卷积神经网络简介

隐藏神经元与输入的连接可以看作一种观察方式。图一中全连接层，第一个隐藏神经元与所有输入相连，相当于一次观察整幅图像；而图二卷积神经网络，一次观察一个小区域。卷积神经网络相对于全连接网络最大的特点就是参数量的大幅减少。

对比CV中的卷积核（经验得出）：

CNN中，w学习得到，可以采用多个卷积核提取特征，增加网络的抽象能力。

卷积神经网络中，一般随着层数增加，w、h下降，channel增加；前面卷积层提取底层特征（颜色、边缘），而后面的层提取的是高层特征（轮子、窗户等）。

（不常用）

卷积神经网络中的梯度下降

池化与采样

pooling

upsample

ReLU

一般，灰度值，白255，黑0。

CIFAR100实战

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)


# conv_layers: 5 units of conv + maxpooling
conv_layers = [

    # 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')
]

fc_layers = [
    layers.Dense(256, activation=tf.nn.relu),
    layers.Dense(128, activation=tf.nn.relu),
    layers.Dense(100, activation=None)

]


def preprocessing(x, y):

    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)
    return x, y



def main():

    # 加载数据
    (x, y), (x_test, y_test) = datasets.cifar100.load_data()
    y = tf.squeeze(y)
    y_test = tf.squeeze(y_test)

    train_db = tf.data.Dataset.from_tensor_slices((x, y))
    train_db = train_db.shuffle(1000).map(preprocessing).batch(64)

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


    # 构建网络
    # conv_net = Sequential(conv_layers)
    # x = tf.random.normal([4, 32, 32, 3])
    # out = conv_net(x)
    # print(out.shape)   ------------(4, 1, 1, 512)

    conv_net = Sequential(conv_layers)
    fc_net = Sequential(fc_layers)

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

    optimizer = optimizers.Adam(lr=1e-4)
    varibles = conv_net.trainable_variables + fc_net.trainable_variables

    for epoch in range(5):

        for step, (x, y) in enumerate(train_db):
            # print(step)

            y_onehot = tf.one_hot(y, depth=100)

            with tf.GradientTape() as tape:
                out = conv_net(x)   # out: [batch, 1, 1, 512]
                out = tf.reshape(out, [-1, 512])
                out = fc_net(out)   # out: [out, 100]

                loss = tf.losses.categorical_crossentropy(y_onehot, out, from_logits=True)
                # print(loss.shape)  # (64,)
                loss = tf.reduce_mean(loss)

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

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

        # 每个epoch做一个测试
        total_num = 0
        correct_num = 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)  # 遍历每个列，即为行
            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]
            correct_num += int(correct)

        acc = correct_num / total_num
        print(epoch, "test_acc:", acc)


if __name__ == '__main__':
    main()