CNN-cr-10-优化-创新互联

import numpy as np

# 序列化和反序列化

import pickle

from sklearn.preprocessing import OneHotEncoder

import warnings

warnings.filterwarnings('ignore')

import tensorflow as tf

数据加载（使用pickle）

def unpickle(file):

import pickle

with open(file, 'rb') as fo:

dict = pickle.load(fo, encoding='ISO-8859-1')

return dict

labels = []

X_train = []

for i in range(1,6):

data = unpickle('./cifar-10-batches-py/data_batch_%d'%(i))

labels.append(data['labels'])

X_train.append(data['data'])

# 将list类型转换为ndarray

X_train = np.array(X_train)

y_train = np.array(labels).reshape(-1)

# reshape

X_train = X_train.reshape(-1,3072)

# 目标值概率

one_hot = OneHotEncoder()

y_train =one_hot.fit_transform(y_train.reshape(-1,1)).toarray()

# 测试数据加载

test = unpickle('./cifar-10-batches-py/test_batch')

X_test = test['data']

y_test = one_hot.transform(np.array(test['labels']).reshape(-1,1)).toarray()

# 从总数据中获取一批数据

index = 0

def next_batch(X,y):

global index

batch_X = X[index*128:(index+1)*128]

batch_y = y[index*128:(index+1)*128]

index+=1

if index == 390:

index = 0

return batch_X,batch_y

构建神经网络

1.生成对应卷积核

2.tf.nn.conv2d进行卷积运算

3.归一化操作tf.layers.batch_normalization

4.激活函数（relu）

5.池化操作

X = tf.placeholder(dtype=tf.float32,shape = [None,3072])

y = tf.placeholder(dtype=tf.float32,shape = [None,10])

kp = tf.placeholder(dtype=tf.float32)

def gen_v(shape,std = 5e-2):

return tf.Variable(tf.truncated_normal(shape = shape,stddev=std))

def conv(input_,filter_,b):

conv = tf.nn.conv2d(input_,filter_,strides=[1,1,1,1],padding='SAME') + b

conv = tf.layers.batch_normalization(conv,training=True)

conv = tf.nn.relu(conv)

return tf.nn.max_pool(conv,[1,3,3,1],[1,2,2,1],'SAME')

def net_work(X,kp):

#     形状改变，4维

input_ = tf.reshape(X,shape = [-1,32,32,3])

#     第一层

filter1 = gen_v(shape = [3,3,3,64])

b1 = gen_v(shape = [64])

pool1 = conv(input_,filter1,b1)

#     第二层

filter2 = gen_v([3,3,64,128])

b2 = gen_v(shape = [128])

pool2 = conv(pool1,filter2,b2)

#     第三层

filter3 = gen_v([3,3,128,256])

b3 = gen_v([256])

pool3 = conv(pool2,filter3,b3)

#     第一层全连接层

dense = tf.reshape(pool3,shape = [-1,4*4*256])

fc1_w = gen_v(shape = [4*4*256,1024])

fc1_b = gen_v([1024])

bn_fc_1 = tf.layers.batch_normalization(tf.matmul(dense,fc1_w) + fc1_b,training=True)

function(){ //智汇返佣 http://www.kaifx.cn/broker/thinkmarkets.html

relu_fc_1 = tf.nn.relu(bn_fc_1)

#     fc1.shape = [-1,1024]

#     dropout

dp = tf.nn.dropout(relu_fc_1,keep_prob=kp)

#     fc2 输出层

out_w = gen_v(shape = [1024,10])

out_b = gen_v(shape = [10])

out = tf.matmul(dp,out_w) + out_b

return out

损失函数准确率&最优化

out = net_work(X,kp)

loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y,logits=out))

# 准确率

y_ = tf.nn.softmax(out)

