Managing Data

Continue reading to learn about:

Data Containers in Lightning

There are a few different data containers used in Lightning:

Data objects




The PyTorch Dataset represents a map from keys to data samples.


The PyTorch IterableDataset represents a stream of data.


The PyTorch DataLoader represents a Python iterable over a DataSet.


A LightningDataModule is simply a collection of: a training DataLoader, validation DataLoader(s), test DataLoader(s) and predict DataLoader(s), along with the matching transforms and data processing/downloads steps required.

Why LightningDataModules?

The LightningDataModule was designed as a way of decoupling data-related hooks from the LightningModule so you can develop dataset agnostic models. The LightningDataModule makes it easy to hot swap different datasets with your model, so you can test it and benchmark it across domains. It also makes sharing and reusing the exact data splits and transforms across projects possible.

Read this for more details on LightningDataModules.

Multiple Datasets

There are a few ways to pass multiple Datasets to Lightning:

  1. Create a DataLoader that iterates over multiple Datasets under the hood.

  2. In the training loop you can pass multiple DataLoaders as a dict or list/tuple and Lightning will automatically combine the batches from different DataLoaders.

  3. In the validation and test loop you have the option to return multiple DataLoaders, which Lightning will call sequentially.

Using LightningDataModule

You can set more than one DataLoader in your LightningDataModule using its dataloader hooks and Lightning will use the correct one under-the-hood.

class DataModule(LightningDataModule):


    def train_dataloader(self):

    def val_dataloader(self):
        return [

    def test_dataloader(self):

    def predict_dataloader(self):

Using LightningModule hooks

Concatenated DataSet

For training with multiple datasets you can create a dataloader class which wraps your multiple datasets (this of course also works for testing and validation datasets).


class ConcatDataset(
    def __init__(self, *datasets):
        self.datasets = datasets

    def __getitem__(self, i):
        return tuple(d[i] for d in self.datasets)

    def __len__(self):
        return min(len(d) for d in self.datasets)

class LitModel(LightningModule):

    def train_dataloader(self):
        concat_dataset = ConcatDataset(

        loader =
        return loader

    def val_dataloader(self):
        # SAME

    def test_dataloader(self):
        # SAME

Return multiple DataLoaders

You can set multiple DataLoaders in your LightningModule, and Lightning will take care of batch combination.

For more details please have a look at multiple_trainloader_mode

class LitModel(LightningModule):

    def train_dataloader(self):

        loader_a =, batch_size=4)
        loader_b =, batch_size=5)

        # pass loaders as a dict. This will create batches like this:
        # {'a': batch from loader_a, 'b': batch from loader_b}
        loaders = {'a': loader_a,
                   'b': loader_b}

        # OR:
        # pass loaders as sequence. This will create batches like this:
        # [batch from loader_a, batch from loader_b]
        loaders = [loader_a, loader_b]

        return loaders

Furthermore, Lightning also supports nested lists and dicts (or a combination).

class LitModel(LightningModule):

    def train_dataloader(self):

        loader_a =, batch_size=4)
        loader_b =, batch_size=2)

        return {'a': loader_a, 'b': loader_b}

    def training_step(self, batch, batch_idx):
        # access a dictionnary with a batch from each DataLoader
        batch_a = batch["a"]
        batch_b = batch["b"]
class LitModel(LightningModule):

    def train_dataloader(self):

        loader_a =, batch_size=4)
        loader_b =, batch_size=4)
        loader_c =, batch_size=4)
        loader_c =, batch_size=4)

        # pass loaders as a nested dict. This will create batches like this:
        loaders = {
            'loaders_a_b': [
            'loaders_c_d': {
                'c': loader_c,
                'd': loader_d
        return loaders

    def training_step(self, batch, batch_idx):
        # access the data
        batch_a_b = batch["loaders_a_b"]
        batch_c_d = batch["loaders_c_d"]

        batch_a = batch_a_b[0]
        batch_b = batch_a_b[1]

        batch_c = batch_c_d["c"]
        batch_d = batch_c_d["d"]

Multiple Validation/Test Datasets

For validation and test DataLoaders, you can pass a single DataLoader or a list of them. This optional named parameter can be used in conjunction with any of the above use cases. You can choose to pass the batches sequentially or simultaneously, as is done for the training step. The default mode for validation and test DataLoaders is sequential.

