def _convert_to_poptorch_loader(
self,
dataloader: DataLoader,
sampler,
mode: Optional[RunningStage] = None) -> "poptorch.DataLoader":
dl_kwargs = _get_dataloader_init_kwargs(dataloader, sampler)
# Override to drop last uneven batch, as IPUs does not support uneven inputs.
dl_kwargs["drop_last"] = True
opts = self.training_opts if mode == RunningStage.TRAINING else self.inference_opts
dataloader = poptorch.DataLoader(**dl_kwargs, options=opts)
return dataloader
def test_dataloader_kwargs_replacement_with_iterable_dataset(mode):
"""Test that DataLoader kwargs are not replaced when using Iterable Dataset."""
dataset = RandomIterableDataset(7, 100)
dataloader = DataLoader(dataset, batch_size=32)
dl_kwargs = _get_dataloader_init_kwargs(dataloader,
dataloader.sampler,
mode=mode)
assert dl_kwargs["sampler"] is None
assert dl_kwargs["batch_sampler"] is None
assert dl_kwargs["batch_size"] is dataloader.batch_size
assert dl_kwargs["dataset"] is dataloader.dataset
assert dl_kwargs["collate_fn"] is dataloader.collate_fn
def _convert_to_poptorch_loader(
self,
dataloader: DataLoader,
sampler,
mode: Optional[RunningStage] = None) -> "poptorch.DataLoader":
if isinstance(dataloader, poptorch.DataLoader):
# the user is returning the `poptorch.DataLoader` directly, don't change anything.
return dataloader
dl_kwargs = _get_dataloader_init_kwargs(dataloader, sampler)
opts = self.training_opts if mode == RunningStage.TRAINING else self.inference_opts
dataloader = poptorch.DataLoader(opts, **dl_kwargs)
return dataloader
