kaggle入门Digit Recognizer

迷南。 2023-06-06 10:37 4阅读 0赞

用的是卷积神经网络CNN，正常的卷积大家应该都有所了解，这里为了提高准确率，针对这个提目加了数据强化的操作，提交kaggle准确率从原来的98.957%达到了99.285%

下面是我的具体代码流程：

## 引入数据： ##

## 初始化权值、偏置值： ##

#初始化权值
    def weight_variable(shape,name):
        initial=tf.truncated_normal(shape,stddev=0.1)#生成一个截断正态分布
        return tf.Variable(initial,name=name)
    
    #初始化偏置
    def bias_variable(shape,name):
        initial=tf.constant(0.1,shape=shape)
        return tf.Variable(initial,name=name)

## 卷积层池化层： ##

#卷积层
    def conv2d(x,W):
        #2d是二维的意思
        #x是一个tensor，形状是[batch,in_height,in_width,in_channels]NHWC关系，分别是批次大小(本例batch_size=100)，图片高度，图片宽度，通道数(黑白照片是1，彩色是3)
        #w是一个滤波器，tensor，形状是[filter_height,filter_width,in_channels,out_channels]，滤波器长，宽，输入和输出通道数
        #步长参数，strides[0]=strides[3]=1,strides[1]代表x方向的步长，strides[2]代表y方向的步长
        #padding:一个字符串，要么是'SAME'要么是'VALID'，对应两种卷积方法，前者补零,后者不会超出平面外部
        return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME')
    
    #池化层
    def max_pool_2x2(x):#这里定义的最大池化的方式，他会取2*2窗口里最大的值
        #ksize[1,x,y,1]
        #ksize是窗口大小，索引0，3对应的值必须是1，因为是2*2，所以索引1，2对应的值为2,步长对应也是2
        return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')

## 模型建立 ##

#命名空间
    with tf.name_scope('input'):
        #定义两个占位符
        x=tf.placeholder(tf.float32,[None,784],name='x-input')#28*28
        y=tf.placeholder(tf.float32,[None,10],name='y-input')
        with tf.name_scope('x_image'):
            # 改变x的格式，转为4d的向量[batch,in_height,in_width,in_channels]
            x_image = tf.reshape(x, [-1, 28, 28, 1],name='x_image')  # -1是指个数不确定 1是通道数
    
    
    with tf.name_scope('Conv1'):
        #初始化第一个卷积层的权值和偏置
        # 多少个卷积核返回多少个矩阵
        with tf.name_scope('W_conv1'):
            W_conv1=weight_variable([5,5,1,32],name='W_conv1')#5*5的采样窗口,32个卷积核从1个平面抽取特征 1的意思是通道数，如果是彩色则为3
        with tf.name_scope('b_conv1'):
            b_conv1=bias_variable([32],name='b_conv1')#每一个卷积核一个偏置值
    
        #把x_image和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数
        with tf.name_scope('conv2d_1'):
            h_conv1=tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1)#[?,28,28,32]
        with tf.name_scope('h_pool1'):
            h_pool1=max_pool_2x2(h_conv1)#进行max—pooling #[?,14,14,32]
    
    with tf.name_scope('Conv2'):
            #初始化第二个卷积层的权值和偏置
        with tf.name_scope('W_conv2'):
            W_conv2=weight_variable([5,5,32,64],name='W_conv2')#5*5的采样窗口,64个卷积核从32个平面抽取特征 因为前面32个卷积核生成了32个平面图。
        with tf.name_scope('b_conv2'):
            b_conv2=bias_variable([64],name='b_conv2')#每一个卷积核一个偏置值
    
            #把x_image和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数
        with tf.name_scope('conv2d_2'):
            h_conv2=tf.nn.relu(conv2d(h_pool1,W_conv2)+b_conv2)
        with tf.name_scope('h_pool2'):
            h_pool2=max_pool_2x2(h_conv2)#进行max—pooling
    
