def compute_loss(self, true_hm, true_wh, true_reg, reg_mask, ind, true_cls):
hm_loss = loss.focal_loss(self.pred_hm, true_hm) * cfgs.HM_LOSS_WEIGHT
wh_loss = loss.reg_l1_loss(self.pred_wh, true_wh, ind, reg_mask) * cfgs.WH_LOSS_WEIGHT
reg_loss = loss.reg_l1_loss(self.pred_reg, true_reg, ind, reg_mask) * cfgs.REG_LOSS_WEIGHT
cls_loss = loss.cross_entropy_loss(self.pred_cls, true_cls, reg_mask) * cfgs.CLS_LOSS_WEIGHT
return hm_loss, wh_loss, reg_loss, cls_loss
def compute_loss(self):
self.cls_loss = loss.focal_loss(self.pred_cls, self.cls_gt)
self.size_loss = loss.reg_l1_loss(self.pred_size, self.size_gt)
self.total_loss = self.cls_loss + 0.1 * self.size_loss
def compute_loss(self, true_hm, true_wh, true_reg, reg_mask, ind):
hm_loss = loss.focal_loss(self.pred_hm, true_hm)
wg_loss = 0.05*loss.reg_l1_loss(self.pred_wh, true_wh, ind, reg_mask)
reg_loss = loss.reg_l1_loss(self.pred_reg, true_reg, ind, reg_mask)
return hm_loss, wg_loss, reg_loss
def compute_loss(self):
self.cls_loss = loss.focal_loss(self.pred_center, self.center_gt)
self.size_loss = loss.reg_l1_loss(self.pred_size, self.size_gt)
self.offset_loss = loss.reg_l1_loss(self.pred_offset, self.offset_gt)
# self.regular_loss=cfg.weight_decay * tf.add_n([tf.nn.l2_loss(var) for var in tf.trainable_variables()])
self.total_loss = self.cls_loss + 0.1 * self.size_loss + self.offset_loss #+self.regular_loss