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])
