51、经典CNN网络VGGNet比赛：17种花的图片识别_笔记

东方耀 · 发表于 2019-1-12 17:15:31

51、经典CNN网络VGGNet比赛：17种花的图片识别_笔记

如果报错：OOM when allocating tensor with shape
If you want to see a list of allocated tensors when OOM happens
tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape
errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape是因为：OOM内存溢出修改代码中 卷积层的深度（核数目调小）计算指标时传入的训练集数据与测试集数据都需要调小

# -*- coding: utf-8 -*-
__author__ = 'dongfangyao'
__date__ = '2019/1/12 下午5:34'
__product__ = 'PyCharm'
__filename__ = 'tf28'
"""
17种花数据分类，是VGG网络初赛时候的数据集，现在网上没有下载；现在唯一一份数据集在tflearn这个框架中默认自带
tflearn这个框架起始是在tensorflow基础上的一个封装，API比较简单(如果代码功底比较好，建议用tensorflow)
tensorflow的执行速度比tflearn要快
基于tensorflow的框架有很多：tflearn Keras 速度上有区别
Keras是一个高层神经网络API，Keras由纯Python编写而成并基Tensorflow、Theano以及CNTK后端。
Keras 为支持快速实验而生，能够把你的idea迅速转换为结果
tflearn安装：pip install tflearn
"""
from tflearn.datasets import oxflower17
from tflearn.datasets import mnist
from tflearn.datasets import cifar10
import tensorflow as tf
from sklearn.model_selection import train_test_split
# import tflearn as tfl
# incoming, nb_filter, filter_size, strides=1, padding='same'
# tfl.conv_2d()
# 第一步：导入数据
X, Y = oxflower17.load_data(dirname="17flowers", one_hot=True)
print(X.shape) # sample_number,224,224,3
print(Y.shape) # sample_number,17
## 数据分割
train_img, test_img, train_label, test_label = train_test_split(X, Y, test_size=0.2, random_state=0)
print("训练数据-图片shape:{}；目标属性shape:{}" .format(train_img.shape, train_label.shape))
print("测试数据-图片shape:{}；目标属性shape:{}" .format(test_img.shape, test_label.shape))
train_sample_number = train_img.shape[0]
print('训练数据样本总数：{}'.format(train_sample_number))
# 第二步：设置超参并定义学习率调整策略
# 学习率，一般学习率设置的比较小
learn_rate_base = 0.1
# 每次迭代的训练样本数量
batch_size = 32
# 展示信息的间隔大小
display_step = 1
def learn_rate_func(epoch):
"""
根据给定的迭代批次，更新产生一个学习率的值均匀分布策略
:param epoch:
:return:
"""
return max(0.001, learn_rate_base * (0.9 ** int(epoch / 10)))
# 第三步：开始构建模型设置输入数据的占位符
# 输出的维度大小信息
n_classes = train_label.shape[1]
x = tf.placeholder(tf.float32, shape=[None, 224, 224, 3], name='x')
y = tf.placeholder(tf.float32, shape=[None, n_classes], name='y')
learn_rate = tf.placeholder(tf.float32, name='learn_rate')
# 第四步：构建VGG Net网络（直接将网络结构翻译成为这个代码）
def get_variable(name, shape=None, dtype=tf.float32, initializer=tf.random_normal_initializer(mean=0, stddev=0.1)):
"""
返回一个对应的变量
:param name:
:param shape:
:param dtype:
:param initializer:
:return:
"""
return tf.get_variable(name, shape, dtype, initializer)
def vgg_network(x, y):
layer1_kernel_size = 8
layer3_kernel_size = 16
layer5_kernal_size_1 = 32
layer5_kernal_size_2 = 32
layer7_kernal_size_1 = 64
layer7_kernal_size_2 = 64
layer9_kernal_size_1 = 64
layer9_kernal_size_2 = 64
layer11_unit_size = 120
layer12_unit_size = 120
layer13_unit_size = 17
# cov3-64 LRN局部响应归一化
with tf.variable_scope('layer1'):
net = tf.nn.conv2d(x, filter=get_variable('w', [3, 3, 3, layer1_kernel_size]), strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b', [layer1_kernel_size]))
net = tf.nn.relu(net)
# lrn(input, depth_radius=5, bias=1, alpha=1, beta=0.5, name=None)
# LRN, 局部响应归一化)，主要是对ReLU激活函数的输出进行局部归一化操作而NB层是对样本进行归一化
# depth_radius ==> 对应公式上的n，bias => 对应公式上的k, alpha => 对应公式上的α, beta=>对应公式上的β
net = tf.nn.lrn(net)
print('layer1之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer2'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer2之后的shape:{}'.format(net.shape))
# conv3-128
with tf.variable_scope('layer3'):
net = tf.nn.conv2d(net, filter=get_variable('w', [3, 3, layer1_kernel_size, layer3_kernel_size]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b', [layer3_kernel_size]))
net = tf.nn.relu(net)
print('layer3之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer4'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer4之后的shape:{}'.format(net.shape))
# conv3-256 conv3-256
with tf.variable_scope('layer5'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, layer3_kernel_size, layer5_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [layer5_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, layer5_kernal_size_1, layer5_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [layer5_kernal_size_2]))
net = tf.nn.relu(net)
