发布时间：2024-02-19 17:00
网络介绍：
https://blog.csdn.net/loveliuzz/article/details/79131131
https://blog.csdn.net/jiaoyangwm/article/details/80011656
1.CNN---LeNet5
https://blog.csdn.net/happyorg/article/details/78274066
https://www.cnblogs.com/chizi15/p/9808330.html
https://blog.csdn.net/Alvin_FZW/article/details/81240247
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import time
# 声明输入图片数据，类别
x = tf.placeholder('float', [None, 784])
y_ = tf.placeholder('float', [None, 10])
# 输入图片数据转化
x_image = tf.reshape(x, [-1, 28, 28, 1])
#第一层卷积层，初始化卷积核参数、偏置值，该卷积层5*5大小，一个通道，共有6个不同卷积核
filter1 = tf.Variable(tf.truncated_normal([5, 5, 1, 6]))
bias1 = tf.Variable(tf.truncated_normal([6]))
conv1 = tf.nn.conv2d(x_image, filter1, strides=[1, 1, 1, 1], padding='SAME')
h_conv1 = tf.nn.sigmoid(conv1 + bias1)
#池化层
maxPool1 = tf.nn.max_pool(h_conv1, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME')
#第二层卷积
filter2 = tf.Variable(tf.truncated_normal([5, 5, 6, 16]))
bias2 = tf.Variable(tf.truncated_normal([16]))
conv2 = tf.nn.conv2d(maxPool1, filter2, strides=[1, 1, 1, 1], padding='SAME')
h_conv2 = tf.nn.sigmoid(conv2 + bias2)
#池化层
maxPool2 = tf.nn.max_pool(h_conv2, ksize=[1, 2, 2, 1],strides=[1, 2, 2, 1], padding='SAME')
#全连接层1：本质是卷积，由于尺寸跟内核一样，跟全连接没区别，因此也叫全连接层
filter3 = tf.Variable(tf.truncated_normal([5, 5, 16, 120]))
bias3 = tf.Variable(tf.truncated_normal([120]))
conv3 = tf.nn.conv2d(maxPool2, filter3, strides=[1, 1, 1, 1], padding='SAME')
h_conv3 = tf.nn.sigmoid(conv3 + bias3)
# 全连接层2
# 权值参数
W_fc1 = tf.Variable(tf.truncated_normal([7 * 7 * 120, 80]))
# 偏置值
b_fc1 = tf.Variable(tf.truncated_normal([80]))
# 将卷积的产出展开
h_pool2_flat = tf.reshape(h_conv3, [-1, 7 * 7 * 120])
# 神经网络计算，并添加sigmoid激活函数
h_fc1 = tf.nn.sigmoid(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
# 输出层3，使用softmax进行多分类
W_fc2 = tf.Variable(tf.truncated_normal([80, 10]))
b_fc2 = tf.Variable(tf.truncated_normal([10]))
y_conv = tf.nn.softmax(tf.matmul(h_fc1, W_fc2) + b_fc2)
# 损失函数
cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))
# 使用GDO优化算法来调整参数
train_step = tf.train.GradientDescentOptimizer(0.001).minimize(cross_entropy)
sess = tf.InteractiveSession()
# 测试正确率
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
# 所有变量进行初始化
sess.run(tf.initialize_all_variables())
# 获取mnist数据
mnist_data_set = input_data.read_data_sets('mnist', one_hot=True)
# 进行训练
start_time = time.time()
for i in range(2000):
# 获取训练数据
batch_xs, batch_ys = mnist_data_set.train.next_batch(200)
# 每迭代10个 batch，对当前训练数据进行测试，输出当前预测准确率
if i % 10 == 0:
train_accuracy = accuracy.eval(feed_dict={x: batch_xs, y_: batch_ys})
print("step %d, training accuracy %g" % (i, train_accuracy))
# 计算间隔时间
end_time = time.time()
print('time: ', (end_time - start_time))
start_time = end_time
# 训练数据
train_step.run(feed_dict={x: batch_xs, y_: batch_ys})
# 关闭会话
sess.close()2.CNN---AlexNet
https://blog.csdn.net/u012679707/article/details/80793916
https://blog.csdn.net/qq_24695385/article/details/80368618
https://blog.csdn.net/qq_28123095/article/details/79776329
# -*- coding:utf-8 -*-
"""
@author:Lisa
@file:alexNet.py
@function:实现Alexnet深度模型
@note:learn from《tensorflow实战》
