原文：https://makeoptim.com/deep-learning/yiai-alexnet-implementation
前言
上一篇笔者使用如何阅读深度学习论文的方法阅读了 AlexNet。为了加深理解，本文带大家使用 TensorFlow 2 Keras 实现 AlexNet CNN 架构。
网络结构
image
从上一篇可以得到 AlexNet 的各层网络结构如下：
- 第
1卷积层使用96个核对224 × 224 × 3的输入图像进行滤波，核大小为11 × 11 × 3，步长是4个像素。然后，使用归一化和池化操作。 - 第
2卷积层使用256个核进行滤波，核大小为5 × 5 × 48。然后，使用归一化和池化操作。 - 第
3，4，5卷积层互相连接，中间没有接入池化层或归一化层。 - 第
3卷积层有384个核，核大小为3 × 3 × 256。 - 第
4卷积层有384个核，核大小为3 × 3 × 192。 - 第
5卷积层有256个核，核大小为3 × 3 × 192。然后，使用池化操作。 - 第
6层先Flatten之后接入4096个神经元的全连接层。 - 第
7层是4096个神经元的全连接层。 - 第
8层是1000个神经元的全连接层，作为输出层。
实现
Sequential
使用 Keras API 可以很快“翻译”出网络结构。
model = keras.models.Sequential([
# layer 1
layers.Conv2D(
filters=96,
kernel_size=(11, 11),
strides=(4, 4),
activation=keras.activations.relu,
padding='valid',
input_shape=(227, 227, 3)),
layers.BatchNormalization(),
layers.MaxPool2D(pool_size=(3, 3), strides=(2, 2)),
# layer 2
layers.Conv2D(
filters=256,
kernel_size=(5, 5),
strides=(1, 1),
activation=keras.activations.relu,
padding='same'
),
layers.BatchNormalization(),
layers.MaxPool2D(pool_size=(3, 3), strides=(2, 2)),
# layer 3
layers.Conv2D(
filters=384,
kernel_size=(3, 3),
strides=(1, 1),
activation=keras.activations.relu,
padding='same'
),
layers.BatchNormalization(),
# layer 4
layers.Conv2D(
filters=384,
kernel_size=(3, 3),
strides=(1, 1),
activation=keras.activations.relu,
padding='same'
),
layers.BatchNormalization(),
# layer 5
layers.Conv2D(
filters=256,
kernel_size=(3, 3),
strides=(1, 1),
activation=keras.activations.relu,
padding='same'
),
layers.BatchNormalization(),
layers.MaxPool2D(pool_size=(3, 3), strides=(2, 2)),
# layer 6
layers.Flatten(),
layers.Dense(units=4096, activation=keras.activations.relu),
layers.Dropout(rate=0.5),
# layer 7
layers.Dense(units=4096, activation=keras.activations.relu),
layers.Dropout(rate=0.5),
# layer 8
layers.Dense(units=1000, activation=keras.activations.softmax)
])
model.summary()
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 55, 55, 96) 34944
_________________________________________________________________
batch_normalization (BatchNo (None, 55, 55, 96) 384
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 27, 27, 96) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 27, 27, 256) 614656
_________________________________________________________________
batch_normalization_1 (Batch (None, 27, 27, 256) 1024
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 13, 13, 256) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 13, 13, 384) 885120
_________________________________________________________________
batch_normalization_2 (Batch (None, 13, 13, 384) 1536
_________________________________________________________________
conv2d_3 (Conv2D) (None, 13, 13, 384) 1327488
_________________________________________________________________
batch_normalization_3 (Batch (None, 13, 13, 384) 1536
_________________________________________________________________
conv2d_4 (Conv2D) (None, 13, 13, 256) 884992
_________________________________________________________________
batch_normalization_4 (Batch (None, 13, 13, 256) 1024
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 6, 6, 256) 0
_________________________________________________________________
flatten (Flatten) (None, 9216) 0
_________________________________________________________________
dense (Dense) (None, 4096) 37752832
_________________________________________________________________
dropout (Dropout) (None, 4096) 0
_________________________________________________________________
dense_1 (Dense) (None, 4096) 16781312
_________________________________________________________________
dropout_1 (Dropout) (None, 4096) 0
_________________________________________________________________
dense_2 (Dense) (None, 1000) 4097000
=================================================================
Total params: 62,383,848
Trainable params: 62,381,096
Non-trainable params: 2,752
_________________________________________________________________
注：
这里使用
BatchNormalization代替了原论文中的Local Response Normalization，关于两者的区别详见 https://towardsdatascience.com/difference-between-local-response-normalization-and-batch-normalization-272308c034ac。当然，如果大家还是想用Local Response Normalization的话，可以使用layers.Lambda(tf.nn.local_response_normalization)。AlexNet 使用重叠池化
并且s=2，z=3，因此，池化层为MaxPool2D(pool_size=(3, 3), strides=(2, 2))