print('layer5之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer6'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer6之后的shape:{}'.format(net.shape))
# conv3-512 conv3-512
with tf.variable_scope('layer7'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, layer5_kernal_size_2, layer7_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [layer7_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, layer7_kernal_size_1, layer7_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [layer7_kernal_size_2]))
net = tf.nn.relu(net)
print('layer7之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer8'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer8之后的shape:{}'.format(net.shape))
# conv3-512 conv3-512
with tf.variable_scope('layer9'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, layer7_kernal_size_2, layer9_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [layer9_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, layer9_kernal_size_1, layer9_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [layer9_kernal_size_2]))
net = tf.nn.relu(net)
print('layer9之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer10'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer10之后的shape:{}'.format(net.shape))
# fc-4096
with tf.variable_scope('layer11'):
# 将四维的数据转换为两维的数据
shape = net.get_shape()
feature_number = shape[1] * shape[2] * shape[3]
net = tf.reshape(net, shape=[-1, feature_number])
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [feature_number, layer11_unit_size])),
get_variable('b', [layer11_unit_size]))
net = tf.nn.relu(net)
print('layer11之后的shape:{}'.format(net.shape))
# fc - 4096
with tf.variable_scope('layer12'):
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [layer11_unit_size, layer12_unit_size])),
get_variable('b', [layer12_unit_size]))
net = tf.nn.relu(net)
print('layer12之后的shape:{}'.format(net.shape))
# fc - 1000
with tf.variable_scope('layer13'):
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [layer12_unit_size, layer13_unit_size])),
get_variable('b', [layer13_unit_size]))
print('layer13之后的shape:{}'.format(net.shape))
return net
act = vgg_network(x, y)
# 第五步：构建模型的损失函数
# softmax_cross_entropy_with_logits: 计算softmax中的每个样本的交叉熵，logits指定预测值，labels指定实际值
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=act, labels=y))
# 第六步：构建梯度下降的优化方法（一般用Adam 动量GD）
# 使用Adam优化方式比较多动量GD
# learning_rate: 要注意，不要过大，过大可能不收敛，也不要过小，过小收敛速度比较慢
train = tf.train.AdadeltaOptimizer(learning_rate=learn_rate).minimize(cost)
# 第七步：计算模型正确率
# tf.argmax:对矩阵按行或列计算最大值对应的下标，和numpy中的一样
# tf.equal:是对比这两个矩阵或者向量的相等的元素，如果是相等的那就返回True，反正返回False，返回的值的矩阵维度和A是一样的
pred = tf.equal(tf.argmax(act, axis=1), tf.argmax(y, axis=1))
# 正确率（True转换为1，False转换为0）
acc = tf.reduce_mean(tf.cast(pred, tf.float32))
# 第八步：会话中执行阶段（模型的训练与迭代）
# 初始化
init = tf.global_variables_initializer()
with tf.Session() as sess:
# 进行数据初始化
sess.run(init)
# 模型保存、持久化
saver = tf.train.Saver()
epoch = 0
while True:
avg_cost = 0
# 计算出总的批次
total_batch = int(train_sample_number / batch_size)
# 迭代更新
for i in range(total_batch):
# 获取x和y
batch_xs = train_img[i * batch_size:(i+1) * batch_size]
batch_ys = train_label[i * batch_size:(i+1) * batch_size]
feeds = {x: batch_xs, y: batch_ys, learn_rate: learn_rate_func(epoch)}
# 模型训练
sess.run(train, feed_dict=feeds)
# 获取损失函数值
avg_cost += sess.run(cost, feed_dict=feeds)
# 重新计算平均损失(相当于计算每个样本的损失值)
avg_cost = avg_cost / total_batch
# DISPLAY 显示误差率和训练集的正确率以此测试集的正确率
if (epoch + 1) % display_step == 0:
print("批次: %03d 损失函数值: %.9f" % (epoch, avg_cost))
# 这里之所以使用train_img[:30]和train_label[:30]，是因为我使用train_img会出现内存不够的情况，直接就会退出
feeds = {x: train_img[:30], y: train_label[:30], learn_rate: learn_rate_func(epoch)}
train_acc = sess.run(acc, feed_dict=feeds)
print("训练集准确率: %.3f" % train_acc)
feeds = {x: test_img[:30], y: test_label[:30], learn_rate: learn_rate_func(epoch)}
test_acc = sess.run(acc, feed_dict=feeds)
print("测试准确率: %.3f" % test_acc)
if train_acc >= 0.99 and test_acc >= 0.98:
saver.save(sess, './data/vggnet/model_{}_{}'.format(train_acc, test_acc), global_step=epoch)
break
epoch += 1
# 模型可视化输出
writer = tf.summary.FileWriter('./data/vggnet/graph', tf.get_default_graph())
writer.close()
print("end....")

