keras上使用alexnet模型来高准确度对mnist数据进行分类

纲要

本文有两个特点：一是直接对本地mnist数据进行读取（假设事先已经下载或从别处拷来）二是基于keras框架（网上多是基于tf）使用alexnet对mnist数据进行分类，并获得较高准确度（约为98%）

本地数据读取和分析

很多代码都是一开始简单调用一行代码来从网站上下载mnist数据，虽然只有10来MB，但是现在下载速度非常慢，而且经常中途出错，要费很大的劲才能拿到数据。

(X_train, y_train), (X_test, y_test) = mnist.load_data() 

 其实可以单独来获得这些数据（一共4个gz包，如下所示），然后调用别的接口来分析它们。

mnist = input_data.read_data_sets("./MNIST_data", one_hot = True) #导入已经下载好的数据集,"./MNIST_data"为存放mnist数据的目录

x_train = mnist.train.images
y_train = mnist.train.labels
x_test = mnist.test.images
y_test = mnist.test.labels

这里面要注意的是，两种接口拿到的数据形式是不一样的。 从网上直接下载下来的数据 其image data值的范围是0~255，且label值为0,1,2,3...9。 而第二种接口获取的数据 image值已经除以255（归一化）变成0~1范围，且label值已经是one-hot形式（one_hot=True时），比如label值2的one-hot code为(0 0 1 0 0 0 0 0 0 0)

所以，以第一种方式获取的数据需要做一些预处理（归一和one-hot）才能输入网络模型进行训练 而第二种接口拿到的数据则可以直接进行训练。

Alexnet模型的微调

按照公开的模型框架，Alexnet只有第1、2个卷积层才跟着BatchNormalization，后面三个CNN都没有（如有说错，请指正）。如果按照这个来搭建网络模型，很容易导致梯度消失，现象就是 accuracy值一直处在很低的值。 如下所示。

在每个卷积层后面都加上BN后，准确度才迭代提高。如下所示

完整代码

 

import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.layers.convolutional import Conv2D, MaxPooling2D, ZeroPadding2D
from keras.layers.normalization import BatchNormalization
from keras.callbacks import ModelCheckpoint
import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data  #tensorflow已经包含了mnist案例的数据

batch_size = 64
num_classes = 10
epochs = 10
img_shape = (28,28,1)

# input dimensions
img_rows, img_cols = 28,28

# dataset input
#(x_train, y_train), (x_test, y_test) = mnist.load_data()
mnist = input_data.read_data_sets("./MNIST_data", one_hot = True) #导入已经下载好的数据集,"./MNIST_data"为存放mnist数据的目录
print(mnist.train.images.shape, mnist.train.labels.shape)
print(mnist.test.images.shape, mnist.test.labels.shape)
print(mnist.validation.images.shape, mnist.validation.labels.shape)

x_train = mnist.train.images
y_train = mnist.train.labels
x_test = mnist.test.images
y_test = mnist.test.labels

# data initialization
x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
input_shape = (img_rows, img_cols, 1)

# Define the input layer
inputs = keras.Input(shape = [img_rows, img_cols, 1])

 #Define the converlutional layer 1
conv1 = keras.layers.Conv2D(filters= 64, kernel_size= [11, 11], strides= [1, 1], activation= keras.activations.relu, use_bias= True, padding= 'same')(inputs)
# Define the pooling layer 1
pooling1 = keras.layers.AveragePooling2D(pool_size= [2, 2], strides= [2, 2], padding= 'valid')(conv1)
# Define the standardization layer 1
stand1 = keras.layers.BatchNormalization(axis= 1)(pooling1)

# Define the converlutional layer 2
conv2 = keras.layers.Conv2D(filters= 192, kernel_size= [5, 5], strides= [1, 1], activation= keras.activations.relu, use_bias= True, padding= 'same')(stand1)
# Defien the pooling layer 2
pooling2 = keras.layers.AveragePooling2D(pool_size= [2, 2], strides= [2, 2], padding= 'valid')(conv2)
# Define the standardization layer 2
stand2 = keras.layers.BatchNormalization(axis= 1)(pooling2)

# Define the converlutional layer 3
conv3 = keras.layers.Conv2D(filters= 384, kernel_size= [3, 3], strides= [1, 1], activation= keras.activations.relu, use_bias= True, padding= 'same')(stand2)
stand3 = keras.layers.BatchNormalization(axis=1)(conv3)

# Define the converlutional layer 4
conv4 = keras.layers.Conv2D(filters= 384, kernel_size= [3, 3], strides= [1, 1], activation= keras.activations.relu, use_bias= True, padding= 'same')(stand3)
stand4 = keras.layers.BatchNormalization(axis=1)(conv4)

# Define the converlutional layer 5
conv5 = keras.layers.Conv2D(filters= 256, kernel_size= [3, 3], strides= [1, 1], activation= keras.activations.relu, use_bias= True, padding= 'same')(stand4)
pooling5 = keras.layers.AveragePooling2D(pool_size= [2, 2], strides= [2, 2], padding= 'valid')(conv5)
stand5 = keras.layers.BatchNormalization(axis=1)(pooling5)

# Define the fully connected layer
flatten = keras.layers.Flatten()(stand5)
fc1 = keras.layers.Dense(4096, activation= keras.activations.relu, use_bias= True)(flatten)
drop1 = keras.layers.Dropout(0.5)(fc1)

fc2 = keras.layers.Dense(4096, activation= keras.activations.relu, use_bias= True)(drop1)
drop2 = keras.layers.Dropout(0.5)(fc2)

fc3 = keras.layers.Dense(10, activation= keras.activations.softmax, use_bias= True)(drop2)

# 基于Model方法构建模型
model = keras.Model(inputs= inputs, outputs = fc3)
# 编译模型
model.compile(optimizer= tf.train.AdamOptimizer(0.001),
              loss= keras.losses.categorical_crossentropy,
              metrics= ['accuracy'])
# 训练配置，仅供参考
model.fit(x_train, y_train, batch_size= batch_size, epochs= epochs, validation_data=(x_test,y_test))