原博客地址：https://blog.csdn.net/wc781708249/article/details/78043099?locationNum=10&fps=1

参考简介博客地址：https://blog.csdn.net/michael_yt/article/details/74737489

简介

当我们使用 tensorflow 训练神经网络的时候，模型持久化对于我们的训练有很重要的作用。

如果我们的神经网络比较复杂，训练数据比较多，那么我们的模型训练就会耗时很长，如果在训练过程中出现某些不可预计的错误，导致我们的训练意外终止，那么我们将会前功尽弃。为了避免这个问题，我们就可以通过模型持久化（保存为CKPT格式）来暂存我们训练过程中的临时数据。
如果我们训练的模型需要提供给用户做离线的预测，那么我们只需要前向传播的过程，只需得到预测值就可以了，这个时候我们就可以通过模型持久化（保存为PB格式）只保存前向传播中需要的变量并将变量的值固定下来，这个时候只需用户提供一个输入，我们就可以通过模型得到一个输出给用户。

0、基础

# 保存图表并保存变量参数from tensorflow.python.framework import graph_util
var_list=tf.global_variables()
constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def,output_node_names=[var_list[i].name for i in range(len(var_list))]) # 保存图表并保存变量参数
tf.train.write_graph(constant_graph, './output', 'expert-graph.pb', as_text=False)# -----方式2-------------------
from tensorflow.python.framework import graph_util
var_list=tf.global_variables()
constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def,output_node_names=[var_list[i].name for i in range(len(var_list))])
with tf.gfile.FastGFile(logdir+'expert-graph.pb', mode='wb') as f:f.write(constant_graph.SerializeToString())

# 只保留图表
graph_def = tf.get_default_graph().as_graph_def()
with gfile.GFile('./output/export.pb', 'wb') as f:f.write(graph_def.SerializeToString())# 或者
tf.train.write_graph(graph_def, './output', 'expert-graph.pb', as_text=False)

1、保存pb文件

import tensorflow as tf
from tensorflow.python.framework import graph_utillogdir='./output/'with tf.variable_scope('conv'):w=tf.get_variable('w',[2,2],tf.float32,initializer=tf.random_normal_initializer)b=tf.get_variable('b',[2],tf.float32,initializer=tf.random_normal_initializer)sess=tf.InteractiveSession()tf.global_variables_initializer().run() # 初始化所有变量# tf.train.write_graph(tf.get_default_graph(),logdir,'expert-graph.pb',False)constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ["conv/w"])
with tf.gfile.FastGFile(logdir+'expert-graph.pb', mode='wb') as f:f.write(constant_graph.SerializeToString())sess.close()

或参考：https://github.com/tensorflow/models/blob/master/research/slim/export_inference_graph.py

# -*- coding: UTF-8 -*-
import tensorflow as tf
from tensorflow.python.framework import graph_utillogdir='./'placeholder = tf.placeholder(name='input', dtype=tf.float32,shape=[None, 28* 28* 1])# tf.train.write_graph(tf.get_default_graph(),logdir,'expert-graph.pb',False)
# constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ["conv/w"])
constant_graph = tf.get_default_graph().as_graph_def()
with tf.gfile.FastGFile(logdir+'expert-graph.pb', mode='wb') as f:f.write(constant_graph.SerializeToString())

2、导入pb文件

import tensorflow as tf
from tensorflow.python.framework import graph_utiloutput_graph_path = logdir+'expert-graph.pb'
with tf.Session() as sess:# with tf.gfile.FastGFile(output_graph_path, 'rb') as f:#     graph_def = tf.GraphDef()#     graph_def.ParseFromString(f.read())#     sess.graph.as_default()#     tf.import_graph_def(graph_def, name='')tf.global_variables_initializer().run()output_graph_def = tf.GraphDef()with open(output_graph_path, "rb") as f:output_graph_def.ParseFromString(f.read())_ = tf.import_graph_def(output_graph_def, name="")input_x = sess.graph.get_tensor_by_name("conv/w:0")print(input_x.eval())output = sess.graph.get_tensor_by_name("conv/b:0")print(output.eval())