@time:2018/6/24 0024下午 5:26
"""
import tensorflow as tf
import time
import math
from datetime import datetime
batch_size=32
num_batch=100
keep_prob=0.5
def print_architecture(t):
"""print the architecture information of the network,include name and size"""
print(t.op.name," ",t.get_shape().as_list())
def inference(images):
""" 构建网络 ：5个conv+3个FC"""
parameters=[] #储存参数
with tf.name_scope('conv1') as scope:
"""
images:227*227*3
kernel: 11*11 *64
stride:4*4
padding:name
#通过with tf.name_scope('conv1') as scope可以将scope内生成的Variable自动命名为conv1/xxx
便于区分不同卷积层的组建
input: images[227*227*3]
middle: conv1[55*55*96]
output: pool1 [27*27*96]
"""
kernel=tf.Variable(tf.truncated_normal([11,11,3,96],
dtype=tf.float32,stddev=0.1),name="weights")
conv=tf.nn.conv2d(images,kernel,[1,4,4,1],padding='SAME')
biases=tf.Variable(tf.constant(0.0, shape=[96], dtype=tf.float32),
trainable=True,name="biases")
bias=tf.nn.bias_add(conv,biases) # w*x+b
conv1=tf.nn.relu(bias,name=scope) # reLu
print_architecture(conv1)
parameters +=[kernel,biases]
#添加LRN层和max_pool层
"""
LRN会让前馈、反馈的速度大大降低（下降1/3），但最终效果不明显，所以只有ALEXNET用LRN，其他模型都放弃了
"""
lrn1=tf.nn.lrn(conv1,depth_radius=4,bias=1,alpha=0.001/9,beta=0.75,name="lrn1")
pool1=tf.nn.max_pool(lrn1,ksize=[1,3,3,1],strides=[1,2,2,1],
padding="VALID",name="pool1")
print_architecture(pool1)
with tf.name_scope('conv2') as scope:
"""
input: pool1[27*27*96]
middle: conv2[27*27*256]
output: pool2 [13*13*256]
"""
kernel = tf.Variable(tf.truncated_normal([5, 5, 96, 256],
dtype=tf.float32, stddev=0.1), name="weights")
conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
trainable=True, name="biases")
bias = tf.nn.bias_add(conv, biases) # w*x+b
conv2 = tf.nn.relu(bias, name=scope) # reLu
parameters += [kernel, biases]
# 添加LRN层和max_pool层
"""
LRN会让前馈、反馈的速度大大降低（下降1/3），但最终效果不明显，所以只有ALEXNET用LRN，其他模型都放弃了
"""
lrn2 = tf.nn.lrn(conv2, depth_radius=4, bias=1, alpha=0.001 / 9, beta=0.75, name="lrn1")
pool2 = tf.nn.max_pool(lrn2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],
padding="VALID", name="pool2")
print_architecture(pool2)
with tf.name_scope('conv3') as scope:
"""
input: pool2[13*13*256]
output: conv3 [13*13*384]
"""
kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 384],
dtype=tf.float32, stddev=0.1), name="weights")
conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.0, shape=[384], dtype=tf.float32),
trainable=True, name="biases")
bias = tf.nn.bias_add(conv, biases) # w*x+b
conv3 = tf.nn.relu(bias, name=scope) # reLu
parameters += [kernel, biases]
print_architecture(conv3)
with tf.name_scope('conv4') as scope:
"""
input: conv3[13*13*384]
output: conv4 [13*13*384]
"""
kernel = tf.Variable(tf.truncated_normal([3, 3, 384, 384],
dtype=tf.float32, stddev=0.1), name="weights")
conv = tf.nn.conv2d(conv3, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.0, shape=[384], dtype=tf.float32),
trainable=True, name="biases")
bias = tf.nn.bias_add(conv, biases) # w*x+b
conv4 = tf.nn.relu(bias, name=scope) # reLu
parameters += [kernel, biases]
print_architecture(conv4)
with tf.name_scope('conv5') as scope:
"""
input: conv4[13*13*384]
output: conv5 [6*6*256]
"""