    #多少个卷积核返回多少个矩阵(feature map)哈哈
    #因为我们用的是SAME，不是VALID，所以会自动填充，所以卷积之后不是24*24而依然是28*28，这样池化后才是14*14
    #28*28的图片第一次卷积后还是28*28，第一次池化后变为14*14
    #第二次卷积后，还是14*14，第二次池化后，7*7
    #通过上面操作后得到64张7*7的平面
    with tf.name_scope('fc1'):
        with tf.name_scope('W_fc1'):
            #初始化第一个全连接层的权值和偏置值
            W_fc1=weight_variable([7*7*64,1024],name='W_fc1')#上一层有7*7*64个神经元，全连接层有1024个神经元 这个数是自己随机定义的
        with tf.name_scope('b_fc1'):
            b_fc1=bias_variable([1024],name='b_fc1')#1024个节点
        with tf.name_scope('h_pool2_flat'):
            #把池化层2的输出扁平化为1维 批次的100所以现在的形状是100*7*7*64，转化成1维
            h_pool2_flat=tf.reshape(h_pool2,[-1,7*7*64],name='h_pool2_flat')
        with tf.name_scope('wx_plus_b1'):
            #求第一个全连接层的输出⭐
            h_fc1=tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1)+b_fc1)
            #用keep_prob来表示神经元的输出概率 就是之前用到的dropout
        with tf.name_scope('keep_prob'):
            keep_prob=tf.placeholder(tf.float32,name='keep_prob')
        with tf.name_scope('h_fc1_drop'):
            h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob,name='h_fc1_drop')
    with tf.name_scope('fc2'):
        with tf.name_scope('W_fc2'):
            #初始化第二个全连接层
            W_fc2=weight_variable([1024,10],name='W_fc2')#10个输出
        with tf.name_scope('b_fc2'):
            b_fc2=bias_variable([10],name='b_fc2')
        with tf.name_scope('softmax'):
            #计算输出
            prediction=tf.matmul(h_fc1_drop,W_fc2)+b_fc2
    #pre_y
    pre_y=tf.arg_max(prediction,1)
    #后面就都一样了，
    with tf.name_scope('cross_entropy'):
        #交叉熵代价函数
        cross_entropy=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=prediction),name='cross_entropy')
        tf.summary.scalar('cross_entropy',cross_entropy)
    with tf.name_scope('train'):
        #使用AdamOptimizer进行优化
        train_step=tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
    with tf.name_scope('accuracy'):
        with tf.name_scope('correct_prediction'):
            #结果存放在一个布尔列表中
            correct_prediction=tf.equal(tf.argmax(prediction,1),tf.argmax(y,1))
        with tf.name_scope('accuracy'):
            #求准确率
            accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
            tf.summary.scalar('accuracy',accuracy)
    #合并所有的summary
    merged=tf.summary.merge_all()

上面就是进行了两次卷积—池化的操作，然后进行了两次全连接操作。

然后我们开始对引入的数据进行一些操作，然后输入到我们上面定义的这个网络里。

## 独热编码 ##

with tf.Session() as sess:
        def dense_to_one_hot(labels_dense, num_classes):
            # 获取个数
            num_labels = labels_dense.shape[0]
            # [0,1*类别数，2*类别数，……]
            index_offset = np.arange(num_labels) * num_classes
            # 空白操作板   zeros[样本数,类别数]
            labels_one_hot = np.zeros((num_labels, num_classes))
            # 语言不好解释，看网址：https://docs.scipy.org/doc/numpy/reference/generated/numpy.ndarray.flat.html
            labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1
            return labels_one_hot
        saver = tf.train.Saver()
        sess.run(tf.global_variables_initializer())
    
        total_y = file['label'].values
        file.drop('label', axis=1, inplace=True)
        total_x = file.values
        total_y = dense_to_one_hot(total_y, 10)

## 归一化  ##

total_x = np.multiply(total_x, 1.0 / 255)

