Question

我尝试构建图然后运行它，但是我仍然得到

RuntimeError：“会话”图为空。向图形添加操作在调用run（）之前。

此图形内置的功能是

W/System.err: org.json.JSONException:

然后，我尝试使用此图运行会话，如下所示：

def init_network(self):
    self.graph = tf.Graph()
    with self.graph.as_default():
        self.lstm = tf.contrib.rnn.LSTMCell(self.state_variables)
        self.state = self.lstm.zero_state(1, dtype=tf.float32)
        self.weights = tf.get_variable("Weights",
                                       shape=[self.state_variables, 1],
                                       dtype=tf.float32)

        self.lstm_output_ph = tf.placeholder(shape=[1, self.state_variables], dtype=tf.float32)
        inner_product = tf.linalg.matmul(self.lstm_output_ph, self.weights)
        q_estimate = tf.nn.softmax(inner_product)
        self.reward_ph = tf.placeholder(shape=[1], dtype=tf.float32)
        self.previous_q_ph = tf.placeholder(shape=[1], dtype=tf.float32)
        loss = tf.subtract(tf.add(self.reward_ph, tf.multiply(self.memory, self.previous_q_ph)), q_estimate)

        optimizer = tf.train.GradientDescentOptimizer(
            self.learning_rate,
            use_locking=False,
            name='SGD'
        )

        self.train_step = optimizer.minimize(loss)

        self.state_ph = tf.placeholder(shape=[1, self.state_variables], dtype=tf.float32)
        self.last_output, self.state = self.lstm(self.state_ph, self.state)
        inner_product_predict = tf.linalg.matmul(self.last_output, self.weights)
        q_estimate_init = tf.nn.softmax(inner_product_predict)
        self.predict_step = q_estimate_init

知道我的图表为何仍为空吗？调试 init_network 方法表明，即使在方法结束时， self.graph 变量仍为空（其 _graph_key 变量仍设置为）。 'grap-key-0 /'）。

Answer 1

这不是一个实际的答案（如果有人发布更好的答案，我会删除它；如果我找到实际的解释，我会予以改善）。我设法通过以下方式更改图形定义来解决此问题：

def build_graph(self, init_state):
    self.graph = tf.Graph()
    with self.graph.as_default():
        self.lstm = tf.keras.layers.CuDNNLSTM(units=4, stateful=True)
        self.state = self.lstm.get_initial_state(inputs=init_state)
        self.weights = tf.get_variable("Weights",
                                       shape=[self.state_variables, 1],
                                       dtype=tf.float32,
                                       collections=[tf.GraphKeys.GLOBAL_VARIABLES,
                                                    tf.GraphKeys.TRAINABLE_VARIABLES])

        self.input_ph = tf.placeholder(
            name="input",
            shape=[1, 1, self.state_variables],
            dtype=tf.float32)

        self.output_step = self.lstm(self.input_ph)

        self.lstm_output_ph = tf.placeholder(
            name="lstm_output",
            shape=[1, self.state_variables],
            dtype=tf.float32)

        self.predict_step = tf.linalg.matmul(self.lstm_output_ph, tf.nn.softmax(self.weights))

        self.reward_ph = tf.placeholder(
            name="reward",
            dtype=tf.float32)

        self.previous_q_ph = tf.placeholder(
            name="previous_q",
            dtype=tf.float32)

        self.loss = tf.losses.mean_squared_error(
            labels=self.reward_ph + self.memory * self.previous_q_ph,
            predictions=self.predict_step
        )

        optimizer = tf.train.GradientDescentOptimizer(
            learning_rate=self.learning_rate,
            use_locking=False,
            name='SGD'
        )

        self.train_step = optimizer.minimize(
            self.loss,
            var_list=[self.weights],
            gate_gradients=optimizer.GATE_OP,
            aggregation_method=tf.AggregationMethod.DEFAULT,
            colocate_gradients_with_ops=False,
            name='GD_optimizer'
        )

        self.var_init = tf.global_variables_initializer()

但是，我不知道确切的问题是什么。

构建后，“会话”图为空

1 个答案: