PyTorch设置num_workers>0时触发RuntimeError问题求助

问题根源分析 #

这个错误本质是Windows系统下多进程启动的机制限制（默认用spawn而非fork），加上你的代码中大部分初始化逻辑都在全局作用域——当num_workers>0时，子进程会重新执行整个脚本的全局代码，从而引发重复初始化的冲突。你之前加了if __name__ == '__main__'但没生效，大概率是没把所有关键初始化逻辑都包裹到这个判断块里。

正确的代码重构方案 #

你需要把所有涉及数据集、数据加载器、模型实例化、训练启动的代码，甚至绘图代码都放到if __name__ == '__main__'保护块中，只把导入、函数定义和类定义留在全局作用域。修改后的代码如下：

import torchvision
from torchvision import transforms
import torchmetrics
import pytorch_lightning as pl
from pytorch_lightning.callbacks import ModelCheckpoint
from pytorch_lightning.loggers import TensorBoardLogger
from tqdm import tqdm
import numpy as np
import matplotlib.pyplot as plt

def load_file(path):
    return np.load(path).astype(np.float32)

# 数据变换定义可留在全局作用域
train_transforms = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(0.49, 0.248),
    transforms.RandomAffine(degrees=(-5, 5), translate=(0, 0.05), scale=(0.9, 1.1)),
    transforms.RandomResizedCrop((224, 224), scale=(0.35, 1))
])

val_transforms = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize([0.49], [0.248]),
])

class PneumoniaModel(pl.LightningModule):
    def __init__(self, weight=1):
        super().__init__()
        self.model = torchvision.models.resnet18()
        self.model.conv1 = torch.nn.Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
        self.model.fc = torch.nn.Linear(in_features=512, out_features=1)
        self.optimizer = torch.optim.Adam(self.model.parameters(), lr=1e-4)
        self.loss_fn = torch.nn.BCEWithLogitsLoss(pos_weight=torch.tensor([weight]))
        self.train_acc = torchmetrics.Accuracy()
        self.val_acc = torchmetrics.Accuracy()

    def forward(self, data):
        return self.model(data)

    def training_step(self, batch, batch_idx):
        x_ray, label = batch
        label = label.float()
        pred = self(x_ray)[:, 0]
        loss = self.loss_fn(pred, label)
        self.log("Train Loss", loss)
        self.log("Step Train Acc", self.train_acc(torch.sigmoid(pred), label.int()))
        return loss

    def training_epoch_end(self, outs):
        self.log("Train Acc", self.train_acc.compute())

    def validation_step(self, batch, batch_idx):
        x_ray, label = batch
        label = label.float()
        pred = self(x_ray)[:, 0]
        loss = self.loss_fn(pred, label)
        self.log("Val Loss", loss)
        self.log("Step Val Acc", self.val_acc(torch.sigmoid(pred), label.int()))
        return loss

    def validation_epoch_end(self, outs):
        self.log("Val Acc", self.val_acc.compute())

    def configure_optimizers(self):
        return [self.optimizer]

# 所有运行时逻辑都放到这里
if __name__ == '__main__':
    # 数据集与加载器初始化
    train_dataset = torchvision.datasets.DatasetFolder(
        "./Processed/train/", loader=load_file, extensions="npy", transform=train_transforms)
    val_dataset = torchvision.datasets.DatasetFolder(
        "./Processed/val/", loader=load_file, extensions="npy", transform=val_transforms)

    # 绘图代码移至主进程执行
    fig, axis = plt.subplots(2, 2, figsize=(9, 9))
    for i in range(2):
        for j in range(2):
            random_index = np.random.randint(0, len(train_dataset))
            x_ray, label = train_dataset[random_index]
            axis[i][j].imshow(x_ray[0], cmap="bone")
            axis[i][j].set_title(f"Label:{label}, id:{random_index}")
    plt.show()

    batch_size = 64
    num_workers = 8
    train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, num_workers=num_workers, shuffle=True, pin_memory=True)
    val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False, pin_memory=True)

    print(f"There are {len(train_dataset)} train images and {len(val_dataset)} val images")
    print("Train class distribution:", np.unique(train_dataset.targets, return_counts=True))
    print("Val class distribution:", np.unique(val_dataset.targets, return_counts=True))

    # 模型与训练器初始化
    model = PneumoniaModel()
    checkpoint_callback = ModelCheckpoint(monitor='Val Acc', save_top_k=10, mode='max')
    gpus = 1
    trainer = pl.Trainer(gpus=gpus, logger=TensorBoardLogger(save_dir="./logs"), log_every_n_steps=1, callbacks=checkpoint_callback, max_epochs=35)
    trainer.fit(model, train_loader, val_loader)

额外的PyCharm配置注意事项 #

如果修改后仍报错，检查PyCharm的运行配置：

• 打开Run/Debug Configurations，找到你的运行脚本
• 取消勾选Emulate terminal in output console选项（该选项会干扰多进程的标准输出重定向）
• 确保使用普通"Run"模式，而非"Run in Python Console"模式

这样调整后，num_workers>0的多进程数据加载就能正常工作，训练速度会显著提升。

内容的提问来源于stack exchange，提问作者Enrique Gil Garcia