## 划分训练集测试集并计算批次 ##

#先将total_x转化为图像格式
        total_x_image=sess.run(x_image,feed_dict={x:total_x})
        total_x, test_x, total_y, test_y = train_test_split(total_x_image, total_y, test_size=0.1)
        #total_x和用于测试的test_x都是四维的了。
        
        # 计算一共多少个批次
        batch_size_test = math.floor(total_x.shape[0] / batch_size)

## 数据增强  ##

# 使用数据增强算法
        datagen = ImageDataGenerator(
            featurewise_center=False,  # set input mean to 0 over the dataset
            samplewise_center=False,  # set each sample mean to 0
            featurewise_std_normalization=False,  # divide inputs by std of the dataset
            samplewise_std_normalization=False,  # divide each input by its std
            zca_whitening=False,  # apply ZCA whitening
            rotation_range=10,  # randomly rotate images in the range (degrees, 0 to 180)
            zoom_range=0.1,  # Randomly zoom image
            width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
            height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
            horizontal_flip=False,  # randomly flip images
            vertical_flip=False)  # randomly flip images
    
    
        # 获取一个batch
        def next_batch(dataX,dataY):
            xx=datagen.flow(dataX, dataY, batch_size=batch_size,shuffle=True, seed=None, save_to_dir=None, save_prefix='',
                                         save_format='png')[0]
            return xx[0],xx[1]
            # for batch_num in range(batch_size_test):
            #     start_index = batch_num * batch_size
            #     end_index = min((batch_num + 1) * batch_size, len(total_x))
            #     yield data[start_index:end_index]
    
        total_x=np.array(total_x)
        total_y=np.array(total_y)    
        test_x = np.array(test_x)
        test_y = np.array(test_y)
    
        train_writer=tf.summary.FileWriter('../logs/train',sess.graph)
        test_writer=tf.summary.FileWriter('../logs/test',sess.graph)

## 训练并测试 ##

for epoch in range(50):
            # batched = next_batch(list(zip(total_x_image, total_y)))
            for i in range(batch_size_test):  # 批次
                #batch_xs, batch_ys = zip(*batch)
                # batch_xs = np.array(batch_xs)
                #batch_ys = np.array(batch_ys)
                batch_xs,batch_ys=next_batch(total_x[i*batch_size:(i+1)*batch_size],total_y[i*batch_size:(i+1)*batch_size])
                # 对于y标签进行独热编码处理
                # batch_ys = dense_to_one_hot(batch_ys, 10)
    
                sess.run(train_step,feed_dict={x_image:batch_xs,y:batch_ys,keep_prob:0.7})
                #记录训练集计算的参数
                summary=sess.run(merged,feed_dict={x_image:batch_xs,y:batch_ys,keep_prob:1.0})
                train_writer.add_summary(summary,epoch)
    
            acc = sess.run(accuracy, feed_dict={x_image: test_x, y: test_y, keep_prob: 1.0})
            print("Iter "+str(epoch)+", Testing Accuracy="+str(acc))
        saver.save(sess,r'../model/final_model')
        file = pd.read_csv("../data/test.csv")
        #28000条数据，4000一次
        results=[]
        for i in range(7):
            total_x = file.values[i*4000:(i+1)*4000]
            pred = sess.run(pre_y, feed_dict={x: total_x, keep_prob: 1.0})
            results+=pred.ravel().tolist()
        data = {
            "ImageId": [i for i in range(1, 28001)],
            "Label": results
        }
        pd.DataFrame(data).to_csv(r"E:\TensorFlow\大数据之路\kaggle\Digit_Recongnizer\result\res1.csv",
                                  index=False)

把上面的代码按照顺序复制就可以得到 完整代码，这里不贴出来了。

其实还可以提高准确率，就是使用MNIST数据集，这样样本就多了两倍！让我们看一下全部代码吧：（其实就是相比上面代码加了几行引入MNIST数据）

import tensorflow as tf
    from sklearn.model_selection import train_test_split
    import pandas as pd
    import numpy as np
    import math
    from tensorflow.python.keras.preprocessing.image import ImageDataGenerator,array_to_img,img_to_array,load_img
    from tensorflow.examples.tutorials.mnist import input_data
    #引入数据
    file=pd.read_csv('../data/train.csv')
    
