tensorflow将训练好的模型freeze,即将权重固化到图里面,并使用该模型进行预测

转载来自：http://blog.csdn.net/lujiandong1/article/details/53385092

ML主要分为训练和预测两个阶段,此教程就是将训练好的模型freeze并保存下来.freeze的含义就是将该模型的图结构和该模型的权重固化到一起了.也即加载freeze的模型之后,立刻能够使用了。

下面使用一个简单的demo来详细解释该过程,


一、首先运行脚本tiny_model.py


[python] view plain copy 
#-*- coding:utf-8 -*-  
import tensorflow as tf  
import numpy as np  
  
  
with tf.variable_scope('Placeholder'):  
    inputs_placeholder = tf.placeholder(tf.float32, name='inputs_placeholder', shape=[None, 10])  
    labels_placeholder = tf.placeholder(tf.float32, name='labels_placeholder', shape=[None, 1])  
  
with tf.variable_scope('NN'):  
    W1 = tf.get_variable('W1', shape=[10, 1], initializer=tf.random_normal_initializer(stddev=1e-1))  
    b1 = tf.get_variable('b1', shape=[1], initializer=tf.constant_initializer(0.1))  
    W2 = tf.get_variable('W2', shape=[10, 1], initializer=tf.random_normal_initializer(stddev=1e-1))  
    b2 = tf.get_variable('b2', shape=[1], initializer=tf.constant_initializer(0.1))  
  
    a = tf.nn.relu(tf.matmul(inputs_placeholder, W1) + b1)  
    a2 = tf.nn.relu(tf.matmul(inputs_placeholder, W2) + b2)  
  
    y = tf.div(tf.add(a, a2), 2)  
  
with tf.variable_scope('Loss'):  
    loss = tf.reduce_sum(tf.square(y - labels_placeholder) / 2)  
  
with tf.variable_scope('Accuracy'):  
    predictions = tf.greater(y, 0.5, name="predictions")  
    correct_predictions = tf.equal(predictions, tf.cast(labels_placeholder, tf.bool), name="correct_predictions")  
    accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float32))  
  
  
adam = tf.train.AdamOptimizer(learning_rate=1e-3)  
train_op = adam.minimize(loss)  
  
# generate_data  
inputs = np.random.choice(10, size=[10000, 10])  
labels = (np.sum(inputs, axis=1) > 45).reshape(-1, 1).astype(np.float32)  
print('inputs.shape:', inputs.shape)  
print('labels.shape:', labels.shape)  
  
  
test_inputs = np.random.choice(10, size=[100, 10])  
test_labels = (np.sum(test_inputs, axis=1) > 45).reshape(-1, 1).astype(np.float32)  
print('test_inputs.shape:', test_inputs.shape)  
print('test_labels.shape:', test_labels.shape)  
  
batch_size = 32  
epochs = 10  
  
batches = []  
print("%d items in batch of %d gives us %d full batches and %d batches of %d items" % (  
    len(inputs),  
    batch_size,  
    len(inputs) // batch_size,  
    batch_size - len(inputs) // batch_size,  
    len(inputs) - (len(inputs) // batch_size) * 32)  
)  
for i in range(len(inputs) // batch_size):  
    batch = [ inputs[batch_size*i:batch_size*i+batch_size], labels[batch_size*i:batch_size*i+batch_size] ]  
    batches.append(list(batch))  
if (i + 1) * batch_size < len(inputs):  
    batch = [ inputs[batch_size*(i + 1):],labels[batch_size*(i + 1):] ]  
    batches.append(list(batch))  
print("Number of batches: %d" % len(batches))  
print("Size of full batch: %d" % len(batches[0]))  
print("Size if final batch: %d" % len(batches[-1]))  
  
global_count = 0  
  
with tf.Session() as sess:  
#sv = tf.train.Supervisor()  
#with sv.managed_session() as sess:  
    sess.run(tf.initialize_all_variables())  
    for i in range(epochs):  
        for batch in batches:  
            # print(batch[0].shape, batch[1].shape)  
            train_loss , _= sess.run([loss, train_op], feed_dict={  
                inputs_placeholder: batch[0],  
                labels_placeholder: batch[1]  
            })  
            # print('train_loss: %d' % train_loss)  
  
