def __init__(self,
layer_balance,
ratio_thresh=4.,
expansion_bias=0.5,
cls_pw=1.0,
obj_pw=1.0,
iou_type="ciou",
coord_type="xywh",
iou_ratio=1.0,
iou_weights=0.05,
cls_weights=0.5,
obj_weights=1.0):
super(YOLOv5Loss, self).__init__()
self.layer_balance = layer_balance
self.iou_ratio = iou_ratio
self.iou_weights = iou_weights
self.cls_weights = cls_weights
self.obj_weights = obj_weights
self.expansion_bias = expansion_bias
self.box_similarity = BoxSimilarity(iou_type, coord_type)
self.target_builder = YOLOv5Builder(ratio_thresh, expansion_bias)
self.cls_bce = nn.BCEWithLogitsLoss(pos_weight=torch.tensor(
data=[cls_pw]))
self.obj_bce = nn.BCEWithLogitsLoss(pos_weight=torch.tensor(
data=[obj_pw]))
def __init__(self,
alpha=0.25,
gamma=2.0,
cls_cost=1,
iou_cost=1,
l1_cost=1):
super(HungarianMatcher, self).__init__()
self.alpha = alpha
self.gamma = gamma
self.cls_cost = cls_cost
self.iou_cost = iou_cost
self.l1_cost = l1_cost
self.similarity = BoxSimilarity(iou_type="giou")
def __init__(self,
ratio_thresh=4.,
expansion_bias=0.5,
layer_balance=None,
cls_pw=1.0,
obj_pw=1.0,
iou_type="ciou",
coord_type="xywh",
iou_ratio=1.0,
iou_weights=0.05,
cls_weights=0.5,
obj_weights=1.0,
fl_gamma=0.,
class_smoothing_eps=0.0):
'''
:param ratio_thresh: 标签的长h宽w/anchor的长h_a宽w_a阈值, 即h/h_a, w/w_a都要在(1/2.26, 2.26)之间
:param expansion_bias: anchor匹配时的centrol point位置偏移，能使同一个anchor匹配到更多的样本
:param layer_balance: objectness loss(lobj)置信度损失在不同的层有不同的权重系数, [4.0, 1.0, 0.4] if np == 3 else [4.0, 1.0, 0.4, 0.1] # P3-5 or P3-6
:param cls_pw: cls BCELoss positive_weight/分类BCELoss中正样本的权重
:param obj_pw: obj BCELoss positive_weight/ 有无物体BCELoss中正样本的权重
:param iou_type:
:param coord_type:
:param iou_ratio: which is used to calcu objectness by calcu iou
:param iou_weights: box loss gain
:param cls_weights: cls loss gain
:param obj_weights: obj loss gain (scale with pixels)/有无物体损失的系数
:param fl_gamma: hyper-parameter gamma in focal loss
:param class_smoothing_eps: hyper-parameter in class smoothing label
'''
super(YOLOv5LossOriginal, self).__init__()
if layer_balance is None:
layer_balance = [4.0, 1.0, 0.4]
self.layer_balance = layer_balance
self.iou_ratio = iou_ratio
self.iou_weights = iou_weights
self.cls_weights = cls_weights
self.obj_weights = obj_weights
self.expansion_bias = expansion_bias
self.box_similarity = BoxSimilarity(iou_type, coord_type) # calculate iou between pred_bbox and gt_bbox
self.target_builder = YOLOv5Builder(ratio_thresh, expansion_bias) # Build targets for compute_loss/ for anchors
self.cls_bce = nn.BCEWithLogitsLoss(pos_weight=torch.tensor(data=[cls_pw])) # classfication loss
self.obj_bce = nn.BCEWithLogitsLoss(pos_weight=torch.tensor(data=[obj_pw])) # predicted objectness loss
### focal loss wraper----------------------------------------------------------------
self.fl_gamma=fl_gamma
if fl_gamma>0: # focal loss gamma
self.cls_bce=FocalLoss(self.cls_bce,fl_gamma)
self.obj_bce=FocalLoss(self.obj_bce,fl_gamma)
# 样本标签平滑
# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
self.cp, self.cn = smooth_BCE(eps=class_smoothing_eps)
def __init__(self,
top_k=9,
alpha=0.25,
gamma=2.0,
iou_thresh=0.1,
allow_low_quality_matches=True,
iou_type="giou",
iou_loss_weight=0.5,
reg_loss_weight=1.3):
self.top_k = top_k
self.alpha = alpha
self.gamma = gamma
self.iou_loss_weight = iou_loss_weight
self.reg_loss_weight = reg_loss_weight
self.matcher = Matcher(iou_thresh=iou_thresh,
ignore_iou=iou_thresh,
allow_low_quality_matches=allow_low_quality_matches)
self.box_coder = BoxCoder()
self.iou_loss = IOULoss(iou_type=iou_type)
self.box_similarity = BoxSimilarity(iou_type="iou")
self.bce = torch.nn.BCEWithLogitsLoss(reduction="sum")
def __init__(self,
ratio_thresh=4.,
expansion_bias=0.5,
layer_balance=None,
obj_pw=1.0,
iou_type="giou",
coord_type="xywh",
iou_ratio=1.0,
iou_weights=0.05,
obj_weights=1.2):
super(PedYOLOv5Loss, self).__init__()
if layer_balance is None:
layer_balance = [4.0, 1.0, 0.4]
self.layer_balance = layer_balance
self.iou_ratio = iou_ratio
self.iou_weights = iou_weights
self.obj_weights = obj_weights
self.expansion_bias = expansion_bias
self.box_similarity = BoxSimilarity(iou_type, coord_type)
self.target_builder = PedYOLOv5Builder(ratio_thresh, expansion_bias)
self.obj_bce = nn.BCEWithLogitsLoss(pos_weight=torch.tensor(
data=[obj_pw]))
def __init__(self,
top_k,
anchor_num_per_loc,
strides,
beta=2.0,
iou_type='giou',
iou_loss_weight=2.0,
reg_loss_weight=0.25,
reg_max=16):
self.topk = top_k
self.beta = beta
self.reg_max = reg_max
self.iou_type = iou_type
self.iou_loss_weight = iou_loss_weight
self.reg_loss_weight = reg_loss_weight
self.matcher = ATSSMatcher(self.topk, anchor_num_per_loc)
self.iou_loss = IOULoss(iou_type=iou_type)
self.box_similarity = BoxSimilarity(iou_type='iou')
self.strides = strides
self.expand_strides = None
self.project = Project(reg_max=reg_max)
self.qfl = QFL(beta=beta)
self.dfl = DFL()