或

# -*- coding: UTF-8 -*-
import tensorflow as tf
from tensorflow.python.framework import graph_util
import numpy as npoutput_graph_path = r"./expert-graph.pb"
with tf.Session() as sess:# with tf.gfile.FastGFile(output_graph_path, 'rb') as f:#     graph_def = tf.GraphDef()#     graph_def.ParseFromString(f.read())#     sess.graph.as_default()#     tf.import_graph_def(graph_def, name='')tf.global_variables_initializer().run()output_graph_def = tf.GraphDef()with open(output_graph_path, "rb") as f:output_graph_def.ParseFromString(f.read())_ = tf.import_graph_def(output_graph_def, name="")input_x = sess.graph.get_tensor_by_name("input:0")print(input_x)y=input_xout=sess.run(y,{input_x:np.random.random([1,28*28])})print(out[:10])

3、mnist例子

保存成.ckpt，然后反复训练，最后将最好的结果保存成.pb文件，用于部署生成测试！

train

from PIL import Image
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import tensorflow as tf
import os
import glob
from skimage import io, transform
from tensorflow.python.framework import graph_util
import collections
from tensorflow.examples.tutorials.mnist import input_datamnist = input_data.read_data_sets("MNIST_data/", one_hot=False)
w = 28
h = 28
c = 1def build_network(height, width, channel):x = tf.placeholder(tf.float32, shape=[None, height,width, channel], name='input')y_ = tf.placeholder(tf.int64, shape=[None,], name='y_')def weight_variable(shape, name="weights"):initial = tf.truncated_normal(shape, dtype=tf.float32, stddev=0.1)*0.001return tf.Variable(initial, name=name)def bias_variable(shape, name="biases"):initial = tf.constant(0.1, dtype=tf.float32, shape=shape)return tf.Variable(initial, name=name)def conv2d(input, w):return tf.nn.conv2d(input, w, [1, 1, 1, 1], padding='SAME')def pool_max(input):return tf.nn.max_pool(input,ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1],padding='SAME',name='pool1')def fc(input, w, b):return tf.matmul(input, w) + b# conv1with tf.name_scope('conv1_1') as scope:kernel = weight_variable([3, 3, 1, 32])biases = bias_variable([32])output_conv1_1 = tf.nn.relu(conv2d(x, kernel) + biases, name=scope)pool1 = pool_max(output_conv1_1) # [n,14,14,32]# conv2with tf.name_scope('conv1_2') as scope:kernel = weight_variable([3, 3, 32, 64])biases = bias_variable([64])output_conv1_2 = tf.nn.relu(conv2d(pool1, kernel) + biases, name=scope)pool2 = pool_max(output_conv1_2) # [n,7,7,64]#fc1with tf.name_scope('fc1') as scope:shape = int(np.prod(pool2.get_shape()[1:]))kernel = weight_variable([shape, 512])biases = bias_variable([512])pool2_flat = tf.reshape(pool2, [-1, shape])output_fc1 = tf.nn.relu(fc(pool2_flat, kernel, biases), name=scope)#fc2with tf.name_scope('fc2') as scope:kernel = weight_variable([512, 10])biases = bias_variable([10])output_fc2 = fc(output_fc1, kernel, biases)y = tf.nn.softmax(output_fc2, name="softmax")cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=y_))optimize = tf.train.AdamOptimizer(learning_rate=1e-1).minimize(cost)prediction_labels = tf.argmax(y, axis=1, name="output")correct_prediction = tf.equal(prediction_labels, y_)accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))correct_times_in_batch = tf.reduce_sum(tf.cast(correct_prediction, tf.int32))return dict(x=x,y_=y_,optimize=optimize,correct_prediction=correct_prediction,correct_times_in_batch=correct_times_in_batch,accuracy=accuracy,cost=cost,)def train_network(graph, batch_size, num_epochs, pb_file_path):init = tf.global_variables_initializer()saver=tf.train.Saver()with tf.Session() as sess:sess.run(init)# 验证之前是否已经保存了检查点文件ckpt = tf.train.get_checkpoint_state('./output/')if ckpt and ckpt.model_checkpoint_path:saver.restore(sess, ckpt.model_checkpoint_path)for epoch_index in range(num_epochs):for i in range(1000):batch_xs, batch_ys = mnist.train.next_batch(batch_size)feed={graph['x']: np.reshape(batch_xs, (-1, h, w, c)),graph['y_']: batch_ys}sess.run([graph['optimize']], feed_dict=feed)if i%100==0:print('step',i,'acc',sess.run(graph['accuracy'],feed),'loss',sess.run(graph['cost'],feed))saver.save(sess,'./output/model.ckpt',global_step=i)constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def,['output'])with tf.gfile.FastGFile(pb_file_path, mode='wb') as f:f.write(constant_graph.SerializeToString())def main():batch_size = 128num_epochs = 2pb_file_path = "mnist.pb"g = build_network(height=h, width=w, channel=c)train_network(g, batch_size, num_epochs, pb_file_path)main()

