LibLinear（SVM包）使用說明之（三）實踐

liblinear（svm包）使用說明之（三）實踐

liblinear（svm包）使用說明之（三）實踐

我們在ufldl的教程中，exercise: convolution and pooling這一章節，已經得到了cnnpooledfeatures.mat特徵。在該練習中，我們使用的是softmax分類器來分類的。在這裡我們修改為用svm來替代softmax分類器。svm由liblinear軟體包來提供。這裡是四分類問題，所以liblinear會根據我們傳入的訓練樣本訓練四個二分類器，以實現四分類。以前由softmax分類器得到的準確率是80.406%。在這裡換成liblinear後，準確率變為80.75%。在這裡差別不是很大。

在本文的例子中，我們增加了scale和cross validation，cross validation是用來選擇乙個最好的引數c的（不知道自己這兩個步驟有沒有正確，如有錯誤，還望大家提醒，謝謝）。

具體的**如下：

[plain]view plain

copy

%// classification by liblinear

%// liblinear:

%// author : zouxy

%// date : 2013-9-2

%// homepage :

%// email : [email protected]

clear; clc;

%%% step1: load data

fprintf(1,'step1: load data...\n');

% pooledfeaturestrain大小為400*2000*3*3

% pooledfeaturestest大小為400*3200*3*3

% 第一維是特徵個數，也就是特徵圖個數，第二維是樣本個數，第三維是特徵圖的寬，

% 第四維是特徵圖的高

load cnnpooledfeatures.mat;

load stltrainsubset.mat % loads numtrainimages, trainimages, trainlabels

load stltestsubset.mat % loads numtestimages, testimages, testlabels

% b = permute(a,order) 按照向量order指定的順序重排a的各維

train_x = permute(pooledfeaturestrain, [1 3 4 2]);

% 將每個樣本的特徵拉成乙個列向量，每個樣本乙個列，矩陣大小為3600*2000

train_x = reshape(train_x, numel(pooledfeaturestrain) / numtrainimages, numtrainimages);

train_y = trainlabels; % 2000*1

test_x = permute(pooledfeaturestest, [1 3 4 2]);

test_x = reshape(test_x, numel(pooledfeaturestest) / numtestimages, numtestimages);

test_y = testlabels;

% release some memory

clear trainimages testimages pooledfeaturestrain pooledfeaturestest;

%%% step2: scale the data

fprintf(1,'step2: scale data...\n');

% using the same scaling factors for training and testing sets,

% we obtain much better accuracy. note: scale each attribute(feature), not sample

% scale to [0 1]

% when a is a vector, b = (a - min(a)) .* (upper - lower) ./ (max(a)-min(a)) + lower

lower = 0;

upper = 1.0;

train_x = train_x';

x_max = max(train_x);

x_min = min(train_x);

train_x = (train_x - repmat(x_min, size(train_x, 1), 1)) .* (upper - lower) ...

./ repmat((x_max - x_min), size(train_x, 1), 1) + lower;

test_x = test_x';

test_x = (test_x - repmat(x_min, size(test_x, 1), 1)) .* (upper - lower) ...

./ repmat((x_max - x_min), size(test_x, 1), 1) + lower;

% note: before scale the accuracy is 80.4688%, after scale it turns to 80.1875%,

% and took more time. so is that my scale operation wrong or other reasons?

% after adding bias, accuracy = 80.75% (2584/3200)

%%% step3: cross validation for choosing parameter

fprintf(1,'step3: cross validation for choosing parameter c...\n');

% the larger c is, more time should be costed

c = [2^-6 2^-5 2^-4 2^-3 2^-2 2^-1 2^0 2^1 2^2 2^3];

max_acc = 0;

tic;

for i = 1 : size(c, 2)

option = ['-b 1 -c ' num2str(c(i)) ' -v 5 -q'];

fprintf(1,'stage: %d/%d: c = %d, ', i, size(c, 2), c(i));

accuracy = train(train_y, sparse(train_x), option);

if accuracy > max_acc

max_acc = accuracy;

best_c = i;

end

fprintf(1,'the best c is c = %d.\n', c(best_c));

toc;

%%% step4: train the model

fprintf(1,'step4: training...\n');

tic;

option = ['-c ' num2str(c(best_c)) ' -b 1 -e 0.001'];

model = train(train_y, sparse(train_x), option);

toc;

%%% step5: test the model

fprintf(1,'step5: testing...\n');

tic;

[predict_label, accuracy, dec_values] = predict(test_y, sparse(test_x), model);

toc;

LibLinear（SVM包）使用說明之（三）實踐

r mysql包 R使用RMySQL包

DBMS RANDOM包的使用

javascript閉包使用

LibLinear（SVM包）使用說明之（三）實踐

r mysql包 R使用RMySQL包

DBMS RANDOM包的使用

javascript閉包使用

相關推薦