split Linear from Layers

cabal.project.local-example

@@ -1,4 +1,3 @@
-
 optimization: False
 debug-info: True
 executable-stripping: False

core/hasktorch-core.cabal

@@ -87,6 +87,9 @@ library
     , Torch.Core.Random
     , Torch.Core.LogAdd
 
+  reexported-modules:
+      Torch.Types.Numeric
+
   exposed-modules:
       Torch.Byte
     , Torch.Byte.Dynamic
@@ -604,6 +607,7 @@ library
      , Torch.Double.NN.Conv2d
      , Torch.Double.NN.Criterion
      , Torch.Double.NN.Layers
+     , Torch.Double.NN.Linear
      , Torch.Double.NN.Math
      , Torch.Double.NN.Padding
      , Torch.Double.NN.Pooling
@@ -681,6 +685,7 @@ library
         ,Torch.Indef.Static.NN.Conv2d        as Torch.Double.NN.Conv2d
         ,Torch.Indef.Static.NN.Criterion     as Torch.Double.NN.Criterion
         ,Torch.Indef.Static.NN.Layers        as Torch.Double.NN.Layers
+        ,Torch.Indef.Static.NN.Linear        as Torch.Double.NN.Linear
         ,Torch.Indef.Static.NN.Math          as Torch.Double.NN.Math
         ,Torch.Indef.Static.NN.Padding       as Torch.Double.NN.Padding
         ,Torch.Indef.Static.NN.Pooling       as Torch.Double.NN.Pooling
@@ -1149,6 +1154,7 @@ library
        , Torch.Cuda.Double.NN.Conv2d
        , Torch.Cuda.Double.NN.Criterion
        , Torch.Cuda.Double.NN.Layers
+       , Torch.Cuda.Double.NN.Linear
        , Torch.Cuda.Double.NN.Math
        , Torch.Cuda.Double.NN.Padding
        , Torch.Cuda.Double.NN.Pooling
@@ -1222,6 +1228,7 @@ library
           ,Torch.Indef.Static.NN.Conv2d        as Torch.Cuda.Double.NN.Conv2d
           ,Torch.Indef.Static.NN.Criterion     as Torch.Cuda.Double.NN.Criterion
           ,Torch.Indef.Static.NN.Layers        as Torch.Cuda.Double.NN.Layers
+          ,Torch.Indef.Static.NN.Linear        as Torch.Cuda.Double.NN.Linear
           ,Torch.Indef.Static.NN.Math          as Torch.Cuda.Double.NN.Math
           ,Torch.Indef.Static.NN.Padding       as Torch.Cuda.Double.NN.Padding
           ,Torch.Indef.Static.NN.Pooling       as Torch.Cuda.Double.NN.Pooling
@@ -1466,6 +1473,7 @@ library hasktorch-indef-floating
     , Torch.Indef.Static.NN.Conv2d
     , Torch.Indef.Static.NN.Criterion
     , Torch.Indef.Static.NN.Layers
+    , Torch.Indef.Static.NN.Linear
     , Torch.Indef.Static.NN.Math
     , Torch.Indef.Static.NN.Padding
     , Torch.Indef.Static.NN.Pooling

core/tests/Spec.hs

@@ -8,6 +8,7 @@ import qualified GarbageCollectionSpec as GS
 import qualified Torch.Core.LogAddSpec as LS
 import qualified Torch.Core.RandomSpec as RS
 import qualified Torch.Static.NN.AbsSpec as AbsNN
+import qualified Torch.Static.NN.LinearSpec as LinearNN
 
 main :: IO ()
 main = hspec $ do
@@ -16,6 +17,7 @@ main = hspec $ do
   describe "GarbageCollectionSpec" GS.spec
   describe "Torch.Core.LogAddSpec" LS.spec
   describe "Torch.Core.RandomSpec" RS.spec
-  describe "Torch.NN.Static.AbsSpec" AbsNN.spec
+  describe "Torch.Static.NN.AbsSpec" AbsNN.spec
+  describe "Torch.Static.NN.LinearSpec" LinearNN.spec
 
 

core/tests/Torch/Static/NN/LinearSpec.hs

@@ -5,6 +5,7 @@ module Torch.Static.NN.LinearSpec where
 import Test.Hspec
 import Torch.Double
 import Numeric.Backprop
+import Torch.Double.NN.Linear
 
 main :: IO ()
 main = hspec spec

indef/hasktorch-indef.cabal

@@ -135,6 +135,7 @@ library
     -- , Torch.Indef.Static.NN.Conv3d
     , Torch.Indef.Static.NN.Criterion
     , Torch.Indef.Static.NN.Layers
+    , Torch.Indef.Static.NN.Linear
     , Torch.Indef.Static.NN.Math
     , Torch.Indef.Static.NN.Padding
     , Torch.Indef.Static.NN.Pooling

indef/src/Torch/Indef/Static/NN/Layers.hs

@@ -6,6 +6,8 @@
 -- Maintainer:  sam@stites.io
 -- Stability :  experimental
 -- Portability: non-portable
+--
+-- Miscellaneous layer functions.
 -------------------------------------------------------------------------------
 {-# LANGUAGE ScopedTypeVariables #-}
 {-# LANGUAGE FlexibleContexts #-}
@@ -24,57 +26,6 @@ import Torch.Indef.Static.Tensor.Math.Blas
 import Torch.Indef.Static.NN.Backprop ()
 import qualified Torch.Indef.Dynamic.NN as Dynamic
 
--- | datatype representing a linear layer with bias. Represents
--- @y = Ax + b@.
-newtype Linear i o
-  = Linear { getTensors :: (Tensor '[i, o], Tensor '[o]) }
-
-instance (KnownDim i, KnownDim o) => Show (Linear i o) where
-  show c = intercalate ","
-    [ "Linear ("
-    ++ "input: "  ++ show (inputSize c)
-    , " output: " ++ show (outputSize c)
-    ++ ")"
-    ]
-
-instance (KnownDim i, KnownDim o) => Backprop (Linear i o) where
-  zero = const . Linear $ (constant 0, constant 0)
-  one  = const . Linear $ (constant 1, constant 1)
-  add c0 c1 = Linear (weights c0 + weights c1, bias c0 + bias c1)
-
--- | the dense weight matrix of a linear layer
-weights :: Linear i o -> Tensor '[i, o]
-weights (Linear (w, _)) = w
-
--- | the bias vector of a linear layer
-bias :: Linear i o -> Tensor '[o]
-bias (Linear (_, b)) = b
-
--- | The input size of a linear layer
-inputSize :: forall i o . KnownDim i => Linear i o -> Int
-inputSize _ = fromIntegral (dimVal (dim :: Dim i))
-
--- | The output size of a linear layer
-outputSize :: forall i o kW dW . KnownDim o => Linear i o -> Int
-outputSize _ = fromIntegral (dimVal (dim :: Dim o))
-
--- ========================================================================= --
-
--- | Backprop linear function without batching
-linear
-  :: forall s i o
-  .  Reifies s W
-  => All KnownDim '[i,o]
-  => BVar s (Linear i o)
-  -> BVar s (Tensor '[i])
-  -> BVar s (Tensor '[o])
-linear = liftOp2 $ op2 $ \l i -> (transpose2d (weights l) `mv` i + bias l, go l i)
-  where
-    go :: Linear i o -> Tensor '[i] -> Tensor '[o] -> (Linear i o, Tensor '[i])
-    go (Linear (w, b)) i gout = (Linear (i `outer` b', b'), w `mv` b')
-      where
-        b' = gout - b
-
 -- | A backpropable 'flatten' operation
 flattenBP
   :: (Reifies s W, KnownDim (Product d), Dimensions (d::[Nat]))

indef/src/Torch/Indef/Static/NN/Linear.hs

@@ -0,0 +1,80 @@
+-------------------------------------------------------------------------------
+-- |
+-- Module    :  Torch.Indef.Static.NN.Linear
+-- Copyright :  (c) Sam Stites 2017
+-- License   :  BSD3
+-- Maintainer:  sam@stites.io
+-- Stability :  experimental
+-- Portability: non-portable
+--
+-- Linear layers
+-------------------------------------------------------------------------------
+{-# LANGUAGE ScopedTypeVariables #-}
+{-# LANGUAGE FlexibleContexts #-}
+{-# LANGUAGE MonoLocalBinds #-}
+module Torch.Indef.Static.NN.Linear where
+
+import Data.List
+import Data.Singletons.Prelude.List hiding (All)
+import Numeric.Backprop
+import Numeric.Dimensions
+
+import Torch.Indef.Types
+import Torch.Indef.Static.Tensor
+import Torch.Indef.Static.Tensor.Math
+import Torch.Indef.Static.Tensor.Math.Blas
+import Torch.Indef.Static.NN.Backprop ()
+import qualified Torch.Indef.Dynamic.NN as Dynamic
+
+-- | datatype representing a linear layer with bias. Represents
+-- @y = Ax + b@.
+newtype Linear i o
+  = Linear { getTensors :: (Tensor '[i, o], Tensor '[o]) }
+
+instance (KnownDim i, KnownDim o) => Show (Linear i o) where
+  show c = intercalate ","
+    [ "Linear ("
+    ++ "input: "  ++ show (inputSize c)
+    , " output: " ++ show (outputSize c)
+    ++ ")"
+    ]
+
+instance (KnownDim i, KnownDim o) => Backprop (Linear i o) where
+  zero = const . Linear $ (constant 0, constant 0)
+  one  = const . Linear $ (constant 1, constant 1)
+  add c0 c1 = Linear (weights c0 + weights c1, bias c0 + bias c1)
+
+-- | the dense weight matrix of a linear layer
+weights :: Linear i o -> Tensor '[i, o]
+weights (Linear (w, _)) = w
+
+-- | the bias vector of a linear layer
+bias :: Linear i o -> Tensor '[o]
+bias (Linear (_, b)) = b
+
+-- | The input size of a linear layer
+inputSize :: forall i o . KnownDim i => Linear i o -> Int
+inputSize _ = fromIntegral (dimVal (dim :: Dim i))
+
+-- | The output size of a linear layer
+outputSize :: forall i o kW dW . KnownDim o => Linear i o -> Int
+outputSize _ = fromIntegral (dimVal (dim :: Dim o))
+
+-- ========================================================================= --
+
+-- | Backprop linear function without batching
+linear
+  :: forall s i o
+  .  Reifies s W
+  => All KnownDim '[i,o]
+  => BVar s (Linear i o)
+  -> BVar s (Tensor '[i])
+  -> BVar s (Tensor '[o])
+linear = liftOp2 $ op2 $ \l i -> (transpose2d (weights l) `mv` i + bias l, go l i)
+  where
+    go :: Linear i o -> Tensor '[i] -> Tensor '[o] -> (Linear i o, Tensor '[i])
+    go (Linear (w, b)) i gout = (Linear (i `outer` b', b'), w `mv` b')
+      where
+        b' = gout - b
+
+