Fix Ref and Build semantics for generated code. (#37)

Also:
- Make TensorFlow.Ops.{variable,assign} be the Core generated versions.
- Make ops take "Shape" as mandatory input.

tensorflow-opgen/src/TensorFlow/OpGen.hs

@@ -16,7 +16,32 @@
 {-# LANGUAGE LambdaCase #-}
 {-# LANGUAGE OverloadedStrings #-}
 {-# LANGUAGE TypeFamilies #-}
--- | Rendering of TensorFlow operations as Haskell functions.
+{- | Rendering of TensorFlow operations as Haskell functions.
+
+The basic type signature generated for each op is:
+
+> {constraints} => {mandatory attrs} -> {input tensors} -> {output tensors}
+
+where:
+
+* @{mandatory attrs}@ is of the form @A_1 -> ... -> A_N@, where each @A@ is an
+ op attribute that doesn't have a default and can't be inferred from other
+ inputs.
+
+* @{constraints}@ restrict the type parameters of the input and output tensors
+ (for example: 'TensorType' or 'OneOf').
+
+* @{input tensors}@ is of the form @T_1 -> ... -> T_N@, where each @T@ is of
+the form @Tensor Ref a@, @Tensor v a@ or @ResourceHandle a@ (or a list of one
+of those types), and @a@ is either a concrete type or a (constrained) type
+variable.
+
+* @{output tensors}@ is of the form @(T_1,...,T_N)@ for "pure" ops, and
+@Build (T_1,...,T_N)@ for "stateful" ops.  An op is considered "stateful" if
+it takes a @Tensor Ref@ or @ResourceHandle@ as input, or if it's explicitly
+marked \"Stateful\" in its @REGISTER_OP@ definition.  (If there are no outputs,
+it is either @ControlNode@ or @Build ControlNode@.)
+-}
 
 module TensorFlow.OpGen
   ( OpGenFlags(..)
@@ -259,15 +284,18 @@ typeSig pOp = constraints
         AttrFloat -> "Float"
         AttrBool -> "Bool"
         AttrType -> "DataType"
-        AttrShape -> "TensorShapeProto"
+        AttrShape -> "Shape"
         AttrTensor -> "TensorProto"
 
     tensorArgAndComment t = tensorArg t <+> hang 0 ("-- ^" <+> argComment t)
     outputs = case parsedOutputs pOp of
-        [] -> "ControlNode"
+        [] -> wrapOutput "ControlNode"
         -- TODO(judahjacobson): To improve indentation: `tensorArgAndComment a`
-        [a] -> tensorArg a <+> "-- ^" <+> argComment a
-        as -> tuple (map tensorArg as) <+/> resultComment as
+        [a] -> wrapOutput (tensorArg a) <+> "-- ^" <+> argComment a
+        as -> wrapOutput (tuple (map tensorArg as)) <+/> resultComment as
+    wrapOutput o
+        | parsedOpIsMonadic pOp = "Build" <+> parens o
+        | otherwise = o
         
 -- | Render an op input or output.
 -- For example: "Tensor Ref Int64", "Tensor v t", "ResourceHandle dtype"

tensorflow-opgen/src/TensorFlow/OpGen/ParsedOp.hs

@@ -41,6 +41,8 @@ import Proto.Tensorflow.Core.Framework.OpDef
     , description
     , name
     , inputArg
+    , isRef
+    , isStateful
     , outputArg
     , summary
     , typeListAttr
@@ -67,6 +69,10 @@ data ParsedOp = ParsedOp
         -- Attributes which are list sizes (ints) that are inferred automatically
         -- from one or more of the input tensors.
         -- Associated with the list of tensors whose size it describes.
+    , parsedOpIsMonadic :: Bool
+        -- ^ Whether this op is stateful or takes a stateful input.  Such ops
+        -- should not be CSE'd and must be monadic in our API (i.e., return a
+        -- Build action).
     }
 
 data Name = Name
@@ -127,6 +133,10 @@ data ArgKind
     | ArgTensorEither Text -- Tensor v a; the Text is the variable `v`
     | ArgResource -- Resource a
 