4、对输出层进行修改

保存pb文件

import numpy as np
import tensorflow as tf
from tensorflow.python.framework import graph_util
import collections
from tensorflow.examples.tutorials.mnist import input_datamnist = input_data.read_data_sets("MNIST_data/", one_hot=False)
w = 28
h = 28
c = 1def build_network(height, width, channel):x = tf.placeholder(tf.float32, shape=[None, height,width, channel], name='input')y_ = tf.placeholder(tf.int64, shape=[None,], name='y_')def weight_variable(shape, name="weights"):initial = tf.truncated_normal(shape, dtype=tf.float32, stddev=0.1)*0.001return tf.Variable(initial, name=name)def bias_variable(shape, name="biases"):initial = tf.constant(0.1, dtype=tf.float32, shape=shape)return tf.Variable(initial, name=name)def conv2d(input, w):return tf.nn.conv2d(input, w, [1, 1, 1, 1], padding='SAME')def pool_max(input):return tf.nn.max_pool(input,ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1],padding='SAME',name='pool1')def fc(input, w, b):return tf.matmul(input, w) + b# conv1with tf.name_scope('conv1_1') as scope:kernel = weight_variable([3, 3, 1, 32])biases = bias_variable([32])output_conv1_1 = tf.nn.relu(conv2d(x, kernel) + biases, name=scope)pool1 = pool_max(output_conv1_1) # [n,14,14,32]# conv2with tf.name_scope('conv1_2') as scope:kernel = weight_variable([3, 3, 32, 64])biases = bias_variable([64])output_conv1_2 = tf.nn.relu(conv2d(pool1, kernel) + biases, name=scope)pool2 = pool_max(output_conv1_2) # [n,7,7,64]#fc1with tf.name_scope('fc1') as scope:shape = int(np.prod(pool2.get_shape()[1:]))kernel = weight_variable([shape, 512])biases = bias_variable([512])pool2_flat = tf.reshape(pool2, [-1, shape])output_fc1 = tf.nn.relu(fc(pool2_flat, kernel, biases), name=scope)#fc2with tf.name_scope('fc2') as scope:kernel = weight_variable([512, 10])biases = bias_variable([10])output_fc2 = fc(output_fc1, kernel, biases)y = tf.nn.softmax(output_fc2, name="softmax")cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=y_))optimize = tf.train.AdamOptimizer(learning_rate=1e-1).minimize(cost)prediction_labels = tf.argmax(y, axis=1, name="output")correct_prediction = tf.equal(prediction_labels, y_)accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))correct_times_in_batch = tf.reduce_sum(tf.cast(correct_prediction, tf.int32))return dict(x=x,y_=y_,optimize=optimize,correct_prediction=correct_prediction,correct_times_in_batch=correct_times_in_batch,accuracy=accuracy,cost=cost,)def train_network(graph, batch_size, num_epochs, pb_file_path):init = tf.global_variables_initializer()saver=tf.train.Saver()with tf.Session() as sess:sess.run(init)# 验证之前是否已经保存了检查点文件ckpt = tf.train.get_checkpoint_state('./output/')if ckpt and ckpt.model_checkpoint_path:saver.restore(sess, ckpt.model_checkpoint_path)for epoch_index in range(num_epochs):for i in range(1000):batch_xs, batch_ys = mnist.train.next_batch(batch_size)feed={graph['x']: np.reshape(batch_xs, (-1, h, w, c)),graph['y_']: batch_ys}sess.run([graph['optimize']], feed_dict=feed)if i%100==0:print('step',i,'acc',sess.run(graph['accuracy'],feed),'loss',sess.run(graph['cost'],feed))saver.save(sess,'./output/model.ckpt',global_step=i)constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def,['input','y_','fc1'])with tf.gfile.FastGFile(pb_file_path, mode='wb') as f:f.write(constant_graph.SerializeToString())def main():batch_size = 128num_epochs = 2pb_file_path = "mnist.pb"g = build_network(height=h, width=w, channel=c)train_network(g, batch_size, num_epochs, pb_file_path)main()

