【图像分类】TensorFlow2.7版本搭建NIN网络

NIN网络结构

注解：这里为了简单起见，只是模拟NIN网络结构，本代码只是采用3个mlpconv层和最终的全局平均池化输出层，每个mlpconv层中包含了3个1*1卷积层

mlpconv层
1*1卷积只是会改变通道维数并不会改变feature map的大小，它可以变向起到一个通道交叉全连接的作用

self.mlpconv1 = Sequential([
            Conv2D(filters=3,
                   kernel_size=1),
            ReLU(),
            Conv2D(filters=3,
                   kernel_size=1),
            ReLU(),
            Conv2D(filters=3,
                   kernel_size=1)]
        )

全局平均池化层
GlobalAveragePooling会将每个feature map所有的值相加取均值然后将这个实数作为该通道的特征值，NIN网络结构采用全局平均池化代替传统输出层使用MLP结构，这样有效防止过拟合，如果我们的任务存在1000个分类，那么我们最终的输出层的feature map的个数也为1000，然后对其进行全局平均池化，每个feature map代表一个类别，会形成一个1*1*1000的特征图，也就是一个维度为1000的特征向量，然后进行softmax操作

self.global_average_pool = GlobalAveragePooling2D()

"""
 * Created with PyCharm
 * 作者: 阿光
 * 日期: 2021/1/12
 * 时间: 22:35
 * 描述: 作者原文中的手写数据集是32*32，这里mnist是28*28，所以在训练前修改了图像尺寸
        还有一种解决方式就是在第一个卷积层使用padding='same'进行填充，这样就保证了使用第一个卷积层后尺寸为28*28
        之后仍可正常进行
"""
import tensorflow as tf
from keras import Sequential
from tensorflow.keras.layers import *


class NIN(tf.keras.Model):
    def __init__(self, output_dim=10):
        super(NIN, self).__init__()
        self.mlpconv1 = Sequential([
            Conv2D(filters=3,
                   kernel_size=1),
            ReLU(),
            Conv2D(filters=3,
                   kernel_size=1),
            ReLU(),
            Conv2D(filters=3,
                   kernel_size=1)]
        )
        self.mlpconv2 = Sequential([
            Conv2D(filters=3,
                   kernel_size=1),
            ReLU(),
            Conv2D(filters=3,
                   kernel_size=1),
            ReLU(),
            Conv2D(filters=3,
                   kernel_size=1)]
        )
        self.mlpconv3 = Sequential([
            Conv2D(filters=3,
                   kernel_size=1),
            ReLU(),
            Conv2D(filters=3,
                   kernel_size=1),
            ReLU(),
            Conv2D(filters=output_dim,
                   kernel_size=1)]
        )
        self.global_average_pool = GlobalAveragePooling2D()

    def call(self, inputs):
        x = self.mlpconv1(inputs)
        x = self.mlpconv2(x)
        x = self.mlpconv3(x)
        x = self.global_average_pool(x)
        x = Softmax()(x)
        return x

调用模型，训练mnist数据集

"""
 * Created with PyCharm
 * 作者: 阿光
 * 日期: 2021/1/12
 * 时间: 22:20
 * 描述:
"""

import tensorflow as tf
from tensorflow.keras import Input

# step1:加载数据集
import model
import model_sequential

(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.mnist.load_data()

# step2:将图像归一化
train_images, test_images = train_images / 255.0, test_images / 255.0

# step3:将图像的维度变为(60000,28,28,1)
train_images = tf.expand_dims(train_images, axis=3)
test_images = tf.expand_dims(test_images, axis=3)

# step5:导入模型
# history = LeNet5()
history = model.NIN(10)

# 让模型知道输入数据的形式
history.build(input_shape=(1, 28, 28, 1))

# 结局Output Shape为 multiple
history.call(Input(shape=(28, 28, 1)))
history.summary()

# step6:编译模型
history.compile(optimizer='adam',
                loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                metrics=['accuracy'])

# 权重保存路径
checkpoint_path = "./weight/cp.ckpt"

# 回调函数，用户保存权重
save_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
                                                   save_best_only=True,
                                                   save_weights_only=True,
                                                   monitor='val_loss',
                                                   verbose=1)
# step7:训练模型
history = history.fit(train_images,
                      train_labels,
                      epochs=10,
                      batch_size=32,
                      validation_data=(test_images, test_labels),
                      callbacks=[save_callback])