【人工智能基础】VGG13与ResNet实验

实验环境：anaconda、jupyter notebook、spyder

实验用到的包：tensorflow

一、卷积神经网络CNN

卷积

建立一个卷积核（一般为c×1×1 或者 c×3×3， c为输入图像的通道数），使用卷积核扫过图像，调整计算每个位置的权重。

池化

主要功能为：特征降维、防止过拟合、下采样等。 Max pooling：取区域的最大值作为本区域对应块的值 Avg pooling：取区域的平均值作为本区域对应块的值

二、VGG13实验

tensorflow安装

实验之前需要安装tensorflow包，tensorflow安装需求可以看官网文档，我使用的是python3.9 + tensorflow2.4.1 + numpy1.19.5

环境准备

import tensorflow as tf
from tensorflow.keras import layers, optimizers, datasets, Sequential
import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
tf.random.set_seed(2345)

卷积层模型构建

conv_layers=[
    # unit 1
    # 输出通道数64 卷积核为2×2矩阵 输入数据边缘填充0 使用relu作为激活函数
    layers.Conv2D(64, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.Conv2D(64, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    # 池化窗口为2×2矩阵 步长为2 输入数据边缘填充0
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # unit 2
    layers.Conv2D(128, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.Conv2D(128, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # unit 3
    layers.Conv2D(256, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.Conv2D(256, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # unit 4
    layers.Conv2D(512, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.Conv2D(512, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # unit 5
    layers.Conv2D(512, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.Conv2D(512, kernel_size=[2, 2],padding='same',activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same')
]

加载cifar100数据集以及数据处理

# 数据预处理函数
def preprocess(x, y):
    # 把入参转为float 再归一化到[0,1]区间
    x = tf.cast(x, dtype=tf.float32) / 255
    # 标签数据转换
    y = tf.cast(y, dtype=tf.int32)
    return x, y

# 加载cifar100数据集
(x, y), (x_test, y_test) = datasets.cifar100.load_data()
# 在二维数组中等同于y[:,1]
y = tf.squeeze(y, axis=1)
y_test = tf.squeeze(y_test, axis=1)
print('原始数据shape：', x.shape, y.shape, x_test.shape, y_test.shape)

# 训练集打包
train_db = tf.data.Dataset.from_tensor_slices((x,y))
# 训练集洗牌，预处理，分为64批次
train_db = train_db.shuffle(100).map(preprocess).batch(64)

test_db = tf.data.Dataset.from_tensor_slices((x_test,y_test))
test_db = test_db.map(preprocess).batch(64)

sample = next(iter(train_db))
print('sample：', sample[0].shape, sample[1].shape, 
     tf.reduce_min(sample[0]),tf.reduce_max(sample[0]))

定义运行函数

def run():
    # 卷积层
    # conv_net是实现输入数据[b,32,32,3]到输出输出[b,1,1,512]的变换
    conv_net = Sequential(conv_layers)
    conv_net.build(input_shape=[None,32,32,3])

    # 全连接层
    fc_net = Sequential([
        layers.Dense(256, activation=tf.nn.relu),
        layers.Dense(128, activation=tf.nn.relu),
        layers.Dense(100, activation=None),
    ])
    fc_net.build(input_shape=[None,512])

    # trainable_variables包含权重矩阵weights和偏执矩阵baises
    variables = conv_net.trainable_variables + fc_net.trainable_variables

    # 梯度下降优化器 学习率为0.0001
    optimizer = optimizers.Adam(lr=1e-4)

    for epoch in range(50):
        for step,(x,y) in enumerate(train_db):
            with tf.GradientTape() as tape:
                # [b,32,32,3]=>[b,1,1,512]
                out = conv_net(x)
                # 变成b个1×512的矩阵
                out = tf.reshape(out,[-1,512])
                # [b,512] => [b,100]
                logits = fc_net(out)
                """ 
                将分类变量转变为向量
                如果y = [0,2,1]
                y_onehot = tf.one_hot(y, depth=5)
                y_onehot= 
                [
                  [1, 0, 0, 0, 0],  # 对应 y 中的 0
                  [0, 0, 1, 0, 0],  # 对应 y 中的 2
                  [0, 1, 0, 0, 0]   # 对应 y 中的 1
                ]
                这里是转变为100个类的分类向量
                """
                y_onehot = tf.one_hot(y, depth=100)

