用TensorFlow实现AlexNet，并完成Kaggle上的Dogs Vs Cats竞赛

最新推荐文章于 2022-12-27 11:45:39 发布

Hugsy19

最新推荐文章于 2022-12-27 11:45:39 发布

阅读量534

点赞数

分类专栏：机器学习深度学习 TensorFlow python 文章标签： TensorFlow AlexNet Kaggle

本文链接：https://blog.csdn.net/sinat_35473930/article/details/89856583

版权

机器学习同时被 3 个专栏收录

14 篇文章 2 订阅

订阅专栏

深度学习

13 篇文章 0 订阅

订阅专栏

python

6 篇文章 0 订阅

订阅专栏

原文链接

总体架构

AlexNet的总体架构如下所示：
AlexNet

整个网络总共有八层，其中前五层为卷积层，后面三层为全连接层；
用了两个GPU（GTX 580，3G内存）进行训练，因此整个架构被平均分成了两部分，除了第二层与第三层是跨GPU进行连接，其他卷积层都只和各自的GPU内的前一层进行连接；
在第一、二个卷积层后进行现在并不常用的局部响应归一化（local response normalization，LRN），LRN的公式如下：

公式中的 $k$ 、 $n$ 、 $\alpha$ 、 $\beta$ 均为超参数，在验证集上实验得到 $k = 2$ 、 $n = 5$ 、 $\alpha = 0.00001$ 、 $\beta = 0.75$ ；
在LRN、第五层后面进行窗口大小和步幅不相等的最大池化，称为最大重叠池化；
批量大小为 $128$ ，并使用ReLU作为激活函数，
用数据扩充和DropOut法来防止过拟合。

各层中的具体过程为：
AlexNet具体过程

输入的大小原为 $224 \times 224 \times 3$ ，为方便后续处理将其调整为 $227 \times 227 \times 3$ ；
第一层用 $96$ 个大小为 $11 \times 11$ 的卷积核进行步幅为 $4$ 的卷积，之后用大小为 $3\times3$ 的窗口进行步幅为 $2$ 的最大重叠池化后，再进行尺度为 $5\times5$ 的LRN；
第二层用 $256$ 个大小为 $\times 5$ 的卷积核进行步幅为 $1$ 的卷积，之后和前一层一样进行重叠池化和LRN；
第三层用 $384$ 个大小为 $\times 3$ 的卷积核连接到第二层的所有输出；
第四、第五层分别用 $384$ 、 $256$ 个大小为 $\times 3$ 的卷积核进行步幅为 $1$ 的卷积,且在第五层后用大小为 $3\times3$ 的窗口进行步幅为 $2$ 的最大重叠池化。
第六层的输入大小为 $\times 6 \times 256$ ，与 $4096$ 个大小为 $\times 6 \times 256$ 的卷积核进行卷积，就得到包含 $4096$ 个节点的全连接层。训练时，需要在该全连接层进行一次drop_prob为 $0.5$ 的Dropout。
第七层的 $4096$ 个节点与上一个全连接层进行全连接，训练时，也需要在该全连接层进行一次drop_prob为 $0.5$ 的Dropout。
第八层进行第三次全连接，并输出最后的结果。

TensorFlow实现

辅助方法

卷积：

def conv(x, filter_height, filter_width, filters_num, stride_x, stride_y, name, padding='SAME', groups=1): # groups: 分成多个部分
    input_channels = int(x.get_shape()[-1]) # 输入通道数
    convolve = lambda i, k: tf.nn.conv2d(i, k, strides=[1, stride_x, stride_y, 1], padding=padding) # 卷积
    
    with tf.variable_scope(name) as scope:
        weights = tf.get_variable('weights', shape=[filter_height, filter_width, input_channels/groups, filters_num])
        bias = tf.get_variable('bias', shape=[filters_num])
        
        if groups == 1:
            conv = convolve(x, weights)
        else:
            input_groups = tf.split(value=x, num_or_size_splits=groups, axis=3) # 切分
            weight_groups = tf.split(value=weights, num_or_size_splits=groups, axis=3)
            output_groups = [convolve(i, k) for i, k in zip(input_groups, weight_groups)] # 分别卷积
            conv = tf.concat(values=output_groups, axis=3) # 拼接
        
        z = tf.reshape(tf.nn.bias_add(conv, bias), conv.get_shape().as_list())
        relu = tf.nn.relu(z, name=scope.name)
        return relu

