{-# LANGUAGE CPP #-} {-# LANGUAGE BangPatterns #-} {-# LANGUAGE DeriveDataTypeable #-} {-# LANGUAGE FlexibleContexts #-} {-# LANGUAGE FlexibleInstances #-} {-# LANGUAGE FunctionalDependencies #-} {-# LANGUAGE MultiParamTypeClasses #-} {-# LANGUAGE TypeFamilies #-} {-# LANGUAGE TypeOperators #-} {-# LANGUAGE UndecidableInstances #-} {-# OPTIONS_GHC -fno-warn-name-shadowing #-} {-# OPTIONS_HADDOCK not-home #-} ----------------------------------------------------------------------------- -- | -- Copyright : (c) Edward Kmett 2010-2015 -- License : BSD3 -- Maintainer : ekmett@gmail.com -- Stability : experimental -- Portability : GHC only -- -- Unsafe and often partial combinators intended for internal usage. -- -- Handle with care. ----------------------------------------------------------------------------- module Numeric.AD.Internal.Sparse ( Monomial(..) , emptyMonomial , addToMonomial , indices , Sparse(..) , apply , vars , d, d', ds , skeleton , spartial , partial , vgrad , vgrad' , vgrads , Grad(..) , Grads(..) , terms , primal ) where import Prelude hiding (lookup) #if __GLASGOW_HASKELL__ < 710 import Control.Applicative hiding ((<**>)) #endif import Control.Comonad.Cofree import Control.Monad (join) import Data.Data import Data.IntMap (IntMap, unionWith, findWithDefault, toAscList, singleton, insertWith, lookup) import qualified Data.IntMap as IntMap import Data.Number.Erf import Data.Traversable import Data.Typeable () import Numeric.AD.Internal.Combinators import Numeric.AD.Jacobian import Numeric.AD.Mode newtype Monomial = Monomial (IntMap Int) emptyMonomial :: Monomial emptyMonomial = Monomial IntMap.empty {-# INLINE emptyMonomial #-} addToMonomial :: Int -> Monomial -> Monomial addToMonomial k (Monomial m) = Monomial (insertWith (+) k 1 m) {-# INLINE addToMonomial #-} indices :: Monomial -> [Int] indices (Monomial as) = uncurry (flip replicate) `concatMap` toAscList as {-# INLINE indices #-} -- | We only store partials in sorted order, so the map contained in a partial -- will only contain partials with equal or greater keys to that of the map in -- which it was found. This should be key for efficiently computing sparse hessians. -- there are only (n + k - 1) choose (k - 1) distinct nth partial derivatives of a -- function with k inputs. data Sparse a = Sparse !a (IntMap (Sparse a)) | Zero deriving (Show, Data, Typeable) vars :: (Traversable f, Num a) => f a -> f (Sparse a) vars = snd . mapAccumL var 0 where var !n a = (n + 1, Sparse a $ singleton n $ auto 1) {-# INLINE vars #-} apply :: (Traversable f, Num a) => (f (Sparse a) -> b) -> f a -> b apply f = f . vars {-# INLINE apply #-} skeleton :: Traversable f => f a -> f Int skeleton = snd . mapAccumL (\ !n _ -> (n + 1, n)) 0 {-# INLINE skeleton #-} d :: (Traversable f, Num a) => f b -> Sparse a -> f a d fs (Zero) = 0 <$ fs d fs (Sparse _ da) = snd $ mapAccumL (\ !n _ -> (n + 1, maybe 0 primal $ lookup n da)) 0 fs {-# INLINE d #-} d' :: (Traversable f, Num a) => f a -> Sparse a -> (a, f a) d' fs Zero = (0, 0 <$ fs) d' fs (Sparse a da) = (a, snd $ mapAccumL (\ !n _ -> (n + 1, maybe 0 primal $ lookup n da)) 0 fs) {-# INLINE d' #-} ds :: (Traversable f, Num a) => f b -> Sparse a -> Cofree f a ds fs Zero = r where r = 0 :< (r <$ fs) ds fs (as@(Sparse a _)) = a :< (go emptyMonomial <$> fns) where fns = skeleton fs -- go :: Monomial -> Int -> Cofree f a go ix i = partial (indices ix') as :< (go ix' <$> fns) where ix' = addToMonomial i ix {-# INLINE ds #-} partialS :: Num a => [Int] -> Sparse a -> Sparse a partialS [] a = a partialS (n:ns) (Sparse _ da) = partialS ns $ findWithDefault Zero n da partialS _ Zero = Zero {-# INLINE partialS #-} partial :: Num a => [Int] -> Sparse a -> a partial [] (Sparse a _) = a partial (n:ns) (Sparse _ da) = partial ns $ findWithDefault (auto 0) n da partial _ Zero = 0 {-# INLINE partial #-} spartial :: Num a => [Int] -> Sparse a -> Maybe a spartial [] (Sparse a _) = Just a spartial (n:ns) (Sparse _ da) = do a' <- lookup n da spartial ns a' spartial _ Zero = Nothing {-# INLINE spartial #-} primal :: Num a => Sparse a -> a primal (Sparse a _) = a primal Zero = 0 (<**>) :: Floating a => Sparse a -> Sparse a -> Sparse a Zero <**> y = auto (0 ** primal y) _ <**> Zero = auto 1 x <**> y@(Sparse b bs) | IntMap.null bs = lift1 (**b) (\z -> b *^ z <**> Sparse (b-1) IntMap.empty) x | otherwise = lift2_ (**) (\z xi yi -> (yi * z / xi, z * log xi)) x y instance Num a => Mode (Sparse a) where type Scalar (Sparse a) = a auto a = Sparse a IntMap.empty zero = Zero Zero ^* _ = Zero Sparse a as ^* b = Sparse (a * b) $ fmap (^* b) as _ *^ Zero = Zero a *^ Sparse b bs = Sparse (a * b) $ fmap (a *^) bs Zero ^/ _ = Zero Sparse a as ^/ b = Sparse (a / b) $ fmap (^/ b) as infixr 6 <+> (<+>) :: Num a => Sparse a -> Sparse a -> Sparse a Zero <+> a = a a <+> Zero = a Sparse a as <+> Sparse b bs = Sparse (a + b) $ unionWith (<+>) as bs -- The instances for Jacobian for Sparse and Tower are almost identical; -- could easily be made exactly equal by small changes. instance Num a => Jacobian (Sparse a) where type D (Sparse a) = Sparse a unary f _ Zero = auto (f 0) unary f dadb (Sparse pb bs) = Sparse (f pb) $ IntMap.map (* dadb) bs lift1 f _ Zero = auto (f 0) lift1 f df b@(Sparse pb bs) = Sparse (f pb) $ IntMap.map (* df b) bs lift1_ f _ Zero = auto (f 0) lift1_ f df b@(Sparse pb bs) = a where a = Sparse (f pb) $ IntMap.map ((df a b) *) bs binary f _ _ Zero Zero = auto (f 0 0) binary f _ dadc Zero (Sparse pc dc) = Sparse (f 0 pc) $ IntMap.map (dadc *) dc binary f dadb _ (Sparse pb db) Zero = Sparse (f pb 0 ) $ IntMap.map (dadb *) db binary f dadb dadc (Sparse pb db) (Sparse pc dc) = Sparse (f pb pc) $ unionWith (<+>) (IntMap.map (dadb *) db) (IntMap.map (dadc *) dc) lift2 f _ Zero Zero = auto (f 0 0) lift2 f df Zero c@(Sparse pc dc) = Sparse (f 0 pc) $ IntMap.map (dadc *) dc where dadc = snd (df zero c) lift2 f df b@(Sparse pb db) Zero = Sparse (f pb 0) $ IntMap.map (* dadb) db where dadb = fst (df b zero) lift2 f df b@(Sparse pb db) c@(Sparse pc dc) = Sparse (f pb pc) da where (dadb, dadc) = df b c da = unionWith (<+>) (IntMap.map (dadb *) db) (IntMap.map (dadc *) dc) lift2_ f _ Zero Zero = auto (f 0 0) lift2_ f df b@(Sparse pb db) Zero = a where a = Sparse (f pb 0) (IntMap.map (fst (df a b zero) *) db) lift2_ f df Zero c@(Sparse pc dc) = a where a = Sparse (f 0 pc) (IntMap.map (* snd (df a zero c)) dc) lift2_ f df b@(Sparse pb db) c@(Sparse pc dc) = a where (dadb, dadc) = df a b c a = Sparse (f pb pc) da da = unionWith (<+>) (IntMap.map (dadb *) db) (IntMap.map (dadc *) dc) #define HEAD Sparse a #include "instances.h" class Num a => Grad i o o' a | i -> a o o', o -> a i o', o' -> a i o where pack :: i -> [Sparse a] -> Sparse a unpack :: ([a] -> [a]) -> o unpack' :: ([a] -> (a, [a])) -> o' instance Num a => Grad (Sparse a) [a] (a, [a]) a where pack i _ = i unpack f = f [] unpack' f = f [] instance Grad i o o' a => Grad (Sparse a -> i) (a -> o) (a -> o') a where pack f (a:as) = pack (f a) as pack _ [] = error "Grad.pack: logic error" unpack f a = unpack (f . (a:)) unpack' f a = unpack' (f . (a:)) vgrad :: Grad i o o' a => i -> o vgrad i = unpack (unsafeGrad (pack i)) where unsafeGrad f as = d as $ apply f as {-# INLINE vgrad #-} vgrad' :: Grad i o o' a => i -> o' vgrad' i = unpack' (unsafeGrad' (pack i)) where unsafeGrad' f as = d' as $ apply f as {-# INLINE vgrad' #-} class Num a => Grads i o a | i -> a o, o -> a i where packs :: i -> [Sparse a] -> Sparse a unpacks :: ([a] -> Cofree [] a) -> o instance Num a => Grads (Sparse a) (Cofree [] a) a where packs i _ = i unpacks f = f [] instance Grads i o a => Grads (Sparse a -> i) (a -> o) a where packs f (a:as) = packs (f a) as packs _ [] = error "Grad.pack: logic error" unpacks f a = unpacks (f . (a:)) vgrads :: Grads i o a => i -> o vgrads i = unpacks (unsafeGrads (packs i)) where unsafeGrads f as = ds as $ apply f as {-# INLINE vgrads #-} isZero :: Sparse a -> Bool isZero Zero = True isZero _ = False -- | -- The value of the derivative of (f*g) of order mi is -- -- @ -- 'sum' [a * 'primal' ('partialS' ('indices' b) f) * 'primal' ('partialS' ('indices' c) g) | (a,b,c) <- 'terms' mi ] -- @ -- -- It is a bit more complicated in 'mul' below, since we build the whole tree of -- derivatives and want to prune the tree with 'Zero's as much as possible. -- The number of terms in the sum for order mi as of differentiation has -- @'sum' ('map' (+1) as)@ terms, so this is *much* more efficient -- than the naive recursive differentiation with @2^'sum' as@ terms. -- The coefficients @a@, which collect equivalent derivatives, are suitable products -- of binomial coefficients. terms :: Monomial -> [(Integer,Monomial,Monomial)] terms (Monomial m) = t (toAscList m) where t [] = [(1,emptyMonomial,emptyMonomial)] t ((k,a):ts) = concatMap (f (t ts)) (zip (bins!!a) [0..a]) where f ps (b,i) = map (\(w,Monomial mf,Monomial mg) -> (w*b,Monomial (IntMap.insert k i mf), Monomial (IntMap.insert k (a-i) mg))) ps bins = iterate next [1] next xs@(_:ts) = 1 : zipWith (+) xs ts ++ [1] next [] = error "impossible" mul :: Num a => Sparse a -> Sparse a -> Sparse a mul Zero _ = Zero mul _ Zero = Zero mul f@(Sparse _ am) g@(Sparse _ bm) = Sparse (primal f * primal g) (derivs 0 emptyMonomial) where derivs v mi = IntMap.unions (map fn [v..kMax]) where fn w | and zs = IntMap.empty | otherwise = IntMap.singleton w (Sparse (sum ds) (derivs w mi')) where mi' = addToMonomial w mi (zs,ds) = unzip (map derVal (terms mi')) derVal (bin,mif,mig) = (isZero fder || isZero gder, fromIntegral bin * primal fder * primal gder) where fder = partialS (indices mif) f gder = partialS (indices mig) g kMax = maybe (-1) (fst.fst) (IntMap.maxViewWithKey am) `max` maybe (-1) (fst.fst) (IntMap.maxViewWithKey bm)