                # 损失函数
                loss = tf.losses.categorical_crossentropy(y_onehot,logits,from_logits=True)
                loss = tf.reduce_mean(loss)
            grads = tape.gradient(loss, variables)
            # 梯度下降
            optimizer.apply_gradients(zip(grads,variables))
            # 打印参数
            if step % 100 == 0:
                print(epoch, step, 'loss:',float(loss))
        # 每次梯度下降结束后测试模型
        total_num = 0
        total_correct = 0
        for x,y in test_db:
            out = conv_net(x)
            out = tf.reshape(out, [-1,512])
            logits = fc_net(out)
            prob = tf.nn.softmax(logits, axis=1)
            pred = tf.argmax(prob, axis=1)
            pred = tf.cast(pred, dtype=tf.int32)
            correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)
            correct = tf.reduce_sum(correct)
            total_num += x.shape[0]
            total_correct += int(correct)
        acc = total_correct / total_num
        print(epoch, 'acc:', acc)

运行结果

run()

看到准确度在一点一点上升

三、ResNet实验

ResNet的原理：神经网络在达到22层后，随着层数的增加表现不一定会更好。ResNet在22层后（可以在一开始就）对每层添加一条短路，如果加上这层表现更好则不短路，如果添加后表现更差则短路。

basic block

把两个卷积单元称为一个basicblock

一个卷积单元包含卷积、归一化、激活函数等功能。

class BasicBlock(layers.Layer):
    """
    * filter_num:通道数量
    * stride:步长
    """
    def __init__(self, filter_num, stride=1):
        super(BasicBlock, self).__init__()
        # 步长大于1进行下采样
        self.conv1 = layers.Conv2D(filter_num, (3, 3), strides=stride, padding='same')
        self.bn1 = layers.BatchNormalization()
        self.relu = layers.Activation('relu')
        # 第二个卷积单元不做下采样
        self.conv2 = layers.Conv2D(filter_num, (3, 3), strides=1, padding='same')
        self.bn2 = layers.BatchNormalization()

        # identity层，短接,如果步长不为1，需要下采样
        if stride != 1:
            self.downsample = Sequential()
            self.downsample.add(layers.Conv2D(filter_num, (1, 1), strides=stride))
        else:
            self.downsample = lambda x:x


    def call(self, inputs, training=None):
        # 卷积
        out = self.conv1(inputs)
        # 归一化
        out = self.bn1(out)
        # 激活函数
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        identity = self.downsample(inputs)
        output = layers.add([out, identity])
        # 这个激活函数放在bn2后也可以
        output = tf.nn.relu(output)

        return output

Res Net

多个basic block构成了ResNet

class ResNet(keras.Model):
    """
    * layer_dims 
    """
    def __init__(self, layer_dims, num_classes=10): 
        super(ResNet, self).__init__()
        # stride=(1,1)代表在w、h两个维度上步长都是1
        self.stem = Sequential([layers.Conv2D(64, (3, 3), strides=(1, 1)),
                                layers.BatchNormalization(),
                                layers.Activation('relu'),
                                layers.MaxPool2D(pool_size=(2, 2), strides=(1, 1), padding='same')
                                ])
        self.layer1 = self.build_res_block(64,  layer_dims[0])
        self.layer2 = self.build_res_block(128, layer_dims[1], stride=2)
        self.layer3 = self.build_res_block(256, layer_dims[2], stride=2)
        self.layer4 = self.build_res_block(512, layer_dims[3], stride=2)
        # output[b,512,h,w]转为[b,512,1,1]
        self.avgpool = layers.GlobalAveragePooling2D()
        self.fc = layers.Dense(num_classes)


    def call(self, inputs, training=None):
        out = self.stem(inputs)
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)

