Numeric.Backprop

Automatic differentation and backpropagation.

Main idea: Write a function computing what you want, and the library automatically provies the gradient of that function as well, for usage with gradient descent and other training methods.

In more detail: instead of working directly with values to produce your result, you work with BVars containing those values. Working with these BVars is made smooth with the usage of lenses and other combinators, and libraries can offer operatons on BVars instead of those on normal types directly.

Then, you can use:

evalBP :: (forall s. Reifies s W. BVar s a -> BVar s b) -> (a -> b)

to turn a BVar function into the function on actual values a -> b. This has virtually zero overhead over writing the actual function directly.

Then, there's:

gradBP :: (forall s. Reifies s W. BVar s a -> BVar s b) -> (a -> a)

to automatically get the gradient, as well, for a given input.

See the README for more information and links to demonstrations and tutorials, or dive striaght in by reading the docs for BVar.

In the original version 0.1, this module required Num instances for methods instead of Backprop instances. This interface is still available in Numeric.Backprop.Num, which has the same API as this module, except with Num constraints on all values instead of Backprop constraints.

See Prelude.Backprop.Explicit for a version allowing you to provide zero, add, and one explicitly, which can be useful when attempting to avoid orphan instances or when mixing both Backprop and Num styles.

Types

data BVar s a

A BVar s a is a value of type a that can be "backpropagated".

Functions referring to BVars are tracked by the library and can be automatically differentiated to get their gradients and results.

For simple numeric values, you can use its Num, Fractional, and Floating instances to manipulate them as if they were the numbers they represent.

If a contains items, the items can be accessed and extracted using lenses. A Lens' b a can be used to access an a inside a b, using ^^. (viewVar):

(^.)  ::        a -> Lens' a b ->        b
(^^.) :: BVar s a -> Lens' a b -> BVar s b

There is also ^^? (previewVar), to use a Prism' or Traversal' to extract a target that may or may not be present (which can implement pattern matching), ^^.. (toListOfVar) to use a Traversal' to extract all targets inside a BVar, and .~~ (setVar) to set and update values inside a BVar.

For more complex operations, libraries can provide functions on BVars using liftOp and related functions. This is how you can create primitive functions that users can use to manipulate your library's values.

For example, the hmatrix library has a matrix-vector multiplication function, #> :: L m n -> R n -> L m.

A library could instead provide a function #> :: BVar (L m n) -> BVar (R n) -> BVar (R m), which the user can then use to manipulate their BVars of L m ns and R ns, etc.

See Numeric.Backprop and documentation for liftOp for more information.

Instances

Eq a => Eq (BVar s a)

Compares the values inside the BVar.

Since: 0.1.5.0

Methods

(==) :: BVar s a -> BVar s a -> Bool

(/=) :: BVar s a -> BVar s a -> Bool

(Floating a, Reifies Type s W) => Floating (BVar s a)

Methods

pi :: BVar s a

exp :: BVar s a -> BVar s a

log :: BVar s a -> BVar s a

sqrt :: BVar s a -> BVar s a

(**) :: BVar s a -> BVar s a -> BVar s a

logBase :: BVar s a -> BVar s a -> BVar s a

sin :: BVar s a -> BVar s a

cos :: BVar s a -> BVar s a

tan :: BVar s a -> BVar s a

asin :: BVar s a -> BVar s a

acos :: BVar s a -> BVar s a

atan :: BVar s a -> BVar s a

sinh :: BVar s a -> BVar s a

cosh :: BVar s a -> BVar s a

tanh :: BVar s a -> BVar s a

asinh :: BVar s a -> BVar s a

acosh :: BVar s a -> BVar s a

atanh :: BVar s a -> BVar s a

log1p :: BVar s a -> BVar s a

expm1 :: BVar s a -> BVar s a

log1pexp :: BVar s a -> BVar s a

log1mexp :: BVar s a -> BVar s a

(Fractional a, Reifies Type s W) => Fractional (BVar s a)

Methods

(/) :: BVar s a -> BVar s a -> BVar s a

recip :: BVar s a -> BVar s a

fromRational :: Rational -> BVar s a

(Num a, Reifies Type s W) => Num (BVar s a)

Methods

(+) :: BVar s a -> BVar s a -> BVar s a

(-) :: BVar s a -> BVar s a -> BVar s a

(*) :: BVar s a -> BVar s a -> BVar s a

negate :: BVar s a -> BVar s a

abs :: BVar s a -> BVar s a

signum :: BVar s a -> BVar s a

fromInteger :: Integer -> BVar s a

Ord a => Ord (BVar s a)

Compares the values inside the BVar.

Since: 0.1.5.0

Methods

compare :: BVar s a -> BVar s a -> Ordering

(<) :: BVar s a -> BVar s a -> Bool

(<=) :: BVar s a -> BVar s a -> Bool

(>) :: BVar s a -> BVar s a -> Bool

(>=) :: BVar s a -> BVar s a -> Bool

max :: BVar s a -> BVar s a -> BVar s a

min :: BVar s a -> BVar s a -> BVar s a

NFData a => NFData (BVar s a)

This will force the value inside, as well.

Methods

rnf :: BVar s a -> ()

data W

An ephemeral Wengert Tape in the environment. Used internally to track of the computational graph of variables.

For the end user, one can just imagine Reifies s W as a required constraint on s that allows backpropagation to work.

class Backprop a where

Class of values that can be backpropagated in general.

For instances of Num, these methods can be given by zeroNum, addNum, and oneNum. There are also generic options given in Numeric.Backprop.Class for functors, IsList instances, and Generic instances.

instance Backprop Double where
    zero = zeroNum
    add = addNum
    one = oneNum

If you leave the body of an instance declaration blank, GHC Generics will be used to derive instances if the type has a single constructor and each field is an instance of Backprop.

To ensure that backpropagation works in a sound way, should obey the laws:

identity

• add x (zero y) = x

• add (zero x) y = y

Also implies preservation of information, making zipWith (+) an illegal implementation for lists and vectors.

This is only expected to be true up to potential "extra zeroes" in x and y in the result.

commutativity

• add x y = add y x

associativity

• add x (add y z) = add (add x y) z

idempotence

• zero . zero = zero

• one . one = one

Note that not all values in the backpropagation process needs all of these methods: Only the "final result" needs one, for example. These are all grouped under one typeclass for convenience in defining instances, and also to talk about sensible laws. For fine-grained control, use the "explicit" versions of library functions (for example, in Numeric.Backprop.Explicit) instead of Backprop based ones.

This typeclass replaces the reliance on Num of the previous API (v0.1). Num is strictly more powerful than Backprop, and is a stronger constraint on types than is necessary for proper backpropagating. In particular, fromInteger is a problem for many types, preventing useful backpropagation for lists, variable-length vectors (like Data.Vector) and variable-size matrices from linear algebra libraries like hmatrix and accelerate.

Since: 0.2.0.0

Methods

zero :: a -> a

"Zero out" all components of a value. For scalar values, this should just be const 0. For vectors and matrices, this should set all components to zero, the additive identity.

Should be idempotent:

• zero . zero = zero

Should be as lazy as possible. This behavior is observed for all instances provided by this library.

See zeroNum for a pre-built definition for instances of Num and zeroFunctor for a definition for instances of Functor. If left blank, will automatically be genericZero, a pre-built definition for instances of Generic whose fields are all themselves instances of Backprop.

add :: a -> a -> a

Add together two values of a type. To combine contributions of gradients, so should be information-preserving:

• add x (zero y) = x

• add (zero x) y = y

Should be as strict as possible. This behavior is observed for all instances provided by this library.

See addNum for a pre-built definition for instances of Num and addFunctor for a definition for instances of Functor. If left blank, will automatically be genericAdd, a pre-built definition for instances of Generic with one constructor whose fields are all themselves instances of Backprop.

one :: a -> a

One all components of a value. For scalar values, this should just be const 1. For vectors and matrices, this should set all components to one, the multiplicative identity.

Should be idempotent:

• one . one = one

Should be as lazy as possible. This behavior is observed for all instances provided by this library.

See oneNum for a pre-built definition for instances of Num and oneFunctor for a definition for instances of Functor. If left blank, will automatically be genericOne, a pre-built definition for instances of Generic whose fields are all themselves instances of Backprop.

zero :: (Generic a, GZero (Rep a)) => a -> a

"Zero out" all components of a value. For scalar values, this should just be const 0. For vectors and matrices, this should set all components to zero, the additive identity.

Should be idempotent:

• zero . zero = zero

Should be as lazy as possible. This behavior is observed for all instances provided by this library.

See zeroNum for a pre-built definition for instances of Num and zeroFunctor for a definition for instances of Functor. If left blank, will automatically be genericZero, a pre-built definition for instances of Generic whose fields are all themselves instances of Backprop.

add :: (Generic a, GAdd (Rep a)) => a -> a -> a

Add together two values of a type. To combine contributions of gradients, so should be information-preserving:

• add x (zero y) = x

• add (zero x) y = y

Should be as strict as possible. This behavior is observed for all instances provided by this library.

See addNum for a pre-built definition for instances of Num and addFunctor for a definition for instances of Functor. If left blank, will automatically be genericAdd, a pre-built definition for instances of Generic with one constructor whose fields are all themselves instances of Backprop.

one :: (Generic a, GOne (Rep a)) => a -> a

One all components of a value. For scalar values, this should just be const 1. For vectors and matrices, this should set all components to one, the multiplicative identity.

Should be idempotent:

• one . one = one

Should be as lazy as possible. This behavior is observed for all instances provided by this library.

See oneNum for a pre-built definition for instances of Num and oneFunctor for a definition for instances of Functor. If left blank, will automatically be genericOne, a pre-built definition for instances of Generic whose fields are all themselves instances of Backprop.

Instances

Backprop Double

Methods

zero :: Double -> Double

add :: Double -> Double -> Double

one :: Double -> Double

Backprop Float

Methods

zero :: Float -> Float

add :: Float -> Float -> Float

one :: Float -> Float

Backprop Int

Methods

zero :: Int -> Int

add :: Int -> Int -> Int

one :: Int -> Int

Backprop Integer

Methods

zero :: Integer -> Integer

add :: Integer -> Integer -> Integer

one :: Integer -> Integer

Backprop Natural

Since: 0.2.1.0

Methods

zero :: Natural -> Natural

add :: Natural -> Natural -> Natural

one :: Natural -> Natural

Backprop ()

add is strict, but zero and one are lazy in their arguments.

Methods

zero :: () -> ()

add :: () -> () -> ()

one :: () -> ()

Backprop Void

Methods

zero :: Void -> Void

add :: Void -> Void -> Void

one :: Void -> Void

Backprop a => Backprop [a]

add assumes the shorter list has trailing zeroes, and the result has the length of the longest input.

Methods

zero :: [a] -> [a]

add :: [a] -> [a] -> [a]

one :: [a] -> [a]

Backprop a => Backprop (Maybe a)

Nothing is treated the same as Just 0. However, zero, add, and one preserve Nothing if all inputs are also Nothing.

Methods

zero :: Maybe a -> Maybe a

add :: Maybe a -> Maybe a -> Maybe a

one :: Maybe a -> Maybe a

Integral a => Backprop (Ratio a)

Methods

zero :: Ratio a -> Ratio a

add :: Ratio a -> Ratio a -> Ratio a

one :: Ratio a -> Ratio a

RealFloat a => Backprop (Complex a)

Methods

zero :: Complex a -> Complex a

add :: Complex a -> Complex a -> Complex a

one :: Complex a -> Complex a

Backprop a => Backprop (NonEmpty a)

add assumes the shorter list has trailing zeroes, and the result has the length of the longest input.

Methods

zero :: NonEmpty a -> NonEmpty a

add :: NonEmpty a -> NonEmpty a -> NonEmpty a

one :: NonEmpty a -> NonEmpty a

Backprop a => Backprop (Identity a)

Methods

zero :: Identity a -> Identity a

add :: Identity a -> Identity a -> Identity a

one :: Identity a -> Identity a

Backprop a => Backprop (IntMap a)

zero and one replace all current values, and add merges keys from both maps, adding in the case of double-occurrences.

Methods

zero :: IntMap a -> IntMap a

add :: IntMap a -> IntMap a -> IntMap a

one :: IntMap a -> IntMap a

Backprop a => Backprop (Seq a)

add assumes the shorter sequence has trailing zeroes, and the result has the length of the longest input.

Methods

zero :: Seq a -> Seq a

add :: Seq a -> Seq a -> Seq a

one :: Seq a -> Seq a

Backprop a => Backprop (I a)

Methods

zero :: I a -> I a

add :: I a -> I a -> I a

one :: I a -> I a

(Unbox a, Backprop a) => Backprop (Vector a)

Methods

zero :: Vector a -> Vector a

add :: Vector a -> Vector a -> Vector a

one :: Vector a -> Vector a

(Storable a, Backprop a) => Backprop (Vector a)

Methods

zero :: Vector a -> Vector a

add :: Vector a -> Vector a -> Vector a

one :: Vector a -> Vector a

(Prim a, Backprop a) => Backprop (Vector a)

Methods

zero :: Vector a -> Vector a

add :: Vector a -> Vector a -> Vector a

one :: Vector a -> Vector a

Backprop a => Backprop (Vector a)

Methods

zero :: Vector a -> Vector a

add :: Vector a -> Vector a -> Vector a

one :: Vector a -> Vector a

Num a => Backprop (NumBP a)

Methods

zero :: NumBP a -> NumBP a

add :: NumBP a -> NumBP a -> NumBP a

one :: NumBP a -> NumBP a

(Backprop a, Backprop b) => Backprop (a, b)

add is strict

Methods

zero :: (a, b) -> (a, b)

add :: (a, b) -> (a, b) -> (a, b)

one :: (a, b) -> (a, b)

Backprop (Proxy * a)

add is strict, but zero and one are lazy in their arguments.

Methods

zero :: Proxy * a -> Proxy * a

add :: Proxy * a -> Proxy * a -> Proxy * a

one :: Proxy * a -> Proxy * a

(Backprop a, Ord k) => Backprop (Map k a)

zero and one replace all current values, and add merges keys from both maps, adding in the case of double-occurrences.

Methods

zero :: Map k a -> Map k a

add :: Map k a -> Map k a -> Map k a

one :: Map k a -> Map k a

(Applicative f, Backprop a) => Backprop (ABP f a)

Methods

zero :: ABP f a -> ABP f a

add :: ABP f a -> ABP f a -> ABP f a

one :: ABP f a -> ABP f a

(Backprop a, Backprop b, Backprop c) => Backprop (a, b, c)

add is strict

Methods

zero :: (a, b, c) -> (a, b, c)

add :: (a, b, c) -> (a, b, c) -> (a, b, c)

one :: (a, b, c) -> (a, b, c)

ListC ((<$>) * Constraint Backprop ((<$>) * * f as)) => Backprop (Prod * f as)

Methods

zero :: Prod * f as -> Prod * f as

add :: Prod * f as -> Prod * f as -> Prod * f as

one :: Prod * f as -> Prod * f as

MaybeC ((<$>) * Constraint Backprop ((<$>) * * f a)) => Backprop (Option * f a)

Methods

zero :: Option * f a -> Option * f a

add :: Option * f a -> Option * f a -> Option * f a

one :: Option * f a -> Option * f a

(Backprop a, Backprop b, Backprop c, Backprop d) => Backprop (a, b, c, d)

add is strict

Methods

zero :: (a, b, c, d) -> (a, b, c, d)

add :: (a, b, c, d) -> (a, b, c, d) -> (a, b, c, d)

one :: (a, b, c, d) -> (a, b, c, d)

(Backprop a, Backprop b, Backprop c, Backprop d, Backprop e) => Backprop (a, b, c, d, e)

add is strict

Methods

zero :: (a, b, c, d, e) -> (a, b, c, d, e)

add :: (a, b, c, d, e) -> (a, b, c, d, e) -> (a, b, c, d, e)

one :: (a, b, c, d, e) -> (a, b, c, d, e)

newtype ABP f a

A newtype wrapper over an f a for Applicative f that gives a free Backprop instance (as well as Num etc. instances).

Useful for performing backpropagation over functions that require some monadic context (like IO) to perform.

Since: 0.2.1.0

Constructors

ABP

Fields

• runABP :: f a

Instances

Monad m => Monad (ABP m)

Methods

(>>=) :: ABP m a -> (a -> ABP m b) -> ABP m b

(>>) :: ABP m a -> ABP m b -> ABP m b

return :: a -> ABP m a

fail :: String -> ABP m a

Functor f => Functor (ABP f)

Methods

fmap :: (a -> b) -> ABP f a -> ABP f b

(<$) :: a -> ABP f b -> ABP f a

Applicative f => Applicative (ABP f)

Methods

pure :: a -> ABP f a

(<*>) :: ABP f (a -> b) -> ABP f a -> ABP f b

liftA2 :: (a -> b -> c) -> ABP f a -> ABP f b -> ABP f c

(*>) :: ABP f a -> ABP f b -> ABP f b

(<*) :: ABP f a -> ABP f b -> ABP f a

Foldable f => Foldable (ABP f)

Methods

fold :: Monoid m => ABP f m -> m

foldMap :: Monoid m => (a -> m) -> ABP f a -> m

foldr :: (a -> b -> b) -> b -> ABP f a -> b

foldr' :: (a -> b -> b) -> b -> ABP f a -> b

foldl :: (b -> a -> b) -> b -> ABP f a -> b

foldl' :: (b -> a -> b) -> b -> ABP f a -> b

foldr1 :: (a -> a -> a) -> ABP f a -> a

foldl1 :: (a -> a -> a) -> ABP f a -> a

toList :: ABP f a -> [a]

null :: ABP f a -> Bool

length :: ABP f a -> Int

elem :: Eq a => a -> ABP f a -> Bool

maximum :: Ord a => ABP f a -> a

minimum :: Ord a => ABP f a -> a

sum :: Num a => ABP f a -> a

product :: Num a => ABP f a -> a

Traversable f => Traversable (ABP f)

Methods

traverse :: Applicative f => (a -> f b) -> ABP f a -> f (ABP f b)

sequenceA :: Applicative f => ABP f (f a) -> f (ABP f a)

mapM :: Monad m => (a -> m b) -> ABP f a -> m (ABP f b)

sequence :: Monad m => ABP f (m a) -> m (ABP f a)

Eq (f a) => Eq (ABP f a)

Methods

(==) :: ABP f a -> ABP f a -> Bool

(/=) :: ABP f a -> ABP f a -> Bool

(Applicative f, Floating a) => Floating (ABP f a)

Methods

pi :: ABP f a

exp :: ABP f a -> ABP f a

log :: ABP f a -> ABP f a

sqrt :: ABP f a -> ABP f a

(**) :: ABP f a -> ABP f a -> ABP f a

logBase :: ABP f a -> ABP f a -> ABP f a

sin :: ABP f a -> ABP f a

cos :: ABP f a -> ABP f a

tan :: ABP f a -> ABP f a

asin :: ABP f a -> ABP f a

acos :: ABP f a -> ABP f a

atan :: ABP f a -> ABP f a

sinh :: ABP f a -> ABP f a

cosh :: ABP f a -> ABP f a

tanh :: ABP f a -> ABP f a

asinh :: ABP f a -> ABP f a

acosh :: ABP f a -> ABP f a

atanh :: ABP f a -> ABP f a

log1p :: ABP f a -> ABP f a

expm1 :: ABP f a -> ABP f a

log1pexp :: ABP f a -> ABP f a

log1mexp :: ABP f a -> ABP f a

(Applicative f, Fractional a) => Fractional (ABP f a)

Methods

(/) :: ABP f a -> ABP f a -> ABP f a

recip :: ABP f a -> ABP f a

fromRational :: Rational -> ABP f a

(Data (f a), Typeable * a, Typeable (* -> *) f) => Data (ABP f a)

Methods

gfoldl :: (forall d b. Data d => c (d -> b) -> d -> c b) -> (forall g. g -> c g) -> ABP f a -> c (ABP f a)

gunfold :: (forall b r. Data b => c (b -> r) -> c r) -> (forall r. r -> c r) -> Constr -> c (ABP f a)

toConstr :: ABP f a -> Constr

dataTypeOf :: ABP f a -> DataType

dataCast1 :: Typeable (* -> *) t => (forall d. Data d => c (t d)) -> Maybe (c (ABP f a))

dataCast2 :: Typeable (* -> * -> *) t => (forall d e. (Data d, Data e) => c (t d e)) -> Maybe (c (ABP f a))

gmapT :: (forall b. Data b => b -> b) -> ABP f a -> ABP f a

gmapQl :: (r -> r' -> r) -> r -> (forall d. Data d => d -> r') -> ABP f a -> r

gmapQr :: (r' -> r -> r) -> r -> (forall d. Data d => d -> r') -> ABP f a -> r

gmapQ :: (forall d. Data d => d -> u) -> ABP f a -> [u]

gmapQi :: Int -> (forall d. Data d => d -> u) -> ABP f a -> u

gmapM :: Monad m => (forall d. Data d => d -> m d) -> ABP f a -> m (ABP f a)

gmapMp :: MonadPlus m => (forall d. Data d => d -> m d) -> ABP f a -> m (ABP f a)

gmapMo :: MonadPlus m => (forall d. Data d => d -> m d) -> ABP f a -> m (ABP f a)

(Applicative f, Num a) => Num (ABP f a)

Methods

(+) :: ABP f a -> ABP f a -> ABP f a

(-) :: ABP f a -> ABP f a -> ABP f a

(*) :: ABP f a -> ABP f a -> ABP f a

negate :: ABP f a -> ABP f a

abs :: ABP f a -> ABP f a

signum :: ABP f a -> ABP f a

fromInteger :: Integer -> ABP f a

Ord (f a) => Ord (ABP f a)

Methods

compare :: ABP f a -> ABP f a -> Ordering

(<) :: ABP f a -> ABP f a -> Bool

(<=) :: ABP f a -> ABP f a -> Bool

(>) :: ABP f a -> ABP f a -> Bool

(>=) :: ABP f a -> ABP f a -> Bool

max :: ABP f a -> ABP f a -> ABP f a

min :: ABP f a -> ABP f a -> ABP f a

Read (f a) => Read (ABP f a)

Methods

readsPrec :: Int -> ReadS (ABP f a)

readList :: ReadS [ABP f a]

readPrec :: ReadPrec (ABP f a)

readListPrec :: ReadPrec [ABP f a]

Show (f a) => Show (ABP f a)

Methods

showsPrec :: Int -> ABP f a -> ShowS

show :: ABP f a -> String

showList :: [ABP f a] -> ShowS

Generic (ABP f a)

Associated Types

type Rep (ABP f a) :: * -> *

Methods

from :: ABP f a -> Rep (ABP f a) x

to :: Rep (ABP f a) x -> ABP f a

NFData (f a) => NFData (ABP f a)

Methods

rnf :: ABP f a -> ()

(Applicative f, Backprop a) => Backprop (ABP f a)

Methods

zero :: ABP f a -> ABP f a

add :: ABP f a -> ABP f a -> ABP f a

one :: ABP f a -> ABP f a

type Rep (ABP f a)

type Rep (ABP f a) = D1 * (MetaData "ABP" "Numeric.Backprop.Class" "backprop-0.2.1.0-3cxmKcBToiX2y9THONyvPW" True) (C1 * (MetaCons "ABP" PrefixI True) (S1 * (MetaSel (Just Symbol "runABP") NoSourceUnpackedness NoSourceStrictness DecidedLazy) (Rec0 * (f a))))

newtype NumBP a

A newtype wrapper over an instance of Num that gives a free Backprop instance.

Useful for things like DerivingVia, or for avoiding orphan instances.

Since: 0.2.1.0

Constructors

NumBP

Fields

• runNumBP :: a

Instances

Monad NumBP

Methods

(>>=) :: NumBP a -> (a -> NumBP b) -> NumBP b

(>>) :: NumBP a -> NumBP b -> NumBP b

return :: a -> NumBP a

fail :: String -> NumBP a

Functor NumBP

Methods

fmap :: (a -> b) -> NumBP a -> NumBP b

(<$) :: a -> NumBP b -> NumBP a

Applicative NumBP

Methods

pure :: a -> NumBP a

(<*>) :: NumBP (a -> b) -> NumBP a -> NumBP b

liftA2 :: (a -> b -> c) -> NumBP a -> NumBP b -> NumBP c

(*>) :: NumBP a -> NumBP b -> NumBP b

(<*) :: NumBP a -> NumBP b -> NumBP a

Foldable NumBP

Methods

fold :: Monoid m => NumBP m -> m

foldMap :: Monoid m => (a -> m) -> NumBP a -> m

foldr :: (a -> b -> b) -> b -> NumBP a -> b

foldr' :: (a -> b -> b) -> b -> NumBP a -> b

foldl :: (b -> a -> b) -> b -> NumBP a -> b

foldl' :: (b -> a -> b) -> b -> NumBP a -> b

foldr1 :: (a -> a -> a) -> NumBP a -> a

foldl1 :: (a -> a -> a) -> NumBP a -> a

toList :: NumBP a -> [a]

null :: NumBP a -> Bool

length :: NumBP a -> Int

elem :: Eq a => a -> NumBP a -> Bool

maximum :: Ord a => NumBP a -> a

minimum :: Ord a => NumBP a -> a

sum :: Num a => NumBP a -> a

product :: Num a => NumBP a -> a

Traversable NumBP

Methods

traverse :: Applicative f => (a -> f b) -> NumBP a -> f (NumBP b)

sequenceA :: Applicative f => NumBP (f a) -> f (NumBP a)

mapM :: Monad m => (a -> m b) -> NumBP a -> m (NumBP b)

sequence :: Monad m => NumBP (m a) -> m (NumBP a)

Eq a => Eq (NumBP a)

Methods

(==) :: NumBP a -> NumBP a -> Bool

(/=) :: NumBP a -> NumBP a -> Bool

Floating a => Floating (NumBP a)

Methods

pi :: NumBP a

exp :: NumBP a -> NumBP a

log :: NumBP a -> NumBP a

sqrt :: NumBP a -> NumBP a

(**) :: NumBP a -> NumBP a -> NumBP a

logBase :: NumBP a -> NumBP a -> NumBP a

sin :: NumBP a -> NumBP a

cos :: NumBP a -> NumBP a

tan :: NumBP a -> NumBP a

asin :: NumBP a -> NumBP a

acos :: NumBP a -> NumBP a

atan :: NumBP a -> NumBP a

sinh :: NumBP a -> NumBP a

cosh :: NumBP a -> NumBP a

tanh :: NumBP a -> NumBP a

asinh :: NumBP a -> NumBP a

acosh :: NumBP a -> NumBP a

atanh :: NumBP a -> NumBP a

log1p :: NumBP a -> NumBP a

expm1 :: NumBP a -> NumBP a

log1pexp :: NumBP a -> NumBP a

log1mexp :: NumBP a -> NumBP a

Fractional a => Fractional (NumBP a)

Methods

(/) :: NumBP a -> NumBP a -> NumBP a

recip :: NumBP a -> NumBP a

fromRational :: Rational -> NumBP a

Data a => Data (NumBP a)

Methods

gfoldl :: (forall d b. Data d => c (d -> b) -> d -> c b) -> (forall g. g -> c g) -> NumBP a -> c (NumBP a)

gunfold :: (forall b r. Data b => c (b -> r) -> c r) -> (forall r. r -> c r) -> Constr -> c (NumBP a)

toConstr :: NumBP a -> Constr

dataTypeOf :: NumBP a -> DataType

dataCast1 :: Typeable (* -> *) t => (forall d. Data d => c (t d)) -> Maybe (c (NumBP a))

dataCast2 :: Typeable (* -> * -> *) t => (forall d e. (Data d, Data e) => c (t d e)) -> Maybe (c (NumBP a))

gmapT :: (forall b. Data b => b -> b) -> NumBP a -> NumBP a

gmapQl :: (r -> r' -> r) -> r -> (forall d. Data d => d -> r') -> NumBP a -> r

gmapQr :: (r' -> r -> r) -> r -> (forall d. Data d => d -> r') -> NumBP a -> r

gmapQ :: (forall d. Data d => d -> u) -> NumBP a -> [u]

gmapQi :: Int -> (forall d. Data d => d -> u) -> NumBP a -> u

gmapM :: Monad m => (forall d. Data d => d -> m d) -> NumBP a -> m (NumBP a)

gmapMp :: MonadPlus m => (forall d. Data d => d -> m d) -> NumBP a -> m (NumBP a)

gmapMo :: MonadPlus m => (forall d. Data d => d -> m d) -> NumBP a -> m (NumBP a)

Num a => Num (NumBP a)

Methods

(+) :: NumBP a -> NumBP a -> NumBP a

(-) :: NumBP a -> NumBP a -> NumBP a

(*) :: NumBP a -> NumBP a -> NumBP a

negate :: NumBP a -> NumBP a

abs :: NumBP a -> NumBP a

signum :: NumBP a -> NumBP a

fromInteger :: Integer -> NumBP a

Ord a => Ord (NumBP a)

Methods

compare :: NumBP a -> NumBP a -> Ordering

(<) :: NumBP a -> NumBP a -> Bool

(<=) :: NumBP a -> NumBP a -> Bool

(>) :: NumBP a -> NumBP a -> Bool

(>=) :: NumBP a -> NumBP a -> Bool

max :: NumBP a -> NumBP a -> NumBP a

min :: NumBP a -> NumBP a -> NumBP a

Read a => Read (NumBP a)

Methods

readsPrec :: Int -> ReadS (NumBP a)

readList :: ReadS [NumBP a]

readPrec :: ReadPrec (NumBP a)

readListPrec :: ReadPrec [NumBP a]

Show a => Show (NumBP a)

Methods

showsPrec :: Int -> NumBP a -> ShowS

show :: NumBP a -> String

showList :: [NumBP a] -> ShowS

Generic (NumBP a)

Associated Types

type Rep (NumBP a) :: * -> *

Methods

from :: NumBP a -> Rep (NumBP a) x

to :: Rep (NumBP a) x -> NumBP a

NFData a => NFData (NumBP a)

Methods

rnf :: NumBP a -> ()

Num a => Backprop (NumBP a)

Methods

zero :: NumBP a -> NumBP a

add :: NumBP a -> NumBP a -> NumBP a

one :: NumBP a -> NumBP a

type Rep (NumBP a)

type Rep (NumBP a) = D1 * (MetaData "NumBP" "Numeric.Backprop.Class" "backprop-0.2.1.0-3cxmKcBToiX2y9THONyvPW" True) (C1 * (MetaCons "NumBP" PrefixI True) (S1 * (MetaSel (Just Symbol "runNumBP") NoSourceUnpackedness NoSourceStrictness DecidedLazy) (Rec0 * a)))

Running

backprop :: (Backprop a, Backprop b) => (forall s. Reifies s W => BVar s a -> BVar s b) -> a -> (b, a)

Turn a function BVar s a -> BVar s b into the function a -> b that it represents, also computing its gradient a as well.

The Rank-N type forall s. Reifies s W => ... is used to ensure that BVars do not leak out of the context (similar to how it is used in Control.Monad.ST), and also as a reference to an ephemeral Wengert tape used to track the graph of references.

evalBP :: (forall s. Reifies s W => BVar s a -> BVar s b) -> a -> b

Turn a function BVar s a -> BVar s b into the function a -> b that it represents.

Benchmarks show that this should have virtually no overhead over directly writing a a -> b. BVar is, in this situation, a zero-cost abstraction, performance-wise.

See documentation of backprop for more information.

gradBP :: (Backprop a, Backprop b) => (forall s. Reifies s W => BVar s a -> BVar s b) -> a -> a

Take a function BVar s a -> BVar s b, interpreted as a function a -> b, and compute its gradient with respect to its input.

The resulting a -> a tells how the input (and its components) affects the output. Positive numbers indicate that the result will vary in the same direction as any adjustment in the input. Negative numbers indicate that the result will vary in the opposite direction as any adjustment in the input. Larger numbers indicate a greater sensitivity of change, and small numbers indicate lower sensitivity.

See documentation of backprop for more information.

If you want to provide an explicit "final gradient" for the end, see backpropWith.

backpropWith

Arguments

:: Backprop a

=> (forall s. Reifies s W => BVar s a -> BVar s b)

-> a

-> (b -> b)

Gradient of final result with respect to output of function

-> (b, a)

A version of backprop that allows you to specify the gradent of your "final result" in with respect to the output of your function.

Typically, this is just the scalar 1, or a value of components that are all 1.

Instead of taking the b gradient, the you may provide a b -> b, which backpropWith calls with the result of your function as the argument. This allows you to return something with the correct "shape", if not a scalar.

backprop is essentially backpropWith with const 1 for scalars and Num instances.

Since: 0.2.0.0

Multiple inputs

backprop2 :: (Backprop a, Backprop b, Backprop c) => (forall s. Reifies s W => BVar s a -> BVar s b -> BVar s c) -> a -> b -> (c, (a, b))

backprop for a two-argument function.

Not strictly necessary, because you can always uncurry a function by passing in all of the argument inside a data type, or just use a tuple. However, this could potentially be more performant.

For 3 and more arguments, consider using backpropN.

evalBP2 :: (forall s. Reifies s W => BVar s a -> BVar s b -> BVar s c) -> a -> b -> c

evalBP for a two-argument function. See backprop2 for notes.

gradBP2 :: (Backprop a, Backprop b, Backprop c) => (forall s. Reifies s W => BVar s a -> BVar s b -> BVar s c) -> a -> b -> (a, b)

gradBP for a two-argument function. See backprop2 for notes.

backpropWith2

Arguments

:: (Backprop a, Backprop b)

=> (forall s. Reifies s W => BVar s a -> BVar s b -> BVar s c)

-> a

-> b

-> (c -> c)

Gradient of final result with respect to output of function

-> (c, (a, b))

backprop2, but allows you to provide the gradient of the "final result" with respect to the output of your function. See backpropWith for more details.

Since: 0.2.0.0

backpropN :: (Every Backprop as, Known Length as, Backprop b) => (forall s. Reifies s W => Prod (BVar s) as -> BVar s b) -> Tuple as -> (b, Tuple as)

backprop generalized to multiple inputs of different types. See the Numeric.Backprop.Op for a mini-tutorial on heterogeneous lists.

Not strictly necessary, because you can always uncurry a function by passing in all of the inputs in a data type containing all of the arguments or a giant tuple. However, this could potentially also be more performant.

A Prod (BVar s) '[Double, Float, Double], for instance, is a tuple of BVar s Double, BVar s Float, and BVar s Double, and can be pattern matched on using :< (cons) and 'Ø' (nil).

Tuples can be built and pattern matched on using ::< (cons) and 'Ø' (nil), as well.

The Every Backprop as in the constraint says that every value in the type-level list as must have a Backprop instance. This means you can use, say, '[Double, Float, Int], but not '[Double, Bool, String].

