这次继续在colab中实现TensorFlow学习的第二个任务:对cifar-10数据集进行图像分类任务的学习。本文采用了VGG-16网路结构,去掉了一层max pooling层,最终测试集上可以达到0.92左右的结果。
cifar-10数据集介绍
CIFAR-10数据集包含10个类别的RGB彩色图片。图片尺寸为32×32,这十个类别包括:飞机、汽车、鸟、猫、鹿、狗、蛙、马、船、卡车。一共有50000张训练图片和10000张测试图片。
首先来实现数据集的处理。
下载cifar10 Python版的数据文件并解压,数据文件共有6个data_batch_1 ~ data_batch_5,以及test_batch,每个文件含有10000张图片。
读取数据集得到一个很大的Dictionary,有dict_keys([‘batch_label’, ‘labels’, ‘data’, ‘filenames’])五个关键字,我们关心其中的两项。labels-标签,data-数据。label是100001的数字,每一个标签为0-9的数字。data是100003072的数组,每个图片是32 * 32 * 3的大小。
将data数据转换从图片数组的时候要注意,data中3072个记录,前1024个条目包含红色通道值,下一个1024个绿色,最后1024个蓝色。图像以行优先顺序存储,以便数组的前32个条目是图像第一行的红色通道值。所以进行reshape时要先reshape成(-1,3,32,32)的向量,然后transpose(0,2,3,1)将其转化为我们需要(-1,32,32,3)的向量。
加载数据集并打印其中三张图片
import pickle
import matplotlib.pyplot as plt
def load_file(filename):
with open(filename, 'rb') as fo:
data = pickle.load(fo, encoding='latin1')
return data
data = load_file('./cifar-10-batches-py/data_batch_1')
print(data.keys())
image_data = data['data']
labels = data['labels']
label_count = len(labels)
print(image_data.shape)
picture_data = image_data.reshape(-1,3,32,32)
picture_data = picture_data.transpose(0,2,3,1)
print(picture_data.shape)
plt.subplot(131)
plt.imshow(picture_data[1])
plt.subplot(132)
plt.imshow(picture_data[2])
plt.subplot(133)
plt.imshow(picture_data[3])
![](https://i-blog.csdnimg.cn/blog_migrate/13f39d19bc0d3cefc8cc1648edf4b77b.png)
导入包 预先设置
实验环境google colab,加速方式GPU加速,全部实现分3个step。
step-1: 导入需要的包,设定参数
import tensorflow as tf
import numpy as np
import time
import random
import pickle
import math
import datetime
from keras.preprocessing.image import ImageDataGenerator
#预先定义的变量
class_num = 10
image_size = 32
img_channels = 3
iterations = 200
batch_size = 250
weight_decay = 0.0003
dropout_rate = 0.5
momentum_rate = 0.9
data_dir = './cifar-10-batches-py/'
log_save_path = './vgg_16_logs'
model_save_path = './model/'
数据处理
这一部分将实现加载数据并处理成可输入神经网络进行训练的格式。
- data数据转化为(-1,32,32,3)的格式 ,并进行归一化处理。
- labels数据转化为(-1,10)的格式。
step-2: 读取数据,得到train_data, train_labels, test_data, test_labels
#========数据处理=========
#读文件
def unpickle(file):
with open(file, 'rb') as fo:
dict = pickle.load(fo, encoding='latin1')
return dict
#从读入的文件中获取图片数据(data)和标签信息(labels)
def load_data_one(file):
batch = unpickle(file)
data = batch['data']
labels = batch['labels']
print("Loading %s : img num %d." % (file, len(data)))
return data, labels
#将从文件中获取的信息进行处理,得到可以输入到神经网络中的数据。
def load_data(files, data_dir, label_count):
global image_size, img_channels
data, labels = load_data_one(data_dir + files[0])
for f in files[1:]:
data_n, labels_n = load_data_one(data_dir + '/' + f)
data = np.append(data, data_n, axis=0)
labels = np.append(labels, labels_n, axis=0)
#标签labels从0-9的数字转化为float类型(-1,10)的标签矩阵
labels = np.array([[float(i == label) for i in range(label_count)] for label in labels])
