Kaggle比赛入门赛题·用CNN做手写体数字识别

Kaggle比赛入门赛题·用CNN做手写体数字识别,准确率为0.98260

Digit Recognizer入口链接：https://www.kaggle.com/c/digit-recognizer
数据集：
链接：https://pan.baidu.com/s/13f3rM_lhNGyu2Rsqbc1AUw
提取码：otoy
复制这段内容后打开百度网盘手机App，操作更方便哦

train.csv和test.csv分别为训练集和测试集数据
sample_submission为官方提供的提交格式
predict.csv是本博文对测试集预测的结果，准确率为0.98260

1.数据加载和预处理

读取训练集和测试集中的csv文件

train_file = pd.read_csv(os.path.join(main_path, "train.csv"))
test_file = pd.read_csv(os.path.join(main_path, "test.csv"))

由于MNIST数据集中文本矩阵范围为（0-255），在这做一个数据标准化，把（0-255）归一化到（0-1）范围中。

ps:为什么进行标准化？可以参考下这个博客

# Normalization
train_file_norm = train_file.iloc[:, 1:] / 255.0
test_file_norm = test_file / 255.0

查看此时数据集的形状

train_file_norm.shape

我们也可以使用matplotlib.pyplot打印下数据集，将样本可视化

rand_indices = np.random.choice(train_file_norm.shape[0], 64, replace=False)
examples = train_file_norm.iloc[rand_indices, :]

fig, ax_arr = plt.subplots(8, 8, figsize=(6, 5))
fig.subplots_adjust(wspace=.025, hspace=.025)

ax_arr = ax_arr.ravel()
for i, ax in enumerate(ax_arr):
    ax.imshow(examples.iloc[i, :].values.reshape(28, 28), cmap="gray")
    ax.axis("off")
plt.show() 

我们需要把数据加工为(42000, 32, 32, 3)的形状便于训练
定义样本形状参数

num_examples_train = train_file.shape[0]
num_examples_test = test_file.shape[0]
n_h = 32
n_w = 32
n_c = 3

初始化样本空间

Train_input_images = np.zeros((num_examples_train, n_h, n_w, n_c))
Test_input_images = np.zeros((num_examples_test, n_h, n_w, n_c))

将数据装入样本空间

for example in range(num_examples_train):
    Train_input_images[example,:28,:28,0] = train_file.iloc[example, 1:].values.reshape(28,28)
    Train_input_images[example,:28,:28,1] = train_file.iloc[example, 1:].values.reshape(28,28)
    Train_input_images[example,:28,:28,2] = train_file.iloc[example, 1:].values.reshape(28,28)
    
for example in range(num_examples_test):
    Test_input_images[example,:28,:28,0] = test_file.iloc[example, :].values.reshape(28,28)
    Test_input_images[example,:28,:28,1] = test_file.iloc[example, :].values.reshape(28,28)
    Test_input_images[example,:28,:28,2] = test_file.iloc[example, :].values.reshape(28,28)

利用cv2.resize进行缩放

for example in range(num_examples_train):
    Train_input_images[example] = cv2.resize(Train_input_images[example], (n_h, n_w))
    
for example in range(num_examples_test):
    Test_input_images[example] = cv2.resize(Test_input_images[example], (n_h, n_w))

提取训练集的标签值

Train_labels = np.array(train_file.iloc[:, 0])

打印出来预处理好的样本数据形状

print("Shape of train input images : ", Train_input_images.shape)
print("Shape of test input images : ", Test_input_images.shape)
print("Shape of train labels : ", Train_labels.shape)

到此我们数据加工和预处理过程就结束了！

2.训练模型和预测结果

在介绍模型之前，先介绍下one-hot编码
one-hot编码可以把分类数据转化为二进制格式，供机器学习使用，实现函数如下

def one_hot(labels):
    onehot_labels = np.zeros(shape=[len(labels), 10])
    for i in range(len(labels)):
        index = labels[i]
        onehot_labels[i][index] = 1
    return onehot_labels

构建一个CNN网络模型

def mnist_cnn(input_shape):
    '''
    构建一个CNN网络模型
    :param input_shape: 指定输入维度
    :return:
    '''
    model = keras.Sequential()
    model.add(keras.layers.Conv2D(filters=32, kernel_size=5, strides=(1, 1),
                                  padding='same', activation=tf.nn.relu, input_shape=input_shape))
    model.add(keras.layers.MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid'))
    model.add(keras.layers.Conv2D(filters=64, kernel_size=3, strides=(1, 1), padding='same', activation=tf.nn.relu))
    model.add(keras.layers.MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid'))
    model.add(keras.layers.Dropout(0.25))
    model.add(keras.layers.Flatten())
    model.add(keras.layers.Dense(units=128, activation=tf.nn.relu))
    model.add(keras.layers.Dropout(0.5))
    model.add(keras.layers.Dense(units=10, activation=tf.nn.softmax))
    return model

训练并保存模型

def trian_model(train_images, train_labels):
    #  re-scale to 0~1.0之间
    print("train_images :{}".format(train_images.shape))
    print(train_labels)
    train_labels = one_hot(train_labels)

    # 建立模型
    model = mnist_cnn(input_shape=(32, 32, 3))
    model.compile(optimizer=tf.optimizers.Adam(), loss="categorical_crossentropy", metrics=['accuracy'])
    model.fit(x=train_images, y=train_labels, epochs=5, batch_size = 256)
    model.save('MYCNN2MNIST.h5')

利用训练好的模型预测测试集的标签，save_path为我们保存模型的路径

def pred(save_path,test_images):#载入模型并生成图片
    model=keras.models.load_model(save_path)
    #  开始预测
    predictions = model.predict(test_images)
#     print(predictions)
#     print(type(predictions))
    targetlist = []
    targetlist.append(0)
    for i in range(len(test_images)):
        target = np.argmax(predictions[i])
        targetlist.append(target)
    print(targetlist)
    predictions = pd.DataFrame(targetlist)
    predictions.to_csv("predict.csv")

将我们的预测结果写入文件中

submission = pd.read_csv('DataSet/sample_submission.csv')

然后把文件提交到服务器中，就可以看到自己准确率和排名了
ps:由于kaggle的服务器在国外，所以需要**(懂得都懂)

源码链接：https://pan.baidu.com/s/1HXt24GiUrRZliUXN-0N06g
提取码：pza9

有问题欢迎在评论区下留言~