在 variable_scope 下引入 Optimizer 变量时会重新创建两次，为什么？

Question

正如标题所暗示的，当查看 tensorboard 内的图表时 - 我在 variable_scope 内创建的变量被重新创建了两次。这是为什么？我做错了什么？

def weights_biases(weights_shape, biases_shape):
    weights = tf.get_variable("weights", weights_shape, initializer = tf.random_normal_initializer())
    biases = tf.get_variable("biases", biases_shape, initializer = tf.random_normal_initializer())

    return weights, biases

def hl_relu(input_tensor, weights_shape, biases_shape):
    weights, biases = weights_biases(weights_shape, biases_shape)
    regression = tf.matmul(input_tensor, weights) + biases
    return tf.nn.relu(regression)

def neural_network_model(x):
    # W = tf.Variable(
    #    tf.truncated_normal([vocab_size, embedding_size], stddev=1 / math.sqrt(vocab_size)),
    #    name="W")

    # embedded = tf.nn.embedding_lookup(W, x)    
    # embedding_aggregated = tf.reduce_sum(embedded, [1])

    with tf.variable_scope("hidden_layer_1"):
        relu1 = hl_relu(x, [max_words_len, n_nodes_hl1], [n_nodes_hl1])

    with tf.variable_scope("hidden_layer_2"):
        relu2 = hl_relu(relu1, [n_nodes_hl1,n_nodes_hl2], [n_nodes_hl2])

    with tf.variable_scope("hidden_layer_3"):
        relu3 = hl_relu(relu2, [n_nodes_hl2,n_nodes_hl3], [n_nodes_hl3])

    with tf.variable_scope("output_layer"):
        weights, biases = weights_biases([n_nodes_hl3, n_classes], [n_classes])
        output_regression = tf.matmul(relu3, weights) + biases

    return output_regression

def train_neural_network(test_x, test_y):
    with tf.device("/cpu:0"):
        custom_runner = CustomRunner()
        x_batch, y_batch = custom_runner.get_inputs()

        with tf.variable_scope("test"):
            testX = tf.constant(test_x, name="testX")
            testX = tf.cast(testX, tf.float32)
            testY = tf.constant(test_y, name="testY")
            testY = tf.cast(testY, tf.float32)

        with tf.variable_scope("nn") as scope:
            global_step = tf.Variable(0, trainable=False, name='global_step')    

            logits = neural_network_model(x_batch)
            scope.reuse_variables()
            test_logits = neural_network_model(testX)

            cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, y_batch), name="cost")
            tf.scalar_summary('cost', cost)
            optimizer = tf.train.AdagradOptimizer(0.01).minimize(cost, global_step = global_step)

产生如下异常：

您可以看到“nn”范围被创建了两次，其中我的隐藏层没有连接到任何输入，但通过不断提供它们的初始化随机权重来影响 Adagrad 优化器。我怀疑这也会减慢训练速度。

我的代码做错了什么？

Answer 1

我相信这就是导致它们被创建两次的原因。

logits = neural_network_model(x_batch)
scope.reuse_variables()
test_logits = neural_network_model(testX)

你能把它改成这样吗：

logits = neural_network_model(x_batch)

没有 test_logits，看看你是否仍然遇到同样的问题？

Answer 2

您确定重新创建了变量吗？我怀疑您看到的只是优化器创建的 Adagrad 变量，这样可以节省计算所需的内容。可以试试最简单的 GradientDescentOptimizer 看看还能不能出现？