Subclassing
除了使用 Sequential 实现模型外，还可以使用子类的形式，详见 Making new Layers and Models via subclassing。
class AlexNet(keras.Model):
def __init__(self, num_classes, input_shape=(227, 227, 3)):
super(AlexNet, self).__init__()
self.input_layer = layers.Conv2D(
filters=96,
kernel_size=(11, 11),
strides=(4, 4),
activation=keras.activations.relu,
padding='valid',
input_shape=input_shape)
self.middle_layers = [
layers.BatchNormalization(),
layers.MaxPool2D(pool_size=(3, 3), strides=(2, 2)),
layers.Conv2D(
filters=256,
kernel_size=(5, 5),
strides=(1, 1),
activation=keras.activations.relu,
padding='same'
),
layers.BatchNormalization(),
layers.MaxPool2D(pool_size=(3, 3), strides=(2, 2)),
layers.Conv2D(
filters=384,
kernel_size=(3, 3),
strides=(1, 1),
activation=keras.activations.relu,
padding='same'
),
layers.BatchNormalization(),
layers.Conv2D(
filters=384,
kernel_size=(3, 3),
strides=(1, 1),
activation=keras.activations.relu,
padding='same'
),
layers.BatchNormalization(),
layers.Conv2D(
filters=256,
kernel_size=(3, 3),
strides=(1, 1),
activation=keras.activations.relu,
padding='same'
),
layers.BatchNormalization(),
layers.MaxPool2D(pool_size=(3, 3), strides=(2, 2)),
layers.Flatten(),
layers.Dense(units=4096, activation=keras.activations.relu),
layers.Dropout(rate=0.5),
layers.Dense(units=4096, activation=keras.activations.relu),
layers.Dropout(rate=0.5),
]
self.out_layer = layers.Dense(
units=num_classes, activation=keras.activations.softmax)
def call(self, inputs):
x = self.input_layer(inputs)
for layer in self.middle_layers:
x = layer(x)
probs = self.out_layer(x)
return probs
model = AlexNet(1000)
model.build((None, 227, 227, 3))
model.summary()
Model: "alex_net"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_5 (Conv2D) multiple 34944
_________________________________________________________________
batch_normalization_5 (Batch multiple 384
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 multiple 0
_________________________________________________________________
conv2d_6 (Conv2D) multiple 614656
_________________________________________________________________
batch_normalization_6 (Batch multiple 1024
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 multiple 0
_________________________________________________________________
conv2d_7 (Conv2D) multiple 885120
_________________________________________________________________
batch_normalization_7 (Batch multiple 1536
_________________________________________________________________
conv2d_8 (Conv2D) multiple 1327488
_________________________________________________________________
batch_normalization_8 (Batch multiple 1536
_________________________________________________________________
conv2d_9 (Conv2D) multiple 884992
_________________________________________________________________
batch_normalization_9 (Batch multiple 1024
_________________________________________________________________
max_pooling2d_5 (MaxPooling2 multiple 0
_________________________________________________________________
flatten_1 (Flatten) multiple 0
_________________________________________________________________
dense_3 (Dense) multiple 37752832
_________________________________________________________________
dropout_2 (Dropout) multiple 0
_________________________________________________________________
dense_4 (Dense) multiple 16781312
_________________________________________________________________
dropout_3 (Dropout) multiple 0
_________________________________________________________________
dense_5 (Dense) multiple 4097000
=================================================================
Total params: 62,383,848
Trainable params: 62,381,096
Non-trainable params: 2,752
_________________________________________________________________
注：自定义 Model，在 summary 之前需要 build，否则会报如下错误：
ValueError: This model has not yet been built. Build the model first by calling
build()or callingfit()with some data, or specify aninput_shapeargument in the first layer(s) for automatic build.
Demo
在 AlexNet 论文中，不但有网络结构，还有数据集、数据增强等细节。为了帮助大家更好地理解 AlexNet，笔者在TensorFlow 案例的基础上替换成自定义的 AlexNet 模型进行训练，详见 https://github.com/CatchZeng/YiAI-examples/blob/master/papers/AlexNet/AlexNet.ipynb。
小结
实践出真知，从阅读到实践，是一个提升的过程。在实践中，不但可以了解到实现的细节，而且还能熟悉 TensorFlow 的生态。强烈推荐大家，多看论文，并实践。
