python tensorflow学习（四）加大难度！AlexNet模型实现
加大难度！AlexNet模型实现
Alex 网络[8]是现代意义上的深度卷积神经网络的起源，是在 2012 年被推出
的一个经典网络模型，取得了同年 ImageNet 比赛的最优成绩。相比于上一章的 LeNet-5，AlexNet 的层次显著加深，参数规模也显著变大。
大纲：
- 猫狗大战数据集处理
- 数据集的加工
- 转换为TFRecord格式
- Alex网络模型实现
- 网络模型架构
- 模型的训练存储及结果可视化
猫狗大战数据集处理
猫狗大战的数据集来源于Kaggle上的一个竞赛：Dogs vs. Cats。该数据集包括12500张猫的图片以及12500张狗的图片，是一个二分类问题。官方提供了免费下载：下载地址点这里.如果不想注册账号，还有微软的版本：下载地址二。
不同于MNIIST数据集，这个数据集均来自于真实的照片，tensorflow中也没有封装好的函数来读取该数据，所以需要对该数据进行预处理。
数据集的加工
首先，AlexNet模型的输入图像大小为227×227×3，所以需要把数据集的分辨率调整为该大小，这里使用opencv进行处理
# 把图片大小转换为227x227x3
def rebuild(dir):
for root, dirs, files in os.walk(dir):
for file in files:
try:
filepath = os.path.join(root, file)
image = cv2.imread(filepath)
dim = (227, 227)
resized = cv2.resize(image, dim)
print(file)
path = "D:/cat_and_dog/Cat/"+file
cv2.imwrite(path, resized)
except IOError:
print(filepath)
os.remove(filepath)
cv2.waitKey(0) # 退出
rebuild("D:/PetImages/Cat")
对于损坏的数据，这里使用os.remove（）直接删除。
转换为TFRecord格式
在第二章介绍过了TFRecord文件的创建和读取，传送门->python tensorflow学习（二） tensorflow数据的生成与读取 这里不再介绍，直接贴源代码：
获取数据集和标签：
# 设置标签
def get_file(file_dir):
images = []
temp = []
for root, sub_folders, files in os.walk(file_dir):
for name in files:
images.append(os.path.join(root, name))
for name in sub_folders:
temp.append(os.path.join(root, name))
labels = []
for one_folder in temp:
n_img = len(os.listdir(one_folder))
letter = one_folder.split("\\")[-1]
if letter=='Cat':
labels = np.append(labels, n_img*[0])
else:
labels = np.append(labels, n_img*[1])
temp = np.array([images, labels])
temp = temp.transpose()
np.random.shuffle(temp)
image_list = list(temp[:, 0])
label_list = list(temp[:, 1])
label_list = [int(float(i)) for i in label_list]
return image_list, label_list
imagelist, labellist = get_file("D:/cat_and_dog")
转换为TFRecord文件：
# 生成TFRecord文件
def int64_feature(value):
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))
def bytes_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
def ToTFRecord(image_list, label_list, save_dir, name):
filename = os.path.join(save_dir, name+'.tfrecords')
n_samples = len(imagelist)
writer = tf.python_io.TFRecordWriter(filename)
print("transform start...")
for i in np.arange(0, n_samples):
try:
image = cv2.imread(image_list[i])
image_raw = image.tostring()
label = [int(label_list[i])]
example = tf.train.Example(features=tf.train.Features(feature={
'label':int64_feature(label),
'image_raw':bytes_feature(image_raw)
}))
writer.write(example.SerializeToString())
except IOError as e:
print('could not read:', image_list[i])
writer.close()
print('transform done!')