#将图片数据从(-1,3072)转化为(-1,3,32,32)
data = data.reshape([-1, img_channels, image_size, image_size])
#将(-1,3,32,32)转化为(-1,32,32,3)的图片标准输入
data = data.transpose([0, 2, 3, 1])
#data数据归一化
data = data.astype('float32')
data[:, :, :, 0] = (data[:, :, :, 0] - np.mean(data[:, :, :, 0])) / np.std(data[:, :, :, 0])
data[:, :, :, 1] = (data[:, :, :, 1] - np.mean(data[:, :, :, 1])) / np.std(data[:, :, :, 1])
data[:, :, :, 2] = (data[:, :, :, 2] - np.mean(data[:, :, :, 2])) / np.std(data[:, :, :, 2])
return data, labels
def prepare_data():
print("======Loading data======")
image_dim = image_size * image_size * img_channels
meta = unpickle(data_dir + 'batches.meta')
print(meta)
label_names = meta['label_names']
#依次读取data_batch_1-5的内容
train_files = ['data_batch_%d' % d for d in range(1, 6)]
train_data, train_labels = load_data(train_files, data_dir, class_num)
test_data, test_labels = load_data(['test_batch'], data_dir, class_num)
print("Train data:", np.shape(train_data), np.shape(train_labels))
print("Test data :", np.shape(test_data), np.shape(test_labels))
print("======Load finished======")
#重新打乱训练集的顺序
indices = np.random.permutation(len(train_data))
train_data = train_data[indices]
train_labels = train_labels[indices]
print("======数据准备结束======")
return train_data, train_labels, test_data, test_labels
train_x, train_y, test_x, test_y = prepare_data()
![](https://i-blog.csdnimg.cn/blog_migrate/82824878ddb101efd30908635e7b7577.png)
VGG-16网络
VGG-16网络结构如下:
VGG-16标准输入是224 * 224,经过5次max pooling和13次卷积后变成7 * 7 * 512,cifar10数据集的图片大小只有32 * 32,所以少进行一次max pooling,最后得到2 * 2 *512的结果传入全连接层。
卷积网络增加了batch normalization。参数初始化采用了he_normal()的方法。
梯度下降采用了adam学习率自适应的优化算法。alpha参数初始值为0.001,训练集精度达到0.97以上后alpha变为0.0001,精度达到0.996以上alpha变为0.00001 。
为了防止过拟合,添加了dropout和L2正则化。
设置迭代次数为40,epcho达到25之后精度基本不再提升。最终train_loss: 0.0319, train_acc: 0.9964, test_loss: 0.5629, test_acc: 0.8680
step-3: 设计网络,开始训练
#=========网络设计=========
#初始化权重,采用正则化随机初始,加入少量的噪声来打破对称性以及避免0梯度
def weight_variable(name, sp):
initial = tf.initializers.he_normal()
#initial = tf.truncated_normal(shape, mean=0.0, stddev=0.05, dtype=tf.float32)
return tf.get_variable(name = name, shape = sp, initializer = initial)
#return tf.Variable(initial)
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def batch_norm(input):
return tf.contrib.layers.batch_norm(input, decay=0.9, center=True, scale=True, epsilon=1e-3,
is_training=train_flag, updates_collections=None)
def conv(name,x,w,b):
#去掉BN
#return tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x,w,strides=[1,1,1,1],padding='SAME'),b),name=name)
return tf.nn.relu(batch_norm(tf.nn.bias_add(tf.nn.conv2d(x,w,strides=[1,1,1,1],padding='SAME'),b)),name=name)
def max_pool(name,x,k):
return tf.nn.max_pool(x,ksize=[1,k,k,1],strides=[1,k,k,1],padding='SAME',name=name)
def fc(name,x,w,b):