kernel = tf.Variable(tf.truncated_normal([3, 3, 384, 256],
dtype=tf.float32, stddev=0.1), name="weights")
conv = tf.nn.conv2d(conv4, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
trainable=True, name="biases")
bias = tf.nn.bias_add(conv, biases) # w*x+b
conv5 = tf.nn.relu(bias, name=scope) # reLu
pool5 = tf.nn.max_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],
padding="VALID", name="pool5")
parameters += [kernel, biases]
print_architecture(pool5)
#全连接层6
with tf.name_scope('fc6') as scope:
"""
input:pool5 [6*6*256]
output:fc6 [4096]
"""
kernel = tf.Variable(tf.truncated_normal([6*6*256,4096],
dtype=tf.float32, stddev=0.1), name="weights")
biases = tf.Variable(tf.constant(0.0, shape=[4096], dtype=tf.float32),
trainable=True, name="biases")
# 输入数据变换
flat = tf.reshape(pool5, [-1, 6*6*256] ) # 整形成m*n,列n为7*7*64
# 进行全连接操作
fc = tf.nn.relu(tf.matmul(flat, kernel) + biases,name='fc6')
# 防止过拟合 nn.dropout
fc6 = tf.nn.dropout(fc, keep_prob)
parameters += [kernel, biases]
print_architecture(fc6)
# 全连接层7
with tf.name_scope('fc7') as scope:
"""
input:fc6 [4096]
output:fc7 [4096]
"""
kernel = tf.Variable(tf.truncated_normal([4096, 4096],
dtype=tf.float32, stddev=0.1), name="weights")
biases = tf.Variable(tf.constant(0.0, shape=[4096], dtype=tf.float32),
trainable=True, name="biases")
# 进行全连接操作
fc = tf.nn.relu(tf.matmul(fc6, kernel) + biases, name='fc7')
# 防止过拟合 nn.dropout
fc7 = tf.nn.dropout(fc, keep_prob)
parameters += [kernel, biases]
print_architecture(fc7)
# 全连接层8
with tf.name_scope('fc8') as scope:
"""
input:fc7 [4096]
output:fc8 [1000]
"""
kernel = tf.Variable(tf.truncated_normal([4096, 1000],
dtype=tf.float32, stddev=0.1), name="weights")
biases = tf.Variable(tf.constant(0.0, shape=[1000], dtype=tf.float32),
trainable=True, name="biases")
# 进行全连接操作
fc8 = tf.nn.xw_plus_b(fc7, kernel, biases, name='fc8')
parameters += [kernel, biases]
print_architecture(fc8)
return fc8,parameters
def time_compute(session,target,info_string):
num_step_burn_in=10 #预热轮数，头几轮迭代有显存加载、cache命中等问题可以因此跳过
total_duration=0.0 #总时间
total_duration_squared=0.0
for i in range(num_batch+num_step_burn_in):
start_time=time.time()
_ = session.run(target)
duration= time.time() -start_time
if i>= num_step_burn_in:
if i%10==0: #每迭代10次显示一次duration
print("%s: step %d,duration=%.5f "% (datetime.now(),i-num_step_burn_in,duration))
total_duration += duration
total_duration_squared += duration *duration
time_mean=total_duration /num_batch
time_variance=total_duration_squared / num_batch - time_mean*time_mean
time_stddev=math.sqrt(time_variance)
#迭代完成，输出
print("%s: %s across %d steps,%.3f +/- %.3f sec per batch "%
(datetime.now(),info_string,num_batch,time_mean,time_stddev))
def main():
with tf.Graph().as_default():
"""仅使用随机图片数据 测试前馈和反馈计算的耗时"""
image_size =224
images=tf.Variable(tf.random_normal([batch_size,image_size,image_size,3],
dtype=tf.float32,stddev=0.1 ) )
fc8,parameters=inference(images)
init=tf.global_variables_initializer()
sess=tf.Session()
sess.run(init)
"""
AlexNet forward 计算的测评
传入的target:fc8（即最后一层的输出）
优化目标：loss
使用tf.gradients求相对于loss的所有模型参数的梯度
AlexNet Backward 计算的测评
target:grad
"""
time_compute(sess,target=fc8,info_string="Forward")
