张量流中的角度损失不会减少

Question

最近我正在实施＆＃34; FC4：具有置信度加权池的完全卷积颜色稳定性＆＃34;张量流中的纸张。在论文中，角度损失定义为＆＃34;角度损失= arccos（cosine_distance（vector1，vector2））＆＃34;。但是，我尝试了几种方法来实现这种损失功能，但在训练中它们几乎没有减少。相关代码如下所示，任何帮助和建议都表示赞赏。

def build_model(self):

    self.learning_rate = tf.placeholder(tf.float32)
    self.input = tf.placeholder(tf.float32, [self.batch_size, None, None, 3])
    self.target = tf.placeholder(tf.float32, [self.batch_size, 3])
    #the shape of result_vec is [self.batch_size, 3]
    self.result_vec = self.build_network(self.input, 'color_net')
    cos_dist = tf.losses.cosine_distance(self.target, self.result_vec, dim=1)
    cos_dist_list,acos_list = tf.unstack(cos_dist), []
    for dist in cos_dist_list:
        acos_list.append(tf.acos(dist)) 
    stack_loss = tf.stack(acos_list)
    self.angular_loss = tf.reduce_mean(stack_loss)
    var_list = tf.trainable_variables()
    self.optim = tf.train.AdamOptimizer(learning_rate = self.learning_rate)\
                                        .minimize(self.angular_loss, var_list = var_list)