        # [b, num_channels]
        out = self.avgpool(out)
        # [b, num_classes]
        out = self.fc(out)

        return out


    def build_res_block(self, filter_num, blocks, stride=1):
        res_blocks = Sequential()
        # 只有第一个basic block可能具有下采样能力
        res_blocks.add(BasicBlock(filter_num, stride))

        for _ in range(1, blocks):
            res_blocks.add(BasicBlock(filter_num, stride=1))
        return res_blocks

封装简化ResNet建立过程

def resnet18():
    return ResNet([2,2,2,2])
def resnet34():
    return ResNet([3,4,6,3])

加载预处理cifar10数据集

# 数据预处理函数
def preprocess(x, y):
    # 把入参转为float 再归一化到[-1,1]区间
    x = 2 * tf.cast(x, dtype=tf.float32) / 255 - 1
    # 标签数据转换
    y = tf.cast(y, dtype=tf.int32)
    return x, y

# 加载cifar100数据集
(x, y), (x_test, y_test) = datasets.cifar10.load_data()
# 在二维数组中等同于y[:,1]
y = tf.squeeze(y, axis=1)
y_test = tf.squeeze(y_test, axis=1)
print('原始数据shape：', x.shape, y.shape, x_test.shape, y_test.shape)

# 训练集打包
train_db = tf.data.Dataset.from_tensor_slices((x,y))
# 训练集洗牌，预处理，分为64批次
train_db = train_db.shuffle(100).map(preprocess).batch(64)

test_db = tf.data.Dataset.from_tensor_slices((x_test,y_test))
test_db = test_db.map(preprocess).batch(64)

sample = next(iter(train_db))
print('sample：', sample[0].shape, sample[1].shape, 
     tf.reduce_min(sample[0]),tf.reduce_max(sample[0]))
# 这里del是因为我的服务器只有2g内存，哈哈哈哈
del x, y, x_test, y_test, sample

运行函数

def run():
    # ResNet18网络构建
    model = resnet18()
    model.build(input_shape=(None,32,32,3))

    # 梯度下降优化器 学习率为0.0001
    optimizer = optimizers.Adam(lr=1e-4)

    for epoch in range(100):
        for step,(x,y) in enumerate(train_db):
            with tf.GradientTape() as tape:
                # [b,32,32,3]=>[b,100]
                logits = model(x)
                """ 
                将分类变量转变为向量
                如果y = [0,2,1]
                y_onehot = tf.one_hot(y, depth=5)
                y_onehot= 
                [
                  [1, 0, 0, 0, 0],  # 对应 y 中的 0
                  [0, 0, 1, 0, 0],  # 对应 y 中的 2
                  [0, 1, 0, 0, 0]   # 对应 y 中的 1
                ]
                这里是转变为100个类的分类向量
                """
                y_onehot = tf.one_hot(y, depth=10)

                # 损失函数
                loss = tf.losses.categorical_crossentropy(y_onehot,logits,from_logits=True)
                loss = tf.reduce_mean(loss)
            grads = tape.gradient(loss, model.trainable_variables)
            # 梯度下降
            optimizer.apply_gradients(zip(grads,model.trainable_variables))
            # 打印参数
            if step % 100 == 0:
                print(epoch, step, 'loss:', float(loss))
        # 每次梯度下降结束后测试模型
        total_num = 0
        total_correct = 0
        for x,y in test_db:
            logits = model(x)
            prob = tf.nn.softmax(logits, axis=1)
            pred = tf.argmax(prob, axis=1)
            pred = tf.cast(pred, dtype=tf.int32)
            correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)
            correct = tf.reduce_sum(correct)
            total_num += x.shape[0]
            total_correct += int(correct)
        acc = total_correct / total_num
        print(epoch, 'acc:', acc)

运行结果

run()