版权声明：小博主大三在读，水平有限，希望大家多多指导，Personal Page：holeungliu.com https://blog.csdn.net/soulmeetliang/article/details/78671517
秉持着打代码学习的习惯，将《TensorFlow实战》实现一讲。
关于AlexNet
代码加注释
Code
'''
建立一个完整的AlexNet卷积神经网络，
然后对它每个batch的前馈计算（forward）和反馈计算（backward）的速度进行测试，
这里使用随机图片数据来计算每轮前馈、反馈的平均耗时。
holeung
2017.11.29
'''
from datetime import datetime
import math
import time
import tensorflow as tf
batch_size = 32
num_batches = 100
# 显示每一层的结构,tensor输入
def print_activations(t):
print(t.op.name, ' ', t.get_shape().as_list()) # t.op.name显示名称，t.get_shape()显示大小
# 函数inference接受images作为输入，返回最后一层pool5(第五个池化层)及parameters,包括了很多卷积和池化层
def inference(images):
parameters = []
###第一个卷积层###
with tf.name_scope('conv1') as scope: # 可将scope内生成的Variable自动命名为conv1/xxx,便于区分不同卷积层之间的组件
kernel = tf.Variable(tf.truncated_normal([11,11,3,64], # 使用tf.truncated_normal截断的正态分布函数（标准差为0.1）初始化卷积层的kernel
dtype=tf.float32,stddev=1e-1), name='weights') # 卷积核尺寸为11*11，颜色通道为3，卷积核数量为64
conv = tf.nn.conv2d(images, kernel, [1,4,4,1], padding='SAME') # 对输入images完成卷积操作，步长设为4*4，每次取样卷积核大小都为11*11
biases = tf.Variable(initial_value=tf.constant(0.0, shape=[64], dtype=tf.float32), trainable=True, name='biases') # biases初始化为 0
bias = tf.nn.bias_add(conv, biases) # 将conv和biases加起来
conv1 = tf.nn.relu(bias, name=scope) # 激活函数RELU对结果进行非线性处理
print_activations(conv1) # 打印最后一层输出的tensor conv1的结构
parameters += [kernel,biases] # 将这一层可训练的参数kernel、biases添加到parameters中
###在第一个卷积层后添加LRN层和最大池化层###
lrn1 = tf.nn.lrn(conv1, depth_radius=4, bias=1.0, alpha=0.001/9, beta=0.75, name='lrn1') # （可选）“相邻神经元抑制”使用LRN会让前馈、反馈的速度大大下降
pool1 = tf.nn.max_pool(lrn1, ksize=[1,3,3,1], strides=[1,2,2,1], # 池化尺寸3*3，即将3*3的大小像素降为1*1
padding='VALID', name='pool1') # padding设定为VALID，取样时不能超过边框，设定为SAME可以填充边界外的点
print_activations(pool1)
###第二个卷积层###
with tf.name_scope('conv2') as scope:
kernel = tf.Variable(initial_value=tf.truncated_normal([5,5,64,192], # 卷积核尺寸5*5，输入通道数64（上层卷积核数量），卷积核数量192
dtype=tf.float32, stddev=1e-1), name='weights')
conv = tf.nn.conv2d(pool1, kernel, [1,1,1,1], padding='SAME') # 卷积步长设置为1*1，即扫描全图像素。
biases = tf.Variable(initial_value=tf.constant(0.0, shape=[192], dtype=tf.float32), trainable=True, name='biases')
bias = tf.nn.bias_add(conv, biases)
conv2 = tf.nn.relu(bias, name=scope)
parameters += [kernel, biases]
print_activations(conv2)
###在第二个卷积层后添加LRN层和最大池化层###
lrn2 = tf.nn.lrn(conv2, depth_radius=4, bias=1.0, alpha=0.001/9, beta=0.75, name='lrn2') # （可选）“相邻神经元抑制”使用LRN会让前馈、反馈的速度大大下降
pool2 = tf.nn.max_pool(lrn2, ksize=[1,3,3,1], strides=[1,2,2,1], # 池化尺寸3*3，即将3*3的大小像素降为1*1
padding='VALID', name='pool2') # padding设定为VALID，取样时不能超过边框，设定为SAME可以填充边界外的点
print_activations(pool2)
###第三个卷积层###
with tf.name_scope('conv3') as scope:
kernel = tf.Variable(initial_value=tf.truncated_normal([3,3,192,384], # 卷积核尺寸3*3，输入通道数192（上层卷积核数量），卷积核数量384
dtype=tf.float32, stddev=1e-1), name='weights')
conv = tf.nn.conv2d(pool2, kernel, [1,1,1,1], padding='SAME') # 卷积步长设置为1*1，即扫描全图像素。
biases = tf.Variable(initial_value=tf.constant(0.0, shape=[384], dtype=tf.float32), trainable=True, name='biases')
bias = tf.nn.bias_add(conv, biases)
conv3 = tf.nn.relu(bias, name=scope)
parameters += [kernel, biases]
print_activations(conv3)
###第四个卷积层###
with tf.name_scope('conv4') as scope:
kernel = tf.Variable(initial_value=tf.truncated_normal([3,3,384,256], # 卷积核尺寸3*3，输入通道数384，卷积核数量下降到256
dtype=tf.float32, stddev=1e-1), name='weights')
conv = tf.nn.conv2d(conv3, kernel, [1,1,1,1], padding='SAME') # 卷积步长设置为1*1，即扫描全图像素。
biases = tf.Variable(initial_value=tf.constant(0.0, shape=[256], dtype=tf.float32), trainable=True, name='biases')
bias = tf.nn.bias_add(conv, biases)
conv4 = tf.nn.relu(bias, name=scope)
parameters += [kernel, biases]
print_activations(conv4)
###第五个卷积层###
with tf.name_scope('conv5') as scope:
kernel = tf.Variable(initial_value=tf.truncated_normal([3,3,256,256], # 卷积核尺寸3*3，输入通道数256，卷积核数量也是256
dtype=tf.float32, stddev=1e-1), name='weights')
conv = tf.nn.conv2d(conv4, kernel, [1,1,1,1], padding='SAME') # 卷积步长设置为1*1，即扫描全图像素。