return tf.nn.relu(batch_norm(tf.matmul(x,w)+b),name=name)
#VGG-16网络,因为输入尺寸小,去掉最后两个个max pooling层
def vgg_net(_X,_weights,_biases,keep_prob):
conv1_1=conv('conv1_1',_X,_weights['wc1_1'],_biases['bc1_1'])
conv1_2=conv('conv1_2',conv1_1,_weights['wc1_2'],_biases['bc1_2'])
pool1=max_pool('pool1',conv1_2,k=2)
conv2_1=conv('conv2_1',pool1,_weights['wc2_1'],_biases['bc2_1'])
conv2_2=conv('conv2_2',conv2_1,_weights['wc2_2'],_biases['bc2_2'])
pool2=max_pool('pool2',conv2_2,k=2)
conv3_1=conv('conv3_1',pool2,_weights['wc3_1'],_biases['bc3_1'])
conv3_2=conv('conv3_2',conv3_1,_weights['wc3_2'],_biases['bc3_2'])
conv3_3=conv('conv3_3',conv3_2,_weights['wc3_3'],_biases['bc3_3'])
pool3=max_pool('pool3',conv3_3,k=2)
conv4_1=conv('conv4_1',pool3,_weights['wc4_1'],_biases['bc4_1'])
conv4_2=conv('conv4_2',conv4_1,_weights['wc4_2'],_biases['bc4_2'])
conv4_3=conv('conv4_3',conv4_2,_weights['wc4_3'],_biases['bc4_3'])
pool4=max_pool('pool4',conv4_3,k=2)
conv5_1=conv('conv5_1',pool4,_weights['wc5_1'],_biases['bc5_1'])
conv5_2=conv('conv5_2',conv5_1,_weights['wc5_2'],_biases['bc5_2'])
conv5_3=conv('conv5_3',conv5_2,_weights['wc5_3'],_biases['bc5_3'])
pool5=max_pool('pool5',conv5_3,k=1)
_shape=pool5.get_shape()
flatten=_shape[1].value*_shape[2].value*_shape[3].value
pool5=tf.reshape(pool5,shape=[-1,flatten])
fc1=fc('fc1',pool5,_weights['fc1'],_biases['fb1'])
fc1=tf.nn.dropout(fc1,keep_prob)
fc2=fc('fc2',fc1,_weights['fc2'],_biases['fb2'])
fc2=tf.nn.dropout(fc2,keep_prob)
output=fc('fc3',fc2,_weights['fc3'],_biases['fb3'])
#fc3=tf.nn.dropout(fc3,keep_prob)
return output
weights={
'wc1_1' : weight_variable('wc1_1', [3,3,3,64]),
'wc1_2' : weight_variable('wc1_2', [3,3,64,64]),
'wc2_1' : weight_variable('wc2_1', [3,3,64,128]),
'wc2_2' : weight_variable('wc2_2', [3,3,128,128]),
'wc3_1' : weight_variable('wc3_1', [3,3,128,256]),
'wc3_2' : weight_variable('wc3_2', [3,3,256,256]),
'wc3_3' : weight_variable('wc3_3', [3,3,256,256]),
'wc4_1' : weight_variable('wc4_1', [3,3,256,512]),
'wc4_2' : weight_variable('wc4_2', [3,3,512,512]),
'wc4_3' : weight_variable('wc4_3', [3,3,512,512]),
'wc5_1' : weight_variable('wc5_1', [3,3,512,512]),
'wc5_2' : weight_variable('wc5_2', [3,3,512,512]),
'wc5_3' : weight_variable('wc5_3', [3,3,512,512]),
'fc1' : weight_variable('fc1', [2*2*512,4096]),
'fc2' : weight_variable('fc2', [4096,4096]),
'fc3' : weight_variable('fc3', [4096,10])
}
biases={
'bc1_1' : bias_variable([64]),
'bc1_2' : bias_variable([64]),
'bc2_1' : bias_variable([128]),
'bc2_2' : bias_variable([128]),
'bc3_1' : bias_variable([256]),
'bc3_2' : bias_variable([256]),
'bc3_3' : bias_variable([256]),
'bc4_1' : bias_variable([512]),
'bc4_2' : bias_variable([512]),
'bc4_3' : bias_variable([512]),
'bc5_1' : bias_variable([512]),
'bc5_2' : bias_variable([512]),
'bc5_3' : bias_variable([512]),
'fb1' : bias_variable([4096]),
