clinicadl.data.dataloader.collate.MergeBatchesCollate¶

class clinicadl.data.dataloader.collate.MergeBatchesCollate(ignore: Sequence[str] | None = None)[source]¶

To merge several batches into a single batch.

This collating mode is typically to get a single batch from the outputs of a Dataset returning a sequence of samples. MergeBatchesCollate will try to merge this sequence of samples by merging each field of the Sample, except those in ignore.

More precisely:

torchio.Images will be concatenated along the channel dimension;
numpy.ndarrays and torch.Tensors will be stacked along a new dimension;
otherwise, the values will be merged in a tuple. If this tuple contains only one unique element (i.e. the value is the same in all the samples), a single value will be returned.

Parameters:

ignore (Optional[Sequence[str]], default=None) –

To ignore some fields in the samples to merge. Thus, the output sample will not have these fields.

Important

The mandatory arguments of Sample cannot be ignored.

Examples

from clinicadl.data.dataloader import MergeBatchesCollate
from clinicadl.data.structures.examples import Colin27Sample
import numpy as np

sample = Colin27Sample(participant_id="sub-001", label=np.array([0, 1]), age=55, sex="M")
sample_bis = Colin27Sample(participant_id="sub-001", label=np.array([1, 2]), age=56, to_ignore="abc")

batch = MergeBatchesCollate(ignore=["to_ignore"])([(sample, sample_bis)])

>>> batch
[Colin27Sample(Keys: ('head', 'sex', 'age', 'label', 'file_type', 'image_path', 'sample_type', 'sample_position', 'image', 'participant_id', 'session_id'); images: 2)]
>>> batch[0].participant_id  # same value in the two samples
'sub-001'
>>> batch[0].age
(55, 56)
>>> batch[0].sex  # only in one sample, so kept as it is
'M'
>>> batch[0].image.shape  # 2 channels now!
(2, 181, 217, 181)
>>> batch[0].label  # arrays are stacked
array([[0, 1],
       [1, 2]])