💿 DataLoaders

DataLoaders are used to pass in training or evaluation data to the Composer Trainer. There are three different ways of doing so:

1. Passing PyTorch torch.utils.data.DataLoader objects directly.

2. Providing a DataSpec, which contains a PyTorch dataloader as well as additional configurations, such as on-device transforms.

3. (For validation) Providing Evaluator objects which contain both a dataloader and relevant metrics for validation.

We walk through each of these ways in detail and provide usage examples below.

Passing a PyTorch DataLoader

Composer dataloaders have type torch.utils.data.DataLoader (see PyTorch documentation) and can be passed directly to the Trainer.

from torch.utils.data import DataLoader
from composer import Trainer

train_dataloader = DataLoader(
    training_data,
    batch_size=2048,
    shuffle=True
)

trainer = Trainer(..., train_dataloader=train_dataloader, ...)

Note

The batch_size to the dataloader should be the per-device overall batch size. For example, if you were using device_train_microbatch_size=1024, a batch_size of 2048 would mean that each microbatch (one forward/backward pass) would have a batch size of 1024.

For performance, we highly recommend:

• num_workers > 0 : usually set this to the number of CPU cores in your machine divided by the number of GPUs.

• pin_memory = true : Pinned memory can speed up copying memory from a CPU to a GPU. Try to use it everywhere possible because the only drawback is the reduced RAM available to the host.

• persistent_workers = true : Persisting workers will reduce the overhead of creating workers but will use some RAM since these workers have some persistent state.

Note

Samplers are used to specify the order of indices in dataloading. When using distributed training, it is important to use the Torch DistributedSampler. so that each process sees a unique shard of the dataset. If the dataset is already sharded, then use a SequentialSampler or RandomSampler.

DataSpec

Sometimes, the data configuration requires more than just the dataloader. Some example additional configurations include:

• Some transforms should be run on the data after it has been moved onto the correct device (e.g. GPU).

• Custom batch types would need a split_batch function that tells our trainer how to split the batches into microbatches for gradient accumulation.

• Optionally tracking the number of tokens (or samples) seen during training so far.

• Providing the length of a dataset when len or a similar function isn’t in the dataloader’s interface.

For these and other potential uses cases, the trainer can also accept the DataSpec object with these additional settings. For example,

from composer import Trainer
from composer.core import DataSpec

data_spec = DataSpec(
    dataloader=my_train_dataloader,
    num_tokens=193820,
    get_num_tokens_in_batch=lambda batch: batch['text'].shape[0]
)

trainer = Trainer(train_dataloader=data_spec, ...)

Examples of how DataSpec is used for popular datasets can be seen in our ImageNet and ADE20k files. For reference, the DataSpec arguments are shown below.

class composer.core.DataSpec(dataloader, num_samples=None, num_tokens=None, device_transforms=None, split_batch=None, get_num_samples_in_batch=None, get_num_tokens_in_batch=None)

Specifications for operating and training on data.

An example of constructing a DataSpec object with a device_transforms callable and then using it with Trainer:

>>> # Construct DataSpec and subtract mean from the batch
>>> device_transform_fn = lambda xs, ys: (xs.sub_(xs.mean()), ys)
>>> train_dspec = DataSpec(train_dataloader, device_transforms=device_transform_fn)
>>> # The same function can be used for eval dataloader as well
>>> eval_dspec = DataSpec(eval_dataloader, device_transforms=device_transform_fn)
>>> # Use this DataSpec object to construct trainer
>>> trainer = Trainer(
...     model=model,
...     train_dataloader=train_dspec,
...     eval_dataloader=eval_dspec,
...     optimizers=optimizer,
...     max_duration="1ep",
... )

Parameters

• dataloader (Union[Iterable, DataLoader]) – The dataloader, which can be any iterable that yields batches.

• num_samples (int, optional) – The total number of samples in an epoch, across all ranks. This field is used by the Timestamp (training progress tracker). If not specified, then len(dataloader.dataset) is used (if this property is available). Otherwise, the dataset is assumed to be unsized.

• num_tokens (int, optional) – The total number of tokens in an epoch. This field is used by the Timestamp (training progress tracker).

• device_transforms ((Batch) -> Batch, optional) – Function called by the Trainer to modify the batch once it has been moved onto the device. For example, this function can be used for GPU-based normalization. It can modify the batch in-place, and it should return the modified batch. If not specified, the batch is not modified.

• split_batch ((Batch, (int | float)) -> Sequence[Batch], optional) – Function called by the Trainer to split a batch (the first parameter) into microbatches of a given size (the second parameter). If the dataloader yields batches not of type torch.Tensor, Mapping, Tuple, or List, then this function must be specified.

• get_num_samples_in_batch ((Batch) -> Union[int, float], optional) –

  Function that is called by the Trainer to get the number of samples in the provided batch.

  By default, if the batch contains tensors that all have the same 0th dim, then the value of the 0th dim will be returned. If the batch contains tensors where the 0th dim differ, then this function must be specified.

• get_num_tokens_in_batch ((Batch) -> int, optional) –

  Function that is called by the Trainer to get the number of tokens in the provided batch.

  By default, it checks for HuggingFace-style dictionary batches with input_ids, and then checks dataset.max_seq_len, and returns 0 if both of those fail, meaning that number of tokens processed will not be tracked as a part of the training progress tracking. Note that the defaults do NOT take padding into account, so if you want the token calculation to exclude padding, you should specify this function. This function must be specified to track the number of tokens processed during training in a non-default way.

Validation

To track training progress, validation datasets can be provided to the Composer Trainer through the eval_dataloader parameter. If there are multiple datasets to use for validation/evaluation, each with their own metrics, Evaluator objects can be used to pass in multiple dataloaders/datasets to the trainer.

For more information, see Evaluation.

Batch Types

For custom batch types (not torch.Tensor, List, Tuple, Mapping), implement and provide the split_batch function to the trainer using DataSpec above. Here’s an example function or when the batch from the dataloader is a tuple of two tensors:

def split_batch(self, batch: Batch, num_microbatches: int) -> List[Batch]:
    x, y = batch
    if isinstance(x, torch.Tensor) and isinstance(y, torch.Tensor):
        return list(zip(x.chunk(num_microbatches), y.chunk(num_microbatches)))

Suppose instead the batch had one input image and several target images, e.g. (Tensor, (Tensor, Tensor, Tensor)). Then the function would be:

def split_batch(self, batch: Batch, num_microbatches: int) -> List[Batch]:
    n = num_microbatches

    x, (y1, y2) = batch
    chunked = (x.chunk(n), (y1.chunk(n), y2.chunk(n)))
    return list(zip(*chunked))