复制代码

东方耀 · 发表于 2019-1-12 18:41:02

# -*- coding: utf-8 -*-
__author__ = 'dongfangyao'
__date__ = '2019/1/12 下午5:34'
__product__ = 'PyCharm'
__filename__ = 'tf28'
import tensorflow as tf
import tflearn as tfl
from tflearn.datasets import oxflower17
from tflearn.datasets import mnist
from tflearn.datasets import cifar10
from sklearn.model_selection import train_test_split
# 第一步：导入数据 tflearn
# tfl.conv_2d()
X, Y = oxflower17.load_data(dirname='17flowers', one_hot=True)
print(X.shape)
print(Y.shape)
train_img, test_img, train_label, test_label = train_test_split(X, Y, test_size=0.2, random_state=0)
print(train_img.shape)
print(train_label.shape)
train_sample_number = train_img.shape[0]
# 第二步：设置超参并定义学习率调整策略
learn_rate_base = 0.1
batch_size = 32
display_step = 1
def learn_rate_func(epoch):
return max(0.001, learn_rate_base * (0.9 ** int(epoch / 10)))
# 第三步：开始构建模型设置输入数据的占位符
n_classes = train_label.shape[1]
x = tf.placeholder(tf.float32, shape=[None, 224, 224, 3], name='x')
y = tf.placeholder(tf.float32, shape=[None, n_classes], name='y')
learn_rate = tf.placeholder(tf.float32, name='learn_rate')
# 第四步：构建VGGNet网络（直接将网络结构翻译成为这个代码）
def get_variable(name, shape=None, dtype=tf.float32, initializer=tf.random_normal_initializer(mean=0, stddev=0.1)):
return tf.get_variable(name, shape, dtype, initializer)
def vgg_network(x, y):
layer1_kernel_size = 8
layer3_kernel_size = 16
layer5_kernel_size_1 = 32
layer5_kernel_size_2 = 32
layer7_kernel_size_1 = 64
layer7_kernel_size_2 = 64
layer9_kernel_size_1 = 64
layer9_kernel_size_2 = 64
layer11_unit_size = 120
layer12_unit_size = 120
layer13_unit_size = 17
# conv3-64 lrn
with tf.variable_scope('layer1'):
net = tf.nn.conv2d(x, filter=get_variable('w', shape=[3, 3, 3, layer1_kernel_size]), strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b', shape=[layer1_kernel_size]))
net = tf.nn.relu(net)
# lrn(input, depth_radius=5, bias=1, alpha=1, beta=0.5, name=None):
net = tf.nn.lrn(net)
print('layer1之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer2'):
net = tf.nn.max_pool(value=net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', data_format='NHWC')
print('layer2之后的shape:{}'.format(net.shape))
with tf.variable_scope('layer3'):
net = tf.nn.conv2d(net, filter=get_variable('w', shape=[3, 3, layer1_kernel_size, layer3_kernel_size]), strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b', shape=[layer3_kernel_size]))
net = tf.nn.relu(net)
print('layer3之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer4'):
net = tf.nn.max_pool(value=net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer4之后的shape:{}'.format(net.shape))
with tf.variable_scope('layer5'):
net = tf.nn.conv2d(net, filter=get_variable('w1', shape=[3, 3, layer3_kernel_size, layer5_kernel_size_1]), strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', shape=[layer5_kernel_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', shape=[3, 3, layer5_kernel_size_1, layer5_kernel_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', shape=[layer5_kernel_size_2]))
net = tf.nn.relu(net)
print('layer5之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer6'):
net = tf.nn.max_pool(value=net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer6之后的shape:{}'.format(net.shape))
with tf.variable_scope('layer7'):
