https://github.com/lijingpeng/kaggle/tree/master/competitions/image_recognize
识别谷歌街景图片中的字母
street-view-getting-started-with-julia 让我们从谷歌街景的图片中鉴定字母，这个题目是让我们学习和使用Julia，Julia有python和R的易用性，有C语言的速度，无奈对Julia不是很熟悉，所以还是想用python来试试。
import cv2
import numpy as np
import sys
import pandas as pd
我们希望所有的图片最后存储在一个numpy的矩阵当中，每一行为图片的像素值。为了得到统一的表达呢，我们将RGB三个通道的值做平均得到的灰度图像作为每个图片的表示:
# typeData 为"train"或者"test"
# labelsInfo 包含每一个图片的ID
# 图片存储在trainResized和testResized文件夹内
def read_data(typeData, labelsInfo, imageSize):
labelsIndex = labelsInfo["ID"]
x = np.zeros((np.size(labelsIndex), imageSize))
for idx, idImage in enumerate(labelsIndex):
# 得到图片文件名并读取
nameFile = typeData + "Resized/" + str(idImage) + ".Bmp"
img = cv2.imread(nameFile)
# 转化为灰度图
temp = np.mean(img, 2)
# 将图片转化为行向量
x[idx, :] = np.reshape(temp, (1, imageSize))
return x
预处理训练集和测试集
imageSize = 400
trainlabels = pd.read_csv("trainLabels.csv")
testlabels = pd.read_csv("sampleSubmission.csv")
# 得到训练集的特征
xTrain = read_data('train', trainlabels, imageSize)
# 得到测试集的特征
xTest = read_data("test", testlabels, imageSize)
预览数据：
print trainlabels.head(2)
print testlabels.head(2)
ID Class
0 1 n
1 2 8
ID Class
0 6284 A
1 6285 A
yTrain = trainlabels["Class"]
yTrain = [ord(x) for x in yTrain]
模型训练
随机森林
使用随机森林进行训练，树的个数和深度需要多次调解寻求最佳值
from sklearn.ensemble import RandomForestClassifier
%time rfc = RandomForestClassifier(n_estimators = 500, max_features = 50, max_depth=None)
rfc.fit(xTrain, yTrain)
CPU times: user 121 µs, sys: 367 µs, total: 488 µs
Wall time: 494 µs
RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
max_depth=None, max_features=50, max_leaf_nodes=None,
min_samples_leaf=1, min_samples_split=2,
min_weight_fraction_leaf=0.0, n_estimators=500, n_jobs=1,
oob_score=False, random_state=None, verbose=0,
warm_start=False)
预测
将训练后的模型应用到测试集上，并保存结果：
predTest = rfc.predict(xTest)
predResult = [chr(x) for x in predTest]
testlabels["Class"] = predResult
testlabels.to_csv("rf_500_50_result.csv",index = None)
结果
使用50颗树进行训练，提交kaggle之后准确率约为0.40
改用300颗树进行训练，提交kaggle之后准确率为0.46695
改用500颗树进行训练，深度为10，提价kaggle后准确率为0.40，估计出现了过拟合
改用500颗树进行训练，不设置深度，提价kaggle后准确率为0.47480
贝叶斯
from sklearn.naive_bayes import GaussianNB as GNB
model_GNB = GNB()
model_GNB.fit(xTrain, yTrain)
predTest = model_GNB.predict(xTest)
predResult = [chr(x) for x in predTest]
testlabels["Class"] = predResult
testlabels.to_csv("gnb_result.csv",index = None)
贝叶斯的训练非常的快，把结果提交kaggle后，得到0.02389的准确率，明显低于随机森林
GBDT
from sklearn.ensemble import GradientBoostingClassifier
%time GBDT = GradientBoostingClassifier(loss='deviance', learning_rate=0.1, n_estimators=100, subsample=1.0, \
