114dab4 · grenade · git.iko.soy

grenade.cabal+1

@@ -14,6 +14,7 @@ library

1414 build-depends:

1515 base >= 4.8 && < 5

1616 , bytestring == 0.10.*

17+ , async

1718 , either == 4.4.*

1819 , exceptions == 0.8.*

1920 , hmatrix

main/feedforward.hs+11−7

@@ -35,14 +35,13 @@ randomNet = do

3535 c :: FullyConnected 10 1 <- randomFullyConnected

3636 return $ a :~> Tanh :~> b :~> Relu :~> c :~> O Logit

3737

38−netTest :: MonadRandom m => Double -> Int -> m String

38+netTest :: MonadRandom m => LearningParameters -> Int -> m String

3939netTest rate n = do

4040 inps <- replicateM n $ do

41− s <- getRandom

41+ s <- getRandom

4242 return $ S1D' $ SA.randomVector s SA.Uniform * 2 - 1

4343 let outs = flip map inps $ \(S1D' v) ->

44− if v `inCircle` (fromRational 0.33, 0.33)

45− || v `inCircle` (fromRational (-0.33), 0.33)

44+ if v `inCircle` (fromRational 0.33, 0.33) || v `inCircle` (fromRational (-0.33), 0.33)

4645 then S1D' $ fromRational 1

4746 else S1D' $ fromRational 0

4847 net0 <- randomNet

@@ -70,11 +69,16 @@ netTest rate n = do

7069 normx (S1D' r) = SA.mean r

7170

7271

73−data FeedForwardOpts = FeedForwardOpts Int Double

72+data FeedForwardOpts = FeedForwardOpts Int LearningParameters

7473

7574feedForward' :: Parser FeedForwardOpts

76−feedForward' = FeedForwardOpts <$> option auto (long "examples" <> short 'e' <> value 1000000)

77− <*> option auto (long "train_rate" <> short 'r' <> value 0.01)

75+feedForward' =

76+ FeedForwardOpts <$> option auto (long "examples" <> short 'e' <> value 1000000)

77+ <*> (LearningParameters

78+ <$> option auto (long "train_rate" <> short 'r' <> value 0.01)

79+ <*> option auto (long "momentum" <> value 0.9)

80+ <*> option auto (long "l2" <> value 0.0001)

81+ )

7882

7983main :: IO ()

8084main = do

main/mnist.hs+10−6

@@ -43,7 +43,7 @@ randomMnistNet = do

4343 f :: FullyConnected 80 10 <- randomFullyConnected

4444 return $ pad :~> a :~> b :~> Relu :~> c :~> d :~> FlattenLayer :~> Relu :~> e :~> Logit :~> f :~> O Logit

4545

46−convTest :: Int -> FilePath -> FilePath -> Double -> IO ()

46+convTest :: Int -> FilePath -> FilePath -> LearningParameters -> IO ()

4747convTest iterations trainFile validateFile rate = do

4848 net0 <- evalRandIO randomMnistNet

4949 fT <- T.readFile trainFile

@@ -52,7 +52,7 @@ convTest iterations trainFile validateFile rate = do

5252 let validateRows = traverse (A.parseOnly p) (T.lines fV)

5353 case (trainRows, validateRows) of

5454 (Right tr', Right vr') -> foldM_ (runIteration tr' vr') net0 [1..iterations]

55− err -> putStrLn $ show err

55+ err -> print err

5656

5757 where

5858 trainEach !rate' !nt !(i, o) = train rate' i o nt

@@ -65,20 +65,24 @@ convTest iterations trainFile validateFile rate = do

6565 return (S2D' $ SA.fromList pixels, S1D' $ SA.fromList lab')

6666

6767 runIteration trainRows validateRows net i = do

68− let trained' = runIdentity $ foldM (trainEach (rate * (0.9 ^ i))) net trainRows

68+ let trained' = runIdentity $ foldM (trainEach rate) net trainRows

6969 let res = runIdentity $ traverse (\(rowP,rowL) -> (rowL,) <$> runNet trained' rowP) validateRows

7070 let res' = fmap (\(S1D' label, S1D' prediction) -> (maxIndex (SA.extract label), maxIndex (SA.extract prediction))) res

71− putStrLn $ show trained'

71+ print trained'

7272 putStrLn $ "Iteration " ++ show i ++ ": " ++ show (length (filter ((==) <$> fst <*> snd) res')) ++ " of " ++ show (length res')

7373 return trained'

7474

75−data MnistOpts = MnistOpts FilePath FilePath Int Double

75+data MnistOpts = MnistOpts FilePath FilePath Int LearningParameters

7676

7777mnist' :: Parser MnistOpts

7878mnist' = MnistOpts <$> (argument str (metavar "TRAIN"))

7979 <*> (argument str (metavar "VALIDATE"))

8080 <*> option auto (long "iterations" <> short 'i' <> value 15)

81− <*> option auto (long "train_rate" <> short 'r' <> value 0.01)

81+ <*> (LearningParameters

82+ <$> option auto (long "train_rate" <> short 'r' <> value 0.01)

83+ <*> option auto (long "momentum" <> value 0.9)

84+ <*> option auto (long "l2" <> value 0.0001)

85+ )

8286

8387main :: IO ()