net = tf.nn.conv2d(net, filter=get_variable('w1', shape=[3, 3, layer5_kernel_size_2, layer7_kernel_size_1]), strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', shape=[layer7_kernel_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', shape=[3, 3, layer7_kernel_size_1, layer7_kernel_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', shape=[layer7_kernel_size_2]))
net = tf.nn.relu(net)
print('layer7之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer8'):
net = tf.nn.max_pool(value=net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer8之后的shape:{}'.format(net.shape))
with tf.variable_scope('layer9'):
net = tf.nn.conv2d(net, filter=get_variable('w1', shape=[3, 3, layer7_kernel_size_2, layer9_kernel_size_1]), strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', shape=[layer9_kernel_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', shape=[3, 3, layer9_kernel_size_1, layer9_kernel_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', shape=[layer9_kernel_size_2]))
net = tf.nn.relu(net)
print('layer9之后的shape:{}'.format(net.shape))
# maxpool
with tf.variable_scope('layer10'):
net = tf.nn.max_pool(value=net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print('layer10之后的shape:{}'.format(net.shape))
# fc-4096
with tf.variable_scope('layer11'):
shape = net.get_shape()
feature_number = shape[1] * shape[2] * shape[3]
net = tf.reshape(net, shape=[-1, feature_number])
# FC
net = tf.add(tf.matmul(net, get_variable('w', shape=[feature_number, layer11_unit_size])), get_variable('b', shape=[layer11_unit_size]))
net = tf.nn.relu(net)
print('layer11之后的shape:{}'.format(net.shape))
with tf.variable_scope('layer12'):
# FC
net = tf.add(tf.matmul(net, get_variable('w', shape=[layer11_unit_size, layer12_unit_size])), get_variable('b', shape=[layer12_unit_size]))
net = tf.nn.relu(net)
print('layer12之后的shape:{}'.format(net.shape))
with tf.variable_scope('layer13'):
# FC
net = tf.add(tf.matmul(net, get_variable('w', shape=[layer12_unit_size, layer13_unit_size])),
get_variable('b', shape=[layer13_unit_size]))
print('layer13之后的shape:{}'.format(net.shape))
return net
act = vgg_network(x, y)
# 第五步：构建模型的损失函数
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=act, labels=y))
# 第六步：构建梯度下降的优化方法（一般用Adam 动量GD）
train = tf.train.AdadeltaOptimizer(learning_rate=learn_rate).minimize(cost)
# 第七步：计算模型正确率
pred = tf.equal(tf.argmax(act, axis=1), tf.argmax(y, axis=1))
acc = tf.reduce_mean(tf.cast(pred, dtype=tf.float32))
# 第八步：会话中执行阶段（模型的训练与迭代）
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
saver = tf.train.Saver()
epoch = 0
while True:
avg_cost = 0
total_batch = int(train_sample_number / batch_size)
for i in range(total_batch):
# batch_xs, batch_ys = mnist.train.next_batch(batch_size)
batch_xs = train_img[i*batch_size: (i+1)*batch_size]
batch_ys = train_label[i*batch_size: (i+1)*batch_size]
feeds = {x: batch_xs, y: batch_ys, learn_rate: learn_rate_func(epoch)}
sess.run(train, feed_dict=feeds)
avg_cost += sess.run(cost, feed_dict=feeds)
avg_cost = avg_cost / total_batch
if (epoch + 1) % display_step == 0:
print('批次: %03d 损失函数值：%.9f' % (epoch, avg_cost))
train_acc = sess.run(acc, feed_dict={x: train_img, y: train_label, learn_rate: learn_rate_func(epoch)})
print('训练集的准确率:%0.3f' % train_acc)
test_acc = sess.run(acc, feed_dict={x: test_img, y: test_label, learn_rate: learn_rate_func(epoch)})
print('测试集的准确率:%0.3f' % test_acc)
if train_acc > 0.99 and test_acc > 0.98:
saver.save(sess, './data/vggnet/model_{}_{}'.format(train_acc, test_acc), global_step=epoch)
break
epoch += 1
writer = tf.summary.FileWriter('./data/vggnet/graph', tf.get_default_graph())
writer.close()
print('end....')