池化、LRN、Dropout：

def max_pool(x, filter_height, filter_width, stride_x, stride_y, name, padding='SAME'):
    return tf.nn.max_pool(x, [1, filter_height, filter_width, 1], strides=[1, stride_x, stride_y, 1], padding=padding, name=name)

# 局部响应归一化
def lrn(x, radius, alpha, beta, name, bias=1.0):
    return tf.nn.lrn(x, depth_radius=radius, alpha=alpha, beta=beta, bias=bias, name=name) # bias对应k, radius对应n/2

# 全连接
def fc(x, num_in, num_out, name, relu=True):
    with tf.variable_scope(name) as scope:
        weights = tf.get_variable('weights', shape=[num_in, num_out])
        bias = tf.get_variable('bias', shape=[num_out])
        z = tf.nn.xw_plus_b(x, weights, bias, name=scope.name)
        
        if relu == True:
            act = tf.nn.relu(z)
        else:
            act = z
    return act

# Dropout
def dropout(x, keep_prob):
    return tf.nn.dropout(x, keep_prob)

用动态图测试上面的各方法：

import tensorflow.contrib.eager as tfe
tfe.enable_eager_execution()
x = tf.truncated_normal(shape=[1, 227, 227, 3], seed = 1)
cnv = conv(x, 11, 11, 96, 4, 4, padding='VALID', name='conv')
pool = max_pool(cnv, 3, 3, 2, 2, padding='VALID', name='pool')
norm = lrn(pool, 2, 2e-05, 0.75, name='norm')
norm.get_shape()

整个模型

建立整个AlenNet：

class AlexNetModel(object):
    
    def __init__(self, num_classes=1000, keep_prob=0.5, skip_layer=[], weights_path='DEFAULT'):
        self.num_classes = num_classes
        self.keep_prob = keep_prob
        self.skip_layer = skip_layer
        
        if weights_path == 'DEFAULT':
            self.weights_path = 'bvlc_alexnet.npy'
        else:
            self.weights_path = weights_path
    
    def inference(self, x, training=False): # 模型
        # conv1: CONV --> POOL --> LRN
        conv1 = conv(x, 11, 11, 96, 4, 4, padding='VALID', name='conv1')
        pool1 = max_pool(conv1, 3, 3, 2, 2, padding='VALID', name='pool1')
        norm1 = lrn(pool1, 2, 2e-05, 0.75, name='norm1')
        
        # conv2: CONV --> POOL --> LRN with 2 Groups
        conv2 = conv(norm1, 5, 5, 256, 1, 1, groups=2, name='conv2')
        pool2 = max_pool(conv2, 3, 3, 2, 2, padding='VALID', name='pool2')
        norm2 = lrn(pool2, 2, 2e-05, 0.75, name='norm2')
        
        # conv3: CONV 
        conv3 = conv(norm2, 3, 3, 384, 1, 1, name='conv3')
        
        # conv4: CONV with 2 Groups
        conv4 = conv(conv3, 3, 3, 384, 1, 1, groups=2, name='conv4')
        
        # conv5: CONV --> PooL with 2 Groups
        conv5 = conv(conv4, 3, 3, 256, 1, 1, groups=2, name='conv5')
        pool5 = max_pool(conv5, 3, 3, 2, 2, padding='VALID', name='pool5')
        
