Kaggle——MNIST数字识别的卷积神经网络Keras实现

MNIST数据集是几乎所有人学习机器学习分类时的入门数据集。但麻雀虽小，五脏俱全，从逻辑回归到卷积神经网络都可以通过该数据集来练习。这里记录下Kaggle入门比赛中的MNIST项目及其99.9%准确率的实现过程。

数据前处理

卷积神经网络CNN对图像的分类十分高效准确，卷积层会将图像分成一块一块的局部接受野，在底层对局部接受野内的像素进行扫描，在高层再将这些底层抽取出来的特征整合起来。具体的原理这里不多谈，但我们要知道的是CNN层接受的输入图片数据格式应该为(n, height, width, channels)这样一个4维张量，分别表示实例数，高度，宽度和通道数。RGB图片的通道数为3，而一般的灰度图片为1。这里，我们应该将数据转换为卷积层接受的格式。

import pandas as pd
import numpy as np
from keras.utils.np_utils import to_categorical
from keras.datasets import mnist


class get_data(object):
    def __init__(self):
        self.train = pd.read_csv('train.csv')
        self.x_test = pd.read_csv('test.csv').values
        self.y_train = self.train['label']
        self.x_train = self.train.drop(['label'], axis=1).values.reshape(-1, 28, 28)

    def extend_data(self):
        (x_train_1, y_train_1), (x_test_1, y_test_1) = mnist.load_data()
        self.x_train = np.concatenate([self.x_train, x_train_1, x_test_1], axis=0)
        self.y_train = np.concatenate([self.y_train.ravel(), y_train_1, y_test_1], axis=0)
        return self.x_train, self.y_train

    def normalize(self):
        self.x_train = self.x_train / 255
        self.x_test = self.x_test / 255
        return self.x_train, self.x_test

    def reshape(self, keras_way=True):
        self.x_train = self.x_train.reshape(-1, 28, 28, 1)
        self.x_test = self.x_test.reshape(-1, 28, 28, 1)
        if keras_way:
            self.y_train = to_categorical(self.y_train, num_classes=10)
        return self.x_train, self.y_train, self.x_test

    def pipeline(self):
        data = get_data()
        data.extend_data()
        data.normalize()
        self.x_train, self.y_train, self.x_test = data.reshape()
        return self.x_train, self.y_train, self.x_test


if __name__ == '__main__':
    x_train, y_train, x_test = get_data().pipeline()
    print(x_train.shape)
    print(x_test.shape)
    print(y_train.shape)

返回处理后数据的形状。

(112000, 28, 28, 1)
(28000, 28, 28, 1)
(112000, 10)

模型训练

Keras构建神经网络模型十分方便，不用了解太多底层细节就能轻松使用。

搭建的模型框架。

from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPool2D, BatchNormalization
from keras.utils import plot_model
from IPython.display import Image

model = Sequential()
model.add(Conv2D(filters=64, kernel_size=(5, 5), padding='Same',
                 activation