src/Gargantext/Text/List/Learn.hs

   1 {-|
   2 Module      : Gargantext.Text.List.Learn
   3 Description : Learn to make lists
   4 Copyright   : (c) CNRS, 2018-Present
   5 License     : AGPL + CECILL v3
   6 Maintainer  : team@gargantext.org
   7 Stability   : experimental
   8 Portability : POSIX
   9
  10 CSV parser for Gargantext corpus files.
  11
  12 -}
  13
  14 {-# OPTIONS_GHC -fno-warn-orphans #-}
  15
  16
  17 module Gargantext.Text.List.Learn
  18   where
  19
  20 import Control.Monad.Reader (MonadReader)
  21 -- TODO remvoe this deps
  22 import Gargantext.API.Admin.Settings
  23 import Data.Map (Map)
  24 import Data.Maybe (maybe)
  25 import Gargantext.Core.Types.Main (ListType(..), listTypeId, fromListTypeId)
  26 import Gargantext.Prelude
  27 import Gargantext.Prelude.Utils
  28 import Gargantext.Text.Metrics.Count (occurrencesWith)
  29 import qualified Data.IntMap as IntMap
  30 import qualified Data.List   as List
  31 import qualified Data.Map    as Map
  32 import qualified Data.SVM    as SVM
  33 import qualified Data.Vector as Vec
  34
  35 ------------------------------------------------------------------------
  36 train :: Double -> Double -> SVM.Problem -> IO SVM.Model
  37 train x y = (SVM.train (SVM.CSvc x) (SVM.RBF y))
  38
  39 predict :: SVM.Model -> [Vec.Vector Double] -> IO [Double]
  40 predict m vs = mapM (predict' m) vs
  41   where
  42     predict' m' vs' = SVM.predict m' (IntMap.fromList $ (zip [1..]) $ Vec.toList vs')
  43
  44 ------------------------------------------------------------------------
  45 trainList :: Double -> Double -> Map ListType [Vec.Vector Double] -> IO SVM.Model
  46 trainList x y = (train x y) . trainList'
  47   where
  48     trainList' :: Map ListType [Vec.Vector Double] -> SVM.Problem
  49     trainList' = mapVec2problem . (Map.mapKeys (fromIntegral . listTypeId))
  50
  51     mapVec2problem :: Map Double [Vec.Vector Double] -> SVM.Problem
  52     mapVec2problem = List.concat . (map (\(a,as) -> zip (repeat a) as)) . Map.toList . (Map.map vecs2maps)
  53
  54     vecs2maps :: [Vec.Vector Double] -> [IntMap.IntMap Double]
  55     vecs2maps = map (IntMap.fromList . (zip [1..]) . Vec.toList)
  56
  57
  58 predictList :: Model -> [Vec.Vector Double] -> IO [Maybe ListType]
  59 predictList (ModelSVM m _ _) vs = map (fromListTypeId . round) <$> predict m vs
  60
  61 ------------------------------------------------------------------------
  62 data Model = ModelSVM { modelSVM :: SVM.Model
  63                       , param1 :: Maybe Double
  64                       , param2 :: Maybe Double
  65                       }
  66 --{-
  67 instance SaveFile Model
  68   where
  69     saveFile' fp (ModelSVM m _ _) = SVM.saveModel m fp
  70
  71 instance ReadFile Model
  72   where
  73     readFile' fp = do
  74       m <- SVM.loadModel fp
  75       pure $ ModelSVM m Nothing Nothing
  76 --}
  77 ------------------------------------------------------------------------
  78 -- | TODO
  79 -- shuffle list
  80 -- split list : train / test
  81 -- grid parameters on best result on test
  82
  83 type Train = Map ListType [Vec.Vector Double]
  84 type Tests = Map ListType [Vec.Vector Double]
  85 type Score = Double
  86 type Param = Double
  87
  88 grid :: (MonadReader env m, MonadBase IO m, HasSettings env)
  89      => Param -> Param -> Train -> [Tests] -> m (Maybe Model)
  90 grid _ _ _ []  = panic "Gargantext.Text.List.Learn.grid : empty test data"
  91 grid s e tr te = do
  92   let
  93     grid' :: (MonadReader env m, MonadBase IO m, HasSettings env)
  94           => Double -> Double
  95           -> Train
  96           -> [Tests]
  97           -> m (Score, Model)
  98     grid' x y tr' te' = do
  99       model'' <- liftBase $ trainList x y tr'
 100
 101       let
 102         model' = ModelSVM model'' (Just x) (Just y)
 103
 104         score' :: [(ListType, Maybe ListType)] -> Map (Maybe Bool) Int
 105         score' = occurrencesWith (\(a,b) -> (==) <$> Just a <*> b)
 106
 107         score'' :: Map (Maybe Bool) Int -> Double
 108         score'' m'' = maybe 0 (\t -> (fromIntegral t)/total) (Map.lookup (Just True) m'')
 109           where
 110             total = fromIntegral $ foldl (+) 0 $ Map.elems m''
 111
 112         getScore m t = do
 113           let (res, toGuess) = List.unzip
 114                              $ List.concat
 115                              $ map (\(k,vs) -> zip (repeat k) vs)
 116                              $ Map.toList t
 117
 118           res' <- liftBase $ predictList m toGuess
 119           pure $ score'' $ score' $ List.zip res res'
 120
 121       score <- mapM (getScore model') te'
 122       pure (mean score, model')
 123
 124   r <- head . List.reverse
 125             . (List.sortOn fst)
 126            <$> mapM (\(x,y) -> grid' x y tr te)
 127                          [(x,y) | x <- [s..e], y <- [s..e]]
 128
 129   printDebug "GRID SEARCH" (map fst r)
 130   --printDebug "file" fp
 131   --fp <- saveFile (ModelSVM model')
 132   --save best result
 133   pure $ snd <$> r
 134
 135
 136