def criterion(
self,
cls: Tensor,
reg: Tensor,
centerness: Tensor,
target: Tensor,
reg_scaling: float = 1.0,
centerness_scaling: float = 1.0,
) -> Dict[str, Tensor]:
# get losses
target_bbox, target_cls = split_box_target(target)
cls_loss, reg_loss, centerness_loss = self._criterion(cls, reg, centerness, target_bbox, target_cls)
# get number of boxes for normalization
num_boxes = (target_cls != -1).sum().clamp_min(1)
cls_loss = cls_loss / num_boxes
reg_loss = reg_loss / num_boxes
centerness_loss = centerness_loss / num_boxes
# compute a total loss
total_loss = cls_loss + reg_loss * reg_scaling + centerness_loss * centerness_scaling
return {
"type_loss": cls_loss,
"centerness_loss": centerness_loss,
"reg_loss": reg_loss,
"total_loss": total_loss,
}
def test_split_box_target_returns_views(self):
torch.random.manual_seed(42)
bbox_target = torch.randint(0, 10, (3, 4, 4))
label_target = torch.randint(0, 10, (3, 4, 2))
target = torch.cat([bbox_target, label_target], dim=-1)
bbox, label = split_box_target(target)
target.mul_(10)
assert torch.allclose(bbox, target[..., :4])
assert torch.allclose(label, target[..., 4:])
def get_pred_target_pairs(
self,
pred: Tensor,
target: Tensor,
) -> Tensor:
r"""Given a predicted CenterNet heatmap and target bounding box label, use box IoU to
create a paring of predicted and target boxes such that each predicted box has
an associated gold standard label.
.. warning::
This method should work with batched input, but such inputs are not thoroughly tested
Args:
pred (:class:`torch.Tensor`):
Predicted heatmap.
target (:class:`torch.Tensor`):
Target bounding boxes in format ``x1, y1, x2, y2, class``.
iou_threshold (float):
Intersection over union threshold for which a prediction can be considered a
true positive.
true_positive_limit (bool):
By default, only one predicted box overlapping a target box will be counted
as a true positive. If ``False``, allow multiple true positive boxes per
target box.
Returns:
A 3-tuple of tensors as follows:
1. Predicted binary score for each box
2. Classification target value for each box
3. Boolean indicating if the prediction was a correct
Shape:
* ``pred`` - :math:`(N_{pred}, 6)`
* ``target`` - :math:`(N_{true}, 5)`
* Output - :math:`(N_o)`, :math:`(N_o)`, :math:`(N_o)`
"""
# get a paring of predicted probability to target labels
# if we didn't detect a target box at any threshold, assume P_pred = 0.0
xform = CategoricalLabelIoU(self.iou_threshold,
self.true_positive_limit)
pred_boxes, pred_scores, pred_cls = split_bbox_scores_class(pred)
target_bbox, target_cls = split_box_target(target)
pred_out, binary_target, target_out = xform(pred_boxes, pred_scores,
pred_cls, target_bbox,
target_cls)
assert pred_out.ndim == 1
assert target_out.ndim == 1
assert pred_out.shape == target_out.shape
return pred_out, target_out.long(), binary_target
def test_split_box_target_result(self, split_label):
torch.random.manual_seed(42)
bbox_target = torch.randint(0, 10, (3, 4, 4))
label_target = torch.randint(0, 10, (3, 4, 2))
target = torch.cat([bbox_target, label_target], dim=-1)
result = split_box_target(target, split_label=split_label)
bbox = result[0]
label = result[1:]
assert torch.allclose(bbox, bbox_target)
if not split_label:
assert torch.allclose(torch.cat(label, dim=-1), label_target)
else:
for i, sub_label in enumerate(label):
assert torch.allclose(sub_label, label_target[..., i:i + 1])