VGGNet
優勢： 探索卷積神經網絡的深度與其性能之間的關系—>反複堆疊3 * 3 kernel 和 2 * 2 最大池化層—>成功建構了16~19層深度卷積神經網絡—>遷移和泛化能力強，網絡結構簡單
[外鍊圖檔轉存失敗(img-JWb9qJR3-1567153752204)(attachment:image.png)]
-
A —> E 六個級别網絡結構
[外鍊圖檔轉存失敗(img-gmPBgLMv-1567153752204)(attachment:%E6%8D%95%E8%8E%B72.PNG)]
#建構VGGNet-16（D級）
import tensorflow as tf
import time
from datetime import datetime
import math
# 定義卷積層函數----->相比于AlexNet網絡有了進一步的簡化
def conv_op(input_op,name,kh,kw,n_out,dh,dw,parameters):
'''
input_op:輸入的tensor
name:tensor名稱
kh：kernel的高
kw：kernel的寬
n_out:tensor輸出通道數
dh：步長的高
dw：步長的寬
parameters：參數清單
'''
n_in = input_op.get_shape()[-1].value#擷取tensor的通道數
with tf.name_scope(name) as scope:
#kernel進行初始化---->tf.contrib.layers.xavier_initializer_conv2d()
kernel = tf.get_variable(scope+'w',
shape=[kh,kw,n_in,n_out],dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer_conv2d())
#biases初始化
biases = tf.Variable(tf.constant(0.0,shape=[n_out],dtype=tf.float32),
trainable=True,name='bias')
#計算卷積并激活relu
conv = tf.nn.relu(
tf.nn.bias_add(tf.nn.conv2d(input_op,kernel,[1,dh,dw,1],padding='SAME'),
biases),name=scope)
parameters += [kernel,biases]
return conv
# 定義全連接配接層函數
def fc_op(input_op,name,n_out,parameters):
n_in = input_op.get_shape()[-1].value
with tf.name_scope(name) as scope:
kerenl = tf.get_variable(scope+'w',
shape=[n_in,n_out],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
biases = tf.Variable(tf.constant(0.1,shape=[n_out],dtype=tf.float32),
name='bias')
activation = tf.nn.relu_layer(input_op,kerenl,biases,name=scope)
parameters += [kerenl,biases]
return activation
# 定義最大池化層函數
def m_max_pool(input_op,name,kh,kw,dh,dw):
return tf.nn.max_pool(input_op,
ksize=[1,kh,kw,1],
strides=[1,dh,dw,1],
padding='SAME',
name=name)
# 定義VGGNet的inference
def inference(input_op,keep_prop):
'''
keep_prop: dropout 政策的保留率
'''
parameters = []
conv1_1 = conv_op(input_op,name='conv1_1',kh=3,kw=3,n_out=64,dh=1,dw=1,parameters=parameters)
conv1_2 = conv_op(conv1_1,name='conv1_2',kh=3,kw=3,n_out=64,dh=1,dw=1,parameters=parameters)
pool1 = m_max_pool(conv1_2,name='pool1',kh=2,kw=2,dh=2,dw=2)
conv2_1 = conv_op(pool1,name='conv2_1',kh=3,kw=3,n_out=128,dh=1,dw=1,parameters=parameters)
conv2_2 = conv_op(conv2_1,name='conv2_2',kh=3,kw=3,n_out=128,dh=1,dw=1,parameters=parameters)
pool2 = m_max_pool(conv2_2,name='pool2',kh=2,kw=2,dh=2,dw=2)
conv3_1 = conv_op(pool2,name='conv3_1',kh=3,kw=3,n_out=256,dh=1,dw=1,parameters=parameters)
conv3_2 = conv_op(conv3_1,name='conv3_2',kh=3,kw=3,n_out=256,dh=1,dw=1,parameters=parameters)
conv3_3 = conv_op(conv3_2,name='conv3_3',kh=3,kw=3,n_out=256,dh=1,dw=1,parameters=parameters)
pool3 = m_max_pool(conv3_3,name='pool3',kh=2,kw=2,dh=2,dw=2)
conv4_1 = conv_op(pool3,name='conv4_1',kh=3,kw=3,n_out=512,dh=1,dw=1,parameters=parameters)
conv4_2 = conv_op(conv4_1,name='conv4_2',kh=3,kw=3,n_out=512,dh=1,dw=1,parameters=parameters)
conv4_3 = conv_op(conv4_2,name='conv4_3',kh=3,kw=3,n_out=512,dh=1,dw=1,parameters=parameters)
