VGG实战

卷积层参数：
卷积层 $conv3$ 表示卷积核尺寸为 $3\times 3$，$conv3-64$ 表示卷积后通道数为 $64$，所有卷积步长 $stride=1$，填充方式 $padding=same$（即输出大小和输入大小一致），这样卷积操作不会改变图像大小，仅改变通道数。

池化层参数：
池化层 $maxpooling$ 的参数均为 $2\times 2$，步长 $stride=2$

import tensorflow as tf
import tensorflow.keras as keras
import matplotlib.pyplot as plt
import numpy as np
from tensorflow.keras import datasets, layers, optimizers, models, regularizers
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

def normalize(x_train, x_test):
    x_train = x_train / 255.
    x_test = x_test / 255.

    mean = np.mean(x_train, axis=(0, 1, 2, 3))
    std = np.std(x_train, axis=(0, 1, 2, 3))
    print('mean:', mean, 'std', std)
    x_train = (x_train -  mean) / (std + 1e-7)
    x_test = (x_test- mean) / (std + 1e-7)
    return x_train, x_test

def preprocess(x, y):
    x = tf.cast(x, tf.float32)
    y = tf.cast(y, tf.int32)
    y = tf.squeeze(y, axis=1)
    y = tf.one_hot(y, depth=10)
    return x, y

(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()
x_train, x_test = normalize(x_train, x_test)

train_db = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_db = train_db.shuffle(50000).batch(128).map(preprocess)

test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_db = test_db.shuffle(10000).batch(128).map(preprocess)

num_classes = 10 # imggenet数据集，这里是1000

model = models.Sequential()

model.add(layers.Conv2D(64, (3, 3), padding='same', activation='relu'))
model.add(layers.Conv2D(64, (3, 3), padding='same', activation='relu'))
model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))

model.add(layers.Conv2D(128, (3, 3), padding='same', activation='relu'))
model.add(layers.Conv2D(128, (3, 3), padding='same', activation='relu'))
model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))

model.add(layers.Conv2D(256, (3, 3), padding='same', activation='relu'))
model.add(layers.Conv2D(256, (3, 3), padding='same', activation='relu'))
model.add(layers.Conv2D(256, (3, 3), padding='same', activation='relu'))
model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))

model.add(layers.Conv2D(512, (3, 3), padding='same', activation='relu'))
model.add(layers.Conv2D(512, (3, 3), padding='same', activation='relu'))
model.add(layers.Conv2D(512, (3, 3), padding='same', activation='relu'))
model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))

model.add(layers.Conv2D(512, (3, 3), padding='same', activation='relu'))
model.add(layers.Conv2D(512, (3, 3), padding='same', activation='relu'))
model.add(layers.Conv2D(512, (3, 3), padding='same', activation='relu'))
model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))

model.add(layers.Flatten())
model.add(layers.Dense(256, activation='relu')) # VGG 16 为4096
model.add(layers.Dense(128, activation='relu')) # VGG 16 为4096
model.add(layers.Dense(num_classes, activation='softmax')) # VGG 16 为1000

model.build(input_shape=(None, 32, 32, 3))
model.compile(optimizer=keras.optimizers.Adam(0.0001),
              loss=keras.losses.CategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])
history = model.fit(train_db, epochs=50)

plt.plot(history.history['loss'])
plt.title("model loss")
plt.ylabel("loss")
plt.xlabel("epoch")
plt.show()

model.evaluate(test_db)

这样的准确率最后仅为0.13左右，是很不理想的，所以，在每一层加入归一化和正则化操作，如下面代码，之后的准确率在0.8左右，就可以接受了，当然还可以再优化。

weight_decay = 0.000
model.add(layers.Conv2D(64, (3, 3),
          padding='same',
          kernel_regularizer=regularizers.l2(weight_decay)))
model.add(layers.Activation('relu'))
model.add(layers.BatchNormalization())
model.add(layers.Dropout(0.3)) #在每一个block的第一次卷积加入dropout