导入pb文件并修改输出层

import tensorflow as tf
import  numpy as np
from tensorflow.examples.tutorials.mnist import input_datamnist = input_data.read_data_sets("MNIST_data/", one_hot=False)def recognize(mnist, pb_file_path):with tf.Graph().as_default():output_graph_def = tf.GraphDef()with open(pb_file_path, "rb") as f:output_graph_def.ParseFromString(f.read())_ = tf.import_graph_def(output_graph_def, name="")def weight_variable(shape, name="weights"):initial = tf.truncated_normal(shape, dtype=tf.float32, stddev=0.1) * 0.001return tf.Variable(initial, name=name)def bias_variable(shape, name="biases"):initial = tf.constant(0.1, dtype=tf.float32, shape=shape)return tf.Variable(initial, name=name)def fc(input, w, b):return tf.matmul(input, w) + bwith tf.Session() as sess:# init = tf.global_variables_initializer()# sess.run(init)input_x = sess.graph.get_tensor_by_name("input:0")input_y=sess.graph.get_tensor_by_name('y_:0')output_fc1=sess.graph.get_tensor_by_name('fc1:0')with tf.name_scope('fc2') as scope:kernel =weight_variable([512,1024])biases = bias_variable([1024])output_fc2 = tf.nn.relu(fc(output_fc1, kernel, biases),name=scope)with tf.name_scope('fc3') as scope:kernel = weight_variable([1024,10])biases = bias_variable([10])output_fc3 = fc(output_fc2, kernel, biases)y = tf.nn.softmax(output_fc3, name="softmax")cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=input_y))optimize = tf.train.AdamOptimizer(learning_rate=1e-1).minimize(cost)prediction_labels = tf.argmax(y, axis=1, name="output")correct_prediction = tf.equal(prediction_labels, input_y)accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))tf.global_variables_initializer().run()saver = tf.train.Saver()# 验证之前是否已经保存了检查点文件ckpt = tf.train.get_checkpoint_state('./output2/')if ckpt and ckpt.model_checkpoint_path:saver.restore(sess, ckpt.model_checkpoint_path)for i in range(100):batch_xs, batch_ys = mnist.train.next_batch(32)feed = {input_x: np.reshape(batch_xs, (-1, 28, 28, 1)),input_y: batch_ys}sess.run(optimize, feed_dict=feed)if i % 100 == 0:print('step', i, 'acc', sess.run(accuracy, feed), 'loss',sess.run(cost, feed))saver.save(sess, './output2/model.ckpt', global_step=i)recognize(mnist, "./mnist.pb")

