使用PyTorch Lightning的validate()函数时遇两类报错求助

报错原因与对应解决逻辑 #

• 第一个报错：ValueError: An invalid dataloader was passed to Trainer.validate(dataloaders=...)
  根源是调用trainer.validate()时，训练器无法获取验证数据集——既没在LightningModule/LightningDataModule中实现val_dataloader()方法，也没给validate()传入dataloader参数。

• 第二个报错：Unwrapping the module did not yield a LightningModule
  根源是调用validate()时只传了dataloader，未传入你的LightningModule实例，训练器找不到要执行验证的模型。

两种正确实现方式 #

方式一：在LightningModule中实现val_dataloader()方法

import pytorch_lightning as pl
from torchvision.datasets import MNIST
from torchvision import transforms
from torch.utils.data import DataLoader

class MyModel(pl.LightningModule):
    def __init__(self):
        super().__init__()
        # 初始化模型层、损失函数等
        self.model = ...
        self.criterion = ...

    def val_dataloader(self):
        transform = transforms.Compose([transforms.ToTensor()])
        mnist_val = MNIST(root='data', train=False, download=True, transform=transform)
        return DataLoader(mnist_val, batch_size=256, num_workers=4)

    def validation_step(self, batch, batch_idx):
        # 定义验证逻辑
        x, y = batch
        preds = self.model(x)
        loss = self.criterion(preds, y)
        self.log('val_loss', loss)

# 初始化模型与训练器
model = MyModel()
trainer = pl.Trainer()

# 直接调用validate，无需额外传dataloader
trainer.validate(model)

方式二：调用validate()时同时传入模型和dataloader

import pytorch_lightning as pl
from torchvision.datasets import MNIST
from torchvision import transforms
from torch.utils.data import DataLoader

class MyModel(pl.LightningModule):
    def __init__(self):
        super().__init__()
        self.model = ...
        self.criterion = ...

    def validation_step(self, batch, batch_idx):
        x, y = batch
        preds = self.model(x)
        loss = self.criterion(preds, y)
        self.log('val_loss', loss)

# 初始化模型
model = MyModel()

# 单独创建验证dataloader
transform = transforms.Compose([transforms.ToTensor()])
mnist_val = MNIST(root='data', train=False, download=True, transform=transform)
mnist_val_loader = DataLoader(mnist_val, batch_size=256, num_workers=4)

trainer = pl.Trainer()
# 必须同时传入模型和dataloader参数
trainer.validate(model, dataloaders=mnist_val_loader)

内容的提问来源于stack exchange，提问作者Stefano Borzò