network_resnet.py 文件源码

def _smooth_l1_loss(self, bbox_pred, bbox_targets, bbox_inside_weights, bbox_outside_weights, sigma=1.0, dim=[1]):
    sigma_2 = sigma ** 2
    box_diff = bbox_pred - bbox_targets
    in_box_diff = bbox_inside_weights * box_diff
    abs_in_box_diff = tf.abs(in_box_diff)
    smoothL1_sign = tf.stop_gradient(tf.to_float(tf.less(abs_in_box_diff, 1. / sigma_2)))
    in_loss_box = tf.pow(in_box_diff, 2) * (sigma_2 / 2.) * smoothL1_sign \
                  + (abs_in_box_diff - (0.5 / sigma_2)) * (1. - smoothL1_sign)
    out_loss_box = bbox_outside_weights * in_loss_box
    loss_box = tf.reduce_mean(tf.reduce_sum(
      out_loss_box,
      axis=dim
    ))
    return loss_box