其中save_dir和name分别是存储路径和文件名。
至此TFRecord文件已经生成，还需要一个能读取该文件的函数来获取数据：
# 读取数据
def read_and_decode(tfrecord_file, batch_size):
filename_queue = tf.train.string_input_producer([tfrecord_file])
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
img_features = tf.parse_single_example(
serialized_example,
features={
'label': tf.FixedLenFeature([], tf.int64),
'image_raw': tf.FixedLenFeature([], tf.string),
})
image = tf.decode_raw(img_features['image_raw'], tf.uint8)
image = tf.reshape(image, [227, 227, 3])
label = tf.cast(img_features['label'], tf.int32)
image_batch, label_batch = tf.train.shuffle_batch([image, label],
batch_size=batch_size,
min_after_dequeue=100,
num_threads=64,
capacity=200
)
return image_batch, tf.reshape(label_batch, [batch_size])
image_batch, label_batch = read_and_decode('cat_and_dog.tfrecords', 25)
该函数每次读取指定数目的数据集以便训练时提供。
网络模型架构
在实现每一层的架构之前，先对参数进行集中管理，这是一个很好的习惯：
# 集中管理参数
learning_rate = 1e-4 # 学习速率
training_iters = 200 # 迭代次数
batch_size = 50 # 每批的大小
n_classes = 2 # 种类
n_fc1 = 4096
n_fc2 = 2048
# 构建模型
x = tf.placeholder(tf.float32, [None, 227, 227, 3])
y = tf.placeholder(tf.int32, [None, n_classes])
W_conv = {
'conv1': tf.Variable(tf.truncated_normal([11, 11, 3, 96], stddev=0.0001)),
'conv2': tf.Variable(tf.truncated_normal([5, 5, 96, 256], stddev=0.01)),
'conv3': tf.Variable(tf.truncated_normal([3, 3, 256, 384], stddev=0.01)),
'conv4': tf.Variable(tf.truncated_normal([3, 3, 384, 384], stddev=0.01)),
'conv5': tf.Variable(tf.truncated_normal([3, 3, 384, 256], stddev=0.01)),
'fc1': tf.Variable(tf.truncated_normal([6*6*256, n_fc1], stddev=0.1)),
'fc2': tf.Variable(tf.truncated_normal([n_fc1, n_fc2], stddev=0.1)),
'fc3': tf.Variable(tf.truncated_normal([n_fc2, n_classes], stddev=0.1))
}
b_conv = {
'conv1': tf.Variable(tf.constant(0.0, dtype=tf.float32, shape=[96])),
'conv2': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[256])),
'conv3': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[384])),
'conv4': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[384])),
'conv5': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[256])),
'fc1': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[n_fc1])),
'fc2': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[n_fc2])),
'fc3': tf.Variable(tf.constant(0.0, dtype=tf.float32, shape=[n_classes]))
}
下面实现各层：
第一层卷积池化
# 第一层卷积
conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, W_conv['conv1'], strides=[1, 4, 4, 1], padding='SAME'), b_conv['conv1']))
pool1 = tf.nn.avg_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID')
# LRN局部相应归一化
norm1 = tf.nn.lrn(pool1, 5, bias=1.0, alpha=0.001/9.0, beta=0.75)
在AlexNet中使用了LRN（局部相应归一化）来缓解过拟合，后来被BN（批归一化）所取代。但在这里尽量还原原模型。
第二层卷积池化
# 第二层卷积
conv2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(norm1, W_conv['conv2'], strides=[1, 1, 1, 1], padding='SAME'), b_conv['conv2']))
pool2 = tf.nn.avg_pool(conv2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID')
# LRN局部相应归一化
norm2 = tf.nn.lrn(pool2, 5, bias=1.0, alpha=0.001/9.0, beta=0.75)
第三层卷积（没有进行池化）
# 第三层卷积
conv3 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(norm2, W_conv['conv3'], strides=[1, 1, 1, 1], padding='SAME'), b_conv['conv3']))
第四层卷积（没有池化）
# 第四层卷积
conv4 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv3, W_conv['conv4'], strides=[1, 1, 1, 1], padding='SAME'), b_conv['conv4']))
第五层卷积池化
# 第五层卷积
conv5 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv4, W_conv['conv5'], strides=[1, 1, 1, 1], padding='SAME'), b_conv['conv5']))
pool5 = tf.nn.avg_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID')
将矩阵展开成一维进行第六层全连接：
reshape = tf.reshape(pool5, [-1, 6*6*256])
# 第六层全连接
fc1 = tf.nn.relu(tf.add(tf.matmul(reshape, W_conv['fc1']), b_conv['fc1']))
# dropout
fc1 = tf.nn.dropout(fc1, 0.5)
这里采用了dropout技术来抑制一部分神经元，以缓解过拟合
第七层全连接
# 第七层全连接
fc2 = tf.nn.relu(tf.add(tf.matmul(fc1, W_conv['fc2']), b_conv['fc2']))
# dropout
fc2 = tf.nn.dropout(fc2, 0.5)
分类：
# 全连接分类
fc3 = tf.add(tf.matmul(fc2, W_conv['fc3']), b_conv['fc3'])
定义损失和精度：
# 定义损失
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=fc3, labels=y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss)
# 模型精度
correct_num = tf.equal(tf.arg_max(fc3, 1), tf.arg_max(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_num, tf.float32))
至此，整个AlexNet模型已经搭建完成！
模型的训练存储及结果可视化
tensorflow提供了方法以存储模型和结构及权重，即tf.sava函数。
下面开始训练，并将训练后的模型保存下来：
# 训练
def train():
with tf.Session() as sess:
sess.run(init)
save_model = ".//model//AlexNetModel.ckpt"
train_writer = tf.summary.FileWriter(".//log", sess.graph)
saver = tf.train.Saver()
losslist = [] # 存储损失
coord = tf.train.Coordinator()
treades = tf.train.start_queue_runners(coord=coord)
step = 0
for i in range(500):
step = i
image, label = sess.run([image_batch, label_batch])
labels = onehot(label)
sess.run(optimizer, feed_dict={x: image, y: labels})
loss_record = sess.run(loss, feed_dict={x: image, y: labels})
print(loss_record)
losslist.append(loss_record)
saver = tf.train.Saver()
saver.save(sess, save_model)
print("model save finished!")