    #每个批次的大小
    batch_size=50
    
    #初始化权值
    def weight_variable(shape,name):
        initial=tf.truncated_normal(shape,stddev=0.1)#生成一个截断正态分布
        return tf.Variable(initial,name=name)
    
    #初始化偏置
    def bias_variable(shape,name):
        initial=tf.constant(0.1,shape=shape)
        return tf.Variable(initial,name=name)
    #卷积层
    def conv2d(x,W):
        #2d是二维的意思
        #x是一个tensor，形状是[batch,in_height,in_width,in_channels]NHWC关系，分别是批次大小(本例batch_size=100)，图片高度，图片宽度，通道数(黑白照片是1，彩色是3)
        #w是一个滤波器，tensor，形状是[filter_height,filter_width,in_channels,out_channels]，滤波器长，宽，输入和输出通道数
        #步长参数，strides[0]=strides[3]=1,strides[1]代表x方向的步长，strides[2]代表y方向的步长
        #padding:一个字符串，要么是'SAME'要么是'VALID'，对应两种卷积方法，前者补零,后者不会超出平面外部
        return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME')
    
    #池化层
    def max_pool_2x2(x):#这里定义的最大池化的方式，他会取2*2窗口里最大的值
        #ksize[1,x,y,1]
        #ksize是窗口大小，索引0，3对应的值必须是1，因为是2*2，所以索引1，2对应的值为2,步长对应也是2
        return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
    
    #命名空间
    with tf.name_scope('input'):
        #定义两个占位符
        x=tf.placeholder(tf.float32,[None,784],name='x-input')#28*28
        y=tf.placeholder(tf.float32,[None,10],name='y-input')
        with tf.name_scope('x_image'):
            # 改变x的格式，转为4d的向量[batch,in_height,in_width,in_channels]
            x_image = tf.reshape(x, [-1, 28, 28, 1],name='x_image')  # -1是指个数不确定 1是通道数
    
    
    with tf.name_scope('Conv1'):
        #初始化第一个卷积层的权值和偏置
        # 多少个卷积核返回多少个矩阵
        with tf.name_scope('W_conv1'):
            W_conv1=weight_variable([5,5,1,32],name='W_conv1')#5*5的采样窗口,32个卷积核从1个平面抽取特征 1的意思是通道数，如果是彩色则为3
        with tf.name_scope('b_conv1'):
            b_conv1=bias_variable([32],name='b_conv1')#每一个卷积核一个偏置值
    
        #把x_image和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数
        with tf.name_scope('conv2d_1'):
            h_conv1=tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1)#[?,28,28,32]
        with tf.name_scope('h_pool1'):
            h_pool1=max_pool_2x2(h_conv1)#进行max—pooling #[?,14,14,32]
    
    with tf.name_scope('Conv2'):
            #初始化第二个卷积层的权值和偏置
        with tf.name_scope('W_conv2'):
            W_conv2=weight_variable([5,5,32,64],name='W_conv2')#5*5的采样窗口,64个卷积核从32个平面抽取特征 因为前面32个卷积核生成了32个平面图。
        with tf.name_scope('b_conv2'):
            b_conv2=bias_variable([64],name='b_conv2')#每一个卷积核一个偏置值
    
            #把x_image和权值向量进行卷积，再加上偏置值，然后应用于relu激活函数
        with tf.name_scope('conv2d_2'):
            h_conv2=tf.nn.relu(conv2d(h_pool1,W_conv2)+b_conv2)
        with tf.name_scope('h_pool2'):
            h_pool2=max_pool_2x2(h_conv2)#进行max—pooling
    
