【TensorFlow系列】【四】基於lenet的fine-tuning

基於TensorFlow作fine-tuning，與基於caffe作fine-tuning本質是同樣的，只是caffe僅須要修改配置文件，而TensorFlow修改也是配置文件，只不過是用python寫的配置文件。

 

fine-tuning的本質就是：移花接木。

本文講述如何基於lenet作fine-tuning。

原始的網絡，FC1與FC2，如今想在在FC1與FC2中間加入一層，即造成FC1--->FC2-->FC3

fine-tuning步驟
1.加載模型pb文件。
2.獲取輸入與label的tensor。
3.獲取須要修改層（fc1)的tensor。
4.定義修改的層及loss、optimization等。
5.灌入數據，開始fine-tuning。

【原始網絡以下】

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from tensorflow.python.framework import graph_util

mnist = input_data.read_data_sets(train_dir=r"E:\mnist_data",one_hot=True)


#定義輸入數據mnist圖片大小28*28*1=784,None表示batch_size
x = tf.placeholder(dtype=tf.float32,shape=[None,28*28],name="input")
#定義標籤數據,mnist共10類
y_ = tf.placeholder(dtype=tf.float32,shape=[None,10],name="y_")
#將數據調整爲二維數據，w*H*c---> 28*28*1,-1表示N張
image = tf.reshape(x,shape=[-1,28,28,1])

#第一層，卷積核={5*5*1*32}，池化核={2*2*1,1*2*2*1}
w1 = tf.Variable(initial_value=tf.random_normal(shape=[5,5,1,32],stddev=0.1,dtype=tf.float32,name="w1"))
b1= tf.Variable(initial_value=tf.zeros(shape=[32]))
conv1 = tf.nn.conv2d(input=image,filter=w1,strides=[1,1,1,1],padding="SAME",name="conv1")
relu1 = tf.nn.relu(tf.nn.bias_add(conv1,b1),name="relu1")
pool1 = tf.nn.max_pool(value=relu1,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME")
#shape={None，14,14,32}
#第二層，卷積核={5*5*32*64}，池化核={2*2*1,1*2*2*1}
w2 = tf.Variable(initial_value=tf.random_normal(shape=[5,5,32,64],stddev=0.1,dtype=tf.float32,name="w2"))
b2 = tf.Variable(initial_value=tf.zeros(shape=[64]))
conv2 = tf.nn.conv2d(input=pool1,filter=w2,strides=[1,1,1,1],padding="SAME")
relu2 = tf.nn.relu(tf.nn.bias_add(conv2,b2),name="relu2")
pool2 = tf.nn.max_pool(value=relu2,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME",name="pool2")
#shape={None，7,7,64}
#FC1
w3 = tf.Variable(initial_value=tf.random_normal(shape=[7*7*64,1024],stddev=0.1,dtype=tf.float32,name="w3"))
b3 = tf.Variable(initial_value=tf.zeros(shape=[1024]))
#關鍵，進行reshape
input3 = tf.reshape(pool2,shape=[-1,7*7*64],name="input3")
fc1 = tf.nn.relu(tf.nn.bias_add(value=tf.matmul(input3,w3),bias=b3),name="fc1")
#shape={None，1024}
#FC2
w4 = tf.Variable(initial_value=tf.random_normal(shape=[1024,10],stddev=0.1,dtype=tf.float32,name="w4"))
b4 = tf.Variable(initial_value=tf.zeros(shape=[10]))
fc2 = tf.nn.bias_add(value=tf.matmul(fc1,w4),bias=b4)
#shape={None，10}
#定義交叉熵損失
# 使用softmax將NN計算輸出值表示爲機率
y = tf.nn.softmax(fc2,name="out")

# 定義交叉熵損失函數
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=fc2,labels=y_)
loss = tf.reduce_mean(cross_entropy)
#定義solver
train = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(loss=loss)

#定義正確值,判斷兩者下標index是否相等
correct_predict = tf.equal(tf.argmax(y,1),tf.argmax(y_,1))
#定義如何計算準確率
accuracy = tf.reduce_mean(tf.cast(correct_predict,dtype=tf.float32),name="accuracy")
#定義初始化op
init = tf.global_variables_initializer()

#訓練NN
with tf.Session() as session:
    session.run(fetches=init)
    for i in range(0,1000):
        xs, ys = mnist.train.next_batch(100)
        session.run(fetches=train,feed_dict={x:xs,y_:ys})
        if i%100 == 0:
            train_accuracy = session.run(fetches=accuracy,feed_dict={x:xs,y_:ys})
            print(i,"accuracy=",train_accuracy)
    #訓練完成後，將網絡中的權值轉化爲常量，造成常量graph，注意：須要x與label
    constant_graph = graph_util.convert_variables_to_constants(sess=session,
                                                            input_graph_def=session.graph_def,
                                                            output_node_names=['out','y_','input'])
    #將帶權值的graph序列化，寫成pb文件存儲起來
    with tf.gfile.FastGFile("lenet.pb", mode='wb') as f:
        f.write(constant_graph.SerializeToString())

【fine-tuning文件以下】

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np

mnist = input_data.read_data_sets(train_dir=r"E:\mnist_data",one_hot=True)
pb_path = r"C:\Users\ThinkPad\PycharmProjects\tf\TensorFlow\fine-tuning\mnist\lenet.pb"
#導入pb文件到graph中
with tf.gfile.FastGFile(pb_path,'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    _ = tf.import_graph_def(graph_def, name='')
with tf.Session() as session:
    #獲取輸入tensor
    x = tf.get_default_graph().get_tensor_by_name("input:0")
    #獲取標籤ensor
    y_ = tf.get_default_graph().get_tensor_by_name("y_:0")
    #進行fine-tuning
    #獲取fc1
    fc1 = tf.get_default_graph().get_tensor_by_name("fc1:0")
    #定義新的FC2
    # FC2
    fc2_w = tf.Variable(initial_value=tf.random_normal(shape=[1024, 512], stddev=0.1, dtype=tf.float32, name="w4"),name="fc2_w")
    fc2_b = tf.Variable(initial_value=tf.zeros(shape=[512]),name="fc2_b")
    fc2 = tf.nn.bias_add(value=tf.matmul(fc1, fc2_w), bias=fc2_b)

    # FC3
    fc3_w = tf.Variable(initial_value=tf.random_normal(shape=[512, 10], stddev=0.1, dtype=tf.float32, name="w4"),name="fc3_w")
    fc3_b = tf.Variable(initial_value=tf.zeros(shape=[10]),name="fc3_b")
    fc3 = tf.nn.bias_add(value=tf.matmul(fc2, fc3_w), bias=fc3_b)

    y = tf.nn.softmax(fc3, name="out")

    # 定義交叉熵損失函數
    cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=fc3, labels=y_)
    loss = tf.reduce_mean(cross_entropy)
    # 定義solver
    train = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(loss=loss)

    # 定義正確值,判斷兩者下標index是否相等
    correct_predict = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
    # 定義如何計算準確率
    accuracy = tf.reduce_mean(tf.cast(correct_predict, dtype=tf.float32), name="accuracy")
    # 定義初始化op
    init = tf.global_variables_initializer()
    session.run(init)

    for i in range(0,1000):
        xs, ys = mnist.train.next_batch(100)
        session.run(fetches=train,feed_dict={x:xs,y_:ys})
        if i%100 == 0:
            train_accuracy = session.run(fetches=accuracy,feed_dict={x:xs,y_:ys})
            print(i,"accuracy=",train_accuracy)

結果以下圖：