kaggle Digit Recognizer（TensorFlow CNN实现）

开始用python写的KNN，但是不支持GPU，跑的太慢了，果断放弃，找了半天，终于找到TensorFlow CNN实现的了。

之前按照莫烦的视频写过一个TensorFlow识别mnist数据集的demo，但是只输出准确率，不知道怎么去读csv文件，把向量喂入tf，还有保存训练出来的权重，说到底还是对TensorFlow不熟，果断去补TensorFlow了。

这个也是用了两个卷积、池化，两个全连接层，然后softmax分类，准确率97.9%，具体的看代码吧^_^

import numpy as np
import tensorflow as tf
import os
import pandas as pd

#读取csv， x_train为训练集矩阵42000*784， x_test为测试集矩阵28000*784， label_train为训练集标签42000*1
train = pd.read_csv('E:/all/train.csv')
test = pd.read_csv('E:/all/test.csv')

x_train = train.iloc[:, 1:].values
x_train = x_train.astype(np.float)

x_test = test.iloc[:, :].values
x_test = x_test.astype(np.float)

x_train = np.multiply(x_train, 1.0/255)
x_test = np.multiply(x_test, 1.0/255)

image_size = x_train.shape[1]
image_height = image_width = np.ceil(np.sqrt(image_size)).astype(np.uint8)

label_train = train.iloc[:, 0].values
label_count = np.unique(label_train).shape[0]


def dense_to_one_hot(label_dense, count):    #变成了（42000,10）的矩阵，第i个数标签为j则（i，j）为1
    num_labels = label_dense.shape[0]
    index_offset = np.arange(num_labels) * count
    labels_one_hot = np.zeros((num_labels, count))
    labels_one_hot.flat[index_offset+label_dense.ravel()] = 1

    return labels_one_hot

#把label_train转成one-hot的labels，为42000*10
labels = dense_to_one_hot(label_train, label_count)
labels = labels.astype(np.uint8)

#分批次，一个批次大小为128，这样就分成了42000/128批
batch_size = 128
n_batch = int(len(x_train)/batch_size)

#定义x，y
x = tf.placeholder(tf.float32, [None,784])
y = tf.placeholder(tf.float32, [None, 10])


def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)

def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)

def conv2d(x, w):
    return tf.nn.conv2d(x, w, strides=[1,1,1,1], padding='SAME')

def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')

#把x重构成28*28的
x_image = tf.reshape(x, [-1, 28, 28, 1])

#两个卷积、池化，两个全连接层，然后softmax分类（其中dropout防止过拟合）
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1)+b_conv1)
h_pool1 = max_pool_2x2(h_conv1)


W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2)+b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])


w_fc1 = weight_variable([7*7*64,1024])
b_fc1 = bias_variable([1024])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1)+b_fc1)
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)


w_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_conv = tf.matmul(h_fc1_drop, w_fc2) + b_fc2
prediction = tf.nn.softmax(y_conv)

#loss梯度下降
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=y_conv))
train_step_1 = tf.train.AdadeltaOptimizer(0.1).minimize(loss)

#计算准确率
correction_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_conv, 1))
accuracy = tf.reduce_mean(tf.cast(correction_prediction, tf.float32))

global_step = tf.Variable(0, name='global_step', trainable=False)

#初始化TensorFlow
init = tf.global_variables_initializer()
saver = tf.train.Saver()

with tf.Session() as sess:
    sess.run(init)

    #迭代20次，每次都把全部数据分成42000/128个批次，分批喂入tf
    for epoch in range(20):
        print('epoch', epoch + 1)
        for batch in range(n_batch):
            batch_x = x_train[(batch)*batch_size:(batch+1)*batch_size]
            batch_y = labels[(batch)*batch_size:(batch+1)*batch_size]

            sess.run(train_step_1, feed_dict={x:batch_x, y:batch_y, keep_prob:0.5})

        batch_x = x_train[n_batch*batch_size:]
        batch_y = labels[n_batch*batch_size:]

        sess.run(train_step_1, feed_dict={x: batch_x, y: batch_y, keep_prob: 0.5})

    #保存训练出来的模型（权重）
    saver.save(sess, 'E:/1/model.ckpt')

    #读取模型，全部喂入tf，预测值，写入result.csv
    saver.restore(sess, 'E:/1/model.ckpt')

    filename = 'E:/1/result3.csv'
    test_batch_x = x_test[:]
    myPrediction = sess.run(prediction, feed_dict={x:test_batch_x, keep_prob:1.0})
    label_test = np.argmax(myPrediction, axis=1)
    pd.DataFrame(label_test).to_csv(filename)