kaggle入门--手写数字识别
前言：
对于像学大数据，机器学习，深度学习相关的初学者往往面临着两个问题：
- 没有合适的训练数据
- 机器不行，没有集群
这里推荐一个不错的平台，相信大家都听过，就是kaggle平台了，这里提供不少数据的的下载，和相关比赛。
View
基本是围绕着competition就是比赛进行的。
打开比赛我们可以看到很多，注意下左面的有颜色的条，绿色是简单的，红色是难，还有蓝色，橙色等。
我们作为初学者橙色以上的就不用看了。。。
Getting started
这里我选择一个入门级别的题，Digit Recognizer，就是数字识别，点开后，我们点击data，
Each image is 28 pixels in height and 28 pixels in width, for a total of 784 pixels in total. Each pixel has a single pixel-value associated with it, indicating the lightness or darkness of that pixel, with higher numbers meaning darker. This pixel-value is an integer between 0 and 255, inclusive.
可以看到描述是28 * 28的图片数据，但是和tensoflow的那个MNISTdemo不同的是这份数据的格式是csv，每一张图片的像素共 784个(28*28),矩阵展开成一维的，我们再看看其他的competition，基本上都是csv格式的数据，不同手动分数据，这里已经给我们分开了test和train数据，我们要做的就是将最后的预测结果还是以csv的形式上传即可，系统为我们评分。
需要注意的是，不上传代码，就是不论我们用什么算法，框架，只要结果对就行，不管是spark mllib ，py sklearn， tensorflow，caffe/caffe2都行~
给一个我实现的随机森林分类树的解法：
#!/usr/bin/python
# -*- coding: utf-8 -*-
from numpy import *
import csv
# array format to int
def toInt(array):
array=mat(array)
m,n=shape(array)
newmat=zeros((m,n))
for i in xrange(m):
for j in xrange(n):
newmat[i,j]=int(array[i,j])
return newmat
def nomalizing(array):
m,n=shape(array)
for i in xrange(m):
for j in xrange(n):
if array[i,j]!=0:
if array[i,j] > 128:
array[i,j] = 2
else:
array[i,j] = 1
return array
def loadTrainData():
l=[]
with open('train.csv','r') as fp:
lines=csv.reader(fp)
for line in lines:
l.append(line) #42001*785
#remove title
l.remove(l[0])
l=array(l)
label=l[:,0]
data=l[:,1:]
return nomalizing(toInt(data)),toInt(label) #label 1*42000 data 42000*784
def loadTestData():
l=[]
with open('test.csv') as file:
lines=csv.reader(file)
for line in lines:
l.append(line)#28001*784
l.remove(l[0])
data=array(l)
return nomalizing(toInt(data)) # data 28000*784
def saveResult(result,csvName):
with open(csvName,'wb') as myFile:
myWriter=csv.writer(myFile)
myWriter.writerow(["ImageId","Label"])
index=0;
for i in result:
tmp=[]
index=index+1
tmp.append(index)
#tmp.append(i)
tmp.append(int(i))
myWriter.writerow(tmp)
from sklearn.ensemble import RandomForestClassifier
def RFClassify(trainData,trainLabel,testData):
nbCF=RandomForestClassifier(n_estimators=200,warm_start = True)
nbCF.fit(trainData,ravel(trainLabel))
testLabel=nbCF.predict(testData)
saveResult(testLabel,'Result.csv')
return testLabel
def dRecognition():
trainData,trainLabel=loadTrainData()
print "load train data finish"
testData=loadTestData()
print "load test data finish"
result=RFClassify(trainData,trainLabel,testData)
print "finish!"