8488main = do

src/Grenade.hs−13

@@ -1,16 +1,3 @@

1−{-# LANGUAGE BangPatterns #-}

2−{-# LANGUAGE DataKinds #-}

3−{-# LANGUAGE GADTs #-}

4−{-# LANGUAGE KindSignatures #-}

5−{-# LANGUAGE ScopedTypeVariables #-}

6−{-# LANGUAGE StandaloneDeriving #-}

7−{-# LANGUAGE TypeOperators #-}

8−{-# LANGUAGE TypeFamilies #-}

9−{-# LANGUAGE PolyKinds #-}

10−{-# LANGUAGE MultiParamTypeClasses #-}

11−{-# LANGUAGE FlexibleContexts #-}

12−{-# LANGUAGE FlexibleInstances #-}

13−

141module Grenade (

152 module X

163 ) where

src/Grenade/Core/Network.hs+26−10

@@ -1,9 +1,7 @@

1−{-# LANGUAGE BangPatterns #-}

21{-# LANGUAGE DataKinds #-}

32{-# LANGUAGE GADTs #-}

43{-# LANGUAGE KindSignatures #-}

54{-# LANGUAGE ScopedTypeVariables #-}

6−{-# LANGUAGE StandaloneDeriving #-}

75{-# LANGUAGE TypeOperators #-}

86{-# LANGUAGE TypeFamilies #-}

97{-# LANGUAGE PolyKinds #-}

@@ -14,31 +12,49 @@

1412module Grenade.Core.Network (

1513 Layer (..)

1614 , Network (..)

15+ , UpdateLayer (..)

16+ , LearningParameters (..)

1717 ) where

1818

19+import Data.Typeable

1920import Grenade.Core.Shape

2021

22+data LearningParameters = LearningParameters {

23+ learningRate :: Double

24+ , learningMomentum :: Double

25+ , learningRegulariser :: Double

26+ } deriving (Eq, Show)

27+

28+-- | Class for updating a layer. All layers implement this, and it is

29+-- shape independent.

30+class UpdateLayer (m :: * -> *) x where

31+ -- | The type for the gradient for this layer.

32+ -- Unit if there isn't a gradient to pass back.

33+ type Gradient x :: *

34+ -- | Update a layer with its gradient and learning parameters

35+ runUpdate :: LearningParameters -> x -> Gradient x -> m x

36+

2137-- | Class for a layer. All layers implement this, however, they don't

2238-- need to implement it for all shapes, only ones which are appropriate.

23−class Layer (m :: * -> *) x (i :: Shape) (o :: Shape) where

39+class UpdateLayer m x => Layer (m :: * -> *) x (i :: Shape) (o :: Shape) where

2440 -- | Used in training and scoring. Take the input from the previous

2541 -- layer, and give the output from this layer.

2642 runForwards :: x -> S' i -> m (S' o)

27− -- | Back propagate a step. Takes a learning rate (move from here?)

28− -- the current layer, the input that the layer gave from the input

29− -- and the back propagated derivatives from the layer above.

30− -- Returns the updated layer and the derivatives to push back further.

31− runBackards :: Double -> x -> S' i -> S' o -> m (x, S' i)

43+ -- | Back propagate a step. Takes the current layer, the input that the

44+ -- layer gave from the input and the back propagated derivatives from

45+ -- the layer above.

46+ -- Returns the gradient layer and the derivatives to push back further.

47+ runBackards :: x -> S' i -> S' o -> m (Gradient x, S' i)

3248

3349-- | Type of a network.

3450-- The [Shape] type specifies the shapes of data passed between the layers.

3551-- Could be considered to be a heterogeneous list of layers which are able to

3652-- transform the data shapes of the network.

3753data Network :: (* -> *) -> [Shape] -> * where

38− O :: (Show x, Layer m x i o, KnownShape o, KnownShape i)

54+ O :: (Typeable x, Show x, Layer m x i o, KnownShape o, KnownShape i)

3955 => !x

4056 -> Network m '[i, o]

41− (:~>) :: (Show x, Layer m x i h, KnownShape h, KnownShape i)

57+ (:~>) :: (Typeable x, Show x, Layer m x i h, KnownShape h, KnownShape i)

4258 => !x

4359 -> !(Network m (h ': hs))

4460 -> Network m (i ': h ': hs)

src/Grenade/Core/Runner.hs+13−9

@@ -1,9 +1,7 @@

11{-# LANGUAGE BangPatterns #-}

22{-# LANGUAGE GADTs #-}

33{-# LANGUAGE DataKinds #-}

4−{-# LANGUAGE KindSignatures #-}

54{-# LANGUAGE ScopedTypeVariables #-}

6−{-# LANGUAGE StandaloneDeriving #-}

75{-# LANGUAGE TypeOperators #-}

86{-# LANGUAGE TypeFamilies #-}

97

@@ -18,7 +16,7 @@ import Grenade.Core.Shape

1816

1917-- | Update a network with new weights after training with an instance.

2018train :: forall m i o hs. (Monad m, Head hs ~ i, Last hs ~ o, KnownShape i, KnownShape o)

21− => Double -- ^ learning rate

19+ => LearningParameters -- ^ learning rate

2220 -> S' i -- ^ input vector

2321 -> S' o -- ^ target vector

2422 -> Network m hs -- ^ network to train

@@ -35,20 +33,26 @@ train rate x0 target = fmap fst . go x0

3533 -- run the rest of the network, and get the layer from above.

