dataloader

[ allennlp.data.dataloader ]

---

TensorDict

TensorDict = Dict[str, Union[torch.Tensor, Dict[str, torch.Tensor]]]

allennlp_collate

def allennlp_collate(instances: List[Instance]) -> TensorDict

DataLoader

class DataLoader(Registrable)

A DataLoader is responsible for generating batches of instances from a Dataset, or another source of data. This is essentially just an abstraction over torch.utils.data.DataLoader.

This class only has one required method, __iter__(), that creates an iterable of TensorDicts. Additionally, this class comes with a __len__() method that just raises a TypeError by default. When possible, this should be overriden to return the number of batches that will be generated by the __iter__() method.

default_implementation

class DataLoader(Registrable):
 | ...
 | default_implementation = "pytorch_dataloader"

PyTorchDataLoader

class PyTorchDataLoader(data.DataLoader,  DataLoader):
 | def __init__(
 |     self,
 |     dataset: data.Dataset,
 |     batch_size: int = 1,
 |     shuffle: bool = False,
 |     sampler: Sampler = None,
 |     batch_sampler: BatchSampler = None,
 |     num_workers: int = 0,
 |     collate_fn=allennlp_collate,
 |     pin_memory: bool = False,
 |     drop_last: bool = False,
 |     timeout: int = 0,
 |     worker_init_fn=None,
 |     multiprocessing_context: str = None,
 |     batches_per_epoch: int = None
 | )

A registrable version of the pytorch DataLoader. Firstly, this class exists is so that we can construct a DataLoader from a configuration file and have a different default collate_fn. You can use this class directly in python code, but it is identical to using pytorch dataloader with allennlp's custom collate function:

from torch.utils.data import DataLoader

from allennlp.data import allennlp_collate
# Construct a dataloader directly for a dataset which contains allennlp
# Instances which have _already_ been indexed.
my_loader = DataLoader(dataset, batch_size=32, collate_fn=allennlp_collate)

Secondly, this class adds a batches_per_epoch parameter which, if given, determines the number of batches after which an epoch ends. If this is None, then an epoch is set to be one full pass through your data. You might use this if you have a very large dataset and want more frequent checkpoints and evaluations on validation data, for instance.

In a typical AllenNLP configuration file, the dataset parameter does not get an entry under the "data_loader", it gets constructed separately.

from_partial_objects

class PyTorchDataLoader(data.DataLoader,  DataLoader):
 | ...
 | @classmethod
 | def from_partial_objects(
 |     cls,
 |     dataset: data.Dataset,
 |     batch_size: int = 1,
 |     shuffle: bool = False,
 |     sampler: Lazy[Sampler] = None,
 |     batch_sampler: Lazy[BatchSampler] = None,
 |     num_workers: int = 0,
 |     pin_memory: bool = False,
 |     drop_last: bool = False,
 |     timeout: int = 0,
 |     worker_init_fn=None,
 |     multiprocessing_context: str = None,
 |     batches_per_epoch: int = None
 | ) -> "PyTorchDataLoader"