coord.request_stop()
coord.join(treades)
plt.plot(losslist)
plt.xlabel('iter')
plt.ylabel('loss')
plt.tight_layout()
plt.savefig('cnn-tf-AlexNet.png', dpi=200)
上述函数用到了onehot函数来把标签转换为onehot形式，该函数的代码如下：
# 转换为one-hot
def onehot(label):
onehot_label = np.zeros([len(label), max(label)+1])
for i in range(len(label)):
onehot_label[i][label[i]] = 1
return onehot_label
与上一张不同的是，这次我们的输出和图表都是以损失为标准（一般精度变化和损失变化都是重要的指标）
训练结果：
完整代码：
import tensorflow as tf
import matplotlib.pyplot as plt
import cv2
import os
import numpy as np
import io
# 集中管理参数
learning_rate = 1e-4 # 学习速率
training_iters = 200 # 迭代次数
batch_size = 50 # 每批的大小
display_step = 5 #
n_classes = 2 #
n_fc1 = 4096 #
n_fc2 = 2048 #
# 构建模型
x = tf.placeholder(tf.float32, [None, 227, 227, 3])
y = tf.placeholder(tf.int32, [None, n_classes])
W_conv = {
'conv1': tf.Variable(tf.truncated_normal([11, 11, 3, 96], stddev=0.0001)),
'conv2': tf.Variable(tf.truncated_normal([5, 5, 96, 256], stddev=0.01)),
'conv3': tf.Variable(tf.truncated_normal([3, 3, 256, 384], stddev=0.01)),
'conv4': tf.Variable(tf.truncated_normal([3, 3, 384, 384], stddev=0.01)),
'conv5': tf.Variable(tf.truncated_normal([3, 3, 384, 256], stddev=0.01)),
'fc1': tf.Variable(tf.truncated_normal([6*6*256, n_fc1], stddev=0.1)),
'fc2': tf.Variable(tf.truncated_normal([n_fc1, n_fc2], stddev=0.1)),
'fc3': tf.Variable(tf.truncated_normal([n_fc2, n_classes], stddev=0.1))
}
b_conv = {
'conv1': tf.Variable(tf.constant(0.0, dtype=tf.float32, shape=[96])),
'conv2': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[256])),
'conv3': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[384])),
'conv4': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[384])),
'conv5': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[256])),
'fc1': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[n_fc1])),
'fc2': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[n_fc2])),
'fc3': tf.Variable(tf.constant(0.0, dtype=tf.float32, shape=[n_classes]))
}
# 第一层卷积
conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, W_conv['conv1'], strides=[1, 4, 4, 1], padding='SAME'), b_conv['conv1']))
pool1 = tf.nn.avg_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID')
# LRN局部相应归一化
norm1 = tf.nn.lrn(pool1, 5, bias=1.0, alpha=0.001/9.0, beta=0.75)
# 第二层卷积
conv2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(norm1, W_conv['conv2'], strides=[1, 1, 1, 1], padding='SAME'), b_conv['conv2']))
pool2 = tf.nn.avg_pool(conv2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID')
# LRN局部相应归一化
norm2 = tf.nn.lrn(pool2, 5, bias=1.0, alpha=0.001/9.0, beta=0.75)
# 第三层卷积
conv3 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(norm2, W_conv['conv3'], strides=[1, 1, 1, 1], padding='SAME'), b_conv['conv3']))
# 第四层卷积
conv4 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv3, W_conv['conv4'], strides=[1, 1, 1, 1], padding='SAME'), b_conv['conv4']))
# 第五层卷积
conv5 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv4, W_conv['conv5'], strides=[1, 1, 1, 1], padding='SAME'), b_conv['conv5']))
pool5 = tf.nn.avg_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID')
reshape = tf.reshape(pool5, [-1, 6*6*256])
# 第六层全连接
fc1 = tf.nn.relu(tf.add(tf.matmul(reshape, W_conv['fc1']), b_conv['fc1']))
# dropout
fc1 = tf.nn.dropout(fc1, 0.5)
# 第七层全连接
fc2 = tf.nn.relu(tf.add(tf.matmul(fc1, W_conv['fc2']), b_conv['fc2']))
# dropout
fc2 = tf.nn.dropout(fc2, 0.5)
# 全连接分类
fc3 = tf.add(tf.matmul(fc2, W_conv['fc3']), b_conv['fc3'])
# 定义损失
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=fc3, labels=y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss)
# 模型精度
correct_num = tf.equal(tf.arg_max(fc3, 1), tf.arg_max(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_num, tf.float32))
# 初始化变量
init = tf.global_variables_initializer()