        # fc6: Flatten --> FC --> Dropout
        flattened  = tf.reshape(pool5, [-1, 6*6*256])
        fc6 = fc(flattened, 6*6*256, 4096, name='fc6')
        if training:
            fc6 = dropout(fc6, self.keep_prob)
        
        # fc7: FC --> Dropout
        fc7 = fc(fc6, 4096, 4096, name='fc7')
        if training:
            fc7 = dropout(fc7, self.keep_prob)
            
        # fc8: FC
        self.score = fc(fc7, 4096, self.num_classes, relu=False, name='fc8')
        
        return self.score
    
    def loss(self, batch_x, batch_y): # 损失
        y_predict = self.inference(batch_x, training=True)
        self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=y_predict, labels=batch_y))
        return self.loss
    
    def optimize(self, learning_rate, train_layers=[]): # 优化
        var_list = [v for v in tf.trainable_variables() if v.name.split('/')[0] in train_layers]
        return tf.train.AdamOptimizer(learing_rate).minimize(self.loss, var_list=var_list)
    
    def load_original_weights(self, session): # 导入训练好的权重
        weights_dict = np.load(self.weights_path, encoding='bytes').item()
        
        for op_name in weights_dict:
            if op_name not in self.skip_layer:     
                with tf.variable_scope(op_name, reuse=True):
                    for data in weights_dict[op_name]:
                        if len(data.shape) == 1:
                            var = tf.get_variable('bias')
                            session.run(var.assign(data))
                        else:
                            var = tf.get_variable('weights')
                            session.run(var.assign(data))

模型测试

用原始的参数值来测试构建好的AlexNet模型，原始参数的文件可从这里下载。

import matplotlib.pyplot as plt
from caffe_classes import class_names
import cv2
import os

img_dir = os.path.join(os.getcwd(), 'images')
img_file = [os.path.join(img_dir, f) for f in os.listdir(img_dir) if f.endswith('.jpeg')]

imgs = []
for f in img_file:
    imgs.append(cv2.imread(f))

imagenet_mean = np.array([104., 117., 124.], dtype=np.float32) # ImageNet中的图片像素均值
x = tf.placeholder(tf.float32, [1, 227, 227, 3])
model = AlexNetModel()
score = model.inference(x)
softmax = tf.nn.softmax(score)

writer = tf.summary.FileWriter('./graph/alexnet', tf.get_default_graph())
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    model.load_original_weights(sess)
    
    fig2 = plt.figure(figsize=(15, 6))
    
    for i, image in enumerate(imgs):
        img = cv2.resize(image.astype(np.float32), (227, 227))
        img -= imagenet_mean
        img = img.reshape((1, 227, 227, 3))
        
        probs = sess.run(softmax, feed_dict={x: img})
        class_name = class_names[np.argmax(probs)]
        
        writer.close()
        fig2.add_subplot(1,3,i+1)
        plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
        plt.title("Class: " + class_name + ", probability: %.4f" %probs[0,np.argmax(probs)])
        plt.axis('off')

得到预测结果：

Dogs VS Cats

下载Kaggle上的Dogs vs. Cats Redux Competetion的数据集，考虑微调上面构建好的AlexNet来完成该竞赛。

解压其中的压缩包后，可以看到其中包含 $25000$ 张带标签的训练数据：
train

和 $15000$ 张不带标签的测试数据：
test

数据处理

首先将数据集划分一下，这里把 $85\%$ 的训练数据也就是 $21250$ 张图片作为训练集，剩下的 $4750$ 张图片作为验证集。

竞赛中要求将猫图的标签设为 $0$ ，狗图的标签设为 $1$ 。要使用这些数据来训练我们的模型，为了方便读取，可考虑获取所有的图片的路径和对应的标签，把它们统一放在一个txt文件中。实现该过程的程序如下：

import os

train_sets_dir = os.path.join(os.getcwd(), 'train')
train_images_file = os.listdir(train_sets_dir)
train_sets_list = []