obj=tf.nn.l2_loss(fc8)
grad=tf.gradients(obj,parameters)
(sess,grad,"Forward-backward")
if __name__=="__main__":
main()
3.CNN---VGG-16
https://blog.csdn.net/u012679707/article/details/80807406
https://blog.csdn.net/gybheroin/article/details/79903044
https://blog.csdn.net/gybheroin/article/details/79806096
from datetime import datetime
import math
import time
import tensorflow as tf
batch_size = 16 #一个批次的数据
num_batches = 100 #测试一百个批次的数据
'''卷积层创建函数，并将本层参数存入参数列表
input_op:输入的tensor name:这一层的名称 kh:kernel height即卷积核的高 kw:kernel width即卷积核的宽
n_out:卷积核数量即输出通道数 dh:步长的高 dw:步长的宽 p:参数列表
'''
def conv_op(input_op, name, kh, kw, n_out, dh, dw, p):
# 获取输入数据的通道数
n_in = input_op.get_shape()[-1].value
with tf.name_scope(name) as scope:
#创建卷积核，shape的值的意义参见alexNet
kernel = tf.get_variable(scope+"w", shape=[kh, kw, n_in, n_out], dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer_conv2d())
#卷积操作
conv = tf.nn.conv2d(input_op, kernel, (1, dh, dw, 1), padding='SAME')
#初始化bias为0
bias_init_val = tf.constant(0.0, shape=[n_out], dtype=tf.float32)
biases = tf.Variable(bias_init_val, trainable=True, name='b')
#将卷积后结果与biases加起来
z = tf.nn.bias_add(conv, biases)
#使用激活函数relu进行非线性处理
activation = tf.nn.relu(z, name=scope)
#将卷积核和biases加入到参数列表
p += [kernel, biases]
# tf.image.resize_images()
#卷积层输出作为函数结果返回
return activation
'''全连接层FC创建函数'''
def fc_op(input_op, name, n_out, p):
#获取input_op的通道数
n_in = input_op.get_shape()[-1].value
with tf.name_scope(name) as scope:
#初始化全连接层权重
kernel = tf.get_variable(scope+"w", shape=[n_in, n_out], dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
#初始化biases为0.1而不为0，避免dead neuron
biases = tf.Variable(tf.constant(0.1, shape=[n_out], dtype=tf.float32), name='b')
#Computes Relu(x * weight + biases)
activation = tf.nn.relu_layer(input_op, kernel, biases, name=scope)
#将权重和biases加入到参数列表
p += [kernel, biases]
#activation作为函数结果返回
return activation
'''最大池化层创建函数'''
def mpool_op(input_op, name, kh, kw, dh, dw):
return tf.nn.max_pool(input_op,
ksize=[1, kh, kw, 1], #池化窗口大小
strides=[1, dh, dw, 1], #池化步长
padding='SAME',
name= name)
'''创建VGGNet-16-D的网络结构
input_op为输入数据，keep_prob为控制dropoout比率的一个placeholder
'''
def inference_op(input_op, keep_prob):
p = []
'''D-第一段'''
#第一段卷积网络第一个卷积层，输出尺寸224*224*64，卷积后通道数(厚度)由3变为64
conv1_1 = conv_op(input_op, name="conv1_1", kh=3, kw=3, n_out=64, dh=1, dw=1, p=p)
#第一段卷积网络第二个卷积层，输出尺寸224*224*64
conv1_2 = conv_op(conv1_1, name="conv1_2", kh=3, kw=3, n_out=64, dh=1, dw=1, p=p)
#第一段卷积网络的最大池化层，经过池化后输出尺寸变为112*112*64
pool1 = mpool_op(conv1_2, name="pool1", kh=2, kw=2, dw=2, dh=2)
'''D-第二段'''
#第二段卷积网络第一个卷积层，输出尺寸112*112*128，卷积后通道数由64变为128
conv2_1 = conv_op(pool1, name="conv2_1", kh=3, kw=3, n_out=128, dh=1, dw=1, p=p)
#第二段卷积网络第二个卷积层，输出尺寸112*112*128
conv2_2 = conv_op(conv2_1, name="conv2_1", kh=3, kw=3, n_out=128, dh=1, dw=1, p=p)
#第二段卷积网络的最大池化层，经过池化后输出尺寸变为56*56*128
pool2 = mpool_op(conv2_2, name="pool2", kh=2, kw=2, dw=2, dh=2)
'''D-第三段'''
#第三段卷积网络第一个卷积层，输出尺寸为56*56*256，卷积后通道数由128变为256
conv3_1 = conv_op(pool2, name="conv3_1", kh=3, kw=3, n_out=256, dh=1, dw=1, p=p)
#第三段卷积网络第二个卷积层，输出尺寸为56*56*256
conv3_2 = conv_op(conv3_1, name="conv3_2", kh=3, kw=3, n_out=256, dh=1, dw=1, p=p)