'fb2' : bias_variable([4096]),
'fb3' : bias_variable([10]),
}
#数据增强
def _random_crop(batch, crop_shape, padding=None):
oshape = np.shape(batch[0])
if padding:
oshape = (oshape[0] + 2*padding, oshape[1] + 2*padding)
new_batch = []
npad = ((padding, padding), (padding, padding), (0, 0))
for i in range(len(batch)):
new_batch.append(batch[i])
if padding:
new_batch[i] = np.lib.pad(batch[i], pad_width=npad,
mode='constant', constant_values=0)
nh = random.randint(0, oshape[0] - crop_shape[0])
nw = random.randint(0, oshape[1] - crop_shape[1])
new_batch[i] = new_batch[i][nh:nh + crop_shape[0],
nw:nw + crop_shape[1]]
return new_batch
def _random_flip_leftright(batch):
for i in range(len(batch)):
if bool(random.getrandbits(1)):
batch[i] = np.fliplr(batch[i])
return batch
def data_augmentation(batch):
batch = _random_flip_leftright(batch)
batch = _random_crop(batch, [32, 32], 4)
return batch
#测试集跑模型,记录精度和损失
def run_test(sess):
acc = 0.0
loss = 0.0
pre_index = 0
#数据量大,采用类似minibatch的方法,分批次测试,在将其汇总
add = 1000
for it in range(10):
batch_x = test_x[pre_index : pre_index+add]
batch_y = test_y[pre_index : pre_index+add]
pre_index = pre_index + add
loss_, acc_ = sess.run([cross_entropy, accuracy],feed_dict={x: batch_x, y_: batch_y, keep_prob: 1.0, train_flag : False})
loss += loss_ / 10.0
acc += acc_ / 10.0
summary = tf.Summary(value=[tf.Summary.Value(tag="test_loss", simple_value=loss),
tf.Summary.Value(tag="test_accuracy", simple_value=acc)])
return acc, loss, summary
#训练训练集完所有50000张图片的过程
#采用mini-batch梯度的方法,batch_size=250, iterations=200
#一个epcho的精度和损失是200个iterations精度和损失的均值
def train_epcho(sess, epcho):
pre_index = 0
acc = 0.0
loss = 0.0
for it in range(1, iterations+1):
batch_x = train_x[pre_index : pre_index+batch_size]
batch_y = train_y[pre_index : pre_index+batch_size]
#使用数据增强策略
batch_x = data_augmentation(batch_x)
_, batch_loss = sess.run([train_step, cross_entropy],
feed_dict={x: batch_x, y_: batch_y, keep_prob: dropout_rate, train_flag : True, learning_rate: alpha})
batch_acc = accuracy.eval(feed_dict={x: batch_x, y_: batch_y, keep_prob: 1.0, train_flag : True})
loss += batch_loss
acc += batch_acc
pre_index += batch_size
if it%20 == 0:
print("epcho: %d, iterations: %d, loss: %.4f, acc: %.4f"%(epcho, it, batch_loss, batch_acc))
loss /= iterations
acc /= iterations
summary = tf.Summary(value=[tf.Summary.Value(tag="train_loss", simple_value=loss),
tf.Summary.Value(tag="train_accuracy", simple_value=acc)])
return acc, loss, summary
if __name__ == '__main__':
x = tf.placeholder(tf.float32,[None, image_size, image_size, 3])
y_ = tf.placeholder(tf.float32, [None, class_num])
keep_prob = tf.placeholder(tf.float32)
train_flag = tf.placeholder(tf.bool)
learning_rate = tf.placeholder(tf.float32)
output = vgg_net(x, weights, biases, keep_prob)