for fn in train_images_file:
    file_label = fn.split('.')[0]
    
    if file_label == 'cat':
        label = '0'
    else:
        label = '1'
        
    path_and_label = os.path.join(train_sets_dir, fn) + ' ' + label + '\n'
    train_sets_list.append(path_and_label)

validate_sets_list = train_sets_list[int(len(train_sets_list)*0.85):] # 15%作为验证集
train_sets_list = train_sets_list[:int(len(train_sets_list)*0.85)]

train_text = open('train.txt', 'w') # 写入txt文件
for img in train_sets_list:
    train_text.writelines(img)    

validate_text = open('validate.txt', 'w') # 写入txt文件
for img in validate_sets_list:
    validate_text.writelines(img)

这样就能得到名为train.txt、validate.txt的两个文件，文件的内容如下：
txt内容

前面是图片的数据路径，后面则是该图片对应的标签。

导入数据

用下面的辅助方法来调用tf.data.Dataset导入数据，并对数据进行一些简单的处理：

# 数据处理
IMAGENET_MEAN = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32) # 用于放缩范围

def parse_image(filename, label):
    img_string = tf.read_file(filename) # 读取
    img_decoded = tf.image.decode_png(img_string, channels=3) # 编码
    img_resized = tf.image.resize_images(img_decoded, [227, 227]) # 裁剪
    img_converted = tf.cast(img_resized, tf.float32) # 数据格式
    img_centered = tf.subtract(img_resized, IMAGENET_MEAN) # 放缩范围
    
    return img_centered, label

def data_generate(txt_file, batch_size, num_classes, shuffle=True):
    
    paths_and_labels = np.loadtxt(txt_file, dtype=str).tolist() # 读取文件，组成列表
        
    if shuffle:
        np.random.shuffle(paths_and_labels) # 打乱

    paths, labels = zip(*[(l[0], int(l[1])) for l in paths_and_labels]) # 将paths和labels分开
    steps_per_epoch = np.ceil(len(labels)/batch_size).astype(np.int32)
            
    paths = tf.convert_to_tensor(paths, dtype=tf.string) # 转换为tensor
    labels = tf.one_hot(labels, num_classes)
    labels = tf.convert_to_tensor(labels, dtype=tf.float32) 
            
    dataset = tf.data.Dataset.from_tensor_slices((paths, labels)) # 创建数据集
    dataset = dataset.map(parse_image) # 调函数进行预处理
    
    if shuffle:
        dataset = dataset.shuffle(buffer_size=batch_size)
        
    dataset = dataset.batch(batch_size) # 小批量
    
    return dataset, steps_per_epoch

微调AlexNet，其中的卷积层仍采用原始参数，而使用现有的数据集来训练其中的全连接层。因此调用上面的方法导入数据，设置并初始化迭代器，并设置一些超参数如下：

train_file = 'train.txt'
validate_file = 'validate.txt'


learning_rate = 0.01 # 超参数
num_epochs = 10
batch_size = 256

num_classes = 2
train_layers = ['fc8', 'fc7', 'fc6']

train_data, train_steps = data_generate(train_file, batch_size=batch_size, num_classes=num_classes)
validate_data, validate_steps = data_generate(validate_file, batch_size=batch_size, num_classes=num_classes)

iterator = tf.data.Iterator.from_structure(train_data.output_types, train_data.output_shapes) # 迭代器
train_init = iterator.make_initializer(train_data)
validate_init = iterator.make_initializer(validate_data)

imgs, labels = iterator.get_next()

建立并训练模型

建立AlexNet模型：

model = AlexNetModel(num_classes=num_classes, skip_layer=train_layers)
loss = model.loss(imgs, labels) # 损失
optimizer = model.optimize(learning_rate=learning_rate)

correct_pred = tf.equal(tf.argmax(model.score, 1), tf.argmax(labels, 1))
accuracy = tf.reduce_sum(tf.cast(correct_pred, tf.float32)) # 识别正确的总个数

