def __cal_loss_per_layer(self, p, p_d, label, bboxes, stride):
"""
(1)The loss of regression of boxes.
GIOU loss is defined in https://arxiv.org/abs/1902.09630.
Note: The loss factor is 2-w*h/(img_size**2), which is used to influence the
balance of the loss value at different scales.
(2)The loss of confidence.
Includes confidence loss values for foreground and background.
Note: The backgroud loss is calculated when the maximum iou of the box predicted
by the feature point and all GTs is less than the threshold.
(3)The loss of classes。
The category loss is BCE, which is the binary value of each class.
:param stride: The scale of the feature map relative to the original image
:return: The average loss(loss_giou, loss_conf, loss_cls) of all batches of this detection layer.
"""
BCE = nn.BCEWithLogitsLoss(reduction="none")
FOCAL = FocalLoss(gamma=2, alpha=1.0, reduction="none")
batch_size, grid = p.shape[:2]
img_size = stride * grid
p_conf = p[..., 4:5]
p_cls = p[..., 5:]
p_d_xywh = p_d[..., :4]
label_xywh = label[..., :4]
label_obj_mask = label[..., 4:5]
label_cls = label[..., 6:]
label_mix = label[..., 5:6]
# loss giou
giou = tools.GIOU_xywh_torch(p_d_xywh, label_xywh).unsqueeze(-1)
# The scaled weight of bbox is used to balance the impact of small objects and large objects on loss.
bbox_loss_scale = 2.0 - 1.0 * label_xywh[..., 2:3] * label_xywh[
..., 3:4] / (img_size**2)
loss_giou = label_obj_mask * bbox_loss_scale * (1.0 - giou) * label_mix
# loss confidence
iou = tools.iou_xywh_torch(
p_d_xywh.unsqueeze(4),
bboxes.unsqueeze(1).unsqueeze(1).unsqueeze(1))
iou_max = iou.max(-1, keepdim=True)[0]
label_noobj_mask = (1.0 - label_obj_mask) * (
iou_max < self.__iou_threshold_loss).float()
loss_conf = (
label_obj_mask * FOCAL(input=p_conf, target=label_obj_mask) +
label_noobj_mask *
FOCAL(input=p_conf, target=label_obj_mask)) * label_mix
# loss classes
loss_cls = label_obj_mask * BCE(input=p_cls,
target=label_cls) * label_mix
loss_giou = (torch.sum(loss_giou)) / batch_size
loss_conf = (torch.sum(loss_conf)) / batch_size
loss_cls = (torch.sum(loss_cls)) / batch_size
loss = loss_giou + loss_conf + loss_cls
return loss, loss_giou, loss_conf, loss_cls
def __cal_loss_per_layer(self, p, p_d, label, bboxes, stride):
"""
(1)The loss of regression of boxes.
GIOU loss is defined in https://arxiv.org/abs/1902.09630.
Note: The loss factor is 2-w*h/(img_size**2), which is used to influence the
balance of the loss value at different scales.
(2)The loss of confidence.
Includes confidence loss values for foreground and background.
Note: The backgroud loss is calculated when the maximum iou of the box predicted
by the feature point and all GTs is less than the threshold.
(3)The loss of classes。
The category loss is BCE, which is the binary value of each class.
:param stride: The scale of the feature map relative to the original image
:return: The average loss(loss_giou, loss_conf, loss_cls) of all batches of this detection layer.
"""
BCE = nn.BCEWithLogitsLoss(reduction="none")
FOCAL = FocalLoss(gamma=2, alpha=1.0, reduction="none")
batch_size, grid = p.shape[:2]
img_size = stride * grid
p_conf = p[..., 4:5]
p_cls = p[..., 5:]
p_d_xywh = p_d[..., :4]
label_xywh = label[..., :4]
label_obj_mask = label[..., 4:5]
label_cls = label[..., 6:]
label_mix = label[..., 5:6]
# loss giou (pai: loss xywh, consider only label_obj_mask)
giou = tools.GIOU_xywh_torch(p_d_xywh, label_xywh).unsqueeze(-1)
# The scaled weight of bbox is used to balance the impact of small objects and large objects on loss.
bbox_loss_scale = 2.0 - 1.0 * label_xywh[..., 2:3] * label_xywh[..., 3:4] / (img_size ** 2)
loss_giou = label_obj_mask * bbox_loss_scale * (1.0 - giou) * label_mix
# loss confidence iot.shape >> torch.Size([4, 36, 36, 3, 150]) p_d >> torch.Size([4, 36, 36, 3, 4])
iou = tools.iou_xywh_torch(p_d_xywh.unsqueeze(4), bboxes.unsqueeze(1).unsqueeze(1).unsqueeze(1))
iou_max = iou.max(-1, keepdim=True)[0]
label_noobj_mask = (1.0 - label_obj_mask) * (iou_max < self.__iou_threshold_loss).float() # find noobj_mask not only -label_obj_mask but also have to anchors that has max iou < thres wil be thinked as no object
# we can panetilze the obj more and noobj less
loss_conf = (label_obj_mask * FOCAL(input=p_conf, target=label_obj_mask) +
label_noobj_mask * FOCAL(input=p_conf, target=label_obj_mask)) * label_mix # there is still a cell that not fall in both obj or noobj and the loss of those will = 0
# doing like this mean we will not panalize(loss = 0) the cell that predict high conf score even if there is the groud thruth is 0 which in fact will mak ethe loss of that cell higt.
#doing this will help the model to predict lot, leading to higher recall
# loss classes
loss_cls = label_obj_mask * BCE(input=p_cls, target=label_cls) * label_mix #torch.Size([4, 36, 36, 3, 80])
loss_giou = (torch.sum(loss_giou)) / batch_size
loss_conf = (torch.sum(loss_conf)) / batch_size
loss_cls = (torch.sum(loss_cls)) / batch_size
loss = loss_giou + loss_conf + loss_cls
return loss, loss_giou, loss_conf, loss_cls