] all_post = [] all_survival = [] all_events = [] print(f"\n[2] 收集数据中...") for batch in tqdm(dataloader): all_pre.append(batch['pre_latent']) all_post.append(batch['post_latent']) all_survival.append(batch['survival_time']) all_events.append(batch['event_indicator']) all_pre = torch.cat(all_pre, dim=0) # [N, 4, 767] all_post = torch.cat(all_post, dim=0) # [N, 4, 767] all_survival = torch.cat(all_survival) # [N] all_events = torch.cat(all_events) # [N] # 3. Latent特征统计 print(f"\n[3] Latent特征统计") print(f" Pre-latent:") print(f" 均值: {all_pre.mean():.6f}") print(f" 标准差: {all_pre.std():.6f}") print(f" 最小值: {all_pre.min():.6f}") print(f" 最大值: {all_pre.max():.6f}") print(f" 范围: [{all_pre.min():.3f}, {all_pre.max():.3f}]") print(f"\n Post-latent:") print(f" 均值: {all_post.mean():.6f}") print(f" 标准差: {all_post.std():.6f}") print(f" 最小值: {all_post.min():.6f}") print(f" 最大值: {all_post.max():.6f}") print(f"\n Pre vs Post差异:") diff = (all_post - all_pre).abs() print(f" 平均L1差异: {diff.mean():.6f}") print(f" 最大差异: {diff.max():.6f}") print(f" 相似度: {1 - diff.mean()/all_pre.abs().mean():.2%}") # 4. 生存数据统计 print(f"\n[4] 生存数据统计") print(f" 生存时间:") print(f" 均值: {all_survival.mean():.2f} 天") print(f" 中位数: {all_survival.median():.2f} 天") print(f" 范围: [{all_survival.min():.1f}, {all_survival.max():.1f}]") print(f"\n 事件指示器:") print(f" 事件发生率: {all_events.float().mean():.2%}") print(f" 事件数量: {all_events.sum().item()}") print(f" 删失数量: {(1-all_events).sum().item()}") # 5. 5年生存统计 five_year_survived = ((all_survival > 1825) & (all_events == 0)) print(f"\n 5年生存:") print(f" 5年生存人数: {five_year_survived.sum().item()}") print(f" 5年生存率: {five_year_survived.float().mean():.2%}") print(f" ⚠️ 5年内死亡: {((all_survival <= 1825) & (all_events == 1)).sum().item()}") # 6. 潜在问题检测 print(f"\n[5] 潜在问题检测") issues = [] if all_pre.std() < 0.01: issues.append("⚠️ Pre-latent标准差过小，特征可能未归一化") if all_pre.abs().max() > 100: issues.append("⚠️ Pre-latent数值范围过大，建议归一化") if diff.mean() < 0.01: issues.append("⚠️ Pre和Post几乎相同，模型难以学习") if all_events.float().mean() < 0.1 or all_events.float().mean() > 0.9: issues.append(f"⚠️ 事件发生率极端 ({all_events.float().mean():.2%})，可能导致训练不稳定") if five_year_survived.sum() < 10: issues.append("⚠️ 5年生存样本过少，AUC可能无法计算") if len(issues) == 0: print(" ✓ 未发现明显问题") else: for issue in issues: print(f" {issue}") print("\n" + "=" * 80) return { 'pre_mean': all_pre.mean().item(), 'pre_std': all_pre.std().item(), 'post_mean': all_post.mean().item(), 'post_std': all_post.std().item(), 'event_rate': all_events.float().mean().item(), 'five_year_survival_rate': five_year_survived.float().mean().item() } def normalize_latent_features(dataset, method='standardize'): """ 归一化latent特征 Args: dataset: GliomaAllPairsTextDataset method: 'standardize' (z-score) or 'minmax' (0-1) Returns: mean, std (for standardize) or min, max (for minmax) """ print(f"计算归一化参数（方法: {method}）...") all_features = [] for i in tqdm(range(len(dataset))): sample = dataset[i] all_features.append(sample['pre_latent']) all_features.append(sample['post_latent']) all_features = torch.stack(all_features, dim=0) # [2N, 4, 767] if method == 'standardize': mean = all_features.mean(dim=(0, 1), keepdim=True) # [1, 1, 767] std = all_features.std(dim=(0, 1), keepdim=True) # [1, 1, 767] std = torch.clamp(std, min=1e-6) # 避免除0 print(f"均值范围: [{mean.min():.4f}, {mean.max():.4f}]") print(f"标准差范围: [{std.min():.4f}, {std.max():.4f}]") return mean.squeeze(), std.squeeze() elif method == 'minmax': min_val = all_features.min(dim=0, keepdim=True)[0].min(dim=1, keepdim=True)[0] max_val = all_features.max(dim=0, keepdim=True)[0].max(dim=1, keepdim=True)[0] print(f"最小值: {min_val.min():.4f}") print(f"最大值: {max_val.max():.4f}") return min_val.squeeze(), max_val.squeeze() class NormalizedDataset(torch.utils.data.Dataset): """ 带归一化的数据集包装器 """ def __init__(self, base_dataset, mean, std, method='standardize'): self.base_dataset = base_dataset self.mean = mean self.std = std self.method = method def __len__(self): return len(self.base_dataset) def __getitem__(self, idx): sample = self.base_dataset[idx] # 归一化 if self.method == 'standardize': sample['pre_latent'] = (sample['pre_latent'] - self.mean) / self.std sample['post_latent'] = (sample['post_latent'] - self.mean) / self.std elif self.method == 'minmax': sample['pre_latent'] = (sample['pre_latent'] - self.mean) / (self.std - self.mean + 1e-6) sample['post_latent'] = (sample['post_latent'] - self.mean) / (self.std - self.mean + 1e-6) return sample # 使用示例 if __name__ == "__main__": from dataset.dataset_glioma_all_pairs_text import GliomaAllPairsTextDataset, Config # 加载数据 cfg = Config( timeline_json="dataset/MU-Glioma-Post/mu_glioma_post_timeline_with_survival.json", features_csv="./dataset/features_output.csv", take_dims=767 ) dataset = GliomaAllPairsTextDataset(cfg) dataloader = DataLoader(dataset, batch_size=32, num_workers=0) # 诊断 stats = diagnose_dataset(dataset, dataloader) # 归一化 mean, std = normalize_latent_features(dataset, method='standardize') normalized_dataset = NormalizedDataset(dataset, mean, std) print("\n✓ 归一化完成！")