tf中的常见卷积网络（残差网络ResNet）

tf中的常见网络

1. tf中常见的卷积网络介绍

• LeNet-5：2卷积层，2下采样层，3全连层
• AlexNet：添加了池化层
• VGG：运算量较大，卷积+池化+全连（如tf中的卷积神经网络实战中网络）
• GoogLeNet：在一个层中使用不同的卷积核
• ResNet：有效解决了因层数增加引起的训练困难问题（层数的加深导致后面层的梯度很难传播到前面的层）
• DenseNet：ResNet的短接线不只是与前面某一层相连，而是与所有层相连

2. 1x1卷积核的作用

• 降/升维，具体只改变通道的维度

3. ResNet（残差网络）

• 网络结构：
  
• 基本思想：随着网络层数的增加，准确率并不会一直增加（在22层左右开始出现此情况），ResNet在超过可训练的最大层数外的层加上一条连接线，当训练效果不好时，遗忘下面多余的层，获得至少准确率一直增加层数的极限；当训练效果好时（即多加的层能有训练效果时），便获得比准确率一直增加层数的极限好的准确率
• 残差的含义：最后的输出与输入x的关系是H(x)，中间卷积层的状态函数是F(x)，则输出输入之间的关系如下，可以理解为残差$$F(x)=H(x)-x$$
• 一般将一个残差网络（如上图结构所示）称为基本残差网路结构（Basic Block），将一组Basic Block称为Res Block。其中Basic Block实现如下

class BasicBlock(layers.Layer):
    def __int__(self, filter_num, stride=1):
        super(BasicBlock, self).__init__()
        # 一般将卷积层，正态层，relu层组成一个unit（单元）
        self.conv1 = layers.Conv2D(filter_num, (3,3), strides=stride, padding='same')
        self.bn1 = layers.BatchNormalization()
        self.relu = layers.Activation('relu')
        self.conv1 = layers.Conv2D(filter_num, (3,3), strides=1 padding='same')
        self.bn1 = layers.BatchNormalization()
        # 保证输出与遗忘线所连处的维度一致
        if stride !=1:
            self.downsample = Sequential()
            self.downsample.add(layers.Conv2D(filter_num, (1,1), strides=strides))
            self.downsample.add(layers.BatchNormalization())
        else:
            self.downsample = lambda x: x    

    def call(self, inputs, training=None):
        residual = self.downsample(inputs)
        conv1 = self.conv1(inputs) 
        bn1 = self.bn1(conv1)
        relu1 = self.relu(bn1)
        conv2 = self.conv2(relu1) 
        bn2 = self.bn2(conv2)
        # 添加遗忘线
        add = layers.add([bn2, residual])
        out = self.relu(add)
        return out    

Res Block实现如下

def _build_resblock(self, block, filter_num, blocks, stride=1):
    res_blocks = keras.Sequential()
    res_blocks.add(block(filter_num, stride))
    for _ in range(1, blocks):
        res_blocks.add(block(filter_num, stride=1))
    return res_blocks

4. CFAR100实战（基于ResNet18）

• strides参数详解
• resnet18代码

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers, Sequential


# 设计BasicBlock
class BasicBlock(layers.Layer):
    def __init__(self, filter_num, stride=1):
        # filter_num指的是卷积核个数
        super(BasicBlock, self).__init__()
        self.conv1 = layers.Conv2D(filter_num, (3, 3), strides=stride, padding='same')
        self.bn1 = layers.BatchNormalization()
        self.relu = layers.Activation('relu')
        self.conv2 = layers.Conv2D(filter_num, (3, 3), strides=1, padding='same')
        self.bn2 = layers.BatchNormalization()
        # 更改短接线数据的维度，保持与输出一致（这样才可做相加）
        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):
        out = self.conv1(inputs)
        out = self.bn1(out, training=training)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        identify = self.downsample(inputs)  # 短接线
        output = layers.add([out, identify])  # 将短接线与输出相连
        output = tf.nn.relu(output)
        return output


# 设计ResBlock
class ResNet(keras.Model):
    # layer_dims指一个BasicBlock包含几个BasicBlock，num_class指分类的类别总数目
    def __init__(self, layer_dims, num_class=100):
        super(ResNet, self).__init__()
        # ResNet18的第一层(第一层不是ResBlock)
        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')
        ])
        # 四层ResBlock
        self.layer1 = self.build_resblock(64, layer_dims[0])
        self.layer2 = self.build_resblock(128, layer_dims[1], stride=2)  # stride=2表示下采样（大于1即可）
        self.layer3 = self.build_resblock(256, layer_dims[2], stride=2)
        self.layer4 = self.build_resblock(512, layer_dims[3], stride=2)
        # ResNet18的最后一层，输出为[b,512,h,w]
        self.avgpool = layers.GlobalAveragePooling2D()
        self.fc = layers.Dense(num_class)

    def call(self, inputs, training=None):
        x = self.stem(inputs, training=training)
        x = self.layer1(x, training=training)
        x = self.layer2(x, training=training)
        x = self.layer3(x, training=training)
        x = self.layer4(x, training=training)
        x = self.avgpool(x)  # 输出为[b,c]
        x = self.fc(x)  # 输出为[b,100]
        return x

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


def resnet18():
    return ResNet([2, 2, 2, 2])  # ResNet18的网络结构，指每组ResBlock含有BasicBlock的数量

• main函数代码

import tensorflow as tf
from tensorflow.keras import layers, optimizers, datasets, Sequential
from resnet import resnet18
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'


def pre_process(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255. - 0.5
    y = tf.cast(y, dtype=tf.int32)
    return x, y


batch_size = 128
(x, y), (x_test, y_test) = datasets.cifar100.load_data()
# 将标签降维
y = tf.squeeze(y, axis=1)
y_test = tf.squeeze(y_test, axis=1)
train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.shuffle(1000).map(pre_process).batch(batch_size)
test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_db = test_db.map(pre_process).batch(batch_size)
sample = next(iter(train_db))


def main():
    model = resnet18()
    model.build(input_shape=(None, 32, 32, 3))
    model.summary()  # 打印参数信息
    optimizer = optimizers.Adam(lr=1e-3)
    for epoch in range(300):
        for step, (x, y) in enumerate(train_db):
            with tf.GradientTape() as tape:
                output = model(x, training=None)
                y_onehot = tf.one_hot(y, depth=100)
                loss = tf.losses.categorical_crossentropy(y_onehot, output, from_logits=True)
                loss = tf.reduce_mean(loss)
            grads = tape.gradient(loss, model.trainable_variables)
            optimizer.apply_gradients(zip(grads, model.trainable_variables))
            if step % 100 == 0:
                print(epoch, step, 'loss:', float(loss))

        # 测试
        total_num = 0
        total_correct = 0
        for x,y in test_db:
            output = model(x, training=False)
            prob = tf.nn.softmax(output, 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)
        accuracy = total_correct / total_num
        accuracy = accuracy * 100
        print(epoch, '准确率', accuracy, "%")


if __name__ == '__main__':
    main()