            if global_count % 100 == 0:  
                acc = sess.run(accuracy, feed_dict={  
                    inputs_placeholder: test_inputs,  
                    labels_placeholder: test_labels  
                })  
                print('accuracy: %f' % acc)  
            global_count += 1  
  
    acc = sess.run(accuracy, feed_dict={  
        inputs_placeholder: test_inputs,  
        labels_placeholder: test_labels  
    })  
    print("final accuracy: %f" % acc)  
    #在session当中就要将模型进行保存  
    saver = tf.train.Saver()  
    last_chkp = saver.save(sess, 'results/graph.chkp')  
    #sv.saver.save(sess, 'results/graph.chkp')  
  
for op in tf.get_default_graph().get_operations():  
    print(op.name)  
说明：saver.save必须在session里面,因为在session里面,整个图才是激活的,才能够将参数存进来,使用save之后能够得到如下的文件：


说明：
.data:存放的是权重参数
.meta:存放的是图和metadata,metadata是其他配置的数据
如果想将我们的模型固化，让别人能够使用，我们仅仅需要的是图和参数，metadata是不需要的
二、综合上述几个文件,生成可以使用的模型的步骤如下：


1、恢复我们保存的图
2、开启一个Session，然后载入该图要求的权重
3、删除对预测无关的metadata
4、将处理好的模型序列化之后保存
运行freeze.py
[python] view plain copy 
#-*- coding:utf-8 -*-  
import os, argparse  
import tensorflow as tf  
from tensorflow.python.framework import graph_util  
  
dir = os.path.dirname(os.path.realpath(__file__))  
  
def freeze_graph(model_folder):  
    # We retrieve our checkpoint fullpath  
    checkpoint = tf.train.get_checkpoint_state(model_folder)  
    input_checkpoint = checkpoint.model_checkpoint_path  
      
    # We precise the file fullname of our freezed graph  
    absolute_model_folder = "/".join(input_checkpoint.split('/')[:-1])  
    output_graph = absolute_model_folder + "/frozen_model.pb"  
  
    # Before exporting our graph, we need to precise what is our output node  
    # this variables is plural, because you can have multiple output nodes  
    #freeze之前必须明确哪个是输出结点,也就是我们要得到推论结果的结点  
    #输出结点可以看我们模型的定义  
    #只有定义了输出结点,freeze才会把得到输出结点所必要的结点都保存下来,或者哪些结点可以丢弃  
    #所以,output_node_names必须根据不同的网络进行修改  
    output_node_names = "Accuracy/predictions"  
  
    # We clear the devices, to allow TensorFlow to control on the loading where it wants operations to be calculated  
    clear_devices = True  
      
    # We import the meta graph and retrive a Saver  
    saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=clear_devices)  
  
    # We retrieve the protobuf graph definition  
    graph = tf.get_default_graph()  
    input_graph_def = graph.as_graph_def()  
  
    #We start a session and restore the graph weights  
    #这边已经将训练好的参数加载进来,也即最后保存的模型是有图,并且图里面已经有参数了,所以才叫做是frozen  
    #相当于将参数已经固化在了图当中   
    with tf.Session() as sess:  
        saver.restore(sess, input_checkpoint)  
  
        # We use a built-in TF helper to export variables to constant  
        output_graph_def = graph_util.convert_variables_to_constants(  
            sess,   
            input_graph_def,   
            output_node_names.split(",") # We split on comma for convenience  
        )   
  
        # Finally we serialize and dump the output graph to the filesystem  
        with tf.gfile.GFile(output_graph, "wb") as f:  
            f.write(output_graph_def.SerializeToString())  
        print("%d ops in the final graph." % len(output_graph_def.node))  
  
  
if __name__ == '__main__':  
    parser = argparse.ArgumentParser()  
    parser.add_argument("--model_folder", type=str, help="Model folder to export")  
    args = parser.parse_args()  
  
    freeze_graph(args.model_folder)  


说明：对于freeze操作,我们需要定义输出结点的名字.因为网络其实是比较复杂的,定义了输出结点的名字,那么freeze的时候就只把输出该结点所需要的子图都固化下来,其他无关的就舍弃掉.因为我们freeze模型的目的是接下来做预测.所以,一般情况下,output_node_names就是我们预测的目标.
三、加载freeze后的模型,注意该模型已经是包含图和相应的参数了.所以,我们不需要再加载参数进来.也即该模型加载进来已经是可以使用了.


