本笔记记录使用ResNet18网络结构，进行CIFAR100数据集的训练和验证。由于参数较多，训练时间会比较长，因此只跑了10个epoch，准确率还没有提升上去。

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

os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
#tf.random.set_seed(12345)
tf.__version__


#关于ResNet的描述，可以参考如下链接：
#https://blog.csdn.net/qq_39770163/article/details/126169080
#代码基于ResNet18结构，有少许不一样
class BasicBlock(layers.Layer):
    def __init__(self, filter_num, strides = 1):
        super(BasicBlock, self).__init__()
        #卷积层1
        self.conv1 = layers.Conv2D(filter_num, (3,3), strides = strides, padding='same')
        #BN层
        self.bn1 = layers.BatchNormalization()
        #Relu层
        self.relu = layers.Activation('relu')

        #卷积层2,BN层2,
        self.conv2 = layers.Conv2D(filter_num, (3,3), strides = 1, padding='same')
        self.bn2 = layers.BatchNormalization()

        #Shortcut
        if strides != 1:
            #如果strides不为1，需要下采样
            self.downsample = Sequential()
            self.downsample.add(layers.Conv2D(filter_num, (1,1), strides=strides))
        else:
            #strides为1， 直接返回原始值即可
            self.downsample = lambda x:x
        
    def call(self, inputs, training = None):
        #经过第一个卷积层,BN和Relu
        out = self.conv1(inputs)
        out = self.bn1(out)
        out = self.relu(out)

        #经过第二个卷积层
        out = self.conv2(out)
        out = self.bn2(out)

        #Shortt处理，out和输入相加
        identity = self.downsample(inputs)
        output = layers.add([out, identity])
        #再经过一个relu
        output = tf.nn.relu(output)
        return output

class ResNet(keras.Model):
    #layer_dims表示对应位置的ResBlock包含了几个BasicBlock
    #比如[2,2,2,2] => 总共4个ResBlock，每个ResBlock包含两个BasicBlock
    #num_classes表示输出的类别的个数
    def __init__(self, layer_dims, num_classes=100):
        super(ResNet, self).__init__()
        #预处理单元
        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')
                               ])
        #创建中间ResBlock层
        self.layer1 = self.buildResBlock(64, layer_dims[0])
        self.layer2 = self.buildResBlock(128, layer_dims[1], strides=2)
        self.layer3 = self.buildResBlock(256, layer_dims[2], strides=2)
        self.layer4 = self.buildResBlock(512, layer_dims[3], strides=2)

        #自适应输出层
        self.avgpool = layers.GlobalAveragePooling2D()
        #全连接层
        self.fc = layers.Dense(num_classes)

    def call(self, inputs, training = None):
        x = self.stem(inputs)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        #经过avgpool => [b, 512]
        x = self.avgpool(x)
        #经过Dense => [b, 100]
        x = self.fc(x)
        return x

    def buildResBlock(self, filter_num, blocks, strides = 1):
        resBlocks = Sequential()
        resBlocks.add(BasicBlock(filter_num, strides))
        #后续的resBlock的strides都设置为1
        for _ in range(1, blocks):
            resBlocks.add(BasicBlock(filter_num))
        return resBlocks;

def ResNet18():
    return ResNet([2, 2, 2 ,2]);

def ResNet34():
    return ResNet([3, 4, 6, 3])


#加载CIFAR100数据集
#如果下载很慢，可以使用迅雷下载到本地，迅雷的链接也可以直接用官网URL：
#      https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz
#下载好后，将cifar-100.python.tar.gz放到 .keras\datasets 目录下（我的环境是C:\Users\Administrator\.keras\datasets）
# 参考：https://blog.csdn.net/zy_like_study/article/details/104219259
(x_train,y_train), (x_test, y_test) = datasets.cifar100.load_data()
print("Train data shape:", x_train.shape)
print("Train label shape:", y_train.shape)
print("Test data shape:", x_test.shape)
print("Test label shape:", y_test.shape)

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

y_train = tf.squeeze(y_train, axis=1)
y_test = tf.squeeze(y_test, axis=1)

batch_size = 128
train_db = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_db = train_db.shuffle(1000).map(preprocess).batch(batch_size)

test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_db = test_db.map(preprocess).batch(batch_size)

sample = next(iter(train_db))
print("Train data sample：", sample[0].shape, sample[1].shape, 
         tf.reduce_min(sample[0]), tf.reduce_max(sample[0]))



def main():
    #创建ResNet
    resNet = ResNet18()
    resNet.build(input_shape=[None, 32, 32, 3])
    resNet.summary()
    
    #设置优化器
    optimizer = optimizers.Adam(learning_rate=1e-3)
    #进行训练
    num_epoches = 10
    for epoch in range(num_epoches):
        for step, (x,y) in enumerate(train_db):
            with tf.GradientTape() as tape:
                #[b, 32, 32, 3] => [b, 100]
                logits = resNet(x)
                #标签做one_hot encoding
                y_onehot = tf.one_hot(y, depth=100)
                #计算损失
                loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
                loss = tf.reduce_mean(loss)
            #计算梯度
            grads = tape.gradient(loss, resNet.trainable_variables)
            #更新参数
            optimizer.apply_gradients(zip(grads, resNet.trainable_variables))

            if (step % 100 == 0):
                print("Epoch[", epoch + 1, "/", num_epoches, "]: step - ", step, " loss:", float(loss))
        #进行验证
        total_samples = 0
        total_correct = 0
        for x,y in test_db:
            logits = resNet(x)
            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_samples += x.shape[0]
            total_correct += int(correct)

        #统计准确率
        acc = total_correct / total_samples
        print("Epoch[", epoch + 1, "/", num_epoches, "]: accuracy:", acc)

if __name__ == '__main__':
    main()

运行结果：