# 把图片大小转换为227x227
def rebuild(dir):
for root, dirs, files in os.walk(dir):
for file in files:
try:
filepath = os.path.join(root, file)
image = cv2.imread(filepath)
dim = (227, 227)
resized = cv2.resize(image, dim)
print(file)
path = "D:/cat_and_dog/Cat/"+file
cv2.imwrite(path, resized)
except IOError:
print(filepath)
os.remove(filepath)
cv2.waitKey(0) # 退出
rebuild("D:/PetImages/Cat")
# 设置标签
def get_file(file_dir):
images = []
temp = []
for root, sub_folders, files in os.walk(file_dir):
for name in files:
images.append(os.path.join(root, name))
for name in sub_folders:
temp.append(os.path.join(root, name))
labels = []
for one_folder in temp:
n_img = len(os.listdir(one_folder))
letter = one_folder.split("\\")[-1]
if letter=='Cat':
labels = np.append(labels, n_img*[0])
else:
labels = np.append(labels, n_img*[1])
temp = np.array([images, labels])
temp = temp.transpose()
np.random.shuffle(temp)
image_list = list(temp[:, 0])
label_list = list(temp[:, 1])
label_list = [int(float(i)) for i in label_list]
return image_list, label_list
imagelist, labellist = get_file("D:/cat_and_dog")
# 生成TFRecord文件
def int64_feature(value):
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))
def bytes_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
def ToTFRecord(image_list, label_list, save_dir, name):
filename = os.path.join(save_dir, name+'.tfrecords')
n_samples = len(imagelist)
writer = tf.python_io.TFRecordWriter(filename)
print("transform start...")
for i in np.arange(0, n_samples):
try:
image = cv2.imread(image_list[i])
image_raw = image.tostring()
label = [int(label_list[i])]
example = tf.train.Example(features=tf.train.Features(feature={
'label':int64_feature(label),
'image_raw':bytes_feature(image_raw)
}))
writer.write(example.SerializeToString())
except IOError as e:
print('could not read:', image_list[i])
writer.close()
print('transform done!')
ToTFRecord(imagelist, labellist, '.', 'cat_and_dog')
# 读取数据
def read_and_decode(tfrecord_file, batch_size):
filename_queue = tf.train.string_input_producer([tfrecord_file])
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
img_features = tf.parse_single_example(
serialized_example,
features={
'label': tf.FixedLenFeature([], tf.int64),
'image_raw': tf.FixedLenFeature([], tf.string),
})
image = tf.decode_raw(img_features['image_raw'], tf.uint8)
image = tf.reshape(image, [227, 227, 3])
label = tf.cast(img_features['label'], tf.int32)
image_batch, label_batch = tf.train.shuffle_batch([image, label],
batch_size=batch_size,
min_after_dequeue=100,
num_threads=64,
capacity=200
)
return image_batch, tf.reshape(label_batch, [batch_size])
image_batch, label_batch = read_and_decode('cat_and_dog.tfrecords', 10)
# 转换为one-hot
def onehot(label):
onehot_label = np.zeros([len(label), max(label)+1])
for i in range(len(label)):
onehot_label[i][label[i]] = 1
return onehot_label
# 训练
def train():
with tf.Session() as sess:
sess.run(init)
save_model = ".//model//AlexNetModel.ckpt"
train_writer = tf.summary.FileWriter(".//log", sess.graph)
saver = tf.train.Saver()
losslist = [] # 存储损失
coord = tf.train.Coordinator()
treades = tf.train.start_queue_runners(coord=coord)
step = 0
for i in range(500):
step = i
image, label = sess.run([image_batch, label_batch])
labels = onehot(label)
sess.run(optimizer, feed_dict={x: image, y: labels})
loss_record = sess.run(loss, feed_dict={x: image, y: labels})
print(loss_record)
losslist.append(loss_record)
saver = tf.train.Saver()
saver.save(sess, save_model)
print("model save finished!")
coord.request_stop()
coord.join(treades)
plt.plot(losslist)
plt.xlabel('iter')
plt.ylabel('loss')
plt.tight_layout()
plt.savefig('cnn-tf-AlexNet.png', dpi=200)
train()
这一章的内容还是非常多的，而且随着数据集大小的增长以及模型越来越复杂，训练时间也会越来越久，明天没空，预计断更一天^ ^。
ps：新学校饭堂的饭真的好吃啊，看来这三年又要无解肥了。。。