ctrl_model.py 文件源码

def compute_loss_reg(self, sim_reg_mat, offset_label):

        sim_score_mat, p_reg_mat, l_reg_mat = tf.split(2, 3, sim_reg_mat)
        sim_score_mat = tf.reshape(sim_score_mat, [self.batch_size, self.batch_size])
        l_reg_mat = tf.reshape(l_reg_mat, [self.batch_size, self.batch_size])
        p_reg_mat = tf.reshape(p_reg_mat, [self.batch_size, self.batch_size])
        # unit matrix with -2
        I_2 = tf.diag(tf.constant(-2.0, shape=[self.batch_size]))
        all1 = tf.constant(1.0, shape=[self.batch_size, self.batch_size])
        #               | -1  1   1...   |

        #   mask_mat =  | 1  -1  -1...   |

        #               | 1   1  -1 ...  |
        mask_mat = tf.add(I_2, all1)
        # loss cls, not considering iou
        I = tf.diag(tf.constant(1.0, shape=[self.batch_size]))
        I_half = tf.diag(tf.constant(0.5, shape=[self.batch_size]))
        batch_para_mat = tf.constant(self.alpha, shape=[self.batch_size, self.batch_size])
        para_mat = tf.add(I,batch_para_mat)
        loss_mat = tf.log(tf.add(all1, tf.exp(tf.mul(mask_mat, sim_score_mat))))
        loss_mat = tf.mul(loss_mat, para_mat)
        loss_align = tf.reduce_mean(loss_mat)
        # regression loss
        l_reg_diag = tf.matmul(tf.mul(l_reg_mat, I), tf.constant(1.0, shape=[self.batch_size, 1]))
        p_reg_diag = tf.matmul(tf.mul(p_reg_mat, I), tf.constant(1.0, shape=[self.batch_size, 1]))
        offset_pred = tf.concat(1, (p_reg_diag, l_reg_diag))
        loss_reg = tf.reduce_mean(tf.abs(tf.sub(offset_pred, offset_label)))

        loss=tf.add(tf.mul(self.lambda_regression, loss_reg), loss_align)
        return loss, offset_pred, loss_reg