5、mnist例子-2

部署生成测试

①保存pb文件

#!/usr/bin/python
# -*- coding: UTF-8 -*-# from __future__ import print_function
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from tensorflow.python.framework import graph_util
# number 1 to 10 data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)def compute_accuracy(v_xs, v_ys):global predictiony_pre = sess.run(prediction, feed_dict={xs: v_xs, keep_prob: 1})correct_prediction = tf.equal(tf.argmax(y_pre,1), tf.argmax(v_ys,1))accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))result = sess.run(accuracy, feed_dict={xs: v_xs, ys: v_ys, keep_prob: 1})return resultdef weight_variable(shape):inital=tf.truncated_normal(shape,stddev=0.1) #产生随机变量return tf.Variable(inital)def bias_variable(shape):inital=tf.constant(0.1,tf.float32,shape)return tf.Variable(inital)def conv2d(x, W):#stride[1,x_movement,y_movement,1]#Must have stride[0]=stride[3]=1return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME')def max_pool_2x2(x):#Must have stride[0]=stride[3]=1return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')# define placeholder for inputs to network
xs = tf.placeholder(tf.float32, [None, 784],name='input') # 28x28
ys = tf.placeholder(tf.float32, [None, 10])
keep_prob = tf.placeholder(tf.float32,name='keep_prob')x_image=tf.reshape(xs,[-1,28,28,1]) # 28x28
# print(x_image.shape)#[n_samples,28,28,1] 1表示黑白图 如果是RGB是3## conv1 layer ##
# patch 5x5 in size 1,out size 32 取样时按5x5个像素 1为图层厚度，因为这里是黑白图所有
# 为1，32为取样压缩后图的厚度
W_conv1=weight_variable([5,5,1,32])
b_conv1=bias_variable([32])
h_conv1=tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1) #output size 28x28x32取样后图片的长宽不变 图层厚度变化
h_pool1=max_pool_2x2(h_conv1) #output size 14x14x32  x,y方向都是跨2歩取 都缩小一半## conv2 layer ##
W_conv2=weight_variable([5,5,32,64]) #图层厚度变成64
b_conv2=bias_variable([64])
h_conv2=tf.nn.relu(conv2d(h_pool1,W_conv2)+b_conv2) #output size 14x14x64
h_pool2=max_pool_2x2(h_conv2)  #output size 7x7x64## func1 layer ##W_fc1=weight_variable([7*7*64,1024])
b_fc1=bias_variable([1024])
#[n_samples,7,7,64]-->[n_sample,7*7*64] 由三维 变成1维
h_pool2_flat=tf.reshape(h_pool2,[-1,7*7*64])
h_fc1=tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1)+b_fc1)h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob)## func2 layer ##
W_fc2=weight_variable([1024,10])
b_fc2=bias_variable([10])
prediction=tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2)+b_fc2,name='softmax')# the error between prediction and real data
cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction),reduction_indices=[1]))       # loss
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)sess = tf.Session()
# important step
sess.run(tf.global_variables_initializer())for i in range(1000):batch_xs, batch_ys = mnist.train.next_batch(100)sess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys, keep_prob: 0.5})if i % 50 == 0:print(compute_accuracy(mnist.test.images, mnist.test.labels))constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def,['input','keep_prob','softmax'])
with tf.gfile.FastGFile('mnist.pb', mode='wb') as f:f.write(constant_graph.SerializeToString())sess.close()

②导入pb文件，部署生成测试

import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_datamnist = input_data.read_data_sets("MNIST_data/", one_hot=True)with tf.Graph().as_default():output_graph_def = tf.GraphDef()with open('mnist.pb', "rb") as f:output_graph_def.ParseFromString(f.read())_ = tf.import_graph_def(output_graph_def, name="")with tf.Session() as sess:init = tf.global_variables_initializer()sess.run(init)input_x = sess.graph.get_tensor_by_name("input:0")keep_prob=sess.graph.get_tensor_by_name("keep_prob:0")out_softmax = sess.graph.get_tensor_by_name("softmax:0")img_out_softmax = sess.run(out_softmax, feed_dict={input_x:mnist.test.images[:10],keep_prob:1.})prediction_labels = np.argmax(img_out_softmax, axis=1)print("label:", prediction_labels)print('true label:', np.argmax(mnist.test.labels[:10], axis=1))

