Tensorflow version 0.11:
Save the model:
import tensorflow as tf
#Prepare to feed input, i.e. feed_dict and placeholders
w1 = tf.placeholder("float", name="w1")
w2 = tf.placeholder("float", name="w2")
b1= tf.Variable(2.0,name="bias")
feed_dict ={w1:4,w2:8}
#Define a test operation that we will restore
w3 = tf.add(w1,w2)
w4 = tf.multiply(w3,b1,name="op_to_restore")
sess = tf.Session()
sess.run(tf.global_variables_initializer())
#Create a saver object which will save all the variables
saver = tf.train.Saver()
#Run the operation by feeding input
print sess.run(w4,feed_dict)
#Prints 24 which is sum of (w1+w2)*b1
#Now, save the graph
saver.save(sess, 'my_test_model',global_step=1000)
Restore the model:
import tensorflow as tf
sess=tf.Session()
#First let's load meta graph and restore weights
saver = tf.train.import_meta_graph('my_test_model-1000.meta')
saver.restore(sess,tf.train.latest_checkpoint('./'))
# Access saved Variables directly
print(sess.run('bias:0'))
# This will print 2, which is the value of bias that we saved
# Now, let's access and create placeholders variables and
# create feed-dict to feed new data
graph = tf.get_default_graph()
w1 = graph.get_tensor_by_name("w1:0")
w2 = graph.get_tensor_by_name("w2:0")
feed_dict ={w1:13.0,w2:17.0}
#Now, access the op that you want to run.
op_to_restore = graph.get_tensor_by_name("op_to_restore:0")
print sess.run(op_to_restore,feed_dict)
#This will print 60 which is calculated
Reference:
- https://stackoverflow.com/questions/33759623/tensorflow-how-to-save-restore-a-model
- https://cv-tricks.com/tensorflow-tutorial/save-restore-tensorflow-models-quick-complete-tutorial/
- https://blog.csdn.net/liuxiao214/article/details/79048136
总结
-
保存模型
- 记得给张量命名，如：tf.placeholder(“float”, name=“w1”) 中的 name
- saver = tf.train.Saver()
- saver.save(sess, ‘my_test_model’,global_step=1000)
- my_test_model是文件名的前缀 ，global_step是保存第几轮的模型
-
读取模型
- saver = tf.train.import_meta_graph(‘my_test_model-1000.meta’)
- saver.restore(sess,tf.train.latest_checkpoint(’./’))
- graph = tf.get_default_graph()
- graph.get_tensor_by_name(“w1:0”) # w1：张量名 => name
- 建议引用name的时候最好把张量输出出来查看它的name，免得会报错
CNN实例
tensorflow代码
def run(self):
print('Params{\n learning_rate => %.5f\n iteration => %d\n divide_rate => %.2f\n the number of units => %d\n early_stopping => %d\n}\n' %
(self.learn_rate, self.round, self.divide_rate, self.units, self.early_stopping))
# image, label = self.load_image()
train_x, train_y, test_x, test_y = self.divide(rate=self.divide_rate, shuffle=self.shuffle)
print('train_image_size:', train_x.shape)
print('train_label_size:', train_y.shape)
print('test_image_size: ', test_x.shape)
print('test_label_size: ', test_y.shape)
self.text = self.text + '\n' + "train_image_size: {}".format(train_x.shape)
self.text = self.text + '\n' + "train_label_size: {}".format(train_y.shape)
self.text = self.text + '\n' + "test_image_size: {}".format(test_x.shape)
self.text = self.text + '\n' + "test_label_size: {}".format(test_y.shape)
input_x = tf.placeholder(tf.float32, [None, 112 * 92], name='input_x') # None表示张量的第一个维度可以是任意长度
output_y = tf.placeholder(tf.int32, [None, 40]) # 输出：40个数字的标签
input_x_images = tf.reshape(input_x, [-1, 112, 92, 1]) # 改变形状之后的输入
"""Construct convolution neural network"""
"""First convolution layer"""
convolution_one = tf.layers.conv2d(
inputs=input_x_images, # 形状是[112, 92, 1]
filters=32, # 32个过滤器，输出的深度是32
kernel_size=[5, 5], # 过滤器在二维的大小是（5，5）
strides=1, # 步长是1
padding='same', # same表示输出的大小不变，因此需要在外围补0两圈
activation=tf.nn.relu # 激活函数是Relu
) # 形状[112, 92, 32]
"""First pooling layer"""
pool_one = tf.layers.max_pooling2d(
inputs=convolution_one, # 形状[112, 92, 32]
pool_size=[2, 2], # 过滤器在二维的大小是（2 * 2）
strides=2 # 步长是2
) # 形状[56, 46, 32]
"""Second convolution layer"""
convolution_two = tf.layers.conv2d(
inputs=pool_one, # 形状是[56, 46, 32]
filters=64, # 64个过滤器，输出的深度是64
kernel_size=[5, 5], # 过滤器在二维的大小是（5，5）
strides=1, # 步长是1
padding='same', # same表示输出的大小不变，因此需要在外围补0两圈
activation=tf.nn.relu # 激活函数是Relu
) # 形状[56, 46, 64]
"""Second pooling layer"""
pool_two = tf.layers.max_pooling2d(
inputs=convolution_two, # 形状[56, 46, 64]
pool_size=[2, 2], # 过滤器在二维的大小是（2 * 2）
strides=2 # 步长是2
) # 形状[28, 23, 64]
"""flatting"""
