我怎样才能摆脱这个 ValueError

Question

权重和偏差

self.weights = { "layer1":tensorflow.Variable(tensorflow.random_normal([self.state_size, self.neurons_layer_1])), "layer2":tensorflow.Variable(tensorflow.random_normal([self.neurons_layer_1, self.neurons_layer_2])), "layerOut":tensorflow.Variable(tensorflow.random_normal([self.neurons_layer_2, self.action_size])) }
self.biases = { "layer1":tensorflow.Variable(tensorflow.random_normal([self.neurons_layer_1])), "layer2":tensorflow.Variable(tensorflow.random_normal([self.neurons_layer_2])), "layerOut":tensorflow.Variable(tensorflow.random_normal([self.action_size])) }

模型函数

def model(self, X, weights, biases):
    layer1 = tensorflow.add(tensorflow.matmul(X, weights["layer1"]),biases["layer1"])
    layer1 = tensorflow.nn.relu(layer1)

    layer2 = tensorflow.add(tensorflow.matmul(layer1, weights["layer2"]), biases["layer2"])
    layer2 = tensorflow.nn.relu(layer2)

    layerOut = tensorflow.add(tensorflow.matmul(layer2, weights["layerOut"]), biases["layerOut"])
    layerOut = tensorflow.nn.softmax(layerOut)

    with tensorflow.Session() as sess:
        init_op = tensorflow.global_variables_initializer()
        sess.run(init_op)
        q_values = sess.run(layerOut)
        return q_values

我这样计算预测值。状态是这样的列表：

predicted = self.model(state, self.weights, self.biases)    # state = [0.4, 0.3, 0.2, 0.9]

这些是我的占位符

X_placeholder = tensorflow.placeholder("float", [None, state_size])
Y_placeholder = tensorflow.placeholder("float", [None, action_size])

这是我的成本函数

cost = tensorflow.reduce_mean(tensorflow.square(right_prediction - predicted))

这是我的优化器

optimizer = tensorflow.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(cost)

然后我初始化上面的变量

init = tensorflow.global_variables_initializer()

并运行会话

with tensorflow.Session() as sess:
                sess.run(init)

                for epoch in range(1):
                    sess.run(optimizer, feed_dict={X_placeholder:X , Y_placeholder:Y})

现在我运行这段代码，它会抛出以下错误。

ValueError: No gradients provided for any variable, check your graph for ops that do not 
support gradients, between variables ["<tf.Variable 'Variable:0' shape=(4, 24) 
dtype=float32_ref>", "<tf.Variable 'Variable_1:0' shape=(24, 24) dtype=float32_ref>", 
"<tf.Variable 'Variable_2:0' shape=(24, 2) dtype=float32_ref>", "<tf.Variable 
'Variable_3:0' shape=(24,) dtype=float32_ref>", "<tf.Variable 'Variable_4:0' shape=(24,) 
dtype=float32_ref>", "<tf.Variable 'Variable_5:0' shape=(2,) dtype=float32_ref>", 
"<tf.Variable 'Variable_6:0' shape=(4, 24) dtype=float32_ref>", "<tf.Variable 
'Variable_7:0' shape=(24, 24) dtype=float32_ref>", "<tf.Variable 'Variable_8:0' 
shape=(24, 2) dtype=float32_ref>", 
"<tf.Variable 'Variable_9:0' shape=(24,) dtype=float32_ref>", 
"<tf.Variable 'Variable_10:0' shape=(24,) dtype=float32_ref>", 
"<tf.Variable 'Variable_11:0' shape=(2,) 
dtype=float32_ref>"] and loss Tensor("Mean:0", shape=(), dtype=float32).

所以请帮帮我。

感谢您的关注。

编辑：好的，这是我的全部代码，我不知道这是否是通过此链接发布的合法，但 here 是。

Answer 1

错误表明从优化器到参数没有连接方式，换句话说，优化器不知道参数（权重和偏差）。

由于您没有发布整个代码，因此您似乎在一个类中定义了参数和模型，并在其外部使用优化器，以相同的方式执行它们，或者定义参数、模型、成本和类内部的优化器或在外部定义所有优化器。