[python] view plain copy 
#-*- coding:utf-8 -*-  
import argparse   
import tensorflow as tf  
  
def load_graph(frozen_graph_filename):  
    # We parse the graph_def file  
    with tf.gfile.GFile(frozen_graph_filename, "rb") as f:  
        graph_def = tf.GraphDef()  
        graph_def.ParseFromString(f.read())  
  
    # We load the graph_def in the default graph  
    with tf.Graph().as_default() as graph:  
        tf.import_graph_def(  
            graph_def,   
            input_map=None,   
            return_elements=None,   
            name="prefix",   
            op_dict=None,   
            producer_op_list=None  
        )  
    return graph  
  
if __name__ == '__main__':  
    parser = argparse.ArgumentParser()  
    parser.add_argument("--frozen_model_filename", default="results/frozen_model.pb", type=str, help="Frozen model file to import")  
    args = parser.parse_args()  
    #加载已经将参数固化后的图  
    graph = load_graph(args.frozen_model_filename)  
  
    # We can list operations  
    #op.values() gives you a list of tensors it produces  
    #op.name gives you the name  
    #输入,输出结点也是operation,所以,我们可以得到operation的名字  
    for op in graph.get_operations():  
        print(op.name,op.values())  
        # prefix/Placeholder/inputs_placeholder  
        # ...  
        # prefix/Accuracy/predictions  
    #操作有:prefix/Placeholder/inputs_placeholder  
    #操作有:prefix/Accuracy/predictions  
    #为了预测,我们需要找到我们需要feed的tensor,那么就需要该tensor的名字  
    #注意prefix/Placeholder/inputs_placeholder仅仅是操作的名字,prefix/Placeholder/inputs_placeholder:0才是tensor的名字  
    x = graph.get_tensor_by_name('prefix/Placeholder/inputs_placeholder:0')  
    y = graph.get_tensor_by_name('prefix/Accuracy/predictions:0')  
          
    with tf.Session(graph=graph) as sess:  
        y_out = sess.run(y, feed_dict={  
            x: [[3, 5, 7, 4, 5, 1, 1, 1, 1, 1]] # < 45  
        })  
        print(y_out) # [[ 0.]] Yay!  
    print ("finish")  
说明：
1、在预测的过程中,当把freeze后的模型加载进来后,我们只需要定义好输入的tensor和目标tensor即可


2、在这里要注意一下tensor_name和ops_name,


注意prefix/Placeholder/inputs_placeholder仅仅是操作的名字,prefix/Placeholder/inputs_placeholder:0才是tensor的名字


x = graph.get_tensor_by_name('prefix/Placeholder/inputs_placeholder:0')一定要使用tensor的名字


3、要获取图中ops的名字和对应的tensor的名字,可用如下的代码：


[python] view plain copy 
# We can list operations  
#op.values() gives you a list of tensors it produces  
#op.name gives you the name  
#输入,输出结点也是operation,所以,我们可以得到operation的名字  
for op in graph.get_operations():  
    print(op.name,op.values())  


=============================================================================================================================
上面是使用了Saver()来保存模型,也可以使用sv = tf.train.Supervisor()来保存模型


