CIF100实战(VGG13)

CIFAR-100 数据集就像CIFAR-10，除了它有100个类，每个类包含600个图像。，每类各有500个训练图像和100个测试图像。CIFAR-100 中的100个类被分成20个超类。每个图像都带有一个精细标签（它所属的类）和一个粗糙标签（它所属的超类）

   

这里使用比较强大的经典网络结构VGG13，根据数据集特点修改部分网络结构，完成 CIFAR100 图片识别。调整后的VGG13网络模型：

一、数据集加载以及数据集预处理

# 预处理
def preprocess(x, y):
    # x :[-1,1]
    x = 2 * tf.cast(x, dtype=tf.float32) / 255 - 1
    y = tf.cast(y, dtype=tf.int32)
    return x, y


# 数据集加载
(x, y), (x_text, y_text) = datasets.cifar100.load_data()
print("y:",y.shape)
# 压缩最后一个维度为1
y = tf.squeeze(y)
y_text = tf.squeeze(y_text, axis=1)
print('squeeze:',x.shape, y.shape, x_text.shape, y_text.shape)


# 创建batch
train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.map(preprocess).shuffle(1000).batch(128)

test_db = tf.data.Dataset.from_tensor_slices((x, y))
test_db = test_db.map(preprocess).batch(128)

# 获取下一个batch
sample = next(iter(test_db))
print('sample:', sample[0].shape, sample[1].shape,
      tf.reduce_min(sample[0]), tf.reduce_max(sample[0]))

数据集的处理和CIF10的处理是一样的，这里也要将y进行维度压缩如下。将维度为1的压缩，为one_hot编码做准备

 上述代码运行后，得到训练集的和形状为：(50000, 32, 32, 3)和(50000)，测试集的和形状为(10000, 32, 32, 3)和(10000)，分别代表了图片大小为32 × 32，彩色图片，训练集样本数为 50000，测试集样本数为 10000

二、网络模型构建与装配

将网络实现为 2 个子网络：卷积子网络和全连接子网络。卷积子网络由 5 个子模块构成，每个子模块包含了 Conv-Conv-MaxPooling 单元结构

conv_layers = [  # 5 units : conv + conv + max pooling
    # units1  64个3x3 卷积核, 输入输出同大小
    layers.Conv2D(64, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(64, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    # 高宽减半
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # units 2 由于上一层池化层减半，下一层将卷积层的卷积核翻一倍，为了弥补信息特征的减少
    layers.Conv2D(128, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(128, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

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

    # units 4
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # units 5
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same')
]

一般在上一个池化层进行最大化采用后(pool_size=[2,2] ,s=2)，降低了网络的参数量，得到的信息特征后减半，在进行下一个卷积层的时会将卷积核的倍数翻倍，以弥补信息的减少。

# [b,32,32,3] => [b,1,1,512]
    conv_net = Sequential(conv_layers)

    fc_net = Sequential([
        layers.Dense(256, activation=tf.nn.relu),
        layers.Dense(128, activation=tf.nn.relu),
        layers.Dense(10, activation=None)
    ])

    conv_net.build(input_shape=[None, 32, 32, 3])
    fc_net.build(input_shape=[None, 512])
    conv_net.summary()
    fc_net.summary()
    optimizer = optimizers.Adam(learning_rate=1e-4)

    # 可训练的变量 两个网络层之和
    variables = conv_net.trainable_variables + fc_net.trainable_variables

全连接子网络包含了 3 个全连接层，每层添加 ReLU 非线性激活函数，最后一层除外。卷积子网层输入的就是一张图片大小的维度[32,32,3], 不像全连接层那样需要打平层一维的。但是在进行两层的连接时，需要将卷积子网层打平成一维的。

设置优化器，注意需要将两层的可训练变量加起来。

三、梯度计算与参数更新

      for step, (x, y) in enumerate(train_db):
            # 梯度求导
            with tf.GradientTape() as tape:
                # [b, 32,32,3 ] =>[b,1,1,512]
                out = conv_net(x)
                out = tf.reshape(out, [-1, 512])

                # [b, 512] => [b, 10]
                logits = fc_net(out)
                # [b,10] =>
                y_onehot = tf.one_hot(y, depth=10)
                # loss
                loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
                loss = tf.reduce_mean(loss)
            # 求导
            grads = tape.gradient(loss, variables)
            # 更新参数
            optimizer.apply_gradients(zip(grads, variables))

            if step % 100 == 0:
                print(epoch, step, 'loss', float(loss))