3634 (n', dWs') <- go y n

3735 -- calculate the gradient for this layer to pass down,

38− (layer', dWs) <- runBackards rate layer x dWs'

39− return (layer' :~> n', dWs)

36+ (layer', dWs) <- runBackards layer x dWs'

37+

38+ -- Update this layer using the gradient

39+ newLayer <- runUpdate rate layer layer'

40+

41+ return (newLayer :~> n', dWs)

4042

4143 -- handle the output layer, bouncing the derivatives back down.

4244 go !x (O layer)

43− = do y <- runForwards layer x

45+ = do y <- runForwards layer x

4446 -- the gradient (how much y affects the error)

45− (layer', dWs) <- runBackards rate layer x (y - target)

46− return (O layer', dWs)

47+ (layer', dWs) <- runBackards layer x (y - target)

48+ newLayer <- runUpdate rate layer layer'

49+

50+ return (O newLayer, dWs)

4751

4852-- | Just forwards propagation with no training.

4953runNet :: forall m hs. (Monad m)

5054 => Network m hs

51− -> (S' (Head hs)) -- ^ input vector

55+ -> S' (Head hs) -- ^ input vector

5256 -> m (S' (Last hs)) -- ^ target vector

5357runNet (layer :~> n) !x = do y <- runForwards layer x

5458 runNet n y

src/Grenade/Layers/Convolution.hs+39−11

@@ -2,6 +2,7 @@

22{-# LANGUAGE DataKinds #-}

33{-# LANGUAGE ScopedTypeVariables #-}

44{-# LANGUAGE StandaloneDeriving #-}

5+{-# LANGUAGE RecordWildCards #-}

56{-# LANGUAGE GADTs #-}

67{-# LANGUAGE TypeOperators #-}

78{-# LANGUAGE TypeFamilies #-}

@@ -65,6 +66,24 @@ data Convolution :: Nat -- ^ Number of channels, for the first layer this could

6566 -> !(L kernelFlattened filters) -- ^ The last kernel update (or momentum)

6667 -> Convolution channels filters kernelRows kernelColumns strideRows strideColumns

6768

69+data Convolution' :: Nat -- ^ Number of channels, for the first layer this could be RGB for instance.

70+ -> Nat -- ^ Number of filters, this is the number of channels output by the layer.

71+ -> Nat -- ^ The number of rows in the kernel filter

72+ -> Nat -- ^ The number of column in the kernel filter

73+ -> Nat -- ^ The row stride of the convolution filter

74+ -> Nat -- ^ The columns stride of the convolution filter

75+ -> * where

76+ Convolution' :: ( KnownNat channels

77+ , KnownNat filters

78+ , KnownNat kernelRows

79+ , KnownNat kernelColumns

80+ , KnownNat strideRows

81+ , KnownNat strideColumns

82+ , KnownNat kernelFlattened

83+ , kernelFlattened ~ (kernelRows * kernelColumns * channels))

84+ => !(L kernelFlattened filters) -- ^ The kernel filter gradient

85+ -> Convolution' channels filters kernelRows kernelColumns strideRows strideColumns

86+

6887instance Show (Convolution c f k k' s s') where

6988 show (Convolution a _) = renderConv a

7089 where

@@ -99,6 +118,22 @@ randomConvolution = do

99118 mm = konst 0

100119 return $ Convolution wN mm

101120

121+instance ( Monad m

122+ , KnownNat kernelRows

123+ , KnownNat kernelCols

124+ , KnownNat channels

125+ , KnownNat filters

126+ , KnownNat strideRows

127+ , KnownNat strideCols

128+ , kernelFlattened ~ (kernelRows * kernelColumns * channels)

129+ ) => UpdateLayer m (Convolution channels filters kernelRows kernelCols strideRows strideCols) where

130+ type Gradient (Convolution channels filters kernelRows kernelCols strideRows strideCols) = (Convolution' channels filters kernelRows kernelCols strideRows strideCols)

131+ runUpdate LearningParameters {..} (Convolution oldKernel oldMomentum) (Convolution' kernelGradient) = do

132+ let newMomentum = konst learningMomentum * oldMomentum - konst learningRate * kernelGradient

133+ regulariser = konst (learningRegulariser * learningRate) * oldKernel

134+ newKernel = oldKernel + newMomentum - regulariser

135+ return $ Convolution newKernel newMomentum

136+

102137-- | A two dimentional image may have a convolution filter applied to it

103138instance ( Monad m

104139 , KnownNat kernelRows

@@ -127,7 +162,7 @@ instance ( Monad m

127162 r = col2vid 1 1 1 1 ox oy mt

128163 rs = fmap (fromJust . create) r

129164 in return . S3D' $ mkVector rs

130− runBackards rate (Convolution kernel momentum) (S2D' input) (S3D' dEdy) =

165+ runBackards (Convolution kernel _) (S2D' input) (S3D' dEdy) =

131166 let ex = extract input

132167 ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)

133168 iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)

