torch_tutorial

-- 自定义dropout的写法，参考torch官方文档 http://torch.ch/docs/developer-docs.html#_
require 'nn'

local NewClass, Parent = torch.class('nn.NewClass', 'nn.Module')

function ()
   Parent.__init(self)
   end

function NewClass:updateOutput(input)
    self.output:resizeAs(input):copy(input)
    return self.output
end

function NewClass:updateGradInput(input, gradOutput)
    self.gradInput:resizeAs(gradOutput):copy(gradOutput)
    return self.gradInput
end


-- parameters
local precision = 1e-5
local jac = nn.Jacobian

-- define inputs and module
local ini = math.random(10,20)
local percentage = 0.5
local input = torch.Tensor(ini):zero()
local module = nn.NewClass()

-- test backprop, with Jacobian
local err = jac.testJacobian(module,input)
print('==> error: ' .. err)
if err<precision then
   print('==> module OK')
else
    print('==> error too large, incorrect implementation')
end

require 'nn';
local matio = require 'matio'
-- Model Construction
nhidden = 20
ninput = 310
noutput = 3

net = nn.Sequential()
net:add(nn.Linear(ninput, nhidden))
net:add(nn.Linear(nhidden, noutput))

criterion = nn.CrossEntropyCriterion()

-- Data Preparation
train_input = matio.load('hw1/train_test/train_data.mat')
train_label = matio.load('hw1/train_test/train_label.mat')
test_input = matio.load('hw1/train_test/test_data.mat')
test_label = matio.load('hw1/train_test/test_label.mat')
train_label.train_label = train_label.train_label + 2
test_label.test_label = test_label.test_label + 2

trainset = {}
trainset.data = train_input.train_data
trainset.label = train_label.train_label
setmetatable(trainset, 
    {__index = function(t, i) 
                    return {t.data[i], t.label[i]} 
                end}
);
trainset.data = trainset.data:double() -- convert the data from a ByteTensor to a DoubleTensor.

function trainset:size() 
    return self.data:size(1) 
end

-- train feature normalization
mean = {}
stdv = {}
for i = 1,trainset.data:size(2) do
    mean[i] = trainset.data[{{},{i}}]:mean()
--     print('mean of feature' .. i .. ':'.. mean[i])
    trainset.data[{{},{i}}]:add(-mean[i])
    
    stdv[i] = trainset.data[{{},{i}}]:std()
--     print('deviation of feature' .. i .. ':' .. stdv[i])
    trainset.data[{{},{i}}]:div(stdv[i])
end
-- training settings
trainer = nn.StochasticGradient(net, criterion)
trainer.learningRate = 0.005
trainer.maxIteration = 10 -- just do 10 epochs of training.

-- training
trainer:train(trainset)

-- evaluation of model
testset = {}
testset.data = test_input.test_data
testset.label = test_label.test_label
setmetatable(testset, 
    {__index = function(t, i) 
                    return {t.data[i], t.label[i]} 
                end}
);
testset.data = testset.data:double() -- convert the data from a ByteTensor to a DoubleTensor.

function testset:size() 
    return self.data:size(1) 
end

-- test feature normalization
mean = {}
stdv = {}
for i = 1,testset.data:size(2) do
    mean[i] = testset.data[{{},{i}}]:mean()
--     print('mean of feature' .. i .. ':'.. mean[i])
    testset.data[{{},{i}}]:add(-mean[i])
    
    stdv[i] = testset.data[{{},{i}}]:std()
--     print('deviation of feature' .. i .. ':' .. stdv[i])
    testset.data[{{},{i}}]:div(stdv[i])
end

correct = 0
for i=1,testset:size() do
    local groundtruth = testset.label[i]
    local prediction = net:forward(testset.data[i])
    local confidences, indices = torch.sort(prediction, true)
    if groundtruth[1] == indices[1] then
        correct = correct + 1
    end
end

print(correct, 100*correct/testset:size() .. ' % ')