pool4 = m_max_pool(conv4_3,name='pool4',kh=2,kw=2,dh=2,dw=2)
conv5_1 = conv_op(pool4,name='conv5_1',kh=3,kw=3,n_out=512,dh=1,dw=1,parameters=parameters)
conv5_2 = conv_op(conv5_1,name='conv5_2',kh=3,kw=3,n_out=512,dh=1,dw=1,parameters=parameters)
conv5_3 = conv_op(conv5_2,name='conv5_3',kh=3,kw=3,n_out=512,dh=1,dw=1,parameters=parameters)
pool5 = m_max_pool(conv5_3,name='pool5',kh=2,kw=2,dh=2,dw=2)
conv_shape = pool5.get_shape()
flatten_shape = conv_shape[1].value * conv_shape[2].value * conv_shape[3].value
resh1 = tf.reshape(pool5,[-1,flatten_shape],name='resh1')
fc_6 = fc_op(resh1,name='fc_6',n_out=4096,parameters=parameters)
fc_6drop = tf.nn.dropout(fc_6,rate=1-keep_prop,name='fc_6drop')
fc_7 = fc_op(fc_6drop,name='fc_7',n_out=4096,parameters=parameters)
fc_7drop = tf.nn.dropout(fc_7,rate=1-keep_prop,name='fc_7drop')
fc_8 = fc_op(fc_7drop,name='fc_8',n_out=1000,parameters=parameters)
softmax = tf.nn.softmax(fc_8)
predictions = tf.argmax(softmax,1)
return predictions,softmax,fc_8,parameters
# 測評
def time_tensorflow_run(session,target,feed,info_string):
'''
session:視窗
target:評測對象
feed：輸入的資料
info_string：對象名稱
'''
num_steps_burn_in = 10#程式熱身
#用于計算方差的兩個參數
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,feed_dict=feed)
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
#計算平均耗時和标準差
mean_time = total_duration / num_batches
sd = math.sqrt(total_duration_squared / num_batches - mean_time * mean_time)
print('%s: %s across %d steps,%.3f +/- %.3f sec / batch'%(
datetime.now(),info_string,num_batches,mean_time,sd))
# 建構一個自己的資料集（使用ImageNet資料集訓練過程十分耗時）
def run_benchmark():
#定義一個新的圖進行計算
with tf.Graph().as_default():
image_size = 224
#構造随機tensor
images = tf.Variable(tf.random_normal([batch_size,
image_size,
image_size,3],
dtype=tf.float32,stddev=0.1))
keep_prop = tf.placeholder(tf.float32)
predictions,softmax,fc_8,parameters = inference(images,keep_prop)
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
time_tensorflow_run(sess,predictions,{keep_prop:1.0},'Forward')
objetive = tf.nn.l2_loss(fc_8)#計算fc_8正則化損失
grad = tf.gradients(objetive,parameters)
time_tensorflow_run(sess,grad,{keep_prop:0.5},'Forward-Backward')
batch_size = 8
num_batches = 10
run_benchmark()
Tensor("pool5:0", shape=(8, 7, 7, 512), dtype=float32)
Tensor("resh1:0", shape=(8, 25088), dtype=float32)
Tensor("fc_6:0", shape=(8, 4096), dtype=float32)
Tensor("fc_7:0", shape=(8, 4096), dtype=float32)
Tensor("fc_8:0", shape=(8, 1000), dtype=float32)
Tensor("fc_8:0", shape=(8, 1000), dtype=float32)
2019-08-22 16:47:58.296828:step 0,duration = 1.367
2019-08-22 16:48:08.086028: Forward across 10 steps,1.115 +/- 0.227 sec / batch
Tensor("L2Loss:0", shape=(), dtype=float32)
2019-08-22 16:48:58.554828:step 0,duration = 4.645
2019-08-22 16:49:39.599828: Forward-Backward across 10 steps,4.569 +/- 0.126 sec / batch
![tensorflow筆記實踐：正則化優化過拟合[圖]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)