model.py 文件源码

def add_loss_op(self, target):
        self.target = target  # for outputting later
        real_target = tf.abs(self.target)

        # mean = tf.concat([stats[0][0], stats[0][1]])
        # stdev = tf.concat([stats[1][0], stats[1][1]])

        # print(mean.get_shape())
        # print(stdev.get_shape())

        # real_target -= mean
        # real_target /= stdev

        delta = self.output - real_target 
        squared_error = tf.reduce_mean(tf.pow(delta, 2)) 

        l2_cost = tf.reduce_mean([tf.norm(v) for v in tf.trainable_variables() if len(v.get_shape().as_list()) == 3])

        self.loss = Config.l2_lambda * l2_cost + squared_error

        tf.summary.scalar("loss", self.loss)

        masked_loss = tf.abs(self.soft_masked_output) - real_target
        self.masked_loss = Config.l2_lambda * l2_cost + tf.reduce_mean(tf.pow(masked_loss, 2))
        tf.summary.scalar('masked_loss', self.masked_loss)
        tf.summary.scalar('regularization_cost', Config.l2_lambda * l2_cost)