@@ -145,14 +180,10 @@ instance ( Monad m

145180 vs = vid2col 1 1 1 1 ox oy eo

146181

147182 kN = fromJust . create $ tr c LA.<> vs

148− mm = momentum * 0.9 - konst rate * kN

149− wd = konst (0.0001 * rate) * kernel

150− rM = kernel + mm - wd

151−

152183 dW = vs LA.<> tr ek

153184

154185 xW = col2im kx ky sx sy ix iy dW

155− in return (Convolution rM mm, S2D' . fromJust . create $ xW)

186+ in return (Convolution' kN, S2D' . fromJust . create $ xW)

156187

157188

158189-- | A three dimensional image (or 2d with many channels) can have

@@ -187,7 +218,7 @@ instance ( Monad m

187218 r = col2vid 1 1 1 1 ox oy mt

188219 rs = fmap (fromJust . create) r

189220 in return . S3D' $ mkVector rs

190− runBackards rate (Convolution kernel momentum) (S3D' input) (S3D' dEdy) =

221+ runBackards (Convolution kernel _) (S3D' input) (S3D' dEdy) =

191222 let ex = vecToList $ fmap extract input

192223 ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)

193224 iy = fromIntegral $ natVal (Proxy :: Proxy inputCols)

@@ -205,14 +236,11 @@ instance ( Monad m

205236 vs = vid2col 1 1 1 1 ox oy eo

206237

207238 kN = fromJust . create $ tr c LA.<> vs

208− mm = momentum * 0.9 - konst rate * kN

209− wd = konst (0.0005 * rate) * kernel

210− rM = kernel + mm - wd

211239

212240 dW = vs LA.<> tr ek

213241

214242 xW = col2vid kx ky sx sy ix iy dW

215− in return (Convolution rM mm, S3D' . mkVector . fmap (fromJust . create) $ xW)

243+ in return (Convolution' kN, S3D' . mkVector . fmap (fromJust . create) $ xW)

216244

217245im2col :: Int -> Int -> Int -> Int -> Matrix Double -> Matrix Double

218246im2col nrows ncols srows scols m =

src/Grenade/Layers/Crop.hs+6−4

@@ -1,7 +1,5 @@

1−{-# LANGUAGE BangPatterns #-}

21{-# LANGUAGE DataKinds #-}

32{-# LANGUAGE ScopedTypeVariables #-}

4−{-# LANGUAGE StandaloneDeriving #-}

53{-# LANGUAGE GADTs #-}

64{-# LANGUAGE TypeOperators #-}

75{-# LANGUAGE TypeFamilies #-}

@@ -39,6 +37,10 @@ data Crop :: Nat

3937instance Show (Crop cropLeft cropTop cropRight cropBottom) where

4038 show Crop = "Crop"

4139

40+instance Monad m => UpdateLayer m (Crop l t r b) where

41+ type Gradient (Crop l t r b) = ()

42+ runUpdate _ x _ = return x

43+

4244-- | A two dimentional image can be cropped.

4345instance ( Monad m

4446 , KnownNat cropLeft

@@ -60,11 +62,11 @@ instance ( Monad m

6062 m = extract input

6163 r = subMatrix (cropt, cropl) (nrows, ncols) m

6264 in return . S2D' . fromJust . create $ r

63− runBackards _ crop _ (S2D' dEdy) =

65+ runBackards _ _ (S2D' dEdy) =

6466 let cropl = fromIntegral $ natVal (Proxy :: Proxy cropLeft)

6567 cropt = fromIntegral $ natVal (Proxy :: Proxy cropTop)

6668 cropr = fromIntegral $ natVal (Proxy :: Proxy cropRight)

6769 cropb = fromIntegral $ natVal (Proxy :: Proxy cropBottom)

6870 eo = extract dEdy

6971 vs = diagBlock [konst 0 (cropt,cropl), eo, konst 0 (cropb,cropr)]

70− in return (crop, S2D' . fromJust . create $ vs)

72+ in return ((), S2D' . fromJust . create $ vs)

src/Grenade/Layers/Dropout.hs+7−7

@@ -1,7 +1,5 @@

1−{-# LANGUAGE BangPatterns #-}

21{-# LANGUAGE DataKinds #-}

32{-# LANGUAGE ScopedTypeVariables #-}

4−{-# LANGUAGE StandaloneDeriving #-}

53{-# LANGUAGE TypeOperators #-}

64{-# LANGUAGE TypeFamilies #-}

75{-# LANGUAGE MultiParamTypeClasses #-}

@@ -32,6 +30,10 @@ import Numeric.LinearAlgebra.Static

3230data Dropout o = Dropout Double (R o)

3331 deriving Show

3432

33+instance (MonadRandom m, KnownNat i) => UpdateLayer m (Dropout i) where

34+ type Gradient (Dropout i) = ()

35+ runUpdate _ (Dropout rate _) _ = randomDropout rate

36+

3537randomDropout :: (MonadRandom m, KnownNat i)

3638 => Double -> m (Dropout i)

3739randomDropout rate = do

@@ -44,8 +46,6 @@ instance (MonadRandom m, MonadState Phase m, KnownNat i) => Layer m (Dropout i)

4446 runForwards (Dropout rate drops) (S1D' x) = isTrainingPhase >>= \case

4547 True -> return . S1D' $ x * drops

4648 False -> return . S1D' $ dvmap (* (1 - rate)) x

47− runBackards _ oldDropout@(Dropout rate drops) _ (S1D' x) = isTrainingPhase >>= \case

48− True -> do

49− newDropout <- randomDropout rate

50− return (newDropout, S1D' $ x * drops)

51− False -> return (oldDropout, S1D' $ dvmap (* (1 - rate)) x)

49+ runBackards (Dropout rate drops) _ (S1D' x) = isTrainingPhase >>= \case

50+ True -> return ((), S1D' $ x * drops)

51+ False -> return ((), S1D' $ dvmap (* (1 - rate)) x)

src/Grenade/Layers/Flatten.hs+8−4

@@ -25,20 +25,24 @@ import Grenade.Core.Network

2525data FlattenLayer = FlattenLayer

2626 deriving Show

2727

28+instance Monad m => UpdateLayer m FlattenLayer where

29+ type Gradient FlattenLayer = ()

30+ runUpdate _ _ _ = return FlattenLayer

31+

2832instance (Monad m, KnownNat a, KnownNat x, KnownNat y, a ~ (x * y)) => Layer m FlattenLayer ('D2 x y) ('D1 a) where

29− runForwards _ (S2D' y) = return $ S1D' . fromList . toList . flatten . extract $ y

30− runBackards _ _ _ (S1D' y) = return (FlattenLayer, S2D' . fromList . toList . unwrap $ y)

33+ runForwards _ (S2D' y) = return $ S1D' . fromList . toList . flatten . extract $ y

34+ runBackards _ _ (S1D' y) = return ((), S2D' . fromList . toList . unwrap $ y)

3135

3236instance (Monad m, KnownNat a, KnownNat x, KnownNat y, KnownNat z, a ~ (x * y * z)) => Layer m FlattenLayer ('D3 x y z) ('D1 a) where

3337 runForwards _ (S3D' y) = return $ S1D' . raiseShapeError . create . vjoin . vecToList . fmap (flatten . extract) $ y

34− runBackards _ _ _ (S1D' o) = do

38+ runBackards _ _ (S1D' o) = do

3539 let x' = fromIntegral $ natVal (Proxy :: Proxy x)

3640 y' = fromIntegral $ natVal (Proxy :: Proxy y)

3741 z' = fromIntegral $ natVal (Proxy :: Proxy z)

3842 vecs = takesV (replicate z' (x' * y')) (extract o)

3943 ls = fmap (raiseShapeError . create . reshape y') vecs

4044 ls' = mkVector ls :: Vector z (L x y)

41− return (FlattenLayer, S3D' ls')

45+ return ((), S3D' ls')

4246

4347raiseShapeError :: Maybe a -> a

4448raiseShapeError (Just x) = x

src/Grenade/Layers/FullyConnected.hs+21−11

@@ -1,7 +1,6 @@

1−{-# LANGUAGE BangPatterns #-}

21{-# LANGUAGE DataKinds #-}

32{-# LANGUAGE ScopedTypeVariables #-}

4−{-# LANGUAGE StandaloneDeriving #-}

3+{-# LANGUAGE RecordWildCards #-}

54{-# LANGUAGE TypeOperators #-}

65{-# LANGUAGE TypeFamilies #-}

76{-# LANGUAGE MultiParamTypeClasses #-}

@@ -25,25 +24,36 @@ import Grenade.Core.Shape

2524data FullyConnected i o = FullyConnected

2625 !(R o) -- Bias neuron weights

2726 !(L o i) -- Activation weights

28− !(L o i) -- Activation momentums

27+ !(L o i) -- Momentum

28+

29+data FullyConnected' i o = FullyConnected'

30+ !(R o) -- Bias neuron gradient

31+ !(L o i) -- Activation gradient

2932

3033instance Show (FullyConnected i o) where

31− show (FullyConnected _ _ _) = "FullyConnected"

34+ show FullyConnected {} = "FullyConnected"

35+

36+instance (Monad m, KnownNat i, KnownNat o) => UpdateLayer m (FullyConnected i o) where

37+ type Gradient (FullyConnected i o) = (FullyConnected' i o)

38+

39+ runUpdate LearningParameters {..} (FullyConnected oldBias oldActivations oldMomentum) (FullyConnected' biasGradient activationGradient) = do

40+ let newBias = oldBias - konst learningRate * biasGradient

41+ newMomentum = konst learningMomentum * oldMomentum - konst learningRate * activationGradient

42+ regulariser = konst (learningRegulariser * learningRate) * oldActivations

43+ newActivations = oldActivations + newMomentum - regulariser

44+ return $ FullyConnected newBias newActivations newMomentum

3245

3346instance (Monad m, KnownNat i, KnownNat o) => Layer m (FullyConnected i o) ('D1 i) ('D1 o) where

3447 -- Do a matrix vector multiplication and return the result.

3548 runForwards (FullyConnected wB wN _) (S1D' v) = return $ S1D' (wB + wN #> v)

3649

3750 -- Run a backpropogation step for a full connected layer.

