深度学习之卷积神经网络（10）CIFAR10与VGG13实战

CIFAR10数据集

 在TensorFlow中，同样地，不需要手动下载、解析和加载CIFAR10数据集，通过datasets.cifar10.load_data()函数就看人义直接加载切割好的训练集和测试集。例如:

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

import os

from Chapter10.CIFAR10 import load_data

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


def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)

    return x, y


# 在线下载，加载CIFAR10数据集
(x, y), (x_test, y_test) = load_data('/Users/xuruihang/.keras/datasets/cifar-10-batches-py')
# 删除y的一个维度，[b,1] => [b]
y = tf.squeeze(y, axis=1)
y_test = tf.squeeze(y_test, axis=1)
# 打印训练集和测试集的形状
print(x.shape, y.shape, x_test.shape, y_test.shape)
# 构建训练集对象，随机打乱，预处理，批量化
train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.shuffle(1000).map(preprocess).batch(128)
# 构建测试集对象，预处理，批量化
test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
train_db = train_db.map(preprocess).batch(128)
# 从训练集中采样一个Batch，并观察
sample = next(iter(train_db))
print('sample: ', sample[0].shape, sample[1].shape,
      tf.reduce_min(sample[0]), tf.reduce_max(sample[0]))

运行结果如下图所示:

注: 这里的load_data()是调用自己写的一段代码，因为直接下载会报错:

import numpy as np
import os

def load_batch(file):
    import pickle
    with open(file, 'rb') as fo:
        d = pickle.load(fo, encoding='bytes')
        d_decoded = {}
        for k, v in d.items():
            d_decoded[k.decode('utf8')] = v
        d = d_decoded
        data = d['data']
        labels = d['labels']
        data = data.reshape(data.shape[0], 3, 32, 32)
    return data, labels

def load_data(path ='data/cifar-10-batches-py'):
    """Loads CIFAR10 dataset.
    # Returns
        Tuple of Numpy arrays: `(x_train, y_train), (x_test, y_test)`.
    """
    from tensorflow.python.keras import backend as K

    num_train_samples = 50000

    x_train = np.empty((num_train_samples, 3, 32, 32), dtype='uint8')
    y_train = np.empty((num_train_samples,), dtype='uint8')

    for i in range(1, 6):
        fpath = os.path.join(path, 'data_batch_' + str(i))
        (x_train[(i - 1) * 10000: i * 10000, :, :, :],
         y_train[(i - 1) * 10000: i * 10000]) = load_batch(fpath)

    fpath = os.path.join(path, 'test_batch')
    x_test, y_test = load_batch(fpath)

    y_train = np.reshape(y_train, (len(y_train), 1))
    y_test = np.reshape(y_test, (len(y_test), 1))

    if K.image_data_format() == 'channels_last':
        x_train = x_train.transpose(0, 2, 3, 1)
        x_test = x_test.transpose(0, 2, 3, 1)

    return (x_train, y_train), (x_test, y_test)


(x_train, y_train), (x_test, y_test) = load_data('/Users/xuruihang/.keras/datasets/cifar-10-batches-py')

详见Keras CIFAR10离线加载

 可以看到，上述代码运行后，得到的训练集的$\mathbf{X}$和$\mathbf{y}$形状为: $(50000,32,32,3)$和$(50000)$，测试集的$\mathbf{X}$和$\mathbf{y}$形状为: $(10000,32,32,3)$和$(10000)$，分别代表了图片大小为$32\times 32$，彩色图片，训练集样本数为50000，测试集样本数为10000。

 CIFAR10图片识别任务并不简单，这主要是由于CIFAR10的图片内容需要大量细节才能呈现，而保存的图片分辨率仅有$32\times 32$，使得主题部分信息较为模糊，甚至人眼都很难分辨。浅层的神经网络表达能力有限，很难训练优化到较好的性能，本节将基于表达能力更强的VGG13网络，根据我们的数据集特点修改部分网络结构，完成CIFAR10图片识别。修改如下:

• 将网络输入调整为32×32。原网络输入为$224\times 224$，导致全连
  接层输入特征维度过大，网络参数量过大。
• 3个全连接层的维度调整为$[256,64,10]$，满足10分类任务的设定。

 下图是调整后的VGG13网络结构，我们统称之为VGG13网络模型。

调整的VGG13模型结构

 我们将网络实现为两个子网络: 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，代码如下:

conv_layers = [  # 先创建包含多网络层的列表
    # Conv-Conv-Pooling单元1
    # 64个3×3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'), 
    
    # Conv-Conv-Pooling单元2，输出通道提升至128，高宽大小减半
    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'), 
    
    # Conv-Conv-Pooling单元3，输出通道提升至256，高宽大小减半
    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'), 
    
    # Conv-Conv-Pooling单元4，输出通道提升至512，高宽大小减半
    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'), 
    
    # Conv-Conv-Pooling单元5，输出通道提升至512，高宽大小减半
    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'), 
]
# 利用前面创建的层列表构建网络容器
conv_net = Sequential(conv_layers)

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。代码如下:

# 创建3层全连接层子网络
fc_net = Sequential([
    layers.Dense(256, activation=tf.nn.relu), 
    layers.Dense(128, activation=tf.nn.relu), 
    layers.Dense(10, activation=None)
])

子网络创建完成后，通过如下代码查看网络的参数量:

conv_net.build(input_shape=[None, 32, 32, 3])
fc_net.build(input_shape=[None, 512])
conv_net.summary()
fc_net.summary()

卷积网络总参数量约为940万个，全连接网络总参数量约为17.7万个，网络总参数量约为950万个，相比于原始版本的VGG13参数量减少了很多。

 由于我们将网络实现为2个子网络，在进行梯度更新时，需要合并2个子网络的待优化参数列表。代码如下:

# 列表合并，合并2个子网络的参数
variables = conv_net.trainable_variables + fc_net.trainable_variables
# 对所有参数求梯度
grads = tape.gradient(loss, variables)
# 自动更新
optimizer.apply_gradients(zip(grads, variables))

运行代码即可开始训练模型，在训练完50个Epoch后，网络的测试准确率达到了77.5%。
完整代码:

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

from Chapter10.CIFAR10 import load_data

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

conv_layers = [ # 5 units of conv + max pooling
    # unit 1
    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),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    # unit 2
    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'),

    # unit 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'),

    # unit 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'),

    # unit 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 preprocess(x, y):
    # [0~1]
    x = 2*tf.cast(x, dtype=tf.float32) / 255.-1
    y = tf.cast(y, dtype=tf.int32)
    return x,y


# 在线下载，加载CIFAR10数据集
(x, y), (x_test, y_test) = load_data('/Users/xuruihang/.keras/datasets/cifar-10-batches-py')
y = tf.squeeze(y, axis=1)
y_test = tf.squeeze(y_test, axis=1)
print(x.shape, y.shape, x_test.shape, y_test.shape)


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

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]))


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(lr=1e-4)

    # [1, 2] + [3, 4] => [1, 2, 3, 4]
    # 列表合并，合并2个子网络的参数
    variables = conv_net.trainable_variables + fc_net.trainable_variables

    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)
                # flatten, => [b, 512]
                out = tf.reshape(out, [-1, 512])
                # [b, 512] => [b, 10]
                logits = fc_net(out)
                # [b] => [b, 10]
                y_onehot = tf.one_hot(y, depth=10)
                # compute loss
                loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
                loss = tf.reduce_mean(loss)

            grads = tape.gradient(loss, variables)
            optimizer.apply_gradients(zip(grads, variables))

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



        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)



if __name__ == '__main__':
    main()

运行结果如下图所示:

可以看到，准确率达到了77.41%。（破程序运行了一晚上，电脑直接起飞了）