在全连接层进行向前计算的时，需要将卷积子层的输出进行打平，其余都和前面的一样

四、测试

  # 测试集
        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)  # 概率化，和为1
            pred = tf.argmax(prob, axis=1)  # 获得最大下标
            pred = tf.cast(pred, dtype=tf.int32)

            correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)  # 测试值与真实值比较
            correct = tf.reduce_sum(correct)        # 统计正确的

            total_num += x.shape[0]  # 样本数
            total_correct += int(correct)
        acc = total_correct / total_num
        print(epoch, 'acc', acc)

数据集的形状十分重要，无论是加载后数据集还是要预处理的数据集，都应确保其 shape 准确，否则无法代入网络进行训练

若三，四步有看不懂的地方，可以参考mnist数据集实战那篇文章哦！
 

五、完整程序

# -*- codeing = utf-8 -*-
# @Time : 16:15
# @Author:Paranipd
# @File : cifar100_test.py
# @Software:PyCharm

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

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

tf.random.set_seed(2345)

# 预处理
def preprocess(x, y):
    # x :[-1,1]
    x = 2 * tf.cast(x, dtype=tf.float32) / 255 - 1
    y = tf.cast(y, dtype=tf.int32)
    return x, y


# 数据集加载
(x, y), (x_text, y_text) = datasets.cifar100.load_data()
print("y:",y.shape)
# 压缩最后一个维度为1
y = tf.squeeze(y)
y_text = tf.squeeze(y_text, axis=1)
print('squeeze:',x.shape, y.shape, x_text.shape, y_text.shape)


# 创建batch
train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.map(preprocess).shuffle(1000).batch(128)

test_db = tf.data.Dataset.from_tensor_slices((x, y))
test_db = test_db.map(preprocess).batch(128)

# 获取下一个batch
sample = next(iter(test_db))
print('sample:', sample[0].shape, sample[1].shape,
      tf.reduce_min(sample[0]), tf.reduce_max(sample[0]))


conv_layers = [  # 5 units : conv + conv + max pooling
    # units1  64个3x3 卷积核, 输入输出同大小
    layers.Conv2D(64, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(64, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    # 高宽减半
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # units 2 由于上一层池化层减半，下一层将卷积层的卷积核翻一倍，为了弥补信息特征的减少
    layers.Conv2D(128, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(128, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

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

    # units 4
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # units 5
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same')
]


def main():

    # [b,32,32,3] => [b,1,1,512]
    conv_net = Sequential(conv_layers)

    fc_net = Sequential([
        layers.Dense(256, activation=tf.nn.relu),
        layers.Dense(128, activation=tf.nn.relu),
        layers.Dense(10, activation=None)
    ])

    conv_net.build(input_shape=[None, 32, 32, 3])
    fc_net.build(input_shape=[None, 512])
    conv_net.summary()
    fc_net.summary()
    optimizer = optimizers.Adam(learning_rate=1e-4)

    # 可训练的变量 两个网络层之和
    variables = conv_net.trainable_variables + fc_net.trainable_variables

    for epoch in range(50):
        for step, (x, y) in enumerate(train_db):
            # 梯度求导
            with tf.GradientTape() as tape:
                # [b, 32,32,3 ] =>[b,1,1,512]
                out = conv_net(x)
                out = tf.reshape(out, [-1, 512])

                # [b, 512] => [b, 10]
                logits = fc_net(out)
                # [b,10] =>
                y_onehot = tf.one_hot(y, depth=10)
                # loss
                loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
                loss = tf.reduce_mean(loss)
            # 求导
            grads = tape.gradient(loss, variables)
            # 更新参数
            optimizer.apply_gradients(zip(grads, variables))

            if step % 100 == 0:
                print(epoch, step, 'loss', float(loss))

        # 测试集
        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)  # 概率化，和为1
            pred = tf.argmax(prob, axis=1)  # 获得最大下标
            pred = tf.cast(pred, dtype=tf.int32)

            correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)  # 测试值与真实值比较
            correct = tf.reduce_sum(correct)        # 统计正确的

            total_num += x.shape[0]  # 样本数
            total_correct += int(correct)
        acc = total_correct / total_num
        print(epoch, 'acc', acc)


if __name__ == '__main__':
    main()