[python] view plain copy 
#-*- coding:utf-8 -*-  
import tensorflow as tf  
import numpy as np  
  
  
with tf.variable_scope('Placeholder'):  
    inputs_placeholder = tf.placeholder(tf.float32, name='inputs_placeholder', shape=[None, 10])  
    labels_placeholder = tf.placeholder(tf.float32, name='labels_placeholder', shape=[None, 1])  
  
with tf.variable_scope('NN'):  
    W1 = tf.get_variable('W1', shape=[10, 1], initializer=tf.random_normal_initializer(stddev=1e-1))  
    b1 = tf.get_variable('b1', shape=[1], initializer=tf.constant_initializer(0.1))  
    W2 = tf.get_variable('W2', shape=[10, 1], initializer=tf.random_normal_initializer(stddev=1e-1))  
    b2 = tf.get_variable('b2', shape=[1], initializer=tf.constant_initializer(0.1))  
  
    a = tf.nn.relu(tf.matmul(inputs_placeholder, W1) + b1)  
    a2 = tf.nn.relu(tf.matmul(inputs_placeholder, W2) + b2)  
  
    y = tf.div(tf.add(a, a2), 2)  
  
with tf.variable_scope('Loss'):  
    loss = tf.reduce_sum(tf.square(y - labels_placeholder) / 2)  
  
with tf.variable_scope('Accuracy'):  
    predictions = tf.greater(y, 0.5, name="predictions")  
    correct_predictions = tf.equal(predictions, tf.cast(labels_placeholder, tf.bool), name="correct_predictions")  
    accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float32))  
  
  
adam = tf.train.AdamOptimizer(learning_rate=1e-3)  
train_op = adam.minimize(loss)  
  
# generate_data  
inputs = np.random.choice(10, size=[10000, 10])  
labels = (np.sum(inputs, axis=1) > 45).reshape(-1, 1).astype(np.float32)  
print('inputs.shape:', inputs.shape)  
print('labels.shape:', labels.shape)  
  
  
test_inputs = np.random.choice(10, size=[100, 10])  
test_labels = (np.sum(test_inputs, axis=1) > 45).reshape(-1, 1).astype(np.float32)  
print('test_inputs.shape:', test_inputs.shape)  
print('test_labels.shape:', test_labels.shape)  
  
batch_size = 32  
epochs = 10  
  
batches = []  
print("%d items in batch of %d gives us %d full batches and %d batches of %d items" % (  
    len(inputs),  
    batch_size,  
    len(inputs) // batch_size,  
    batch_size - len(inputs) // batch_size,  
    len(inputs) - (len(inputs) // batch_size) * 32)  
)  
for i in range(len(inputs) // batch_size):  
    batch = [ inputs[batch_size*i:batch_size*i+batch_size], labels[batch_size*i:batch_size*i+batch_size] ]  
    batches.append(list(batch))  
if (i + 1) * batch_size < len(inputs):  
    batch = [ inputs[batch_size*(i + 1):],labels[batch_size*(i + 1):] ]  
    batches.append(list(batch))  
print("Number of batches: %d" % len(batches))  
print("Size of full batch: %d" % len(batches[0]))  
print("Size if final batch: %d" % len(batches[-1]))  
  
global_count = 0  
  
#with tf.Session() as sess:  
sv = tf.train.Supervisor()  
with sv.managed_session() as sess:  
    #sess.run(tf.initialize_all_variables())  
    for i in range(epochs):  
        for batch in batches:  
            # print(batch[0].shape, batch[1].shape)  
            train_loss , _= sess.run([loss, train_op], feed_dict={  
                inputs_placeholder: batch[0],  
                labels_placeholder: batch[1]  
            })  
            # print('train_loss: %d' % train_loss)  
  
            if global_count % 100 == 0:  
                acc = sess.run(accuracy, feed_dict={  
                    inputs_placeholder: test_inputs,  
                    labels_placeholder: test_labels  
                })  
                print('accuracy: %f' % acc)  
            global_count += 1  
  
    acc = sess.run(accuracy, feed_dict={  
        inputs_placeholder: test_inputs,  
        labels_placeholder: test_labels  
    })  
    print("final accuracy: %f" % acc)  
    #在session当中就要将模型进行保存  
    #saver = tf.train.Saver()  
    #last_chkp = saver.save(sess, 'results/graph.chkp')  
    sv.saver.save(sess, 'results/graph.chkp')  
  
for op in tf.get_default_graph().get_operations():  
    print(op.name)  


注意：使用了sv = tf.train.Supervisor(),就不需要再初始化了,将sess.run(tf.initialize_all_variables())注释掉,否则会报错.