noether.data.pipeline.collators

Submodules

Classes

ConcatSparseTensorCollator

Concatenates a sparse tensor along its first axis. Additionally, creates the batch_idx tensor which maps samples

DefaultCollator

Applies torch.utils.data.default_collate to the specified items.

FieldDecoderCollator

Collates a field to be used in a UPT-style decoder. It requires:

SparseTensorOffsetCollator

Collates sparse tensors by concatenating them along the first axis and creating an offset tensor that maps

Package Contents

class noether.data.pipeline.collators.ConcatSparseTensorCollator(items, create_batch_idx=False, batch_idx_key='batch_idx')

Concatenates a sparse tensor along its first axis. Additionally, creates the batch_idx tensor which maps samples to their respective index in a batch.

Parameters:

  • items (list[str]) – Which pointcloud items (i.e., keys in the sample dictionaries) should be collated.

  • create_batch_idx (bool) – If true, creates a batch_idx tensor that maps samples to their index in the batch. Defaults to False.

  • batch_idx_key (str) – How the generated batch_idx tensor should be called. Defaults to “batch_idx”. If multiple batch_idx tensors are generated, this needs to be set to distinguish between them. For example, if a subsampled pointcloud is loaded and also the raw pointcloud without subsampling one could use batch_idx for the subsampled pointcloud and batch_idx_raw for the raw pointcloud.

items
create_batch_idx = False
batch_idx_key = 'batch_idx'
class noether.data.pipeline.collators.DefaultCollator(items=None, optional_items=None)

Applies torch.utils.data.default_collate to the specified items.

Parameters:

  • items (list[str] | None) – Items (i.e, keys) to apply default_collate to.

  • optional_items (list[str] | None) – Items (i.e, keys) to apply default_collate to if they are present.

items = None
optional_items = None
class noether.data.pipeline.collators.FieldDecoderCollator(position_item, target_items)

Collates a field to be used in a UPT-style decoder. It requires: - Positions as dense tensor (used as query for the Perceiver decoder) - Targets as sparse tensor (e.g., used for calculating a loss) - Unbatch mask to convert the dense output of the Perceiver decoder into a sparse tensor to compare it to the

targets (i.e., calculate a loss)

Parameters:

  • position_item (str) – Identifier for the position.

  • target_items (list[str]) – Identifiers for the position, can use multiple target_items if multiple values are predicted with

  • decoder (the same)

position_item
target_items
class noether.data.pipeline.collators.SparseTensorOffsetCollator(item, offset_key)

Bases: noether.data.pipeline.collator.Collator

Collates sparse tensors by concatenating them along the first axis and creating an offset tensor that maps each sample to its respective index in the batch.

Parameters:
item
offset_key