#第三段卷积网络第三个卷积层，输出尺寸为56*56*256
conv3_3 = conv_op(conv3_2, name="conv3_3", kh=3, kw=3, n_out=256, dh=1, dw=1, p=p)
#第三段卷积网络的最大池化层，池化后输出尺寸变为28*28*256
pool3 = mpool_op(conv3_3, name="pool3", kh=2, kw=2, dh=2, dw=2)
'''D-第四段'''
#第四段卷积网络第一个卷积层，输出尺寸为28*28*512，卷积后通道数由256变为512
conv4_1 = conv_op(pool3, name="conv4_1", kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
# 第四段卷积网络第二个卷积层，输出尺寸为28*28*512
conv4_2 = conv_op(conv4_1, name="conv4_2", kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
# 第四段卷积网络第三个卷积层，输出尺寸为28*28*512
conv4_3 = conv_op(conv4_2, name="conv4_3", kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
#第四段卷积网络的最大池化层，池化后输出尺寸为14*14*512
pool4 = mpool_op(conv4_3, name="pool4", kh=2, kw=2, dh=2, dw=2)
'''D-第五段'''
#第五段卷积网络第一个卷积层，输出尺寸为14*14*512
conv5_1 = conv_op(pool4, name="conv5_1", kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
# 第五段卷积网络第二个卷积层，输出尺寸为14*14*512
conv5_2 = conv_op(conv5_1, name="conv5_2", kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
# 第五段卷积网络第三个卷积层，输出尺寸为14*14*512
conv5_3 = conv_op(conv5_2, name="conv5_3", kh=3, kw=3, n_out=512, dh=1, dw=1, p=p)
#第五段卷积网络的最大池化层，池化后尺寸为7*7*512
pool5 = mpool_op(conv5_3, name="conv5_3", kh=2, kw=2, dh=2, dw=2)
'''对卷积网络的输出结果进行扁平化，将每个样本化为一个长度为25088的一维向量'''
shp = pool5.get_shape()
flattened_shape = shp[1].value * shp[2].value * shp[3].value #图像的长、宽、厚度相乘，即7*7*512=25088
# -1表示该样本有多少个是自动计算得出的，得到一个矩阵，准备传入全连接层
resh1 = tf.reshape(pool5, [-1,flattened_shape], name="resh1")
'''全连接层,共三个'''
fc6 = fc_op(resh1, name="fc6", n_out=4096, p=p)
fc6_drop = tf.nn.dropout(fc6, keep_prob, name="fc6_drop") #dropout层，keep_prob数据待外部传入
fc7 = fc_op(fc6_drop, name="fc7", n_out=4096, p=p)
fc7_drop = tf.nn.dropout(fc7, keep_prob, name="fc7_drop")
#最后一个全连接层
fc8 = fc_op(fc7_drop, name="fc8", n_out=1000, p=p)
softmax = tf.nn.softmax(fc8) #使用softmax进行处理得到分类输出概率
predictions = tf.argmax(softmax,1) #求概率最大的类别
#返回参数
return predictions, softmax, fc8, p
'''评测函数'''
def time_tensorflow_run(session, target, feed, info_string): #target:需要评测的运算算字， info_string:测试的名称
num_steps_burn_in = 10 #给程序热身，头几轮迭代有显存的加载、cache命中等问题因此可以跳过，我们只考量10轮迭代之后的计算时间
total_duration = 0.0 #总时间
total_duration_squared = 0.0 #平方和
for i in range(num_batches + num_steps_burn_in):
start_time = time.time()
_ = session.run(target, feed_dict=feed)
duration = time.time() - start_time
if i>= num_steps_burn_in: #程序热身完成后，记录时间
if not i % 10: #每10轮 显示 当前时间，迭代次数(不包括热身)，用时
print('%s: step %d, duration = %.3f' % (datetime.now(), i-num_steps_burn_in, duration))
# 累加total_duration和total_duration_squared
total_duration += duration
total_duration_squared += duration * duration
# 循环结束后，计算每轮迭代的平均耗时mn和标准差sd，最后将结果显示出来
mn = total_duration / num_batches
vr = total_duration_squared / num_batches - mn * mn
sd = math.sqrt(vr)
print('%s: %s across %d steps, %.3f +/- %.3f sec / batch' %
(datetime.now(), info_string, num_batches, mn, sd))
'''评测的主函数,不使用ImageNet数据集来训练，只使用随机图片测试前馈和反馈计算的耗时'''
def run_benchmaek():
with tf.Graph().as_default():
image_size = 224
#利用tf.random_normal()生成随机图片
images = tf.Variable(tf.random_normal([batch_size, #每轮迭代的样本数
