deeplearning-hs 0.1.0.1 → 0.1.0.2

raw patch · 2 files changed

+42/−30 lines, 2 filesdep ~basePVP: major bump suggested

API removals or changes: PVP suggests a major version bump

Dependency ranges changed: base

API changes (from Hackage documentation)

Files

• DeepLearning/ConvNet.hs +41/−29
• deeplearning-hs.cabal +1/−1

@@ -14,18 +14,28 @@ -} module DeepLearning.ConvNet     (-     (>->),+     -- ** Main Types+     Vol,      DVol,-     Forward,+     Label,++     -- ** Layers+     Layer,      InnerLayer,-     SoftMaxLayer(..),      TopLayer,-     Vol,+     SoftMaxLayer(..),+     FullyConnectedLayer(..),++     -- ** Composing layers+     (>->),+     Forward,+     withActivations,++     -- ** Network building helpers      flowNetwork,      net1,      net2,      newFC,-     withActivations,     ) where  import           Control.Monad as CM@@ -35,9 +45,6 @@ import qualified Data.Vector.Unboxed                  as V import           Prelude                              as P hiding (map, zipWith) ---- ** Helper Types- -- |Activation matrix type Vol sh = Array U sh Double -- |Delayed activation matrix@@ -46,18 +53,24 @@ -- |Label for supervised learning type Label = Int --- **  Top Layers+-- |'Layer' reprsents a layer that can pass activations forward.+-- 'TopLayer' and 'InnerLayer' are derived layers that can be+-- backpropagated through.+class (Shape sh, Shape sh') => Layer a sh sh' | a -> sh, a -> sh' where+    forward :: (Monad m) => a -> Vol sh -> m (DVol sh')+ -- |'TopLayer' is a top level layer that can initialize a -- backpropagation pass.-class TopLayer a where-    topForward :: (Monad m) => a -> Vol DIM1 -> m (DVol DIM1)+class Layer a DIM1 DIM1 => TopLayer a where     topBackward :: (Monad m) => a -> Label -> Vol DIM1 -> Vol DIM1 -> m (DVol DIM1)  -- |'SoftMaxLayer' computes the softmax activation function.-data SoftMaxLayer = SoftMaxLayer -- +data SoftMaxLayer = SoftMaxLayer -- +instance Layer SoftMaxLayer DIM1 DIM1 where+    forward _ = softMaxForward+ instance TopLayer SoftMaxLayer where-    topForward _ = softMaxForward     topBackward _ = softMaxBackward  softMaxForward :: (Shape sh, Monad m) => Vol sh -> m (DVol sh)@@ -80,11 +93,9 @@               indicator = label == target               bool2Double x = if x then 1.0 else 0.0 --- ** Inner Layers -- |'InnerLayer' represents an inner layer of a neural network that -- can accept backpropagation input from higher layers-class (Shape sh, Shape sh') => InnerLayer a sh sh' | a -> sh, a -> sh' where-    innerForward :: Monad m => a -> Vol sh -> m (DVol sh')+class (Layer a sh sh', Shape sh, Shape sh') => InnerLayer a sh sh' | a -> sh, a -> sh' where     innerBackward :: Monad m => a -> Vol sh' -> Vol sh -> m (DVol sh)  -- |'FullyConnectedLayer' represents a fully-connected input layer@@ -93,8 +104,10 @@       _bias    :: Vol DIM1     } +instance (Shape sh) => Layer (FullyConnectedLayer sh) sh DIM1 where+    forward = fcForward+ instance (Shape sh) => InnerLayer (FullyConnectedLayer sh) sh DIM1 where-    innerForward = fcForward     innerBackward = fcBackward  fcForward :: (Shape sh, Monad m)@@ -118,12 +131,10 @@       prod lv rv = V.sum $ V.zipWith (*) lv rv  --- ** Composing Layers- -- |The 'Forward' function represents a single forward pass through a layer. type Forward m sh sh' = (Vol sh -> WriterT [V.Vector Double] m (DVol sh')) --- |'(>->)' composes two forward activation functions+-- |'>->' composes two forward activation functions (>->) :: (Monad m, Shape sh, Shape sh', Shape sh'')         => Forward m sh sh' -> Forward m sh' sh'' -> Forward m sh sh'' (f >-> g) input = do@@ -137,8 +148,7 @@ net1   :: (Monad m, InnerLayer a sh DIM1, TopLayer a1) =>      a -> a1 -> Forward m sh DIM1-net1 bottom top = innerForward bottom >-> topForward top-+net1 bottom top = forward bottom >-> forward top  -- |'net1' constructs a two-layer fully connected MLP with -- softmax output.@@ -146,7 +156,7 @@   :: (Monad m, InnerLayer a sh sh', InnerLayer a1 sh' DIM1,       TopLayer a2) =>      a -> a1 -> a2 -> Forward m sh DIM1-net2 bottom middle top = innerForward bottom >-> net1 middle top+net2 bottom middle top = forward bottom >-> net1 middle top  -- |'withActivations' computes the output activation, along with the -- intermediate activations@@ -177,9 +187,11 @@ flowNetwork :: (Monad m, Shape sh) => sh -> Int -> Int -> Int -> Forward m sh DIM1 flowNetwork inputShape numHiddenLayers numHiddenNodes numClasses =     inputLayer >-> innerLayers >-> preTopLayer >-> topLayer-    where-      inputLayer = innerForward $ newFC inputShape numHiddenNodes-      innerLayers = flatInner $ P.fmap (\_ -> newFC (Z :. numHiddenNodes) numHiddenNodes) [1..numHiddenLayers]-      preTopLayer = innerForward $ newFC (Z :. numHiddenNodes) numClasses-      topLayer = topForward SoftMaxLayer-      flatInner layers = P.foldl1 (>->) (P.fmap innerForward layers)+        where+          flatInner layers = P.foldl1 (>->) (P.fmap forward layers)+          innerLayer = newFC (Z :. numHiddenNodes) numHiddenNodes+          innerLayers = flatInner $ P.fmap (const innerLayer)+                                           [1..numHiddenLayers]+          inputLayer = forward $ newFC inputShape numHiddenNodes+          preTopLayer = forward $ newFC (Z :. numHiddenNodes) numClasses+          topLayer = forward SoftMaxLayer

@@ -3,7 +3,7 @@   name:                deeplearning-hs-version:             0.1.0.1+version:             0.1.0.2 description:         Implements type-safe deep neural networks synopsis:            Deep Learning in Haskell homepage:            https://github.com/ajtulloch/deeplearning-hs