flat = tf.reshape(pool_two, [-1, 28 * 23 * 64]) # 形状[7 * 7 * 64, ]
"""The number of neural cell in fully-connected layer"""
dense = tf.layers.dense(inputs=flat, units=self.units, activation=tf.nn.relu)
# Dropout:50%，rate=0.5
dropout = tf.layers.dropout(inputs=dense, rate=0.5, name='dropout')
"""Output layer, there needn't activation function"""
logit = tf.layers.dense(inputs=dropout, units=40, name='logit') # 输出。形状[1, 1, 10]
"""Calculation error - Cross entropy then compute percent probability with Softmax"""
loss = tf.losses.softmax_cross_entropy(onehot_labels=output_y, logits=logit)
# use Adam Optimizer to minimize error with learning_rate
train_op = tf.train.AdamOptimizer(learning_rate=self.learn_rate).minimize(loss) # learn_rate
"""Compute accuracy of test data set"""
# accuracy = tf.metrics.accuracy(labels=tf.argmax(output_y), predictions=tf.argmax(logit))[1]
correct_prediction = tf.equal(tf.argmax(output_y), tf.argmax(logit)) # return a string of bool value
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
epoch = 0 # for counting
with tf.Session() as sess:
"""Initial global and local variables"""
init = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
sess.run(init)
print('\n-------------start to train model------------')
self.text = self.text + '\n' + "\n------------start to train model------------"
count, formal_score = 0, 0
for i in range(self.round):
"""start graph computation"""
epoch = i
train_loss, train_op_ = sess.run([loss, train_op], feed_dict={input_x: train_x, output_y: train_y})
if i % 1 == 0:
test_accuracy = sess.run(accuracy, feed_dict={input_x: test_x, output_y: test_y})
print("Step=%d, Train loss=%.3f, [Test accuracy=%.3f]" % (i+1, train_loss, test_accuracy))
self.text = self.text + '\n' + "Step={}, Loss={:.4f}, [Accuracy={:.3f}]".format(i+1, train_loss, test_accuracy)
"""Test a sample output
test = self.load_predict_sample('image/10.png')
output = sess.run(logit, feed_dict={input_x: test})
print('prediction: ', output)
"""
print(logit)
if test_accuracy <= formal_score:
count += 1
if count >= self.early_stopping:
print('early stopping.')
self.text = self.text + '\n' + "early stopping."
# if self.__store_model is True:
# self.store_model(sess, epoch)
break
else:
continue
else:
formal_score = test_accuracy
count = 0
if self.__store_model is True:
self.store_model(sess, epoch)
self.text = self.text + '\n' + "over."
return self.text
保存模型
def store_model(self, sess, epoch):
absolute_path = 'face_recognition/'
saver = tf.train.Saver() # Save model, maximun to save: 5
my_file = absolute_path + 'model'
if os.path.exists(my_file) is False:
os.mkdir(my_file)
print('make directory successful.')
self.text = self.text + '\n' + "make directory successful."
else:
file_name = os.listdir(my_file)
if len(file_name) != 0:
for string in file_name:
os.remove(my_file + '/' + string) # delete files
saver.save(sess, my_file + "/cnn_multi_classifier", global_step=epoch) # global_step
print('model store successfully.')
self.text = self.text + '\n' + "model store successfully."
保存结果
读取模型
def load_model(self, test):
absolute_path = 'face_recognition/'
if os.path.exists(absolute_path + 'model') is True:
dir_list = os.listdir(absolute_path + 'model')
if len(dir_list) == 0:
print('model is empty.')
exit(0)
else:
with tf.Session() as sess:
path = absolute_path + 'model/cnn_multi_classifier-19.meta'
for file in dir_list:
if 'meta' in file:
path = absolute_path + 'model/' + file
saver = tf.train.import_meta_graph(path)
saver.restore(sess, tf.train.latest_checkpoint(absolute_path + "model/"))
print('load model successfully.')
graph = tf.get_default_graph()
input_x = graph.get_tensor_by_name('input_x:0')
logit = graph.get_tensor_by_name('logit/BiasAdd:0')
feed_dict = {input_x: test}
result = sess.run(logit, feed_dict=feed_dict)
max_index = sess.run(tf.argmax(result, axis=1)) # find the indices of maximum value
print('Prediction result:', result)
print('This image is %d th people.' % (max_index[0] + 1))
return max_index[0] + 1
else:
print('directory is not exist.')
exit(0)
特别提醒
建议引用name的时候最好把张量输出出来查看它的name，免得会报错，就像我上面的案例一样。因为我的logit张量的name是‘logit’，但是它在图里面name确是’logit/BiasAdd:0’。
扫一扫
8万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