+isRefKind :: ArgKind -> Bool
+isRefKind ArgTensorRef = True
+isRefKind ArgResource = True
+isRefKind _ = False
 
 makeName :: Text -> Name
 makeName n = Name
@@ -197,6 +207,8 @@ parseOp o = ParsedOp
     { parsedOpName = makeName $ o ^. name
     , parsedOpSummary = o ^. summary
     , parsedOpDescription = o ^. description
+    , parsedOpIsMonadic = o ^. isStateful
+                            || any (isRefKind . parsedArgKind) parsedInputs
     , ..
     }
   where
@@ -218,11 +230,13 @@ parseOp o = ParsedOp
 inputTensorKind :: OpDef'ArgDef -> Text -> ArgKind
 inputTensorKind a v
     | a ^. type' == DT_RESOURCE = ArgResource
+    | a ^. isRef = ArgTensorRef
     | otherwise = ArgTensorEither v
 
 outputTensorKind :: OpDef'ArgDef -> ArgKind
 outputTensorKind a
     | a ^. type' == DT_RESOURCE = ArgResource
+    | a ^. isRef = ArgTensorRef
     | otherwise = ArgTensorValue
 
 getExplicitInputAttr :: Set.Set TFName -> OpDef'AttrDef -> Maybe AttrType
@@ -230,7 +244,7 @@ getExplicitInputAttr implicitAttrs a
     | TFName (a ^. name) `Set.notMember` implicitAttrs
     , a ^. maybe'defaultValue == Nothing
     , t <- parseAttrType (a ^. type')
-    , t `elem` map AttrSingle [AttrBool, AttrInt64, AttrFloat] = Just t
+    , t `elem` map AttrSingle [AttrBool, AttrInt64, AttrFloat, AttrShape] = Just t
     | otherwise = Nothing
 
 getInferredTypeAttr :: OpDef'AttrDef -> Maybe [DataType]

tensorflow-ops/src/TensorFlow/Ops.hs

@@ -60,7 +60,7 @@ module TensorFlow.Ops
     , CoreOps.abs
     , CoreOps.addN
     , CoreOps.argMax
-    , assign
+    , CoreOps.assign
     , CoreOps.broadcastGradientArgs
     , CoreOps.cast
     , CoreOps.concat
@@ -97,7 +97,7 @@ module TensorFlow.Ops
     , CoreOps.sum
     , CoreOps.transpose
     , truncatedNormal
-    , variable
+    , CoreOps.variable
     , vector
     , zeros
     , CoreOps.zerosLike
@@ -154,31 +154,18 @@ matTranspose :: forall a v . TensorType a
              => Tensor v a -> Tensor Value a
 matTranspose = flip CoreOps.transpose (vector [1, 0 :: Int32])
 
--- | Create a new, uninitialized stateful Tensor of the given shape.
-variable :: forall a . TensorType a => Shape -> Build (Tensor Ref a)
-variable shape' = buildOp $ opDef "Variable"
-                          & opAttr "shape" .~ shape'
-                          & opAttr "dtype" .~ tensorType (undefined :: a)
-
 placeholder :: forall a . TensorType a => Shape -> Build (Tensor Value a)
 placeholder shape' =
     buildOp $ opDef "Placeholder"
             & opAttr "dtype" .~ tensorType (undefined :: a)
             & opAttr "shape" .~ shape'
 
--- Assign returns the input ref.
-assign :: forall a v . TensorType a
-       => Tensor Ref a -> Tensor v a -> Build (Tensor Ref a)
-assign = buildOp $ opDef "Assign"
-                      & opAttr "T" .~ tensorType (undefined :: a)
-                      & opAttr "use_locking" .~ True
-
 -- | Creates a variable initialized to the given value.
 -- Initialization happens next time session runs.
 initializedVariable :: forall a . TensorType a
                     => Tensor Value a -> Build (Tensor Ref a)
 initializedVariable initializer = do
-    v <- variable []  -- The shape is not known initially.
+    v <- CoreOps.variable []  -- The shape is not known initially.
     (i :: Tensor Ref a) <-
         buildOp (opDef "Assign"
                  & opAttr "T" .~ tensorType (undefined :: a)
@@ -220,7 +207,8 @@ restoreFromName :: forall a . TensorType a
 restoreFromName path name x = do
     let restoreOp = buildOp $ opDef "Restore"
                             & opAttr "dt" .~ tensorType (undefined :: a)
-    group =<< assign x (restoreOp (scalar path) (scalar name) :: Tensor Value a)
+    group =<< CoreOps.assign x
+                (restoreOp (scalar path) (scalar name) :: Tensor Value a)
 
 -- | Restore a tensor's value from a checkpoint file.
 restore :: forall a . TensorType a