复制代码

东方耀 · 发表于 2019-1-12 21:31:11

另外一种写法大家看看

# -- encoding:utf-8 --
from tflearn.datasets import oxflower17
import tensorflow as tf
# 读取数据
X, Y = oxflower17.load_data(dirname="17flowers", one_hot=True)
print(X.shape) # sample_number,224,224,3
print(Y.shape) # sample_number,17
# 相关的参数、超参数的设置
# 学习率，一般学习率设置的比较小
learn_rate = 0.1
# 每次迭代的训练样本数量
batch_size = 32
# 训练迭代次数(每个迭代次数中必须训练完一次所有的数据集)
train_epoch = 10000
# 样本数量
total_sample_number = X.shape[0]
# 模型构建
# 1. 设置数据输入的占位符
x = tf.placeholder(tf.float32, shape=[None, 224, 224, 3], name='x')
y = tf.placeholder(tf.float32, shape=[None, 17], name='y')
def get_variable(name, shape=None, dtype=tf.float32, initializer=tf.random_normal_initializer(mean=0, stddev=0.1)):
"""
返回一个对应的变量
:param name:
:param shape:
:param dtype:
:param initializer:
:return:
"""
return tf.get_variable(name, shape, dtype, initializer)
# 网络的构建
def vgg_network(x, y):
net1_kernel_size = 8
net3_kernel_size = 16
net5_kernal_size_1 = 32
net5_kernal_size_2 = 32
net7_kernal_size_1 = 64
net7_kernal_size_2 = 64
net9_kernal_size_1 = 64
net9_kernal_size_2 = 64
net11_unit_size = 120
net12_unit_size = 120
net13_unit_size = 17
# cov3-64 lrn
with tf.variable_scope('net1'):
net = tf.nn.conv2d(x, filter=get_variable('w', [3, 3, 3, net1_kernel_size]), strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b', [net1_kernel_size]))
net = tf.nn.relu(net)
net = tf.nn.lrn(net)
# maxpool
with tf.variable_scope('net2'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# conv3-128
with tf.variable_scope('net3'):
net = tf.nn.conv2d(net, filter=get_variable('w', [3, 3, net1_kernel_size, net3_kernel_size]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b', [net3_kernel_size]))
net = tf.nn.relu(net)
# maxpool
with tf.variable_scope('net4'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# conv3-256 conv3-256
with tf.variable_scope('net5'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, net3_kernel_size, net5_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [net5_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, net5_kernal_size_1, net5_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [net5_kernal_size_2]))
net = tf.nn.relu(net)
# maxpool
with tf.variable_scope('net6'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# conv3-512 conv3-512
with tf.variable_scope('net7'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, net5_kernal_size_2, net7_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [net7_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, net7_kernal_size_1, net7_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [net7_kernal_size_2]))
net = tf.nn.relu(net)
# maxpool
with tf.variable_scope('net8'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# conv3-512 conv3-512