训练并保存模型：

from datetime import datetime 

saver = tf.train.Saver()

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    writer.add_graph(sess.graph)
    
    model.load_original_weights(sess)
    
    print("{} Start training...".format(datetime.now()))
  
    for epoch in range(num_epochs): # 开始训练
        
        sess.run(train_init) # 训练数据初始化
        total_loss = 0
        n_batches = 0
        total_acc = 0
        try:
          while True:
            _, l, ac = sess.run([optimizer, loss, accuracy])
            total_loss += l
            total_acc += ac
            n_batches += 1
        except tf.errors.OutOfRangeError:
          pass
        
        print('Average loss epoch {0}: {1}'.format(epoch, total_loss/n_batches)) # 平均损失
    
        print("{} Training Accuracy = {:.4f}".format(datetime.now(), total_acc/21250.0) # 训练集准确率
        
    print("{} Start validation".format(datetime.now()))
    sess.run(validate_init) # 初始化验证集
    total_correct_preds = 0
    
    try:
      while True:
        accuracy_batch = sess.run(accuracy)
        total_correct_preds += accuracy_batch
    except tf.errors.OutOfRangeError:
      pass
    
    print("{} Validation Accuracy = {:.4f}".format(datetime.now(), total_correct_preds/4750.0)) # 验证集准确率
    
    print("{} Saving checkpoint of model...".format(datetime.now()))

    model_name = os.path.join(os.getcwd() + '/model', 'model_epoch'+str(epoch+1)+'.ckpt')
    save_path = saver.save(sess, model_name) # 保存模型

    print("{} Model checkpoint saved at {}".format(datetime.now(), model_name))

使用前面设置的超参数，本人的Colab上训练的结果如下：

训练结果

精确度还不够高，可以尝试继续调整超参数。

测试模型

在测试集上测试训练好的模型：

import os
import pandas as pd
import tensorflow as tf

test_sets_dir = os.path.join(os.getcwd(), 'test')
test_images_file = os.listdir(test_sets_dir)
test_images_file.sort(key=lambda x:int(x[:-4]))

test_sets_list = []

for fn in test_images_file:
    path = os.path.join(test_sets_dir, fn) + '\n'
    test_sets_list.append(path)

test_text = open('test.txt', 'w') # 写入txt文件
for img in test_sets_list:
    test_text.writelines(img) 

IMAGENET_MEAN = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32) # 用于放缩范围

def parse_test_image(filename):
    img_string = tf.read_file(filename)
    img_decoded = tf.image.decode_png(img_string, channels=3)
    img_resized = tf.image.resize_images(img_decoded, [227, 227])
    img_converted = tf.cast(img_resized, tf.float32)
    img_centered = tf.subtract(img_resized, IMAGENET_MEAN)
    
    return img_centered

images_path = np.loadtxt('./test.txt', dtype=str).tolist()
images_path = tf.convert_to_tensor(images_path, dtype=tf.string) 
test_dataset = tf.data.Dataset.from_tensor_slices((images_path))
test_dataset = test_dataset.map(parse_test_image) 
test_dataset = test_dataset.batch(1000)
test_iterator = test_dataset.make_one_shot_iterator() 
test_image = test_iterator.get_next()

model = AlexNetModel(num_classes=2)
score = model.inference(test_image)

predicts = []
saver=tf.train.Saver()
with tf.Session() as sess:
    saver.restore(sess, './model/model_epoch10.ckpt')
    
    try:
      while True:
        scores = sess.run(score)
        predicts.extend(tf.argmax(scores, 1).eval())
    except tf.errors.OutOfRangeError:
      pass

# 生成测试结果并写入cvs文件中
results = pd.Series(predicts, name="label") 
submission = pd.concat([pd.Series(range(1,12501),name = "id"), results],axis = 1)
submission.to_csv("sample_submission.csv",index=False)