深入浅出的TensorFlow可视化工具TensorBoard用法教程（二）

这次，让我们用一个真实的程序说明到底如何把TensorBoard应用到我们的真实生产环境中。这次我们的任务是搭建一个CNN神经网络，解决MNIST数字识别问题。

首先，把必要的Python库加载进来：

import tensorflow as tf
from tensorflow.contrib import slim
from tensorflow.examples.tutorials.mnist import input_data

slim是一个使构建，训练，评估神经网络变得简单的库。它可以消除原生tensorflow里面很多重复的模板性的代码，让代码更紧凑，更具备可读性。另外slim提供了很多计算机视觉方面的著名模型（VGG, AlexNet等），我们不仅可以直接使用，甚至能以各种方式进行扩展。

with tf.Graph().as_default() as graph:
    mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
    x = tf.placeholder(tf.float32, shape=[None, 784])
    y = tf.placeholder(tf.float32, shape=[None, 10])
    x_train = tf.reshape(x, [-1, 28, 28, 1])
    tf.summary.image('input', x_train, 10)

    with slim.arg_scope([slim.conv2d, slim.fully_connected],
            normalizer_fn=slim.batch_norm,
            activation_fn=tf.nn.relu):
        with slim.arg_scope([slim.max_pool2d], padding='SAME'):
            conv1 = slim.conv2d(x_train, 32, [5, 5], name='conv1')
            conv_vars = tf.get_collection(tf.GraphKeys.MODEL_VARIABLES, 'Conv')
            tf.summary.histogram('conv_weights', conv_vars[0])
            pool1 = slim.max_pool2d(conv1, [2, 2])
            conv2 = slim.conv2d(pool1, 64, [5, 5])
            pool2 = slim.max_pool2d(conv2, [2, 2])
            flatten = slim.flatten(pool2)
            fc = slim.fully_connected(flatten, 1024)

    logits = slim.fully_connected(fc, 10, activation_fn=None)
    softmax = tf.nn.softmax(logits, name='output')

    with tf.name_scope('loss'):
        loss = slim.losses.softmax_cross_entropy(logits, y)
        tf.summary.scalar('loss', loss)

    train_op = slim.optimize_loss(loss, slim.get_or_create_global_step(),
            learning_rate=0.01,
            optimizer='Adam')

    correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(y, 1))
    with tf.name_scope('accuracy'):
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        tf.summary.scalar('accuracy', accuracy)

    summary = tf.summary.merge_all()

被忘了把我们的模型启动起来，然后进行summary信息汇总，最后使用FileWriter输出到本地磁盘（这里是./cnn这个目录）。

with tf.Session() as sess:
    writer = tf.summary.FileWriter('cnn', sess.graph)
    sess.run(tf.global_variables_initializer())
    sess.run(tf.local_variables_initializer())
    for i in range(500):
        batch = mnist.train.next_batch(50)
        if i % 100 == 0:
            train_accuracy = accuracy.eval(feed_dict={x:batch[0], y:batch[1]})
            print("step %d, training accuracy %g"%(i, train_accuracy))
        _, w_summary = sess.run([train_op, summary], feed_dict={x:batch[0], y:batch[1]})
        writer.add_summary(w_summary, i)

最后，我们就可以使用tensorboard命令进行可视化分析了：

$ tensorboard --logdir=./cnn

打开浏览器，访问localhost:6006，一切都清晰明了。

调参

下面，我们来展示如何利用TensorBoard来进行模型超参数的选择（这里以学习速率为例）。

我们只用对代码进行少量的修改，把学习速率0.1、0.001、0.0001的日志信息保存到不同的子目录下即可。

import os
import tensorflow as tf
from tensorflow.contrib import slim
from tensorflow.examples.tutorials.mnist import input_data

def cnn_model(learning_rate):
    with tf.Graph().as_default() as graph:
        mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
        x = tf.placeholder(tf.float32, shape=[None, 784])
        y = tf.placeholder(tf.float32, shape=[None, 10])
        x_train = tf.reshape(x, [-1, 28, 28, 1])
        tf.summary.image('input', x_train, 10)

        with slim.arg_scope([slim.conv2d, slim.fully_connected],
                normalizer_fn=slim.batch_norm,
                activation_fn=tf.nn.relu):
            with slim.arg_scope([slim.max_pool2d], padding='SAME'):
                conv1 = slim.conv2d(x_train, 32, [5, 5])
                conv_vars = tf.get_collection(tf.GraphKeys.MODEL_VARIABLES, 'Conv')
                tf.summary.histogram('conv_weights', conv_vars[0])
                pool1 = slim.max_pool2d(conv1, [2, 2])
                conv2 = slim.conv2d(pool1, 64, [5, 5])
                pool2 = slim.max_pool2d(conv2, [2, 2])
                flatten = slim.flatten(pool2)
                fc = slim.fully_connected(flatten, 1024)

        logits = slim.fully_connected(fc, 10, activation_fn=None)
        softmax = tf.nn.softmax(logits, name='output')

        with tf.name_scope('loss'):
            loss = slim.losses.softmax_cross_entropy(logits, y)
            tf.summary.scalar('loss', loss)

        train_op = slim.optimize_loss(loss, slim.get_or_create_global_step(),
                learning_rate=learning_rate,
                optimizer='Adam')

        with tf.name_scope('accuracy'):
            correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(y, 1))
            accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
            tf.summary.scalar('accuracy', accuracy)
        summary = tf.summary.merge_all()

    return {'x': x, 'y': y, 'accuracy': accuracy, 'summary': summary, 'mnist': mnist}, train_op, graph

for learning_rate in [0.1, 0.01, 0.001]:
    vars, train_op, graph = cnn_model(learning_rate)
    with tf.Session(graph=graph) as sess:
        log_dir = os.path.join('cnn', 'cnn-lr-{}'.format(learning_rate))
        writer = tf.summary.FileWriter(log_dir, sess.graph)
        sess.run(tf.global_variables_initializer())
        sess.run(tf.local_variables_initializer())
        for i in range(500):
            batch = vars['mnist'].train.next_batch(50)
            if i % 100 == 0:
                train_accuracy = vars['accuracy'].eval(feed_dict={vars['x']:batch[0],
                    vars['y']:batch[1]})
                print("step %d, training accuracy %g"%(i, train_accuracy))
            _, w_summary = sess.run([train_op, vars['summary']],
                    feed_dict={vars['x']:batch[0], vars['y']:batch[1]})
            writer.add_summary(w_summary, i)

然后跟之前一样，使用tensorboard命令

$ tensorboard --logdir=./cnn

打开浏览器，访问localhost:6006

显然，橙色曲线的loss最低，也最快收敛，由此可知把最优学习速率应该在0.001附近。