38− runBackards rate (FullyConnected wB wN mm) (S1D' x) (S1D' dEdy) =

39− let wB' = wB - konst rate * dEdy

40− mm' = 0.9 * mm - konst rate * (dEdy `outer` x)

41− wd' = konst (0.0001 * rate) * wN

42− wN' = wN + mm' - wd'

43− w' = FullyConnected wB' wN' mm'

51+ runBackards (FullyConnected _ wN _) (S1D' x) (S1D' dEdy) =

52+ let wB' = dEdy

53+ mm' = dEdy `outer` x

4454 -- calcluate derivatives for next step

4555 dWs = tr wN #> dEdy

46− in return (w', S1D' dWs)

56+ in return (FullyConnected' wB' mm', S1D' dWs)

4757

4858randomFullyConnected :: (MonadRandom m, KnownNat i, KnownNat o)

4959 => m (FullyConnected i o)

src/Grenade/Layers/Fuse.hs+15−9

@@ -3,7 +3,6 @@

33{-# LANGUAGE GADTs #-}

44{-# LANGUAGE KindSignatures #-}

55{-# LANGUAGE ScopedTypeVariables #-}

6−{-# LANGUAGE StandaloneDeriving #-}

76{-# LANGUAGE TypeOperators #-}

87{-# LANGUAGE TypeFamilies #-}

98{-# LANGUAGE PolyKinds #-}

@@ -23,23 +22,30 @@ import Grenade.Core.Shape

2322-- This can be used to simplify a network if a complicated repeated structure is used.

2423-- This does however have a trade off, internal incremental states in the Wengert tape are

2524-- not retained during reverse accumulation. So less RAM is used, but more compute is required.

26−data Fuse :: (* -> *) -> Shape -> Shape -> Shape -> * where

25+data Fuse :: (* -> *) -> * -> * -> Shape -> Shape -> Shape -> * where

2726 (:$$) :: (Show x, Show y, Layer m x i h, Layer m y h o, KnownShape h, KnownShape i, KnownShape o)

2827 => !x

2928 -> !y

30− -> Fuse m i h o

29+ -> Fuse m x y i h o

3130infixr 5 :$$

3231

33−instance Show (Fuse m i h o) where

32+instance Show (Fuse m x y i h o) where

3433 show (x :$$ y) = "(" ++ show x ++ " :$$ " ++ show y ++ ")"

3534

36−instance (Monad m, KnownShape i, KnownShape h, KnownShape o) => Layer m (Fuse m i h o) i o where

35+instance (Monad m, KnownShape i, KnownShape h, KnownShape o) => UpdateLayer m (Fuse m x y i h o) where

36+ type Gradient (Fuse m x y i h o) = (Gradient x, Gradient y)

37+ runUpdate lr (x :$$ y) (x', y') = do

38+ newX <- runUpdate lr x x'

39+ newY <- runUpdate lr y y'

40+ return (newX :$$ newY)

41+

42+instance (Monad m, KnownShape i, KnownShape h, KnownShape o) => Layer m (Fuse m x y i h o) i o where

3743 runForwards (x :$$ y) input = do

3844 yInput :: S' h <- runForwards x input

3945 runForwards y yInput

4046

41− runBackards rate (x :$$ y) input backGradient = do

47+ runBackards (x :$$ y) input backGradient = do

4248 yInput :: S' h <- runForwards x input

43− (y', yGrad) <- runBackards rate y yInput backGradient

44− (x', xGrad) <- runBackards rate x input yGrad

45− return (x' :$$ y', xGrad)

49+ (y', yGrad) <- runBackards y yInput backGradient

50+ (x', xGrad) <- runBackards x input yGrad

51+ return ((x', y'), xGrad)

src/Grenade/Layers/Logit.hs+8−5

@@ -1,7 +1,5 @@

1−{-# LANGUAGE BangPatterns #-}

21{-# LANGUAGE DataKinds #-}

32{-# LANGUAGE ScopedTypeVariables #-}

4−{-# LANGUAGE StandaloneDeriving #-}

53{-# LANGUAGE TypeOperators #-}

64{-# LANGUAGE TypeFamilies #-}

75{-# LANGUAGE MultiParamTypeClasses #-}

@@ -11,6 +9,7 @@ module Grenade.Layers.Logit (

119 Logit (..)

1210 ) where

1311

12+

1413import Data.Singletons.TypeLits

1514import Grenade.Core.Network

1615import Grenade.Core.Vector

@@ -23,17 +22,21 @@ import Grenade.Core.Shape

2322data Logit = Logit

2423 deriving Show

2524

25+instance Monad m => UpdateLayer m Logit where

26+ type Gradient Logit = ()

27+ runUpdate _ _ _ = return Logit

28+

2629instance (Monad m, KnownNat i) => Layer m Logit ('D1 i) ('D1 i) where

2730 runForwards _ (S1D' y) = return $ S1D' (logistic y)

28− runBackards _ _ (S1D' y) (S1D' dEdy) = return (Logit, S1D' (logistic' y * dEdy))

31+ runBackards _ (S1D' y) (S1D' dEdy) = return ((), S1D' (logistic' y * dEdy))

2932

3033instance (Monad m, KnownNat i, KnownNat j) => Layer m Logit ('D2 i j) ('D2 i j) where

3134 runForwards _ (S2D' y) = return $ S2D' (logistic y)

32− runBackards _ _ (S2D' y) (S2D' dEdy) = return (Logit, S2D' (logistic' y * dEdy))

35+ runBackards _ (S2D' y) (S2D' dEdy) = return ((), S2D' (logistic' y * dEdy))

3336

3437instance (Monad m, KnownNat i, KnownNat j, KnownNat k) => Layer m Logit ('D3 i j k) ('D3 i j k) where

