# @Copyright: CEA-LIST/DIASI/SIALV/ (2023- )
# @Author: CEA-LIST/DIASI/SIALV/ <julien.denize@cea.fr>
# @License: CECILL-C
from __future__ import annotations
from typing import Any, List, Mapping, Optional, Sequence, Tuple, Union
import torch
from torch import Tensor
from torch.utils._pytree import tree_flatten
from ._bounding_boxes import BoundingBoxes, BoundingBoxFormat
from ._ta_tensor import TATensor
[docs]
def convert_bboxes_to_batch_bboxes(
bboxes: List[BoundingBoxes],
) -> BatchBoundingBoxes:
"""Convert a list of :class:`~torchaug.ta_tensors.BoundingBoxes` objects to a
:class:`~torchaug.ta_tensors.BatchBoundingBoxes` object.
Assumes all bboxes are valid.
"""
attrs = [
"canvas_size",
"format",
"requires_grad",
"device",
"dtype",
]
if not all(getattr(bbox, attr) == getattr(bboxes[0], attr) for bbox in bboxes for attr in attrs):
raise ValueError("All bounding boxes must have the same attributes.")
canvas_size, format = (
bboxes[0].canvas_size,
bboxes[0].format,
)
bboxes_data = torch.cat([bbox.data for bbox in bboxes])
idx_sample = torch.tensor([0] + [bbox.shape[0] for bbox in bboxes], dtype=torch.long).cumsum(0).tolist()
batch_bboxes = BatchBoundingBoxes(
bboxes_data,
canvas_size=canvas_size,
format=format,
idx_sample=idx_sample,
)
return batch_bboxes
[docs]
def convert_batch_bboxes_to_bboxes(
bboxes: BatchBoundingBoxes,
) -> List[BoundingBoxes]:
"""Convert :class:`~torchaug.ta_tensors.BatchBoundingBoxes` object to a list of
:class:`~torchaug.ta_tensors.BoundingBoxes` objects.
"""
canvas_size, format, idx_sample = (
bboxes.canvas_size,
bboxes.format,
bboxes.idx_sample,
)
list_bboxes = [
BoundingBoxes(
bboxes[idx_sample[i] : idx_sample[i + 1]],
canvas_size=canvas_size,
format=format,
)
for i in range(len(idx_sample) - 1)
]
return list_bboxes
[docs]
class BatchBoundingBoxes(TATensor):
""":class:`torch.Tensor` subclass for batch of bounding boxes.
.. note::
There should be only one :class:`~torchaug.ta_tensors.BatchBoundingBoxes`
instance per sample e.g. ``{"img": img, "bbox": BatchBoundingBoxes(...)}``.
Args:
data: Any data that can be turned into a tensor with :func:`torch.as_tensor`.
format: Format of the bounding box.
canvas_size: Height and width of the corresponding batch of images or videos.
idx_sample: Each element is the index of the first bounding box of the corresponding sample in
the batch of N samples. Contains N+1 elements whose last value is the number of bounding boxes.
dtype: Desired data type of the bounding box. If omitted, will be inferred from
``data``.
device: Desired device of the bounding box. If omitted and ``data`` is a
:class:`torch.Tensor`, the device is taken from it. Otherwise, the bounding box is constructed on the CPU.
requires_grad: Whether autograd should record operations on the bounding box. If omitted and
``data`` is a :class:`torch.Tensor`, the value is taken from it. Otherwise, defaults to ``False``.
"""
format: BoundingBoxFormat
canvas_size: Tuple[int, int]
idx_sample: List[int]
@property
def batch_size(self) -> int:
return len(self.idx_sample) - 1
@property
def num_boxes(self) -> int:
return self.data.shape[0]
def get_num_boxes_sample(self, idx: int) -> int:
return self.idx_sample[idx + 1] - self.idx_sample[idx]
[docs]
@classmethod
def cat(cls, bounding_boxes_batches: Sequence[BatchBoundingBoxes]) -> BatchBoundingBoxes:
"""Concatenates the given sequence of :class:`~torchaug.ta_tensors._batch_bounding_boxes.BatchBoundingBoxes`
along the first dimension.