with tf.variable_scope('net9'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, net7_kernal_size_2, net9_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [net9_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, net9_kernal_size_1, net9_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [net9_kernal_size_2]))
net = tf.nn.relu(net)
# maxpool
with tf.variable_scope('net10'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# fc
with tf.variable_scope('net11'):
# 将四维的数据转换为两维的数据
shape = net.get_shape()
feature_number = shape[1] * shape[2] * shape[3]
net = tf.reshape(net, shape=[-1, feature_number])
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [feature_number, net11_unit_size])),
get_variable('b', [net11_unit_size]))
net = tf.nn.relu(net)
# fc
with tf.variable_scope('net12'):
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [net11_unit_size, net12_unit_size])),
get_variable('b', [net12_unit_size]))
net = tf.nn.relu(net)
# fc
with tf.variable_scope('net13'):
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [net12_unit_size, net13_unit_size])),
get_variable('b', [net13_unit_size]))
return net
# 获取网络
act = vgg_network(x, y)
# 构建损失函数，优化器，准确率评估
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=act)) # 损失函数
# 优化器
# AdamOptimizer通过使用动量（参数的移动平均数）来改善传统梯度下降，促进超参数动态调整。
# 我们可以通过创建标签错误率的摘要标量来跟踪丢失和错误率
# 一个寻找全局最优点的优化算法，引入了二次方梯度校正。
# 相比于基础SGD算法，1.不容易陷于局部优点。2.速度更快
train = tf.train.GradientDescentOptimizer(learning_rate=learn_rate).minimize(cost)
# 正确率
# y [[1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0],[0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]] ->arg_max(y,1)->[0,1]
# net y [[1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0],[0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0] ]
correct_pred = tf.equal(tf.argmax(y, axis=1), tf.argmax(act, axis=1))
acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
# 训练
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
# 2.初始化所有的变量（一定在变量构建之后调用初始化操作）
tf.global_variables_initializer().run()
# 3.迭代训练
for epoch in range(train_epoch):
# 4.计算一次迭代batch执行的次数
total_batch = int(total_sample_number / batch_size) - 5
# 迭代更新
for step in range(total_batch):
# 获取当前批次的数据
train_x = X[step * batch_size:step * batch_size + batch_size]
train_y = Y[step * batch_size:step * batch_size + batch_size]
# 模型训练
sess.run(train, feed_dict={x: train_x, y: train_y})
# 每更新10次，输出一下
if step % 10 == 0:
loss, accuracy = sess.run([cost, acc], feed_dict={x: train_x, y: train_y})
print('迭代次数:{}, 步骤：{}, 训练集损失函数：{}, 训练集准确率：{}'.format(epoch, step, loss, accuracy))
# 展示一下
if epoch % 2 == 0:
# 获取测试集数据
test_x = X[step * batch_size:]
test_y = Y[step * batch_size:]
loss, accuracy = sess.run([cost, acc], feed_dict={x: test_x, y: test_y})
print('*' * 100)
print('步骤:', epoch)
print('测试集损失函数值：{}，测试集准确率：{}'.format(loss, accuracy))
loss, accuracy = sess.run([cost, acc], feed_dict={x: train_x, y: train_y})
print('训练集损失函数：{}, 训练集准确率：{}'.format(loss, accuracy))
print("End！！")

复制代码

wrt789 · 发表于 2020-5-30 11:25:34

学习学习厉害

		自动登录	找回密码
密码			立即注册

[课堂笔记] 51、经典CNN网络VGGNet比赛：17种花的图片识别_笔记