5ab5db26ff2cf0f09804b34b5533c8d59f53667b,frcnn/rcnn.py,RCNN,loss,#RCNN#Any#,73

Before Change

                    bbox_offset_cleaned, bbox_offsets_target_labeled)

                // Hack to avoid having nan loss.
                reg_loss = tf.cond(tf.is_nan(reg_loss), lambda: tf.constant(0.0, dtype=tf.float32), lambda: reg_loss)

                return {
                    "rcnn_cls_loss": cls_loss,
                    "rcnn_reg_loss": reg_loss,

After Change

                    labels=cls_target_one_hot, logits=cls_score_labeled
                )

                prediction_dict["cross_entropy_per_proposal"] = cross_entropy_per_proposal

                // Second we need to calculate the smooth l1 loss between
                // `bbox_offsets` and `bbox_offsets_target`.

Instances