Main.hs

module Main where

import ML.NN
import System.Random
import System.Environment
import Data.Matrix                          as M
import qualified Data.Vector                as V
import qualified Data.Sequence              as S
import qualified Data.Foldable              as F
import qualified Data.ByteString.Char8 as BS

-- 랜덤 값 생성
makeRand x = replicate x $ randomRIO (0.0, 1.0)

-- 행렬을 받아 가장 큰 값만 1로 만들고 나머지를 0으로 만드는 함수
oneHot x = fromList len 1 $ replicate r 0 ++ [1] ++ iterate (\x -> x) 0
  where 
    r = snd . maximum $ zip (toList x) [0 .. length x - 1]
    len = length x

main = do
  -- 학습용 데이터를 불러옴
  rawdata <- fmap BS.lines $ BS.readFile ".\\resource\\wine_train.csv"

  let set = fmap (fmap (\x -> read x :: Float) . fmap BS.unpack . V.fromList . BS.split ',') $ V.fromList rawdata

  [stepstr] <- getArgs

  let step = read stepstr :: Int

  -- 입력 리스트
  let x = fmap (colVector . V.take 13) set
  
  -- 목표출력 리스트
  let y = fmap (colVector . V.drop 13) set

  -- 뉴런 네트워크 생성
  seed1 <- sequence $ makeRand 52
  seed2 <- sequence $ makeRand 12
  let s1 = fromList 4 13 seed1
  let b1 = fromList 4 1 $ replicate 4 0
  let s2 = fromList 3 4 $ seed2
  let b2 = fromList 3 1 $ replicate 3 0
  let h1 = (Layer sigmoid s1 b1)
  let h2 = (Layer sigmoid s2 b2)
  let dataset = V.zip x y

  -- 랜덤하게 샘플 선정
--  samplesindex <- sequence . replicate 160000 $ randomRIO (0, V.length dataset - 1)

  -- 샘플 데이터 리스트 생성
--  let samples = fmap (\x -> dataset !! x) $ V.fromList samplesindex

  -- 확률적 경사 하강법으로 학습
--  let result = sdg 0.1 dataset (V.fromList samplesindex) (V.fromList [h1,h2])

  gen <- getStdGen
  
  let result = sdg2 gen 0.1 step dataset (V.fromList [h1,h2])

  -- 테스트용 데이터를 불러옴
  testdata <- fmap BS.lines $ BS.readFile ".\\resource\\wine_test.csv"

  let testSet = fmap (fmap (\x -> read x :: Float) . fmap BS.unpack . V.fromList . BS.split ',') $ V.fromList testdata

  -- 입력 리스트
  let testX = fmap (colVector . V.take 13) testSet
  -- 목표 출력 리스트
  let testY = fmap (colVector . V.drop 13) testSet

  putStr "테스트 오차율 : "

  putStr . show . (*100) . (/ fromIntegral (V.length testY)) . V.sum
    . fmap (\x -> if x then 0 else 1)
    . V.zipWith (==) testY $ fmap (oneHot . executeLayer result) testX
  putStrLn "%"