数据集
Cifar-10由60000张32*32的RGB彩色图片构成,一共包含有飞机、汽车、鸟、猫、鹿、狗、青蛙、马、船、卡车这10个类别。50000张训练,10000张测试。
比较知名的模型如AlexNet、NIN、ResNet等都曾在Cifar-10数据集上来评价自己的性能。
它还有一姐妹级的数据集Cifar-100,顾名思义就是包含100个类别,数据更加复杂。关于Cifar数据集的相关介绍以及数据的下载可见官网。
正是因为Cifar-10数据集不大、类别明确、获取方便、训练简单,同时模型的可参照性强,因此作为深度学习的初学者作为一个进阶的内容,再适合不过了
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
def unpickle(file):
import pickle
with open(file, 'rb') as fo:
dict = pickle.load(fo, encoding='ISO-8859-1')
return dict
# def unpickle(file):
# import pickle
# with open(file, 'rb') as fo:
# dict = pickle.load(fo, encoding='bytes') #读进来的是二进制
# 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
y_train = np.array(labels).reshape(-1) #0-9
X_train = np.array(X_train)
# reshape
X_train = X_train.reshape(-1,3072) #32*32*3=3072 50000个数据
# 目标值概率
one_hot = OneHotEncoder()
y_train =one_hot.fit_transform(y_train.reshape(-1,1)).toarray()
display(X_train.shape,y_train.shape)
#(50000, 3072)
#(50000, 10)
网络结构
conv1->relu1->pool1->conv2->relu2->pool2->conv3->relu3>pool3->fc1->dropout1->fc2->out. 三层网络 两个全连接层
所有的卷积核采用的都是3x3大小, pool的核也是3x3,步长为2x2.
中间加入bn(batch_ normalization)层.batch_size设置成100, 迭代500次, 之后减小学习率. 归一化
构建神经网络
X = tf.placeholder(dtype=tf.float32,shape = [None,3072])
y = tf.placeholder(dtype=tf.float32,shape = [None,10])
kp = tf.placeholder(dtype=tf.float32)
### !!! 给成常量了
def gen_v(shape):
return tf.Variable(tf.truncated_normal(shape = shape))
#生成改进版的正态分布 只要中间概率大的部分
def conv(input_,filter_,b):
conv = tf.nn.relu(tf.nn.conv2d(input_,filter_,strides=[1,1,1,1],padding='SAME') + b)
return tf.nn.max_pool(conv,[1,3,3,1],[1,2,2,1],'SAME')
def net_work(input_,kp):
# 形状改变,4维
input_ = tf.reshape(input_,shape = [-1,32,32,3])
# 第一层
filter1 = gen_v(shape = [3,3,3,64])
b1 = gen_v(shape = [64])
conv1 = conv(input_,filter1,b1)
# 归一化
conv1 = tf.layers.batch_normalization(conv1,training=True)
#layer 层 batch一批
# 第二层
filter2 = gen_v([3,3,64,128])
b2 = gen_v(shape = [128])
conv2 = conv(conv1,filter2,b2)
conv2 = tf.layers.batch_normalization(conv2,training=True)
# 第三层
filter3 = gen_v([3,3,128,256])
b3 = gen_v([256])
conv3 = conv(conv2,filter3,b3)
conv3 = tf.layers.batch_normalization(conv3,training=True)
# 第一层全连接层
dense = tf.reshape(conv3,shape = [-1,4*4*256]) #数据形状改变一下
fc1_w = gen_v(shape = [4*4*256,1024])
fc1_b = gen_v([1024])
fc1 = tf.matmul(dense,fc1_w) + fc1_b
fc1 = tf.layers.batch_normalization(fc1,training=True)
fc1 = tf.nn.relu(fc1)
# fc1.shape = [-1,1024]
# dropout
dp = tf.nn.dropout(fc1,keep_prob=kp) #保留比例
# 第二层全连接层
fc2_w = gen_v(shape = [1024,1024])
fc2_b = gen_v(shape = [1024])
fc2 = tf.nn.relu(tf.layers.batch_normalization(tf.matmul(dp,fc2_w) + fc2_b,training=True))
# 输出层
out_w = gen_v(shape = [1024,10])
out_b = gen_v(shape = [10])
out = tf.matmul(fc2,out_w) + out_b
return out
损失函数准确率
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 ==
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(y,axis = -1),tf.argmax(y_,axis = 1)),tf.float16))
#tf.cast([True,True,True,False,False,True])
#[1,1,1,0,1]
accuracy #准确率
#<tf.Tensor 'Mean_8:0' shape=() dtype=float16>
最后化
opt = tf.train.AdamOptimizer().minimize(loss)
opt
#<tf.Operation 'Adam_3' type=NoOp>
开启训练
epoches = 50000
saver = tf.train.Saver()
X_train.shape
y_train.shape
#(50000, 10)
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: #50000//128=390 一批是128个数据
index = 0
return batch_X,batch_y
test = unpickle('./cifar-10-batches-py/test_batch')
y_test = test['labels']
y_test = np.array(y_test)
X_test = test['data']
y_test = one_hot.transform(y_test.reshape(-1,1)).toarray()
y_test[:10]
#array([[0., 0., 0., 1., 0., 0., 0., 0., 0., 0.],
# [0., 0., 0., 0., 0., 0., 0., 0., 1., 0.],
# [0., 0., 0., 0., 0., 0., 0., 0., 1., 0.]])
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_ = sess.run([opt,loss],feed_dict = {X:batch_X,y:batch_y,kp:0.5})
print('----------------------------',loss_)
if i % 100 == 0:
score_test = sess.run(accuracy,feed_dict = {X:X_test,y:y_test,kp:1.0})
score_train = sess.run(accuracy,feed_dict = {X:batch_X,y:batch_y,kp:1.0})
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))