min_samples_split=2, min_samples_leaf=1, min_weight_fraction_leaf=0.0, max_depth=3, init=None, \
random_state=None, max_features=None, verbose=0, max_leaf_nodes=None, warm_start=False, presort='auto')
%time GBDT.fit(xTrain, yTrain)
%time predTest = GBDT.predict(xTest)
predResult = [chr(x) for x in predTest]
testlabels["Class"] = predResult
testlabels.to_csv("gbdt_result.csv",index = None)
CPU times: user 91 µs, sys: 738 µs, total: 829 µs
Wall time: 2.93 ms
CPU times: user 40min 16s, sys: 52.3 s, total: 41min 9s
Wall time: 2h 55min 22s
CPU times: user 1.75 s, sys: 44.5 ms, total: 1.8 s
Wall time: 1.79 s
使用GBDT仅得到了0.31937的准确率，可能是我的默认参数没有调节好，关键是GBDT的训练时间太长，调试成本也比较高
神经网络
import os
from skimage.io import imread
from lasagne import layers
from lasagne.nonlinearities import softmax
from nolearn.lasagne import NeuralNet, BatchIterator
# Define functions
def read_datax(typeData, labelsInfo, imageSize, path):
x = np.zeros((labelsInfo.shape[0], imageSize))
for (index, idImage) in enumerate(labelsInfo['ID']):
# use specially created 32 x 32 images
nameFile = '{0}/{1}Resized32/{2}.Bmp'.format(path,
typeData, idImage)
img = imread(nameFile, as_grey = True)
x[index, :] = np.reshape(img, (1, imageSize))
return x
def fit_model(reshaped_train_x, y, image_width,
image_height, reshaped_test_x):
net = NeuralNet(
layers = [
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('pool1', layers.MaxPool2DLayer),
('dropout1', layers.DropoutLayer),
('conv2', layers.Conv2DLayer),
('pool2', layers.MaxPool2DLayer),
('dropout2', layers.DropoutLayer),
('conv3', layers.Conv2DLayer),
('hidden4', layers.DenseLayer),
('output', layers.DenseLayer),
],
input_shape = (None, 1, 32, 32),
conv1_num_filters=32, conv1_filter_size=(5, 5),
pool1_pool_size=(2, 2),
dropout1_p=0.2,
conv2_num_filters=64, conv2_filter_size=(5, 5),
pool2_pool_size=(2, 2),
dropout2_p=0.2,
conv3_num_filters = 128, conv3_filter_size = (5, 5),
hidden4_num_units=500,
output_num_units = 62, output_nonlinearity = softmax,
update_learning_rate = 0.01,
update_momentum = 0.9,
batch_iterator_train = BatchIterator(batch_size = 100),
batch_iterator_test = BatchIterator(batch_size = 100),
use_label_encoder = True,
regression = False,
max_epochs = 100,
verbose = 1,
)
net.fit(reshaped_train_x, y)
prediction = net.predict(reshaped_test_x)
return prediction
# 预处理数据，首先将图片保存为32*32的小图片
imageSize = 1024 # 32 x 32
image_width = image_height = int(imageSize ** 0.5)
labelsInfoTrain = pd.read_csv\
('trainLabels.csv'.format(path))
labelsInfoTest = pd.read_csv\
('sampleSubmission.csv'.format(path))
# Load dataset
nnxTrain = read_datax('train', labelsInfoTrain, imageSize, '.')
nnxTest = read_datax('test', labelsInfoTest, imageSize, '.')
nnyTrain = map(ord, labelsInfoTrain['Class'])
nnyTrain = np.array(yTrain)
# 归一化数据
nnxTrain /= nnxTrain.std(axis = None)
nnxTrain -= nnxTrain.mean()
nnxTest /= nnxTest.std(axis = None)
nnxTest -= nnxTest.mean()
# Reshape data
train_x_reshaped = nnxTrain.reshape(nnxTrain.shape[0], 1,
image_height, image_width).astype('float32')
test_x_reshaped = nnxTest.reshape(nnxTest.shape[0], 1,
image_height, image_width).astype('float32')
# 进行训练和测试
predict = fit_model(train_x_reshaped, nnyTrain, image_width, image_height, test_x_reshaped)
# Neural Network with 352586 learnable parameters
## Layer information
# name size
--- -------- --------
0 input 1x32x32
1 conv1 32x28x28
2 pool1 32x14x14
3 dropout1 32x14x14
4 conv2 64x10x10
5 pool2 64x5x5
6 dropout2 64x5x5
7 conv3 128x1x1
8 hidden4 500
9 output 62
epoch trn loss val loss trn/val valid acc dur
------- ---------- ---------- --------- ----------- ------
1 [36m4.08201[0m [32m4.01012[0m 1.01793 0.07254 16.55s
2 [36m3.87688[0m [32m3.84326[0m 1.00875 0.04836 17.72s
3 [36m3.82788[0m [32m3.79976[0m 1.00740 0.04914 16.58s
4 [36m3.78741[0m [32m3.78872[0m 0.99965 0.07254 16.14s
5 [36m3.78030[0m [32m3.78600[0m 0.99850 0.07254 16.37s
6 [36m3.77679[0m [32m3.78520[0m 0.99778 0.07254 16.56s
7 [36m3.77487[0m 3.78537 0.99723 0.07254 16.30s
8 [36m3.77411[0m [32m3.78468[0m 0.99721 0.07254 16.51s
9 [36m3.77257[0m 3.78518 0.99667 0.07254 15.92s
10 [36m3.77202[0m [32m3.78459[0m 0.99668 0.07254 16.55s
11 [36m3.76948[0m [32m3.78458[0m 0.99601 0.07254 16.25s
12 [36m3.76882[0m [32m3.78414[0m 0.99595 0.07254 16.31s
13 [36m3.76717[0m [32m3.78411[0m 0.99552 0.07254 15.70s
14 [36m3.76606[0m 3.78469 0.99508 0.07254 16.04s
15 [36m3.76419[0m 3.78671 0.99405 0.07176 15.70s
16 [36m3.76277[0m [32m3.78392[0m 0.99441 0.07176 16.05s
17 [36m3.76014[0m 3.78821 0.99259 0.07176 15.71s
18 3.78179 3.78606 0.99887 0.07254 16.11s
19 3.76928 [32m3.78321[0m 0.99632 0.07254 15.75s
20 3.76688 3.78358 0.99559 0.07254 16.05s
21 3.76434 [32m3.78255[0m 0.99519 0.07254 17.36s
22 3.76186 [32m3.78174[0m 0.99474 0.07254 18.12s
23 [36m3.75829[0m 3.78184 0.99377 0.07878 17.90s
24 [36m3.75370[0m 3.78545 0.99161 0.07488 18.19s
25 [36m3.74749[0m [32m3.77908[0m 0.99164 0.07098 17.81s
26 [36m3.73650[0m [32m3.77806[0m 0.98900 0.07020 18.08s
27 [36m3.71592[0m [32m3.77626[0m 0.98402 0.06474 18.03s
28 [36m3.67805[0m [32m3.74531[0m 0.98204 0.07176 18.04s
29 [36m3.59550[0m 3.79802 0.94668 0.07566 18.12s
30 [36m3.44086[0m [32m3.35483[0m 1.02564 0.19111 18.06s
31 [36m3.14160[0m [32m3.00021[0m 1.04713 0.29251 17.41s
32 [36m2.73389[0m [32m2.89130[0m 0.94556 0.31903 16.19s
33 [36m2.61587[0m [32m2.53098[0m 1.03354 0.38144 15.73s
34 [36m2.25316[0m [32m2.26086[0m 0.99660 0.43994 16.14s
35 [36m1.95499[0m [32m2.03661[0m 0.95993 0.48206 15.76s
36 [36m1.75483[0m [32m1.94987[0m 0.89997 0.49610 16.01s
37 [36m1.60276[0m [32m1.78637[0m 0.89722 0.52106 15.60s
38 [36m1.47862[0m [32m1.73524[0m 0.85211 0.54524 15.98s
39 [36m1.35049[0m [32m1.65705[0m 0.81500 0.55694 15.62s
40 [36m1.27458[0m [32m1.65253[0m 0.77129 0.57254 16.01s
