About the RPN loss (smooth_L1_LOSS)

[hehualin-tut]

`class WeightedSmoothL1Loss(Loss): """Smooth L1 localization loss function. The smooth L1_loss is defined elementwise as .5 x^2 if |x|<1 and |x|-.5 otherwise, where x is the difference between predictions and target. See also Equation (3) in the Fast R-CNN paper by Ross Girshick (ICCV 2015) """

def _compute_loss(self, prediction_tensor, target_tensor, weight):
    """Compute loss function.
    Args:
        prediction_tensor: A float tensor of shape [num_anchors,
            code_size] representing the (encoded) predicted
            locations of objects.
        target_tensor: A float tensor of shape [num_anchors,
            code_size] representing the regression targets
    Returns:
      loss: an anchorwise tensor of shape [num_anchors] representing
        the value of the loss function
    """
    diff = prediction_tensor - target_tensor
    abs_diff = tf.abs(diff)
    abs_diff_lt_1 = tf.less(abs_diff, 1)

    anchorwise_smooth_l1norm = tf.reduce_sum(
        tf.where(abs_diff_lt_1, 0.5 * tf.square(abs_diff), abs_diff - 0.5),
        axis=1) * weight
    return anchorwise_smooth_l1norm`

Question: Could u tell me What is the dimension of this “code_size” in the prediction_tensor and target_tensor? I want to use Giou loss to replace the Initial L1 loss function， I SAW this "boxes_4c: (N, 10) [x1, x2, x3, x4, z1, z2, z3, z4, h1, h2]" in the wiki, SO the prediction_tensor and target_tensor is [N,10] ? Am I wrong? I created the loss function based on 10 dimensions.Can you help me see if that's right? “ Ap=(prediction_tensor[:, 0]-prediction_tensor[:, 2])(prediction_tensor[:, 4]-prediction_tensor[:, 6])(prediction_tensor[:, 9]-prediction_tensor[:, 8])#the volume of the Predicted box Ag=(target_tensor[:, 0]-target_tensor[:, 2])(target_tensor[:, 4]-target_tensor[:, 6])(target_tensor[:, 9]-target_tensor[:, 8])#the volume of the gt box x1=tf.maximum(prediction_tensor,target_tensor,name=None) x2=tf.minimum(prediction_tensor,target_tensor,name=None) I=(x2[:, 0]-x1[:, 2])(x2[:, 4]-x1[:, 6])(x2[:, 9]-x1[:, 8])#Two BOX coincident partial volume calculations Ac=(x1[:, 0]-x2[:, 2])(x1[:, 4]-x2[:, 6])(x1[:, 9]-x2[:, 8])#The minimum bracketing volume of two boxes is calculated U=Ap+Ag-I IOU=I/U#IOU GIOU=IOU-((Ac-U)/Ac)#GIOU anchorwise_smooth_l1norm=(1-GIOU)*weight#loss return anchorwise_smooth_l1norm ” If I'm not right, do you know how to fix it?I would be very grateful if you could give me some advice.