# equal 相当于==

equal = tf.equal(tf.argmax(y,axis = -1),tf.argmax(y_,axis = 1))

accuracy = tf.reduce_mean(tf.cast(equal,tf.float32))

opt = tf.train.AdamOptimizer(0.01).minimize(loss)

opt

开启训练

saver = tf.train.Saver()

epoches = 100

with tf.Session() as sess:

sess.run(tf.global_variables_initializer())

for i in range(epoches):

batch_X,batch_y = next_batch(X_train,y_train)

opt_,loss_ ,score_train= sess.run([opt,loss,accuracy],feed_dict = {X:batch_X,y:batch_y,kp:0.5})

print('iter count:%d。mini_batch loss:%0.4f。训练数据上的准确率：%0.4f。测试数据上准确率：%0.4f'%

(i+1,loss_,score_train,score_test))

if score_train > 0.6:

saver.save(sess,'./model/estimator',i+1)

saver.save(sess,'./model/estimator',i+1)

score_test = sess.run(accuracy,feed_dict = {X:X_test,y:y_test,kp:1.0})

print('测试数据上的准确率：',score_test)

iter count:1。mini_batch loss:3.1455。训练数据上的准确率：0.0938。测试数据上准确率：0.2853

iter count:2。mini_batch loss:3.9139。训练数据上的准确率：0.2891。测试数据上准确率：0.2853

iter count:3。mini_batch loss:5.1961。训练数据上的准确率：0.1562。测试数据上准确率：0.2853

iter count:4。mini_batch loss:3.9102。训练数据上的准确率：0.2344。测试数据上准确率：0.2853

iter count:5。mini_batch loss:4.1278。训练数据上的准确率：0.1719。测试数据上准确率：0.2853

.....

iter count:97。mini_batch loss:1.5752。训练数据上的准确率：0.4844。测试数据上准确率：0.2853

iter count:98。mini_batch loss:1.8480。训练数据上的准确率：0.3906。测试数据上准确率：0.2853

iter count:99。mini_batch loss:1.5662。训练数据上的准确率：0.5391。测试数据上准确率：0.2853

iter count:100。mini_batch loss:1.7489。训练数据上的准确率：0.4141。测试数据上准确率：0.2853

测试数据上的准确率： 0.4711

epoches = 1100

with tf.Session() as sess:

saver.restore(sess,'./model/estimator-100')

for i in range(100,epoches):

batch_X,batch_y = next_batch(X_train,y_train)

opt_,loss_ ,score_train= sess.run([opt,loss,accuracy],feed_dict = {X:batch_X,y:batch_y,kp:0.5})

print('iter count:%d。mini_batch loss:%0.4f。训练数据上的准确率：%0.4f。测试数据上准确率：%0.4f'%

(i+1,loss_,score_train,score_test))

if score_train > 0.6:

saver.save(sess,'./model/estimator',i+1)

saver.save(sess,'./model/estimator',i+1)

if (i%100 == 0) and (i != 100):

score_test = sess.run(accuracy,feed_dict = {X:X_test,y:y_test,kp:1.0})

print('----------------测试数据上的准确率：---------------',score_test)

iter count:101。mini_batch loss:1.4157。训练数据上的准确率：0.5234。测试数据上准确率：0.4711

iter count:102。mini_batch loss:1.6045。训练数据上的准确率：0.4375。测试数据上准确率：0.4711

....

iter count:748。mini_batch loss:0.6842。训练数据上的准确率：0.7734。测试数据上准确率：0.4711

iter count:749。mini_batch loss:0.6560。训练数据上的准确率：0.8203。测试数据上准确率：0.4711

iter count:750。mini_batch loss:0.7151。训练数据上的准确率：0.7578。测试数据上准确率：0.4711

iter count:751。mini_batch loss:0.8092。训练数据上的准确率：0.7344。测试数据上准确率：0.4711

iter count:752。mini_batch loss:0.7394。训练数据上的准确率：0.7422。测试数据上准确率：0.4711

iter count:753。mini_batch loss:0.8732。训练数据上的准确率：0.7188。测试数据上准确率：0.4711

iter count:754。mini_batch loss:0.8762。训练数据上的准确率：0.6953。测试数据上准确率：0.4711

准确率高达80%，博主亲测，以上准确率数据部分展示，大家可以多训练几次。哈哈哈~~~

另外有需要云服务器可以了解下创新互联cdcxhl.cn，海内外云服务器15元起步，三天无理由+7*72小时售后在线，公司持有idc许可证，提供“云服务器、裸金属服务器、高防服务器、香港服务器、美国服务器、虚拟主机、免备案服务器”等云主机租用服务以及企业上云的综合解决方案，具有“安全稳定、简单易用、服务可用性高、性价比高”等特点与优势，专为企业上云打造定制，能够满足用户丰富、多元化的应用场景需求。