import tensorflow as tf
import tensorflow.examples.tutorials.mnist.input_data as input_data
mnist=input_data.read_data_sets('MNIST_data/',one_hot=True)

x=tf.placeholder(tf.float32,[None,784],name='x')
y=tf.placeholder(tf.float32,[None,10],name='y')

h1_num=256                        #設置隱藏層神經(jīng)元數(shù)量
W1=tf.Variable(tf.random_normal([784,h1_num]),name='W1')
b1=tf.Variable(tf.zeros([h1_num]),name='b1')
Y1=tf.nn.relu(tf.matmul(x,W1)+b1)             #激活函數(shù)

W2=tf.Variable(tf.random_normal([h1_num,10]),name='W2')
b2=tf.Variable(tf.zeros([10]),name='b2')
Y2=tf.matmul(Y1,W2)+b2
pred=tf.nn.softmax(Y2)

train_epochs=20                    #訓練輪數(shù)
batch_size=50                     #每個批次的樣本數(shù)
batch_num=int(mnist.train.num_examples/batch_size)  #一輪需要訓練多少批
learning_rate=0.01
#定義損失函數(shù)、優(yōu)化器
loss_function=tf.reduce_mean(             #softmax交叉熵損失函數(shù)
       tf.nn.softmax_cross_entropy_with_logits(logits=Y2,labels=y)) 
optimizer=tf.train.AdamOptimizer(learning_rate).minimize(loss_function)
#定義準確率
correct_prediction=tf.equal(tf.argmax(pred,1),tf.argmax(y,1))
accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))

ss=tf.Session()
ss.run(tf.global_variables_initializer())           #進行全部變量的初始化
 
for epoch in range(train_epochs):
  for batch in range(batch_num):              #分批次讀取數(shù)據(jù)進行訓練
    xs,ys=mnist.train.next_batch(batch_size)
    ss.run(optimizer,feed_dict={x:xs,y:ys})
  loss,acc=ss.run([loss_function,accuracy],\
          feed_dict={x:mnist.validation.images,y:mnist.validation.labels})
  print('第%2d輪訓練：損失為：%9f，準確率：%.4f'%(epoch+1,loss,acc))
 
ss.close()

test_res=ss.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels})
print('測試集的準確率為：%.4f'%(test_res))

#構建輸入層
x=tf.placeholder(tf.float32,[None,784],name='x')
y=tf.placeholder(tf.float32,[None,10],name='y')
#構建第一個隱藏層
h1_num=256                            #第一隱藏層神經(jīng)元數(shù)量256
W1=tf.Variable(tf.truncated_normal([784,h1_num],stddev=0.1),name='W1')
b1=tf.Variable(tf.zeros([h1_num]),name='b1')
Y1=tf.nn.relu(tf.matmul(x,W1)+b1)
#構建第二個隱藏層
h2_num=64                             #第二隱藏層神經(jīng)元數(shù)量64
W2=tf.Variable(tf.random_normal([h1_num,h2_num],stddev=0.1),name='W2')
b2=tf.Variable(tf.zeros([h2_num]),name='b2')
Y2=tf.nn.relu(tf.matmul(Y1,W2)+b2)
#構建輸出層
W3=tf.Variable(tf.random_normal([h2_num,10],stddev=0.1),name='W3')
b3=tf.Variable(tf.zeros([10]),name='b3')
Y3=tf.matmul(Y2,W3)+b3
pred=tf.nn.softmax(Y3)

#定義一個通用的全連接層函數(shù)模型
def fcn_layer(inputs,in_dim,out_dim,activation=None):
  W=tf.Variable(tf.truncated_normal([in_dim,out_dim],stddev=0.1))
  b=tf.Variable(tf.zeros([out_dim]))
  Y=tf.matmul(inputs,W)+b
  if activation==None:
    output=Y
  else:
    output=activation(Y)
  return output
#構建第一個隱藏層
Y1=fcn_layer(x,784,256,tf.nn.relu)
#構建第二個隱藏層
Y2=fcn_layer(Y1,256,64,tf.nn.relu)
#構建輸出層
Y3=fcn_layer(Y2,64,10)
pred=tf.nn.softmax(Y3)

import os
save_dir='D:/Temp/MachineLearning/ModelSaving/'    #定義模型的保存路徑
if not os.path.exists(save_dir):            #如果不存在該路徑則創(chuàng)建
  os.makedirs(save_dir)

save_step=5            #定義存儲粒度
 
saver=tf.train.Saver()      #定義saver

if epoch%save_step==0:
    saver.save(ss,os.path.join(save_dir,'mnist_fcn_{:02d}.ckpt'.format(epoch+1)))

saver.save(ss,os.path.join(save_dir,'mnist_fcn.ckpt'))

save_dir='D:/Temp/MachineLearning/ModelSaving/'    #定義模型的保存路徑
saver=tf.train.Saver()                 #定義saver
 
ss=tf.Session()
ss.run(tf.global_variables_initializer())
 
ckpt=tf.train.get_checkpoint_state(save_dir)      #讀取存盤點
if ckpt and ckpt.model_checkpoint_path:
  saver.restore(ss,ckpt.model_checkpoint_path)    #從存盤中恢復參數(shù)到當前的session
  print('數(shù)據(jù)恢復從',ckpt.model_checkpoint_path)
 
test_res=accuracy.eval(session=ss,feed_dict={x:mnist.test.images,y:mnist.test.labels})
print('測試集的準確率為：%.4f'%(test_res))

NotFoundError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint.

import tensorflow as tf
 
v=tf.Variable(1.0,'new_var')
with tf.Session() as ss:
  tf.train.write_graph(ss.graph_def,'D:\Temp\MachineLearning\ModelSaving\Graph',
            'test_graph.pb',as_text=False)

with tf.Session() as ss:
  with tf.gfile.GFile('D:/Temp\MachineLearning/ModelSaving/Graph/test_graph.pb','rb') as pb_file:
    graph_def=tf.GraphDef()
    graph_def.ParseFromString(pb_file.read())
    ss.graph.as_default()
    tf.import_graph_def(graph_def)
    print(graph_def)