改写输出层

①保存pb文件

#!/usr/bin/python
# -*- coding: UTF-8 -*-# from __future__ import print_function
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from tensorflow.python.framework import graph_util
# number 1 to 10 data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)def compute_accuracy(v_xs, v_ys):global predictiony_pre = sess.run(prediction, feed_dict={xs: v_xs, keep_prob: 1})correct_prediction = tf.equal(tf.argmax(y_pre,1), tf.argmax(v_ys,1))accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))result = sess.run(accuracy, feed_dict={xs: v_xs, ys: v_ys, keep_prob: 1})return resultdef weight_variable(shape):inital=tf.truncated_normal(shape,stddev=0.1) #产生随机变量return tf.Variable(inital)def bias_variable(shape):inital=tf.constant(0.1,tf.float32,shape)return tf.Variable(inital)def conv2d(x, W):#stride[1,x_movement,y_movement,1]#Must have stride[0]=stride[3]=1return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME')def max_pool_2x2(x):#Must have stride[0]=stride[3]=1return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')# define placeholder for inputs to network
xs = tf.placeholder(tf.float32, [None, 784],name='input') # 28x28
ys = tf.placeholder(tf.float32, [None, 10],name='ys')
keep_prob = tf.placeholder(tf.float32,name='keep_prob')x_image=tf.reshape(xs,[-1,28,28,1]) # 28x28
# print(x_image.shape)#[n_samples,28,28,1] 1表示黑白图 如果是RGB是3## conv1 layer ##
# patch 5x5 in size 1,out size 32 取样时按5x5个像素 1为图层厚度，因为这里是黑白图所有
# 为1，32为取样压缩后图的厚度
W_conv1=weight_variable([5,5,1,32])
b_conv1=bias_variable([32])
h_conv1=tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1) #output size 28x28x32取样后图片的长宽不变 图层厚度变化
h_pool1=max_pool_2x2(h_conv1) #output size 14x14x32  x,y方向都是跨2歩取 都缩小一半## conv2 layer ##
W_conv2=weight_variable([5,5,32,64]) #图层厚度变成64
b_conv2=bias_variable([64])
h_conv2=tf.nn.relu(conv2d(h_pool1,W_conv2)+b_conv2) #output size 14x14x64
h_pool2=max_pool_2x2(h_conv2)  #output size 7x7x64## func1 layer ##W_fc1=weight_variable([7*7*64,1024])
b_fc1=bias_variable([1024])
#[n_samples,7,7,64]-->[n_sample,7*7*64] 由三维 变成1维
h_pool2_flat=tf.reshape(h_pool2,[-1,7*7*64])
h_fc1=tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1)+b_fc1,name='fc1')h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob)## func2 layer ##
W_fc2=weight_variable([1024,10])
b_fc2=bias_variable([10])
prediction=tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2)+b_fc2,name='softmax')# the error between prediction and real data
cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction),reduction_indices=[1]))       # loss
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)sess = tf.Session()
# important step
sess.run(tf.global_variables_initializer())for i in range(1000):batch_xs, batch_ys = mnist.train.next_batch(100)sess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys, keep_prob: 0.5})if i % 50 == 0:print(compute_accuracy(mnist.test.images, mnist.test.labels))constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def,['input','ys','keep_prob','fc1'])
with tf.gfile.FastGFile('mnist.pb', mode='wb') as f:f.write(constant_graph.SerializeToString())sess.close()

②导入pb文件并改写输出层

附加：通过改写输出层可以实现模型迁移

import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_datamnist = input_data.read_data_sets("MNIST_data/", one_hot=True)with tf.Graph().as_default():output_graph_def = tf.GraphDef()with open('mnist.pb', "rb") as f:output_graph_def.ParseFromString(f.read())_ = tf.import_graph_def(output_graph_def, name="")def weight_variable(shape):inital = tf.truncated_normal(shape, stddev=0.1)  # 产生随机变量return tf.Variable(inital)def bias_variable(shape):inital = tf.constant(0.1, tf.float32, shape)return tf.Variable(inital)with tf.Session() as sess:# init = tf.global_variables_initializer()# sess.run(init)input_x = sess.graph.get_tensor_by_name("input:0")ys=sess.graph.get_tensor_by_name("ys:0")keep_prob=sess.graph.get_tensor_by_name("keep_prob:0")fc1=sess.graph.get_tensor_by_name("fc1:0")h_fc1_drop = tf.nn.dropout(fc1, keep_prob)## func2 layer ##W_fc2 = weight_variable([1024, 1024])b_fc2 = bias_variable([1024])h_fc2 = tf.nn.relu(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)h_fc2_drop = tf.nn.dropout(h_fc2, keep_prob)## func3 layer ##W_fc3 = weight_variable([1024, 10])b_fc3 = bias_variable([10])prediction = tf.nn.softmax(tf.matmul(h_fc2_drop, W_fc3) + b_fc3)# the error between prediction and real datacross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction),reduction_indices=[1]))  # losstrain_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)correct_prediction = tf.equal(tf.argmax(prediction, 1), tf.argmax(ys, 1))accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))init = tf.global_variables_initializer()sess.run(init)for step in range(200):batch_xs, batch_ys = mnist.train.next_batch(100)sess.run(train_step, feed_dict={input_x: batch_xs, ys: batch_ys, keep_prob: 0.5})if step % 20== 0:print('step',step,'acc',accuracy.eval({input_x: mnist.test.images, ys: mnist.test.labels, keep_prob: 1.}))

