def loss(self, prediction_dict):
# Note: The loss should be using mini-batch values only
loss_dict, rpn_loss = self._rpn_model.loss(prediction_dict)
losses_output = avod_loss_builder.build(self, prediction_dict)
classification_loss = \
losses_output[avod_loss_builder.KEY_CLASSIFICATION_LOSS]
final_reg_loss = losses_output[avod_loss_builder.KEY_REGRESSION_LOSS]
avod_loss = losses_output[avod_loss_builder.KEY_AVOD_LOSS]
offset_loss_norm = \
losses_output[avod_loss_builder.KEY_OFFSET_LOSS_NORM]
loss_dict.update({self.LOSS_FINAL_CLASSIFICATION: classification_loss})
loss_dict.update({self.LOSS_FINAL_REGRESSION: final_reg_loss})
# Add localization and orientation losses to loss dict for plotting
loss_dict.update({self.LOSS_FINAL_LOCALIZATION: offset_loss_norm})
ang_loss_loss_norm = losses_output.get(
avod_loss_builder.KEY_ANG_LOSS_NORM)
if ang_loss_loss_norm is not None:
loss_dict.update({self.LOSS_FINAL_ORIENTATION: ang_loss_loss_norm})
with tf.variable_scope('model_total_loss'):
total_loss = rpn_loss + avod_loss
return loss_dict, total_loss
def loss(self, prediction_dict):
# Note: The loss should be using mini-batch values only
loss_dict, rpn_loss, rpn_acc_all, rpn_acc_pos, \
rpn_acc_score_neg, rpn_acc_score_pos = self._rpn_model.loss(prediction_dict)
losses_output, avod_acc_all, avod_acc_pos = avod_loss_builder.build(
self, prediction_dict)
classification_loss = \
losses_output[avod_loss_builder.KEY_CLASSIFICATION_LOSS]
final_reg_loss = losses_output[avod_loss_builder.KEY_REGRESSION_LOSS]
avod_loss = losses_output[avod_loss_builder.KEY_AVOD_LOSS]
offset_loss_norm = \
losses_output[avod_loss_builder.KEY_OFFSET_LOSS_NORM]
loss_dict.update({self.LOSS_FINAL_CLASSIFICATION: classification_loss})
loss_dict.update({self.LOSS_FINAL_REGRESSION: final_reg_loss})
# Add localization and orientation losses to loss dict for plotting
loss_dict.update({self.LOSS_FINAL_LOCALIZATION: offset_loss_norm})
ang_loss_loss_norm = losses_output.get(
avod_loss_builder.KEY_ANG_LOSS_NORM)
if ang_loss_loss_norm is not None:
loss_dict.update({self.LOSS_FINAL_ORIENTATION: ang_loss_loss_norm})
with tf.variable_scope('model_total_loss'):
if self.TRAIN_REFINE:
total_loss = rpn_loss + avod_loss
else:
total_loss = rpn_loss
rpn_score_2d_loss = loss_dict[self._rpn_model.LOSS_RPN_SCORE_2D]
rpn_class_loss = loss_dict[self._rpn_model.LOSS_RPN_OBJECTNESS]
rpn_reg_loss = loss_dict[self._rpn_model.LOSS_RPN_REGRESSION]
refine_class_loss = classification_loss
refine_reg_loss = final_reg_loss
return loss_dict, \
total_loss, \
rpn_score_2d_loss, \
rpn_acc_score_neg, \
rpn_acc_score_pos, \
rpn_class_loss, \
rpn_reg_loss, \
rpn_acc_all, \
rpn_acc_pos, \
refine_class_loss, \
refine_reg_loss, \
avod_acc_all, \
avod_acc_pos