Args:
bounding_boxes_batches: The sequence of
:class:`~torchaug.ta_tensors._batch_bounding_boxes.BatchBoundingBoxes` to concatenate.
Returns:
BatchBoundingBoxes: The concatenated batch of bounding boxes.
"""
attrs = [
"canvas_size",
"format",
"requires_grad",
"device",
"dtype",
]
for batch_bounding_boxes in bounding_boxes_batches:
if not isinstance(batch_bounding_boxes, BatchBoundingBoxes):
raise ValueError("All elements in the sequence must be instances of BatchBoundingBoxes.")
for attr in attrs:
if getattr(batch_bounding_boxes, attr) != getattr(bounding_boxes_batches[0], attr):
raise ValueError(f"All batches of masks must have the same {attr} attribute.")
idx_sample = (
torch.tensor(
[0]
+ [
batch_bounding_boxes.get_num_boxes_sample(i)
for batch_bounding_boxes in bounding_boxes_batches
for i in range(batch_bounding_boxes.batch_size)
]
)
.cumsum(0)
.tolist()
)
data = torch.cat([mask.data for mask in bounding_boxes_batches], 0)
return cls(
data,
idx_sample=idx_sample,
format=bounding_boxes_batches[0].format,
canvas_size=bounding_boxes_batches[0].canvas_size,
)
@classmethod
def _wrap(
cls,
tensor: Tensor,
*,
format: Union[BoundingBoxFormat, str],
canvas_size: Tuple[int, int],
idx_sample: List[int],
check_dims: bool = True,
) -> BatchBoundingBoxes: # type: ignore[override]
if check_dims and tensor.ndim != 2:
raise ValueError(f"Expected a 2D tensor, got {tensor.ndim}D.")
if isinstance(format, str):
format = BoundingBoxFormat[format.upper()] # type: ignore[misc]
batch_bounding_boxes = tensor.as_subclass(cls)
batch_bounding_boxes.format = format
batch_bounding_boxes.canvas_size = canvas_size
batch_bounding_boxes.idx_sample = idx_sample
return batch_bounding_boxes
def __new__(
cls,
data: Any,
*,
format: Union[BoundingBoxFormat, str],
canvas_size: Tensor,
idx_sample: List[int],
dtype: Optional[torch.dtype] = None,
device: Optional[Union[torch.device, str, int]] = None,
requires_grad: Optional[bool] = None,
) -> BatchBoundingBoxes:
tensor = cls._to_tensor(data, dtype=dtype, device=device, requires_grad=requires_grad)
return cls._wrap(tensor, format=format, canvas_size=canvas_size, idx_sample=idx_sample)
@classmethod
def _wrap_output(
cls,
output: torch.Tensor,
args: Sequence[Any] = (),
kwargs: Optional[Mapping[str, Any]] = None,
) -> BatchBoundingBoxes:
# If there are BatchBoundingBoxes instances in the output, their metadata got lost when we called
# super().__torch_function__. We need to restore the metadata somehow, so we choose to take
# the metadata from the first bbox in the parameters.
# This should be what we want in most cases. When it's not, it's probably a mis-use anyway, e.g.
# something like some_xyxy_bbox + some_xywh_bbox; we don't guard against those cases.
flat_params, _ = tree_flatten(args + (tuple(kwargs.values()) if kwargs else ())) # type: ignore[operator]
first_batch_bboxes_from_args = next(x for x in flat_params if isinstance(x, BatchBoundingBoxes))
format, canvas_size, idx_sample = (
first_batch_bboxes_from_args.format,
first_batch_bboxes_from_args.canvas_size,
first_batch_bboxes_from_args.idx_sample,
)
idx_sample = idx_sample.copy() # clone the list.