3538 runForwards _ (S3D' y) = return $ S3D' (fmap logistic y)

36− runBackards _ _ (S3D' y) (S3D' dEdy) = return (Logit, S3D' (vectorZip (\y' dEdy' -> logistic' y' * dEdy') y dEdy))

39+ runBackards _ (S3D' y) (S3D' dEdy) = return ((), S3D' (vectorZip (\y' dEdy' -> logistic' y' * dEdy') y dEdy))

3740

3841

3942logistic :: Floating a => a -> a

src/Grenade/Layers/Pad.hs+6−4

@@ -1,7 +1,5 @@

1−{-# LANGUAGE BangPatterns #-}

21{-# LANGUAGE DataKinds #-}

32{-# LANGUAGE ScopedTypeVariables #-}

4−{-# LANGUAGE StandaloneDeriving #-}

53{-# LANGUAGE GADTs #-}

64{-# LANGUAGE TypeOperators #-}

75{-# LANGUAGE TypeFamilies #-}

@@ -39,6 +37,10 @@ data Pad :: Nat

3937instance Show (Pad padLeft padTop padRight padBottom) where

4038 show Pad = "Pad"

4139

40+instance Monad m => UpdateLayer m (Pad l t r b) where

41+ type Gradient (Pad l t r b) = ()

42+ runUpdate _ x _ = return x

43+

4244-- | A two dimentional image can be padped.

4345instance ( Monad m

4446 , KnownNat padLeft

@@ -60,11 +62,11 @@ instance ( Monad m

6062 m = extract input

6163 r = diagBlock [konst 0 (padt,padl), m, konst 0 (padb,padr)]

6264 in return . S2D' . fromJust . create $ r

63− runBackards _ pad _ (S2D' dEdy) =

65+ runBackards Pad _ (S2D' dEdy) =

6466 let padl = fromIntegral $ natVal (Proxy :: Proxy padLeft)

6567 padt = fromIntegral $ natVal (Proxy :: Proxy padTop)

6668 nrows = fromIntegral $ natVal (Proxy :: Proxy inputRows)

6769 ncols = fromIntegral $ natVal (Proxy :: Proxy inputColumns)

6870 m = extract dEdy

6971 vs = subMatrix (padt, padl) (nrows, ncols) m

70− in return (pad, S2D' . fromJust . create $ vs)

72+ in return ((), S2D' . fromJust . create $ vs)

src/Grenade/Layers/Pooling.hs+8−4

@@ -51,6 +51,10 @@ instance Show (Pooling k k' s s') where

5151 show Pooling = "Pooling"

5252

5353

54+instance Monad m => UpdateLayer m (Pooling kernelRows kernelColumns strideRows strideColumns) where

55+ type Gradient (Pooling kr kc sr sc) = ()

56+ runUpdate _ Pooling _ = return Pooling

57+

5458-- | A two dimentional image can be pooled.

5559instance ( Monad m

5660 , KnownNat kernelRows

@@ -75,7 +79,7 @@ instance ( Monad m

7579 r = poolForward kx ky sx sy ox oy $ ex

7680 rs = fromJust . create $ r

7781 in return . S2D' $ rs

78− runBackards _ Pooling (S2D' input) (S2D' dEdy) =

82+ runBackards Pooling (S2D' input) (S2D' dEdy) =

7983 let kx = fromIntegral $ natVal (Proxy :: Proxy kernelRows)

8084 ky = fromIntegral $ natVal (Proxy :: Proxy kernelColumns)

8185 sx = fromIntegral $ natVal (Proxy :: Proxy strideRows)

@@ -83,7 +87,7 @@ instance ( Monad m

8387 ex = extract input

8488 eo = extract dEdy

8589 vs = poolBackward kx ky sx sy ex eo

86− in return (Pooling, S2D' . fromJust . create $ vs)

90+ in return ((), S2D' . fromJust . create $ vs)

8791

8892

8993-- | A three dimensional image can be pooled on each layer.

@@ -112,7 +116,7 @@ instance ( Monad m

112116 r = poolForwardList kx ky sx sy ix iy ox oy ex

113117 rs = fmap (fromJust . create) r

114118 in return . S3D' $ rs

115− runBackards _ Pooling (S3D' input) (S3D' dEdy) =

119+ runBackards Pooling (S3D' input) (S3D' dEdy) =

116120 let ix = fromIntegral $ natVal (Proxy :: Proxy inputRows)

117121 iy = fromIntegral $ natVal (Proxy :: Proxy inputColumns)

118122 kx = fromIntegral $ natVal (Proxy :: Proxy kernelRows)

@@ -123,7 +127,7 @@ instance ( Monad m

123127 eo = fmap extract dEdy

124128 ez = vectorZip (,) ex eo

125129 vs = poolBackwardList kx ky sx sy ix iy ez

126− in return (Pooling, S3D' . fmap (fromJust . create) $ vs)

130+ in return ((), S3D' . fmap (fromJust . create) $ vs)

127131

128132poolForward :: Int -> Int -> Int -> Int -> Int -> Int -> Matrix Double -> Matrix Double