41 [36m1.18548[0m [32m1.60550[0m 0.73839 0.58112 15.61s
42 [36m1.11862[0m 1.62259 0.68940 0.58268 16.51s
43 [36m1.05698[0m 1.68044 0.62899 0.58112 16.24s
44 [36m1.01350[0m 1.64642 0.61558 0.59126 16.50s
45 [36m0.93587[0m 1.62059 0.57749 0.59906 15.81s
46 [36m0.87893[0m 1.65983 0.52953 0.59984 16.54s
47 [36m0.83695[0m 1.66309 0.50325 0.60452 16.42s
48 1.72887 2.92194 0.59169 0.54446 16.31s
49 3.85830 3.39520 1.13640 0.21373 15.84s
50 2.26598 1.97743 1.14592 0.46724 18.41s
51 2.11105 1.89927 1.11150 0.49298 18.02s
52 1.66393 1.75705 0.94700 0.51794 17.99s
53 1.48332 1.65795 0.89467 0.54212 17.94s
54 1.38197 [32m1.60296[0m 0.86214 0.55928 17.73s
55 1.28419 [32m1.56050[0m 0.82293 0.56318 17.94s
56 1.21078 [32m1.54983[0m 0.78123 0.57176 17.70s
57 1.13885 1.55330 0.73318 0.55616 17.93s
58 1.10488 [32m1.53462[0m 0.71997 0.57956 17.71s
59 1.03479 1.54234 0.67092 0.58502 17.70s
60 0.98439 [32m1.52492[0m 0.64554 0.59984 17.95s
61 0.93277 [32m1.49128[0m 0.62548 0.59204 17.67s
62 1.03055 1.58280 0.65109 0.57878 18.01s
63 0.89008 1.54904 0.57460 0.59750 17.69s
64 0.83698 1.59463 0.52487 0.58346 17.92s
65 [36m0.79801[0m 1.59534 0.50021 0.60452 17.80s
66 [36m0.77752[0m 1.56702 0.49618 0.60842 17.91s
67 [36m0.73901[0m 1.61821 0.45668 0.59594 17.81s
68 [36m0.71108[0m 1.56703 0.45377 0.61154 17.98s
69 [36m0.67279[0m 1.61497 0.41659 0.61154 17.81s
70 [36m0.64651[0m 1.66452 0.38841 0.60530 17.97s
71 [36m0.61597[0m 1.65828 0.37145 0.62012 17.84s
72 [36m0.59188[0m 1.69796 0.34858 0.60296 17.92s
73 [36m0.57862[0m 1.72392 0.33564 0.60686 17.73s
74 [36m0.56451[0m 1.75449 0.32175 0.60062 17.56s
75 [36m0.53835[0m 1.74351 0.30877 0.62090 17.77s
76 [36m0.53288[0m 1.80642 0.29499 0.60842 18.08s
77 [36m0.49975[0m 1.76941 0.28244 0.61700 17.76s
78 [36m0.48489[0m 1.75930 0.27561 0.60998 17.92s
79 [36m0.45688[0m 1.81943 0.25111 0.61622 17.78s
80 0.46801 1.80187 0.25974 0.62480 17.96s
81 [36m0.45527[0m 1.88136 0.24199 0.61310 17.84s
82 [36m0.43178[0m 1.93961 0.22261 0.61622 18.56s
83 [36m0.41726[0m 1.90341 0.21922 0.61856 16.52s
84 [36m0.38590[0m 1.91029 0.20201 0.61778 15.59s
85 [36m0.38510[0m 1.93524 0.19900 0.61778 16.00s
86 [36m0.37565[0m 1.92514 0.19513 0.61466 15.56s
87 [36m0.36222[0m 1.99870 0.18123 0.61544 15.88s
88 0.38495 2.08839 0.18433 0.61466 15.55s
89 [36m0.34101[0m 1.94872 0.17499 0.62559 15.97s
90 [36m0.33575[0m 2.01506 0.16662 0.61856 15.63s
91 [36m0.32353[0m 2.05956 0.15709 0.62090 16.03s
92 [36m0.30422[0m 2.12548 0.14313 0.64041 15.66s
93 [36m0.29631[0m 2.10645 0.14067 0.63495 16.02s
94 0.32050 2.11861 0.15128 0.62168 15.73s
95 0.30140 2.14516 0.14050 0.62871 15.99s
96 [36m0.28195[0m 2.09292 0.13472 0.63339 15.67s
97 0.30323 2.20744 0.13737 0.62246 16.07s
98 [36m0.27107[0m 2.15645 0.12570 0.63729 16.32s
99 0.27947 2.22565 0.12557 0.62637 16.51s
100 [36m0.26500[0m 2.22825 0.11893 0.64431 16.52s
# 保存结果
yTest = map(chr, predict)
labelsInfoTest['Class'] = yTest
labelsInfoTest.to_csv('nnresult.csv'.format(path), index = False)
提交kaggle之后的准确率：0.64562