[FIXED] Pytorch lightning: 'CIFAR10DataModule' object has no attribute 'train_loader'

Issue

Would you tell me why I failed to import CUFAR10DataModule()?

At first, I run the code on GoogleColab,

from pl_bolts.datamodules import CIFAR10DataModule
dm = CIFAR10DataModule()

then, the code was performed for the confirmation

from torch.optim import Adam
optimizer = Adam(finetune_layer.parameters(), lr=1e-4)

for epoch in range(10):
  for batch in dm.train_loader:
    x, y = batch
    with torch.no_grad():
      features = backbone(x)

    preds = finetune_layer(features)
    loss = cross_entropy(preds, y)

    loss.backward()
    optimizer.step()
    optimizer.zero_grad()
    print(loss.item())

However, the message AttributeError: 'CIFAR10DataModule' object has no attribute 'train_loader' was returned after running the code.

When the code was run to confirm the dm,

for batch in dm.train_dataloader:
  x, y = batch
  print(x.shape, y.shape)
  break

The error says TypeError: 'method' object is not iterable.

The code looks the same with an example, but I wonder why such an error was generated?

Solution

Two problems with your code:

First, the way you get the underlying PyTorch dataloader is dm.train_dataloader() not dm.train_loader. It is a function, not a property.

for batch in dm.train_dataloader():
    x, y = batch
    ...

Secondly, since you are trying to use a LightningDataModule without a Trainer, you need to manually invoke

dm.prepare_data()
dm.setup()

.. in order for the dataloader to be available via .train_dataloader().

Answered By - ayandas