image_size, image_size, #图片的size：image_size x image_size
3], #图片的通道数
dtype=tf.float32,
stddev=1e-1))
#创建keep_prob的placeholder
keep_prob = tf.placeholder(tf.float32)
predictions, softmax, fc8, p = inference_op(images, keep_prob)
#创建Session并初始化全局参数
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
#前向计算测评
time_tensorflow_run(sess, predictions, {keep_prob:1.0}, "Forward")
#前向和反向计算测评
objective = tf.nn.l2_loss(fc8)
grad = tf.gradients(objective, p)
time_tensorflow_run(sess, grad, {keep_prob:0.5}, "Forward-backward")
run_benchmaek()4.CNN---ResNet
https://blog.csdn.net/m0_37917271/article/details/82346233
https://blog.csdn.net/qq1483661204/article/details/79244051
https://www.jianshu.com/p/3fd3df2e1830
https://blog.csdn.net/qq_23981335/article/details/103469497
https://github.com/tensorflow/models/blob/master/research/slim/nets/resnet_v2.py
from datetime import datetime
import time
import math
import collections
import tensorflow as tf
slim = tf.contrib.slim
# 使用collections.namedtuple设计ResNet的Block模块
# scope参数是block的名称
# unit_fn是功能单元（如残差单元）
# args是一个列表，如([256, 64, 1]) X 2 + [256, 64, 2]),代表两个（256, 64, 1）单元
# 和一个（256, 64, 2）单元
Block = collections.namedtuple("Block", ['scope', 'unit_fn', 'args'])
# 定义下采样的方法，通过max_pool2d实现
def subsample(inputs, factor, scope=None):
if factor == 1:
return inputs
else:
return slim.max_pool2d(inputs, [1, 1], stride=factor, scope=scope)
# 定义一个创建卷积层的函数
def conv2d_same(inputs, num_outputs, kernel_size, stride, scope=None):
if stride == 1:
return slim.conv2d(inputs, num_outputs, kernel_size, stride=1,
padding='SAME', scope=scope)
else:
pad_total = kernel_size - 1
pad_beg = pad_total // 2
pad_end = pad_total - pad_beg
inputs = tf.pad(inputs, [[0, 0], [pad_beg, pad_end],
[pad_beg, pad_end], [0, 0]])
return slim.conv2d(inputs, num_outputs, kernel_size, stride=stride,
padding='VALID', scope=scope)
# 定义堆叠的block函数
@slim.add_arg_scope
def stack_blocks_dense(net, blocks, outputs_collections=None):
for block in blocks:
with tf.variable_scope(block.scope, 'block', [net]) as sc:
for i, unit in enumerate(block.args):
with tf.variable_scope('unit_%d' % (i + 1), values=[net]):
unit_depth, unit_depth_bottleneck, unit_stride = unit
net = block.unit_fn(net,
depth=unit_depth,
depth_bottleneck=unit_depth_bottleneck,
stride=unit_stride)
net = slim.utils.collect_named_outputs(outputs_collections, sc.name,net)
return net
# 用于设定默认值
def resnet_arg_scope(is_training=True,
weight_decay=0.0001,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
batch_norm_params = {
'is_training': is_training,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': tf.GraphKeys.UPDATE_OPS,
}
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params
):
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
# 定义残差学习单元
@slim.add_arg_scope
def bottleneck(inputs, depth, depth_bottleneck, stride,
outputs_collections=None, scope=None):
with tf.variable_scope(scope, 'bottleneck_v2', [inputs]) as sc:
depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
preact = slim.batch_norm(inputs, activation_fn=tf.nn.relu,
scope='preact')
# shortcut为直连的X
if depth == depth_in:
shortcut = subsample(inputs, stride, 'shortcut')
else:
shortcut = slim.conv2d(preact, depth, [1, 1], stride=stride,
normalizer_fn=None, activation_fn=None,
scope='shortcut')
residual = slim.conv2d(preact, depth_bottleneck, [1, 1], stride=1,
scope='conv1')
residual = conv2d_same(residual, depth_bottleneck, 3, stride,
scope='conv2')
residual = slim.conv2d(residual, depth, [1, 1], stride=1,
normalizer_fn=None, activation_fn=None,
scope='conv3')
# 将直连的X加到残差上，得到output
output = shortcut + residual
return slim.utils.collect_named_outputs(outputs_collections,
sc.name, output)
# 定义ResNet的主函数
def resnet_v2(inputs,
blocks,
num_classes=None,
global_pool=True,
include_root_block=True,
reuse=None,
scope=None):
with tf.variable_scope(scope, 'resnet_v2', [inputs], reuse=reuse) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
with slim.arg_scope([slim.conv2d, bottleneck,stack_blocks_dense],outputs_collections=end_points_collection):
net = inputs
if include_root_block:
with slim.arg_scope([slim.conv2d], activation_fn=None,normalizer_fn=None):
net = conv2d_same(net, 64, 7, stride=2, scope='conv1')
net = slim.max_pool2d(net, [3, 3], stride=2, scope='pool1')
net = stack_blocks_dense(net, blocks)
net = slim.batch_norm(net, activation_fn=tf.nn.relu, scope='postnorm')
if global_pool:
net = tf.reduce_mean(net, [1, 2], name='pool5', keep_dims=True)
if num_classes is not None:
net = slim.conv2d(net, num_classes, [1, 1], activation_fn=None,normalizer_fn=None, scope='logits')
end_points = slim.utils.convert_collection_to_dict(end_points_collection)
if num_classes is not None:
end_points['predictions'] = slim.softmax(net, scope='predictions')
return net, end_points
# 定义50层的ResNet
def resnet_v2_50(inputs,
num_classes=None,
global_pool=True,
reuse=None,
scope='resnet_v2_152'):
blocks = [
Block('block1', bottleneck, [(256, 64, 1)] * 2 + [(256, 64, 2)]),
Block('block2', bottleneck, [(512, 128, 1)] * 4 + [(512, 128, 2)]),
Block('block3', bottleneck, [(1024, 256, 1)] * 6 + [(1024, 256, 2)]),
Block('block4', bottleneck, [(2048, 512, 1)] * 3)
]
return resnet_v2(inputs, blocks, num_classes, global_pool,
include_root_block=True, reuse=reuse, scope=scope)
num_batches = 100
# 测评性能
def time_tensorflow_run(session, target, info_string):
num_steps_burn_in = 10
total_duration = 0.0
total_duration_squared = 0.0
for i in range(num_batches + num_steps_burn_in):
start_time = time.time()
_ = session.run(target)
duration = time.time() - start_time
if i >= num_steps_burn_in:
if not i % 10:
print('%s: step %d, duration = %.3f' %
(datetime.now(), i - num_steps_burn_in, duration))
total_duration += duration
total_duration_squared += duration * duration
mn = total_duration / num_batches
vr = total_duration_squared / num_batches - mn * mn
sd = math.sqrt(vr)
print('%s: %s across %d step, %.3f +/- %.3f sec / batch' %
(datetime.now(), info_string, num_batches, mn, sd))
batch_size = 32
height, width = 224, 224
inputs = tf.random_uniform([batch_size, height, width, 3])
with slim.arg_scope(resnet_arg_scope(is_training=False)):
net, end_points = resnet_v2_50(inputs, 1000)
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
num_batches = 100
time_tensorflow_run(sess, net, "Forward")