if __name__ == '__main__':
dRecognition()
开始跑，如果机器好的话，也就2-3min就跑完了，毕竟不是DNN(深度神经网络)嘛
然后将得到的文件上传。
可以看到我提交了4次（每次优化下系数，比如树的棵树，是否是纯净叶子节点等。。）最好的accuracy rate是0.96771,之前我就明白，什么都可以用随机森林，但都不够准确。。
我们看看上面排名的：
厉害了，100%的都有，说明题目本身是可以获得精确匹配的。我们看不到人家的代码，也就不知道用了什么方法，或者外星人、、、额，扯远了。
神经网络解法：
这里还有个比较好，就是提供了comment，我们打开一个，根据他的note我大概给出一个TF的版本，但由于还在上班，我没什么环境跑，（公司提供的tensoflow on Yarn，我还没研究明白。。。）还是回校后，没啥事的时候试试“单机单卡”吧，毕竟没钱买好的GPU，唉~~
代码：
import numpy as np
import pandas as pd
import scipy
import tensorflow as tf
# The competition datafiles are in the directory ../input
# Read competition data files:
filedir = os.listdir()
train = pd.read_csv("/home/hdp-like/bigame_gbw/TF28_28/train.csv")
test = pd.read_csv("/home/hdp-like/bigame_gbw/TF28_28/test.csv")
# Write to the log:
print("Training set has {0[0]} rows and {0[1]} columns".format(train.shape))
print("Test set has {0[0]} rows and {0[1]} columns".format(test.shape))
learning_rate = 0.01
training_iteration = 30
batch_size = 300
display_step = 2
trainfv = train.drop(['label'], axis=1).values.astype(dtype=np.float32)
trainLabels = train['label'].tolist()
ohtrainLabels = tf.one_hot(trainLabels, depth=10)
ohtrainLabelsNdarray = tf.Session().run(ohtrainLabels).astype(dtype=np.float64)
trainfv = np.multiply(trainfv, 1.0 / 255.0)
testData = test.values
testData = np.multiply(testData, 1.0 / 255.0)
#train_x,train_y = make_the_data_ready_conv(train_x,train_y)
#valid_x,valid_y = make_the_data_ready_conv(valid_x,valid_y)
x = tf.placeholder("float",[None,784])
y = tf.placeholder("float",[None,10])
W = tf.Variable(tf.zeros([784,10]))
b = tf.Variable(tf.zeros([10]))
with tf.name_scope("Wx_b") as scope:
model = tf.nn.softmax(tf.matmul(x,W) + b)
w_h = tf.summary.histogram("weights",W)
b_h = tf.summary.histogram("biases",b)
#loss function with maximum likelihood
with tf.name_scope("cost_function") as scope:
cost_function = -tf.reduce_sum(y * tf.log(model))
tf.summary.scalar("cost_function",cost_function)
#optimization with SGD
with tf.name_scope("train") as scope:
op = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost_function)
init = tf.initialize_all_variables()
merged_summary_op = tf.summary.merge_all()
import math
from random import randint
# for faster convergence
def random_batch(data,labels,size):
value = math.floor(len(data) / size)
intervall = randint(0,value-1)
return data[intervall*size:intervall*(size+1)],labels[intervall*size:intervall*(size+1)]
with tf.Session() as sess:
sess.run(init)
#run as single machine multiply GPU
# view the stuff on tensorboard
for iteration in range(training_iteration):
avg_cost = 0
total_batch = int(trainfv.shape[0]/batch_size)
for i in range(total_batch):
batch_xs,batch_ys = random_batch(trainfv,ohtrainLabelsNdarray,total_batch)
sess.run(op,feed_dict={x: batch_xs, y: batch_ys})
avg_cost += sess.run(cost_function,feed_dict={x: batch_xs, y: batch_ys}) / total_batch
if iteration % display_step == 0:
print "Iteration:", '%04d' % (iteration + 1), "cost=", "{:.9f}".format(avg_cost)
print "Training Finished"
correct_prediction = tf.equal(tf.argmax(model, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
print "\nAccuracy of the current model: ",sess.run(accuracy, feed_dict={x: trainfv[0:10000], y: ohtrainLabelsNdarray[0:10000]})
prob = sess.run(tf.argmax(model,1), feed_dict = {x: testData})
which = 1
print 'predicted labe: {}'.format(str(prob[which]))
print(prob)
#import csv
#outputFile_dir = '../input/output.csv'
#header = ['ImageID','Label']
#with open(outputFile_dir, 'w', newline='') as csvFile:
# writer = csv.writer(csvFile, delimiter = ',')
# writer.writerow(header)
# for i, p in enumerate(prob):
# writer.writerow([str(i+1), str(p)])
哦，还有就是注释啥的别用中文，生成出来的文件是Utf8，这可能影响检查的准确性。。。还有就是，中文注释多low呀~
结束了
之前有个大二的学弟or学妹（我不知道男女）问我，我想入门大数据，做了几个featured的,几个。。还是featured的，当时我回复他微信，默默地看了看朋友圈，，，，人家把我屏蔽了。看来是遇到大神了，要不就是外星人~我只求，永远不拉黑我就好
