直接上程式碼：

# -*- coding: utf-8 -*-
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
max_step=1000
learning_rate=0.001
dropout=0.9
data_dir='../dataset/mnist/'
log_dir='./tfboard3'
mnist=input_data.read_data_sets(data_dir,one_hot=True)
sess=tf.InteractiveSession()

with tf.name_scope('input'):
    x=tf.placeholder(tf.float32,[None,784],name='X_input')
    y=tf.placeholder(tf.float32,[None,10],name='y_input')
with tf.name_scope('input_reshape'):
    image_shaped_input=tf.reshape(x,[-1,28,28,1])
    tf.summary.image('input',image_shaped_input)

def weight_variable(shape):
    initial=tf.truncated_normal(shape,stddev=0.1)
    return tf.Variable(initial)
def bias_variable(shape):
    initial=tf.constant(0.1,shape=shape)
    return tf.Variable(initial)

def variable_summaries(var):
    with tf.name_scope('summeries'):
        mean=tf.reduce_mean(var)
        tf.summary.scalar('mean',mean)
        with tf.name_scope('stddev'):
            stddev=tf.sqrt(tf.reduce_mean(tf.square(var-mean)))
        tf.summary.scalar('stddev',stddev)
        tf.summary.scalar('max',tf.reduce_max(var))
        tf.summary.scalar('min',tf.reduce_min(var))
        tf.summary.histogram('histogram',var)

def nn_layer(input_tensor,input_dim,output_dim,layer_name,act=tf.nn.relu):
    with tf.name_scope(layer_name):
        with tf.name_scope('weight'):
            weights=weight_variable([input_dim,output_dim])
            variable_summaries(weights)
        with tf.name_scope('biases'):#same layer ,all iamges share one biase
            biases=bias_variable([output_dim])
            variable_summaries(biases)
        with tf.name_scope('wx_plus_b'):
            preactivate=tf.matmul(input_tensor,weights)+biases
            tf.summary.histogram('preactivate',preactivate)
        activations=act(preactivate,name='activate')
        return activations #not really calcute
hidden1=nn_layer(x,784,500,'layer1')

with tf.name_scope('dropout'):
    keep_prob=tf.placeholder(tf.float32)
    tf.summary.scalar('keep_prob',keep_prob)
    droped=tf.nn.dropout(hidden1,keep_prob)
  #calcute relu,print(name,value) to see
y1=nn_layer(droped,500,10,'layer2',act=tf.identity)

with tf.name_scope('cross_entropy'):
    diff=tf.nn.softmax_cross_entropy_with_logits(logits=y1,labels=y)
    with tf.name_scope('total'):
        cross_entropy=tf.reduce_mean(diff)
tf.summary.scalar('cross_entropy',cross_entropy)

with tf.name_scope('train'):
    train_step=tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy)
with tf.name_scope('accuracy_prediction'):
    correct_prediction=tf.equal(tf.arg_max(y1,1),tf.arg_max(y,1))
    with tf.name_scope('accuracy'):
        accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
tf.summary.scalar('accuracy',accuracy)

merged=tf.summary.merge_all()                      #learning this write way
train_writer=tf.summary.FileWriter(log_dir+'/trian',sess.graph)
test_writer=tf.summary.FileWriter(log_dir+'/test')#save events,but not tfboard
tf.global_variables_initializer().run()

def feed_dict(train):
    if train:
        xs,ys=mnist.train.next_batch(100)
        k=dropout
    else:
        xs,ys=mnist.test.images,mnist.test.labels
        k=1.0#key=placeholder_name ,learn 3 elements
    return {x:xs,y:ys,keep_prob:k}

saver=tf.train.Saver()
for i in range(max_step):
    if i%10==0:
        summary,acc=sess.run([merged,accuracy],feed_dict=feed_dict(False))
        test_writer.add_summary(summary,i)#learn %s can value int,float
        print('accuracy at step %s:%s' %(i,acc))
    else:
        if i%100==99:#???
            run_options=tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
            run_metadata=tf.RunMetadata()#???
            summary,_=sess.run([merged,train_step],feed_dict=feed_dict(True),
                               options=run_options,run_metadata=run_metadata)
            train_writer.add_run_metadata(run_metadata,'step%03d' %i)#??
            train_writer.add_summary(summary,i)
            saver.save(sess,log_dir+'/model.ckpt',global_step=i)
            print('adding run metadata for %s' %i)
        else:
            summary,_=sess.run([merged,train_step],feed_dict=feed_dict(True))
            train_writer.add_summary(summary,i)
train_writer.close()
test_writer.close()
            
    
 在tensorboard的顯示結果

參考連結：

https://blog.csdn.net/smf0504/article/details/56369758