Tensorboard不创建网络图（Python）

Question

我真的不明白为什么张量板没有显示我的网络图。我已经按照Tensorboard Website上的教程和网络上的其他内容，这些都不允许显示图表。

我正在嵌入与网络相关的代码部分。我试图删除所有其他部分，但我不想减少太多，否则它可能会造成混乱。它在图表部分显示的唯一内容是global_step。

#tf.reset_default_graph()

# create a glboal step variable
global_step = tf.Variable(0, name='global_step', trainable=False)

with tf.Session() as sess:

    # Writes Tensorboard summaries to disk
    summary_writer = None
        with tf.variable_scope(scope):
            # Build the network

            # Our input are 4 RGB frames of shape ?, ? each
            X_pl = tf.placeholder(shape=[None, 50, 50, 4], dtype=tf.uint8, name="X")
            # The TD target value
            y_pl = tf.placeholder(shape=[None], dtype=tf.float32, name="y")
            # Integer id of which action was selected
            actions_pl = tf.placeholder(shape=[None], dtype=tf.int32, name="actions")

            X = tf.to_float(X_pl) / 255.0
            batch_size = tf.shape(X_pl)[0]

            # three convolutional layers
            with tf.name_scope('Convolutional_Layer_1') as scope:
                conv1 = tf.contrib.layers.conv2d(X, 32, 8, 4, activation_fn=tf.nn.relu)

            # other conv layers

            # flattened layer
            with tf.name_scope('Flattened_Layer') as scope:
                    flattened = tf.contrib.layers.flatten(conv3)

            # fully connected layer
            with tf.name_scope('Fully_Connected_Layer') as scope:
                fc1 = tf.contrib.layers.fully_connected(flattened, 512)
                predictions = tf.contrib.layers.fully_connected(fc1, len(VALID_ACTIONS))

            # other stuff

            # reduce input losses, return tensor with single element
            with tf.name_scope('Loss') as scope:
                    loss = tf.reduce_mean(losses)

            # optimizer parameters from original paper
            with tf.name_scope('Optimization') as scope:
                optimizer = tf.train.RMSPropOptimizer(0.00025, 0.99, 0.0, 1e-6)
                train_op = optimizer.minimize(loss, global_step=tf.contrib.framework.get_global_step())

            # summaries for Tensorboard
            tf.summary.scalar("Loss", loss)
            tf.summary.scalar("Max_Q_Value", tf.reduce_max(predictions))
            tf.summary.histogram("Loss_Hist", losses)
            tf.summary.histogram("Q_Values_Hist", predictions)

            summaries = tf.summary.merge_all()

            if summaries_dir:
            summary_dir = os.path.join(summaries_dir, "summaries_{}".format(scope))
            # create directory if does not exist
            if not os.path.exists(summary_dir):
                os.makedirs(summary_dir)
                    summary_writer = tf.summary.FileWriter(summary_dir, graph=sess.graph)

            sess.run(tf.global_variables_initializer())

            # other stuff not important

            total_t = sess.run(tf.contrib.framework.get_global_step())

            # other stuff not important

            for i_episode in range(num_episodes):

            # other stuff not important

            # one step in the environment
            for t in itertools.count():

                # other stuff not important

                # add epsilon to Tensorboard
                episode_summary = tf.Summary()
                episode_summary.value.add(simple_value=epsilon, tag="epsilon")
                q_estimator.summary_writer.add_summary(episode_summary, total_t)

                # other stuff not important
                break

            # add summaries to tensorboard
            episode_summary = tf.Summary()
            episode_summary.value.add(simple_value=stats.episode_rewards[i_episode], node_name="episode_reward", tag="Episode_Reward")
            episode_summary.value.add(simple_value=stats.episode_lengths[i_episode], node_name="episode_length", tag="Episode_Length")
            q_estimator.summary_writer.add_summary(episode_summary, total_t)
            q_estimator.summary_writer.flush()

我刚开始使用Tensorboard，但我能够按照相同的过程显示其他图形。我不明白现在有什么不对。

提前致谢。

Answer 1

我运行你的代码，直到行sess.run(tf.global_variables_initializer())（太多未定义的变量更进一步）之后，预期的图形显示在Tensorboard中（见下图）。

你确定：

• 你启动/重新启动Tensorboard指向正确的logdir（tensorboard --logdir=[path_to_summary_dir]）？
• 您之后的代码中没有以某种方式覆盖/删除您的tfevents文件（例如，使用您的q_estimator.summary_writer）？