    #多少个卷积核返回多少个矩阵(feature map)哈哈
    #因为我们用的是SAME，不是VALID，所以会自动填充，所以卷积之后不是24*24而依然是28*28，这样池化后才是14*14
    #28*28的图片第一次卷积后还是28*28，第一次池化后变为14*14
    #第二次卷积后，还是14*14，第二次池化后，7*7
    #通过上面操作后得到64张7*7的平面
    with tf.name_scope('fc1'):
        with tf.name_scope('W_fc1'):
            #初始化第一个全连接层的权值和偏置值
            W_fc1=weight_variable([7*7*64,1024],name='W_fc1')#上一层有7*7*64个神经元，全连接层有1024个神经元 这个数是自己随机定义的
        with tf.name_scope('b_fc1'):
            b_fc1=bias_variable([1024],name='b_fc1')#1024个节点
        with tf.name_scope('h_pool2_flat'):
            #把池化层2的输出扁平化为1维 批次的100所以现在的形状是100*7*7*64，转化成1维
            h_pool2_flat=tf.reshape(h_pool2,[-1,7*7*64],name='h_pool2_flat')
        with tf.name_scope('wx_plus_b1'):
            #求第一个全连接层的输出⭐
            h_fc1=tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1)+b_fc1)
            #用keep_prob来表示神经元的输出概率 就是之前用到的dropout
        with tf.name_scope('keep_prob'):
            keep_prob=tf.placeholder(tf.float32,name='keep_prob')
        with tf.name_scope('h_fc1_drop'):
            h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob,name='h_fc1_drop')
    with tf.name_scope('fc2'):
        with tf.name_scope('W_fc2'):
            #初始化第二个全连接层
            W_fc2=weight_variable([1024,10],name='W_fc2')#10个输出
        with tf.name_scope('b_fc2'):
            b_fc2=bias_variable([10],name='b_fc2')
        with tf.name_scope('softmax'):
            #计算输出
            prediction=tf.matmul(h_fc1_drop,W_fc2)+b_fc2
    #pre_y
    pre_y=tf.arg_max(prediction,1)
    #后面就都一样了，
    with tf.name_scope('cross_entropy'):
        #交叉熵代价函数
        cross_entropy=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=prediction),name='cross_entropy')
        tf.summary.scalar('cross_entropy',cross_entropy)
    with tf.name_scope('train'):
        #使用AdamOptimizer进行优化
        train_step=tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
    with tf.name_scope('accuracy'):
        with tf.name_scope('correct_prediction'):
            #结果存放在一个布尔列表中
            correct_prediction=tf.equal(tf.argmax(prediction,1),tf.argmax(y,1))
        with tf.name_scope('accuracy'):
            #求准确率
            accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
            tf.summary.scalar('accuracy',accuracy)
    #合并所有的summary
    merged=tf.summary.merge_all()
    
    with tf.Session() as sess:
        def dense_to_one_hot(labels_dense, num_classes):
            # 获取个数
            num_labels = labels_dense.shape[0]
            # [0,1*类别数，2*类别数，……]
            index_offset = np.arange(num_labels) * num_classes
            # 空白操作板   zeros[样本数,类别数]
            labels_one_hot = np.zeros((num_labels, num_classes))
            # 语言不好解释，看网址：https://docs.scipy.org/doc/numpy/reference/generated/numpy.ndarray.flat.html
            labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1
            return labels_one_hot
        saver = tf.train.Saver()
        sess.run(tf.global_variables_initializer())
    
        total_y = file['label'].values
        file.drop('label', axis=1, inplace=True)
        total_x = file.values
        total_y = dense_to_one_hot(total_y, 10)
        # print(total_x.shape)
        # print(total_y.shape)
        mnist = input_data.read_data_sets("MNIST_data", one_hot=True)  # 独热编码
        total_x1=mnist.train.images
        total_y1 = mnist.train.labels
        # print(total_x1.shape)
        # print(total_y1.shape)
        #合并MNIST_data和题中所给数据集
        total_x=np.concatenate((total_x,total_x1),axis=0)
        total_y=np.concatenate((total_y,total_y1),axis=0)
        # print(total_x.shape)
        # print(total_y.shape)
        total_x = np.multiply(total_x, 1.0 / 255)
    
