def __init__(self, opt):
super(CtdetLoss, self).__init__()
self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss()
self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else \
RegLoss() if opt.reg_loss == 'sl1' else None
self.crit_wh = torch.nn.L1Loss(reduction='sum') if opt.dense_wh else \
NormRegL1Loss() if opt.norm_wh else \
RegWeightedL1Loss() if opt.cat_spec_wh else self.crit_reg
self.opt = opt
def __init__(self, opt):
super(McMotLoss, self).__init__()
self.opt = opt
self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss()
self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else \
RegLoss() if opt.reg_loss == 'sl1' else None # L1 loss or smooth l1 loss
self.crit_wh = torch.nn.L1Loss(reduction='sum') if opt.dense_wh else \
NormRegL1Loss() if opt.norm_wh else \
RegWeightedL1Loss() if opt.cat_spec_wh else self.crit_reg # box size loss
# @even: Test additional loss functions for re-id
self.circle_loss = CircleLoss(m=0.25, gamma=80)
self.ghm_c = GHMC(
) # GHM_C loss for multi-class classification(For ReID)
if opt.id_weight > 0:
self.emb_dim = opt.reid_dim
# @even: 用nID_dict取代nID, 用于MCMOT(multi-class multi-object tracking)训练
self.nID_dict = opt.nID_dict
# 包含可学习参数的层: 用于Re-ID的全连接层
# @even: 为每个需要ReID的类别定义一个分类器
self.classifiers = nn.ModuleDict(
) # 使用ModuleList或ModuleDict才可以自动注册参数
self.focal_loss_dict = nn.ModuleDict()
for cls_id, nID in self.nID_dict.items():
# 选择一: 使用普通的全连接层
self.classifiers[str(cls_id)] = nn.Linear(self.emb_dim,
nID) # FC layers
# 选择二: 使用Arc margin全连接层
# self.classifiers[str(cls_id)] = ArcMarginFc(in_features=self.emb_dim,
# out_features=nID,
# device=self.opt.device,
# m=0.4)
# 选择三: 使用Focal loss
# self.focal_loss_dict[str(cls_id)] = McFocalLoss(nID, self.opt.device)
# using CE loss to do ReID classification
self.ce_loss = nn.CrossEntropyLoss(ignore_index=-1)
# self.TriLoss = TripletLoss()
# @even: 为每个需要ReID的类别定义一个embedding scale
self.emb_scale_dict = dict()
for cls_id, nID in self.nID_dict.items():
self.emb_scale_dict[cls_id] = math.sqrt(2) * math.log(nID - 1)
# track reid分类的损失缩放系数
self.s_id = nn.Parameter(-1.05 * torch.ones(1)) # -1.05
# scale factor of detection loss
self.s_det = nn.Parameter(-1.85 * torch.ones(1))
def __init__(self, opt):
super(McMotLoss, self).__init__()
self.opt = opt
self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss()
self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else \
RegLoss() if opt.reg_loss == 'sl1' else None # L1 loss or smooth l1 loss
self.crit_wh = torch.nn.L1Loss(reduction='sum') if opt.dense_wh else \
NormRegL1Loss() if opt.norm_wh else \
RegWeightedL1Loss() if opt.cat_spec_wh else self.crit_reg # box size loss
self.circle_loss = CircleLoss(m=0.25, gamma=80)
self.emb_dim = opt.reid_dim
# @even: 用nID_dict取代nID, 用于MCMOT(multi-class multi-object tracking)训练
self.nID_dict = opt.nID_dict
# 包含可学习参数的层: 用于Re-ID的全连接层
# @even: 为每个需要ReID的类别定义一个分类器
self.classifiers = nn.ModuleDict() # 使用ModuleList或ModuleDict才可以自动注册参数
for cls_id, nID in self.nID_dict.items():
self.classifiers[str(cls_id)] = nn.Linear(self.emb_dim,
nID) # 全连接层
# self.classifiers[str(cls_id)] = ArcMarginFc(in_features=self.emb_dim, # 使用Arc margin全连接层
# out_features=nID,
# device=self.opt.device,
# m=0.4)
self.IDLoss = nn.CrossEntropyLoss(
ignore_index=-1) # 不同的track id分类用交叉熵损失
# self.TriLoss = TripletLoss()
# 定义每个(y, x)位置的分类损失
self.ClsLoss = nn.CrossEntropyLoss() # 对feature map上每个位置(y, x)检测类别分类
# @even: 为每个需要ReID的类别定义一个embedding scale
self.emb_scale_dict = dict()
for cls_id, nID in self.nID_dict.items():
self.emb_scale_dict[cls_id] = math.sqrt(2) * math.log(nID - 1)
self.s_det = nn.Parameter(-1.85 * torch.ones(1)) # 检测的损失缩放系数
self.s_id = nn.Parameter(-1.05 * torch.ones(1)) # track id分类的损失缩放系数
def __init__(self, opt):
super(MotLoss, self).__init__()
self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss()
self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else \
RegLoss() if opt.reg_loss == 'sl1' else None
self.crit_wh = torch.nn.L1Loss(reduction='sum') if opt.dense_wh else \
NormRegL1Loss() if opt.norm_wh else \
RegWeightedL1Loss() if opt.cat_spec_wh else self.crit_reg
# classification (re-id)
self.opt = opt
self.emb_dim = opt.reid_dim
self.nID = opt.nID
self.classifier = nn.Linear(self.emb_dim, self.nID)
self.IDLoss = nn.CrossEntropyLoss(ignore_index=-1)
# self.TriLoss = TripletLoss()
self.emb_scale = math.sqrt(2) * math.log(self.nID - 1)
self.s_det = nn.Parameter(-1.85 * torch.ones(1))
self.s_id = nn.Parameter(-1.05 * torch.ones(1))
def __init__(self, opt):
super(MotLoss, self).__init__()
self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss()
self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else \
RegLoss() if opt.reg_loss == 'sl1' else None # L1 loss or smooth l1 loss
self.crit_wh = torch.nn.L1Loss(reduction='sum') if opt.dense_wh else \
NormRegL1Loss() if opt.norm_wh else \
RegWeightedL1Loss() if opt.cat_spec_wh else self.crit_reg # box size loss
self.opt = opt
self.emb_dim = opt.reid_dim
self.nID = opt.nID
# 唯一包含可学习参数的层: 用于Re-ID的全连接层
self.classifier = nn.Linear(self.emb_dim, self.nID)
self.IDLoss = nn.CrossEntropyLoss(ignore_index=-1)
# self.TriLoss = TripletLoss()
self.emb_scale = math.sqrt(2) * math.log(self.nID - 1)
self.s_det = nn.Parameter(-1.85 * torch.ones(1)) # 检测的损失缩放系数
self.s_id = nn.Parameter(-1.05 * torch.ones(1)) # track id分类的损失缩放系数