See the following for more details for the default sequential option:

  • val_dataloader()

  • test_dataloader()

def val_dataloader(self):
    loader_1 = DataLoader()
    loader_2 = DataLoader()
    return [loader_1, loader_2]

To combine batches of multiple test and validation DataLoaders simultaneously, one needs to wrap the DataLoaders with CombinedLoader.

from pytorch_lightning.trainer.supporters import CombinedLoader

def val_dataloader(self):
    loader_1 = DataLoader()
    loader_2 = DataLoader()
    loaders = {'a': loader_a,'b': loader_b}
    combined_loaders = CombinedLoader(loaders, "max_size_cycle")
    return combined_loaders

Test with additional data loaders

You can run inference on a test set even if the test_dataloader() method hasn’t been defined within your LightningModule instance. For example, this would be the case if your test data set is not available at the time your model was declared. Simply pass the test set to the test() method:

# setup your data loader
test = DataLoader(...)

# test (pass in the loader)

Sequential Data

Lightning has built in support for dealing with sequential data.

Packed sequences as inputs

When using PackedSequence, do 2 things:

  1. Return either a padded tensor in dataset or a list of variable length tensors in the DataLoader collate_fn (example shows the list implementation).

  2. Pack the sequence in forward or training and validation steps depending on use case.

# For use in DataLoader
def collate_fn(batch):
    x = [item[0] for item in batch]
    y = [item[1] for item in batch]
    return x, y

# In module
def training_step(self, batch, batch_nb):
    x = rnn.pack_sequence(batch[0], enforce_sorted=False)
    y = rnn.pack_sequence(batch[1], enforce_sorted=False)

Truncated Backpropagation Through Time (TBPTT)

There are times when multiple backwards passes are needed for each batch. For example, it may save memory to use Truncated Backpropagation Through Time when training RNNs.

Lightning can handle TBPTT automatically via this flag.

from pytorch_lightning import LightningModule

class MyModel(LightningModule):

    def __init__(self):
        # Important: This property activates truncated backpropagation through time
        # Setting this value to 2 splits the batch into sequences of size 2
        self.truncated_bptt_steps = 2

    # Truncated back-propagation through time
    def training_step(self, batch, batch_idx, hiddens):
        # the training step must be updated to accept a ``hiddens`` argument
        # hiddens are the hiddens from the previous truncated backprop step
        out, hiddens = self.lstm(data, hiddens)
        return {
            "loss": ...,
            "hiddens": hiddens


If you need to modify how the batch is split, override tbptt_split_batch().

Iterable Datasets

Lightning supports using IterableDatasets as well as map-style Datasets. IterableDatasets provide a more natural option when using sequential data.


When using an IterableDataset you must set the val_check_interval to 1.0 (the default) or an int (specifying the number of training batches to run before validation) when initializing the Trainer. This is because the IterableDataset does not have a __len__ and Lightning requires this to calculate the validation interval when val_check_interval is less than one. Similarly, you can set limit_{mode}_batches to a float or an int. If it is set to 0.0 or 0 it will set num_{mode}_batches to 0, if it is an int it will set num_{mode}_batches to limit_{mode}_batches, if it is set to 1.0 it will run for the whole dataset, otherwise it will throw an exception. Here mode can be train/val/test.

# IterableDataset
class CustomDataset(IterableDataset):

    def __init__(self, data):

    def __iter__(self):
        return iter(self.data_source)

# Setup DataLoader
def train_dataloader(self):
    seq_data = ['A', 'long', 'time', 'ago', 'in', 'a', 'galaxy', 'far', 'far', 'away']
    iterable_dataset = CustomDataset(seq_data)

    dataloader = DataLoader(dataset=iterable_dataset, batch_size=5)
    return dataloader
# Set val_check_interval
trainer = Trainer(val_check_interval=100)

# Set limit_val_batches to 0.0 or 0
trainer = Trainer(limit_val_batches=0.0)

# Set limit_val_batches as an int
trainer = Trainer(limit_val_batches=100)