        #先将total_x转化为图像格式
        total_x_image=sess.run(x_image,feed_dict={x:total_x})
        total_x, test_x, total_y, test_y = train_test_split(total_x_image, total_y, test_size=0.1)
        #total_x和用于测试的test_x都是四维的了。
    
        # 计算一共多少个批次
        batch_size_test = math.floor(total_x.shape[0] / batch_size)
        # print(batch_size_test)
        # 对训练数据total_x,total_y增强数据
        # 使用数据增强算法
        datagen = ImageDataGenerator(
            featurewise_center=False,  # set input mean to 0 over the dataset
            samplewise_center=False,  # set each sample mean to 0
            featurewise_std_normalization=False,  # divide inputs by std of the dataset
            samplewise_std_normalization=False,  # divide each input by its std
            zca_whitening=False,  # apply ZCA whitening
            rotation_range=10,  # randomly rotate images in the range (degrees, 0 to 180)
            zoom_range=0.1,  # Randomly zoom image
            width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
            height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
            horizontal_flip=False,  # randomly flip images
            vertical_flip=False)  # randomly flip images
    
    
        # 获取一个batch
        def next_batch(dataX,dataY):
            xx=datagen.flow(dataX, dataY, batch_size=batch_size,shuffle=True, seed=None, save_to_dir=None, save_prefix='',
                                         save_format='png')[0]
            return xx[0],xx[1]
            # for batch_num in range(batch_size_test):
            #     start_index = batch_num * batch_size
            #     end_index = min((batch_num + 1) * batch_size, len(total_x))
            #     yield data[start_index:end_index]
        total_x=np.array(total_x)
        total_y=np.array(total_y)
        test_x = np.array(test_x)
        test_y = np.array(test_y)
    
        train_writer=tf.summary.FileWriter('../logs/train',sess.graph)
        test_writer=tf.summary.FileWriter('../logs/test',sess.graph)
        # print(total_x.shape)
        # print(total_y.shape)
        # print(total_x[0*128:(0+1)*128].shape)
        for epoch in range(50):
            # batched = next_batch(list(zip(total_x_image, total_y)))
            for i in range(batch_size_test):  # 批次
                #batch_xs, batch_ys = zip(*batch)
                # batch_xs = np.array(batch_xs)
                #batch_ys = np.array(batch_ys)
                batch_xs,batch_ys=next_batch(total_x[i*batch_size:(i+1)*batch_size],total_y[i*batch_size:(i+1)*batch_size])
                # 对于y标签进行独热编码处理
                # batch_ys = dense_to_one_hot(batch_ys, 10)
    
                sess.run(train_step,feed_dict={x_image:batch_xs,y:batch_ys,keep_prob:0.7})
                #记录训练集计算的参数
                summary=sess.run(merged,feed_dict={x_image:batch_xs,y:batch_ys,keep_prob:1.0})
                train_writer.add_summary(summary,epoch)
    
            acc = sess.run(accuracy, feed_dict={x_image: test_x, y: test_y, keep_prob: 1.0})
            print("Iter "+str(epoch)+", Testing Accuracy="+str(acc))
        saver.save(sess,r'../model/final_model')
        file = pd.read_csv("../data/test.csv")
        #28000条数据，4000一次
        results=[]
        for i in range(7):
            total_x = file.values[i*4000:(i+1)*4000]
            pred = sess.run(pre_y, feed_dict={x: total_x, keep_prob: 1.0})
            results+=pred.ravel().tolist()
        data = {
            "ImageId": [i for i in range(1, 28001)],
            "Label": results
        }
        pd.DataFrame(data).to_csv(r"E:\TensorFlow\大数据之路\kaggle\Digit_Recongnizer\result\res1.csv",
                                  index=False)