5、查看op.name

import tensorflow as tf
import osdef create_graph():"""Creates a graph from saved GraphDef file and returns a saver."""# Creates graph from saved graph_def.pb.with tf.gfile.FastGFile(os.path.join('mnist.pb'), 'rb') as f:graph_def = tf.GraphDef()graph_def.ParseFromString(f.read())_ = tf.import_graph_def(graph_def, name='')# print all op names 、type 、shape
def print_ops():create_graph()with tf.Session() as sess:ops = sess.graph.get_operations()for op in ops:print(op.name,op.outputs)print_ops()

Tensorflow-pb保存与导入相关推荐

TensorFlow模型保存pb或ckpt
Tensorflow的保存分为三种:1. checkpoint模式:2. pb模式:3. saved_model模式. https://www.zhihu.com/collection/6445044 ...
[tensorflow] 模型保存、加载与转换详解
TensorFlow模型加载与转换详解本次讲解主要涉及到TensorFlow框架训练时候模型文件的管理以及转换. 首先我们需要明确TensorFlow模型文件的存储格式以及文件个数: model_f ...
TensorFlow模型保存和加载方法
TensorFlow模型保存和加载方法模型保存 import tensorflow as tfw1 = tf.Variable(tf.constant(2.0, shape=[1]), name=& ...
TensorFlow模型保存和提取方法
2019独角兽企业重金招聘Python工程师标准>>> 一.TensorFlow模型保存和提取方法 1. TensorFlow通过tf.train.Saver类实现神经网络模型的保存 ...
TensorFlow模型保存和提取方法（含滑动平均模型）
一.TensorFlow模型保存和提取方法 1. TensorFlow通过tf.train.Saver类实现神经网络模型的保存和提取.tf.train.Saver对象saver的save方法将Tens ...
深度学习之格式转换笔记(三)：keras(.hdf5)模型转TensorFlow(.pb) 转TensorRT(.uff)格式
环境: tensorflow1.15,cuda10.0,cudnn7.6.4 将keras训练好保存的.hdf5格式模型转为tensorflow的.pb模型,然后转为tensorrt支持的uff格式. ...
Tensorflow 模型保存与恢复（2）使用SavedModel
使用SavedModel 保存和恢复模型本篇介绍使用SavedModel进行模型的保存与恢复. 其他相关: Tensorflow 模型保存与恢复(1)使用tf.train.Saver() Tenso ...
查看tensorflow pb模型文件的节点信息
查看tensorflow pb模型文件的节点信息: import tensorflow as tf with tf.Session() as sess:with open('./quantized_m ...
TensorFlow 模型保存/载入的两种方法
TensorFlow 模型保存/载入我们在上线使用一个算法模型的时候,首先必须将已经训练好的模型保存下来.tensorflow保存模型的方式与sklearn不太一样,sklearn很直接,一个skl ...
pb 保存变量文件名_【Tensorflow 2.0 正式版教程】模型的保存、加载与迁移
模型的保存和加载可以直接通过Model类的save_weights和load_weights实现.默认的保存格式为tensorflow的checkpoint格式,也可以手动设置保存为h5文件. mo ...

Tensorflow-pb保存与导入