def loss(self, inputs, truth_boxes, truth_labels, truth_instances):
self.rpn_cls_loss, self.rpn_reg_loss = \
rpn_loss(self.rpn_logits_flat, self.rpn_deltas_flat, self.rpn_labels, self.rpn_label_weights,
self.rpn_targets, self.rpn_target_weights)
self.rcnn_cls_loss, self.rcnn_reg_loss = \
rcnn_loss(self.rcnn_logits, self.rcnn_deltas, self.rcnn_labels, self.rcnn_targets)
if self.train_box_only:
self.total_loss = \
self.rpn_cls_loss + \
self.rpn_reg_loss + \
self.rcnn_cls_loss + \
self.rcnn_reg_loss
else:
self.mask_cls_loss = \
mask_loss(self.mask_logits, self.mask_labels, self.mask_instances)
self.total_loss = \
self.rpn_cls_loss + \
self.rpn_reg_loss + \
self.rcnn_cls_loss + \
self.rcnn_reg_loss + \
self.mask_cls_loss
# self.total_loss = self.rpn_cls_loss + self.rpn_reg_loss + self.rcnn_cls_loss + self.rcnn_reg_loss
# self.total_loss = self.rpn_cls_loss + self.rpn_reg_loss + self.mask_cls_loss
# self.total_loss = self.rcnn_cls_loss + self.rcnn_reg_loss + self.mask_cls_loss
return self.total_loss
def loss_train_rcnn(self, inputs, truth_boxes, truth_labels, truth_instances):
self.rpn_cls_loss, self.rpn_reg_loss = \
rpn_loss(self.rpn_logits_flat, self.rpn_deltas_flat, self.rpn_labels, self.rpn_label_weights,
self.rpn_targets, self.rpn_target_weights)
self.rcnn_cls_loss, self.rcnn_reg_loss = \
rcnn_loss(self.rcnn_logits, self.rcnn_deltas, self.rcnn_labels, self.rcnn_targets)
self.total_loss = self.rpn_cls_loss + self.rpn_reg_loss + self.rcnn_cls_loss + self.rcnn_reg_loss
return self.total_loss
def loss(self, inputs, truth_boxes, truth_labels, truth_instances):
cfg = self.cfg
self.rpn_cls_loss, self.rpn_reg_loss = \
rpn_loss( self.rpn_logits_flat, self.rpn_deltas_flat, self.rpn_labels, self.rpn_label_weights, self.rpn_targets, self.rpn_target_weights)
self.rcnn_cls_loss, self.rcnn_reg_loss = \
rcnn_loss(self.rcnn_logits, self.rcnn_deltas, self.rcnn_labels, self.rcnn_targets)
## self.mask_cls_loss = Variable(torch.cuda.FloatTensor(1).zero_()).sum()
self.mask_cls_loss = \
mask_loss( self.mask_logits, self.mask_labels, self.mask_instances )
self.total_loss = self.rpn_cls_loss + self.rpn_reg_loss \
+ self.rcnn_cls_loss + self.rcnn_reg_loss \
+ self.mask_cls_loss
return self.total_loss