Gaussian Naive Bayers Classifier

naive.lua

require 'torch'
_ = require 'moses'
csv2tensor = require 'csv2tensor'

local dataset = csv2tensor.load('pima-indians-diabetes.data')
local trainRatio = 0.67

local dataSize = dataset:size(1)
local trainSize = torch.floor(dataSize * trainRatio)
local testSize = dataSize - trainSize

local trainData = dataset:narrow(1, 1, trainSize)
local testData = dataset:narrow(1, trainSize+1, testSize)

local featuresSize = trainData:size(2)-1
local featuresTrain = trainData:narrow(2,1,featuresSize)
local labelsTrain = trainData:select(2, featuresSize+1)
local featuresTest = testData:narrow(2,1,featuresSize)
local labelsTest = testData:select(2, featuresSize+1)
local classCount = {}
local classFeatures = {}
local pXC = {}
local pC = {}
local numClasses = 0

print(string.format('split %d rows into train=%d and test=%d rows', dataSize, trainSize, testSize))

function calcProbability(x, mean, std)
  return torch.cdiv(torch.exp(-torch.cdiv(torch.pow(x-mean, 2),torch.pow(std,2)*2)), std*torch.sqrt(2*math.pi))
end

for i=1,trainSize do
  local class = labelsTrain[i]
  classCount[class] = (classCount[class] or 0) + 1
end

for class,classSize in pairs(classCount) do
  classFeatures[class] = torch.Tensor(classSize, featuresSize)
  classCount[class] = 1
  numClasses = numClasses + 1
end

for i=1,trainSize do
  local class = labelsTrain[i]
  local classIndex = classCount[class]
  classFeatures[class][classIndex] = featuresTrain[i]
  classCount[class] = classIndex + 1
end

for class,features in pairs(classFeatures) do
  local n = features:size(1)
  pXC[class] = {
    mean = torch.mean(features,1):select(1,1),
    std = torch.std(features,1):select(1,1)
  }
  pC[class] = n/trainSize
end

local accuracy = 0
for i=1,featuresTest:size(1) do
  --local p = _.clone(pC)
  local p = _.map(pC, function(k,v) return 1 end)

  for class,pp in pairs(p) do
    local fp = pXC[class]
    p[class] = pp * torch.prod(calcProbability(featuresTest[i], fp.mean, fp.std))
  end

  local outP = 0
  local out = nil
  local totalP = 0
  for class,pp in pairs(p) do
    totalP = totalP + pp
    if pp > outP then
      out = class
      outP = pp
    end
  end

  outP = outP / totalP
  if out == labelsTest[i] then
    accuracy = accuracy + 1
  end
end
accuracy = 100 * accuracy / testSize

print('accuracy:', accuracy)