if isinstance(output, torch.Tensor) and not isinstance(output, BatchBoundingBoxes):
output = BatchBoundingBoxes._wrap(
output,
format=format,
canvas_size=canvas_size,
idx_sample=idx_sample,
check_dims=False,
)
elif isinstance(output, (tuple, list)):
output = type(output)(
BatchBoundingBoxes._wrap(
part,
format=format,
canvas_size=canvas_size,
idx_sample=idx_sample,
check_dims=False,
)
for part in output
)
return output
[docs]
def get_sample(self, idx: int) -> BoundingBoxes:
"""Get the bounding boxes for a sample in the batch.
Args:
idx: The index of the sample to get.
Returns:
The bounding boxes for the sample.
"""
boxes = self[self.idx_sample[idx] : self.idx_sample[idx + 1]]
return BoundingBoxes(
boxes,
format=self.format,
canvas_size=self.canvas_size,
device=self.device,
requires_grad=self.requires_grad,
)
[docs]
def get_chunk(self, chunk_indices: torch.Tensor) -> BatchBoundingBoxes:
"""Get a chunk of the batch of bounding boxes.
Args:
chunk_indices (torch.Tensor): The indices of the chunk to get.
Returns:
BatchBoundingBoxes: The chunk of the batch bounding boxes.
"""
chunk_idx_sample = torch.tensor(
[0] + [self.idx_sample[chunk_indice + 1] - self.idx_sample[chunk_indice] for chunk_indice in chunk_indices]
)
chunk_idx_sample = chunk_idx_sample.cumsum(0).tolist()
return BatchBoundingBoxes(
self[chunk_indices],
format=self.format,
canvas_size=self.canvas_size,
idx_sample=chunk_idx_sample,
device=self.device,
requires_grad=self.requires_grad,
)
[docs]
def update_chunk_(self, chunk: BatchBoundingBoxes, chunk_indices: torch.Tensor) -> BatchBoundingBoxes:
"""Update a chunk of the batch of bounding boxes.
Args:
chunk (BatchBoundingBoxes): The chunk update.
chunk_indices (torch.Tensor): The indices of the chunk to update.
Returns:
BatchBoundingBoxes: The updated batch of bounding boxes.
"""
if chunk.format != self.format:
raise ValueError("The format of the chunk must be the same as the format of the batch of bounding boxes.")
if chunk.canvas_size != self.canvas_size:
raise ValueError(
"The canvas size of the chunk must be the same as the canvas size of the batch of bounding boxes."
)
self[chunk_indices] = chunk
return self
[docs]
@classmethod
def masked_remove(cls, bboxes: BatchBoundingBoxes, mask: torch.Tensor) -> BatchBoundingBoxes:
"""Remove boxes from the batch of bounding boxes.
Args:
bboxes (BatchBoundingBoxes): The batch of bounding boxes to remove boxes from.
mask (torch.Tensor): A boolean mask to keep boxes.
Returns:
BatchBoundingBoxes: The updated batch of bounding boxes.
"""
# Remove boxes
old_idx_sample = bboxes.idx_sample
data = bboxes.data[~mask]
cpu_mask = mask.cpu()
num_delete_per_sample = [
cpu_mask[old_idx_sample[i] : old_idx_sample[i + 1]].sum().item() for i in range((len(old_idx_sample) - 1))
]
new_idx_sample = [old_idx_sample[i] - sum(num_delete_per_sample[: i + 1]) for i in range(len(old_idx_sample))]
return cls._wrap(
data,
format=bboxes.format,
canvas_size=bboxes.canvas_size,
idx_sample=new_idx_sample,
)
def __repr__(self, *, tensor_contents: Any = None) -> str: # type: ignore[override]
return self._make_repr(
format=self.format,
canvas_size=self.canvas_size,
idx_sample=self.idx_sample,
)