129133poolForward nrows ncols srows scols outputRows outputCols m =

src/Grenade/Layers/Relu.hs+7−5

@@ -1,7 +1,5 @@

1−{-# LANGUAGE BangPatterns #-}

21{-# LANGUAGE DataKinds #-}

32{-# LANGUAGE ScopedTypeVariables #-}

4−{-# LANGUAGE StandaloneDeriving #-}

53{-# LANGUAGE TypeOperators #-}

64{-# LANGUAGE TypeFamilies #-}

75{-# LANGUAGE MultiParamTypeClasses #-}

@@ -24,11 +22,15 @@ import qualified Numeric.LinearAlgebra.Static as LAS

2422data Relu = Relu

2523 deriving Show

2624

25+instance Monad m => UpdateLayer m Relu where

26+ type Gradient Relu = ()

27+ runUpdate _ _ _ = return Relu

28+

2729instance (Monad m, KnownNat i) => Layer m Relu ('D1 i) ('D1 i) where

2830 runForwards _ (S1D' y) = return $ S1D' (relu y)

2931 where

3032 relu = LAS.dvmap (\a -> if a <= 0 then 0 else a)

31− runBackards _ _ (S1D' y) (S1D' dEdy) = return (Relu, S1D' (relu' y * dEdy))

33+ runBackards _ (S1D' y) (S1D' dEdy) = return ((), S1D' (relu' y * dEdy))

3234 where

3335 relu' = LAS.dvmap (\a -> if a <= 0 then 0 else 1)

3436

@@ -36,7 +38,7 @@ instance (Monad m, KnownNat i, KnownNat j) => Layer m Relu ('D2 i j) ('D2 i j) w

3638 runForwards _ (S2D' y) = return $ S2D' (relu y)

3739 where

3840 relu = LAS.dmmap (\a -> if a <= 0 then 0 else a)

39− runBackards _ _ (S2D' y) (S2D' dEdy) = return (Relu, S2D' (relu' y * dEdy))

41+ runBackards _ (S2D' y) (S2D' dEdy) = return ((), S2D' (relu' y * dEdy))

4042 where

4143 relu' = LAS.dmmap (\a -> if a <= 0 then 0 else 1)

4244

@@ -44,6 +46,6 @@ instance (Monad m, KnownNat i, KnownNat j, KnownNat k) => Layer m Relu ('D3 i j

4446 runForwards _ (S3D' y) = return $ S3D' (fmap relu y)

4547 where

4648 relu = LAS.dmmap (\a -> if a <= 0 then 0 else a)

47− runBackards _ _ (S3D' y) (S3D' dEdy) = return (Relu, S3D' (vectorZip (\y' dEdy' -> relu' y' * dEdy') y dEdy))

49+ runBackards _ (S3D' y) (S3D' dEdy) = return ((), S3D' (vectorZip (\y' dEdy' -> relu' y' * dEdy') y dEdy))

4850 where

4951 relu' = LAS.dmmap (\a -> if a <= 0 then 0 else 1)

src/Grenade/Layers/Tanh.hs+7−5

@@ -1,7 +1,5 @@

1−{-# LANGUAGE BangPatterns #-}

21{-# LANGUAGE DataKinds #-}

32{-# LANGUAGE ScopedTypeVariables #-}

4−{-# LANGUAGE StandaloneDeriving #-}

53{-# LANGUAGE TypeOperators #-}

64{-# LANGUAGE TypeFamilies #-}

75{-# LANGUAGE MultiParamTypeClasses #-}

@@ -21,17 +19,21 @@ import Grenade.Core.Shape

2119data Tanh = Tanh

2220 deriving Show

2321

22+instance Monad m => UpdateLayer m Tanh where

23+ type Gradient Tanh = ()

24+ runUpdate _ _ _ = return Tanh

25+

2426instance (Monad m, KnownNat i) => Layer m Tanh ('D1 i) ('D1 i) where

2527 runForwards _ (S1D' y) = return $ S1D' (tanh y)

26− runBackards _ _ (S1D' y) (S1D' dEdy) = return (Tanh, S1D' (tanh' y * dEdy))

28+ runBackards _ (S1D' y) (S1D' dEdy) = return ((), S1D' (tanh' y * dEdy))

2729

2830instance (Monad m, KnownNat i, KnownNat j) => Layer m Tanh ('D2 i j) ('D2 i j) where

2931 runForwards _ (S2D' y) = return $ S2D' (tanh y)

30− runBackards _ _ (S2D' y) (S2D' dEdy) = return (Tanh, S2D' (tanh' y * dEdy))

32+ runBackards _ (S2D' y) (S2D' dEdy) = return ((), S2D' (tanh' y * dEdy))

3133

3234instance (Monad m, KnownNat i, KnownNat j, KnownNat k) => Layer m Tanh ('D3 i j k) ('D3 i j k) where

3335 runForwards _ (S3D' y) = return $ S3D' (fmap tanh y)

34− runBackards _ _ (S3D' y) (S3D' dEdy) = return (Tanh, S3D' (vectorZip (\y' dEdy' -> tanh' y' * dEdy') y dEdy))

36+ runBackards _ (S3D' y) (S3D' dEdy) = return ((), S3D' (vectorZip (\y' dEdy' -> tanh' y' * dEdy') y dEdy))

3537

3638tanh' :: (Floating a) => a -> a

3739tanh' t = 1 - s ^ (2 :: Int) where s = tanh t