correct_prediction = tf.equal(tf.argmax(output, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
with tf.name_scope('loss'):
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_, logits=output))
l2 = tf.add_n([tf.nn.l2_loss(var) for var in tf.trainable_variables()])
with tf.name_scope('train_op'):
train_step = tf.train.AdamOptimizer(learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-8).minimize(cross_entropy + l2 * weight_decay)
#train_step = tf.train.MomentumOptimizer(learning_rate, momentum_rate, use_nesterov=True).minimize(cross_entropy + l2 * weight_decay)
saver = tf.train.Saver()
TIMESTAMP = "{0:%Y-%m-%dT%H-%M-%S/}".format(datetime.now())
log_save_path_now = log_save_path+'/'+TIMESTAMP
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(log_save_path_now,sess.graph)
alpha = 0.001
for ep in range(1, 41):
start_time = time.time()
print("========================epcho: %d============================="%ep)
train_acc, train_loss, train_summary = train_epcho(sess, ep)
val_acc, val_loss, test_summary = run_test(sess)
print("epcho: %d, cost_time: %ds, train_loss: %.4f, train_acc: %.4f, test_loss: %.4f, test_acc: %.4f"
% (ep, int(time.time()-start_time), train_loss, train_acc, val_loss, val_acc))
if train_acc>0.97 and train_acc<0.993:
alpha = 0.0001
elif train_acc>=0.993 and train_acc<0.996:
alpha = 0.00001
elif train_acc>=0.996 :
alpha = 0.000001
summary_writer.add_summary(train_summary, ep)
summary_writer.add_summary(test_summary, ep)
summary_writer.flush()
print("finish!")
save_path = saver.save(sess, model_save_path)
print("Model saved in file: %s" % save_path)
训练结果
未采用数据增强策略之前,红色和灰色是添加了BN和dropout的结果,蓝色曲线是去掉BN和dropout的训练曲线,可以看到去掉BN收敛速度变慢了,去掉dropout和BN之后,过拟合的程度更高,在测试集上的精度下降了5个百分点。
采用了数据增强策略后,train_loss收敛速度变慢,但最终测试集上的精度有了较大的提升。
问题总结
第一次完整实现较大的任务。过程中总有意想不到的问题发生。这里做个总结,希望大家可以不要犯和我一样的错误吧。
-
W参数初始化。起初我采取了上次MNIST任务随机正则化初始的策略,stddev=0.5 ,事实上这是一个很糟糕的初始化策略,初始的W设置的过大,开始学习之后参数值基本不会发生变化。后边采取了tf.keras.initializers.he_normal()的初始化方法,这个方法设置如下stddev = sqrt(2 / fan_in),fan_in是输入张量的个数。
-
tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_, logits=output)
这个函数的实现两步任务:第一步是先对网络最后一层的输出做一个softmax,第二步是softmax的输出向量[Y1,Y2,Y3…]和样本的实际标签做一个交叉熵。
第一次在实现的时候我在VGG网络中定义了softmax部分,输出output就是softmax输出的结果,然后又采用了tf.nn.softmax_cross_entropy_with_logits_v2的方式来计算交叉熵,结果精度到达0.65左右就不在提升了。关于tf.nn.softmax_cross_entropy_with_logits_v2实现的内参考这个文章。 -
数据增强。采用了随机翻转和随机平移两种策略来进行数据增强,采用数据增强后train_loss收敛速度变慢,但测试集上的效果有了明显的提升。
-
学习率的问题。采用adam算法后train_loss很快就可以收敛到一个不错的值,此时精度变化将很小,将学习率调整到一个较小的值之后,梯度下降可以朝着一个更加精确的方向前进,test_loss指标也有了很大的提升,这点在tensorboard的训练图像中表现的很明显。adam算法可以参考这篇博客。
这次采用VGG网络实现cifar10分类任务测试集上最终的精度大约在0.92,epcho=40,训练时长为50分钟左右。相较于别人采用VGG实现的精度还有一点差距,最终的train_loss还在提升,精度也略有提升,后续可以继续增加训练次数来提升效果(对比https://github.com/kuangliu/pytorch-cifar),本文的许多方案和代码来自于https://blog.csdn.net/GOGO_YAO/article/details/80348200。