DataSpec

class composer.DataSpec(dataloader, num_samples=None, num_tokens=None, device_transforms=None, batch_transforms=None, microbatch_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 batch_transforms callable and then using it with Trainer:

>>> # Construct DataSpec and subtract mean from the batch
>>> batch_transform_fn = lambda xs, ys: (xs.sub_(xs.mean()), ys)
>>> train_dspec = DataSpec(train_dataloader, batch_transforms=batch_transform_fn)
>>> # The same function can be used for eval dataloader as well
>>> eval_dspec = DataSpec(eval_dataloader, batch_transforms=batch_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) – Deprecated argument. Please use batch_transforms for batch level transformations on CPU and microbatch_transforms for microbatch level transformations on target device.

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

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