biases = tf.Variable(initial_value=tf.constant(0.0, shape=[256], dtype=tf.float32), trainable=True, name='biases')
bias = tf.nn.bias_add(conv, biases)
conv5 = tf.nn.relu(bias, name=scope)
parameters += [kernel, biases]
print_activations(conv4)
###五个卷积之后加一个最大池化层###
pool5 = tf.nn.max_pool(conv5, ksize=[1,3,3,1], strides=[1,2,2,1], padding='VALID', name='pool5')
print_activations(pool5)
return pool5,parameters
###评估AlexNet每轮计算时间的函数###
def time_tensorflow_run(session,target,info_string): # (TensorFlow的Session, 需要评测的运算算子, 测试的名称)
num_steps_burn_in = 10 # 给程序热身，头几轮迭代有显存加载、cache命中等问题因此可以跳过，只考量10轮之后的计算时间
total_duration = 0.0
total_duration_squared = 0.0 # 用以计算方差
for i in range(num_batches + num_steps_burn_in):
start_time = time.time() # 记录时间
_ = session.run(target) # 执行迭代
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
# 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 step, %.3f +/- %.3f sec / batch' % (datetime.now(), info_string, num_batches, mn, sd))
###主函数###
def run_benchmark():
with tf.Graph().as_default(): # 定义默认Graph
image_size = 224
# 并不使用ImageNet数据集来训练，只使用随机图片数据测试前馈和反馈计算的耗时，tf.random_normal 函数构造正态分布的随机tensor
images = tf.Variable(tf.random_normal([batch_size, # 每轮迭代的样本数
image_size, # 图片尺寸
image_size, 3], # 颜色通道数
dtype=tf.float32,
stddev=1e-1))
# # 使用inference函数构建整个AlexNet网络，得到最后一个池化层的输出pool5和网络中需要训练的参数集合parameters
pool5, parameters = inference(images)
init = tf.global_variables_initializer() # 初始化所有参数
sess = tf.Session() # 创建新的Session
sess.run(init)
time_tensorflow_run(sess, pool5, "Forward") # 使用time_tensorflow_run统计运算时间，传入的target就是pool5，即卷积网络最后一个池化层的输出
objective = tf.nn.l2_loss(pool5) # 与forward计算有些不同，需要给最后输出的pool5设置一个优化目标tf.nn.l2_loss计算pool5的loss
grad = tf.gradients(objective, parameters) # 用tf.gradients求相对于loss的所有模型参数的梯度，这样就模拟了一个训练过程
time_tensorflow_run(sess, grad, "Forward-backward") # 使用time_tensorflow_run统计backward的运算时间，target就是求整个网络梯度gard的操作
###执行主函数###
run_benchmark()Result
conv1 [32, 56, 56, 64]
pool1 [32, 27, 27, 64]
conv2 [32, 27, 27, 192]
pool2 [32, 13, 13, 192]
conv3 [32, 13, 13, 384]
conv4 [32, 13, 13, 256]
conv4 [32, 13, 13, 256]
pool5 [32, 6, 6, 256]
<built-in method now of type object at 0x000000006F551CD0>: step 0, duration = 1.723
<built-in method now of type object at 0x000000006F551CD0>: step 10, duration = 1.388
<built-in method now of type object at 0x000000006F551CD0>: step 20, duration = 1.369
<built-in method now of type object at 0x000000006F551CD0>: step 30, duration = 1.411
<built-in method now of type object at 0x000000006F551CD0>: step 40, duration = 1.366
<built-in method now of type object at 0x000000006F551CD0>: step 50, duration = 1.398
<built-in method now of type object at 0x000000006F551CD0>: step 60, duration = 1.734
<built-in method now of type object at 0x000000006F551CD0>: step 70, duration = 1.384
<built-in method now of type object at 0x000000006F551CD0>: step 80, duration = 1.393
<built-in method now of type object at 0x000000006F551CD0>: step 90, duration = 1.423
2017-11-29 23:51:17.738062: Forward across 100 step, 1.434 +/- 0.131 sec / batch
<built-in method now of type object at 0x000000006F551CD0>: step 0, duration = 5.585
<built-in method now of type object at 0x000000006F551CD0>: step 10, duration = 5.500
<built-in method now of type object at 0x000000006F551CD0>: step 20, duration = 5.530
<built-in method now of type object at 0x000000006F551CD0>: step 30, duration = 5.645
<built-in method now of type object at 0x000000006F551CD0>: step 40, duration = 5.746
<built-in method now of type object at 0x000000006F551CD0>: step 50, duration = 5.529
<built-in method now of type object at 0x000000006F551CD0>: step 60, duration = 5.559
<built-in method now of type object at 0x000000006F551CD0>: step 70, duration = 5.659
<built-in method now of type object at 0x000000006F551CD0>: step 80, duration = 5.927
<built-in method now of type object at 0x000000006F551CD0>: step 90, duration = 6.229
2017-11-30 00:01:51.473134: Forward-backward across 100 step, 5.751 +/- 0.383 sec / batch