PyTorch中数据加载器错误的报错与修复指南

FileNotFoundError: [Errno 2] No such file or directory: 'data/train'

import os

# 使用绝对路径
data_dir = os.path.abspath("data/train")
if not os.path.exists(data_dir):
    raise FileNotFoundError(f"路径 {data_dir} 不存在")

# 动态路径构建
base_dir = os.path.dirname(os.path.abspath(__file__))
data_path = os.path.join(base_dir, "data", "train")

RuntimeError: DataLoader worker (pid 4499) is killed by signal: Segmentation fault

import torch
from torch.utils.data import DataLoader

# 方法1：禁用多进程
dataloader = DataLoader(dataset, batch_size=32, num_workers=0)

# 方法2：设置进程启动方式
import multiprocessing as mp
mp.set_start_method('spawn')
dataloader = DataLoader(dataset, batch_size=32, num_workers=4)

RuntimeError: Expected 4-dimensional input for 4-dimensional weight [64, 3, 7, 7]

from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize(256),
    transforms.ToTensor(),  # 转换为CHW格式的Tensor
    transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                         std=[0.229, 0.224, 0.225])
])

dataset = MyDataset(transform=transform)

# 方法1：分块加载
from torch.utils.data import IterableDataset

class LargeDataset(IterableDataset):
    def __iter__(self):
        for i in range(1000):
            # 动态加载单个样本
            yield torch.randn(3, 224, 224)

# 方法2：使用内存映射
import numpy as np
data = np.memmap("large_data.dat", dtype='float32', mode='r')

from torch.utils.data import get_worker_info

def __getitem__(self, idx):
    worker_info = get_worker_info()
    if worker_info is not None:
        print(f"Worker {worker_info.id} 加载索引 {idx}")
    return self.data[idx]

RuntimeError: Cannot re-initialize CUDA in forked subprocess

# 主程序入口保护
if __name__ == '__main__':
    # 禁用CUDA多进程初始化
    torch.multiprocessing.set_sharing_strategy('file_system')
    
    # 显式指定设备
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    
    # 加载数据
    dataloader = DataLoader(dataset, batch_size=32, num_workers=4)

RuntimeError: version_ <= kMaxSupportedFileFormatVersion INTERNAL ASSERT FAILED

# 方法1：指定map_location
model = torch.load('model.pth', map_location=torch.device('cpu'))

# 方法2：转换模型版本
import torch

with open('legacy_model.pth', 'rb') as f:
    legacy_state = torch.load(f, map_location='cpu')

new_model = NewModel()
new_model.load_state_dict(legacy_state)
torch.save(new_model.state_dict(), 'converted_model.pth')

DataLoader(
    dataset,
    batch_size=32,
    num_workers=4,
    pin_memory=True,  # 加速GPU传输
    persistent_workers=True  # PyTorch 1.8+
)

from torch.utils.data import DataLoader

class SafeDataLoader(DataLoader):
    def __iter__(self):
        try:
            yield from super().__iter__()
        except Exception as e:
            print(f"数据加载异常: {str(e)}")
            raise