If you stick to concerete, monomorphic usage of this (with specific types, typed into source code, known at compile-time), then Every Backprop as should be fulfilled automatically.

evalBPN :: forall as b. (forall s. Reifies s W => Prod (BVar s) as -> BVar s b) -> Tuple as -> b

evalBP generalized to multiple inputs of different types. See documentation for backpropN for more details.

gradBPN :: (Every Backprop as, Known Length as, Backprop b) => (forall s. Reifies s W => Prod (BVar s) as -> BVar s b) -> Tuple as -> Tuple as

gradBP generalized to multiple inputs of different types. See documentation for backpropN for more details.

backpropWithN

Arguments

:: (Every Backprop as, Known Length as)

=> (forall s. Reifies s W => Prod (BVar s) as -> BVar s b)

-> Tuple as

-> (b -> b)

Gradient of final result with respect to output of function

-> (b, Tuple as)

backpropN, but allows you to provide the gradient of the "final result" with respect to the output of your function. See backpropWith for more details.

Since: 0.2.0.0

class EveryC k c as => Every k (c :: k -> Constraint) (as :: [k])

Minimal complete definition

every

Instances

Every k c (Ø k)

Associated Types

type EveryC c (Ø k :: c -> Constraint) (as :: [c]) :: Constraint

Methods

every :: Index c as a -> Wit (Ø k a)

(c a2, Every a1 c as) => Every a1 c ((:<) a1 a2 as)

Associated Types

type EveryC c ((a1 :< a2) as :: c -> Constraint) (as :: [c]) :: Constraint

Methods

every :: Index c as a -> Wit ((a1 :< a2) as a)

Manipulating BVar

constVar :: a -> BVar s a

Lift a value into a BVar representing a constant value.

This value will not be considered an input, and its gradients will not be backpropagated.

auto :: a -> BVar s a

Shorter alias for constVar, inspired by the ad library.

Since: 0.2.0.0

coerceVar :: Coercible a b => BVar s a -> BVar s b

Coerce a BVar contents. Useful for things like newtype wrappers.

Since: 0.1.5.2

(^^.) :: forall a b s. (Reifies s W, Backprop a) => BVar s b -> Lens' b a -> BVar s a infixl 8

An infix version of viewVar, meant to evoke parallels to ^. from lens.

With normal values, you can extract something from that value with a lens:

x ^. myLens

would extract a piece of x :: b, specified by myLens :: Lens' b a. The result has type a.

xVar ^^. myLens

would extract a piece out of xVar :: BVar s b (a BVar holding a b), specified by myLens :: Lens' b a. The result has type BVar s a (a BVar holding a a)

This is the main way to pull out values from BVar of container types.

WARNING: Do not use with any lenses that operate "numerically" on the contents (like multiplying).

(.~~) :: forall a b s. (Reifies s W, Backprop a, Backprop b) => Lens' b a -> BVar s a -> BVar s b -> BVar s b infixl 8

An infix version of setVar, meant to evoke parallels to .~ from lens.

With normal values, you can set something in a value with a lens: a lens:

x & myLens .~ y

would "set" a part of x :: b, specified by myLens :: Lens' a b, to a new value y :: a.

xVar & myLens .~~ yVar

would "set" a part of xVar :: BVar s b (a BVar holding a b), specified by myLens :: Lens' a b, to a new value given by yVar :: BVar s a. The result is a new (updated) value of type BVar s b.

This is the main way to set values inside BVars of container types.

(^^?) :: forall b a s. (Backprop a, Reifies s W) => BVar s b -> Traversal' b a -> Maybe (BVar s a)

An infix version of previewVar, meant to evoke parallels to ^? from lens.

With normal values, you can (potentially) extract something from that value with a lens:

x ^? myPrism

would (potentially) extract a piece of x :: b, specified by myPrism :: Traversal' b a. The result has type Maybe a.

xVar ^^? myPrism

would (potentially) extract a piece out of xVar :: BVar s b (a BVar holding a b), specified by myPrism :: Prism' b a. The result has type Maybe (BVar s a) (Maybe a BVar holding a a).

This is intended to be used with Prism's (which hits at most one target), but will actually work with any Traversal'. If the traversal hits more than one target, the first one found will be extracted.

This can be used to "pattern match" on BVars, by using prisms on constructors.

(^^..) :: forall b a s. (Backprop a, Reifies s W) => BVar s b -> Traversal' b a -> [BVar s a]

An infix version of toListOfVar, meant to evoke parallels to ^.. from lens.

With normal values, you can extract all targets of a Traversal from that value with a:

x ^.. myTraversal

would extract all targets inside of x :: b, specified by myTraversal :: Traversal' b a. The result has type [a].

xVar ^^.. myTraversal

would extract all targets inside of xVar :: BVar s b (a BVar holding a b), specified by myTraversal :: Traversal' b a. The result has type [BVar s a] (A list of BVars holding as).

(^^?!) :: forall b a s. (Backprop a, Reifies s W) => BVar s b -> Traversal' b a -> BVar s a

An *UNSAFE* version of previewVar assuming that it is there.

Is undefined if the Traversal hits no targets.

Is essentially ^^? with fromJust, or ^^.. with head.

Since: 0.2.1.0

viewVar :: forall a b s. (Reifies s W, Backprop a) => Lens' b a -> BVar s b -> BVar s a

Using a Lens', extract a value inside a BVar. Meant to evoke parallels to view from lens.

See documentation for ^^. for more information.

setVar :: forall a b s. (Reifies s W, Backprop a, Backprop b) => Lens' b a -> BVar s a -> BVar s b -> BVar s b

Using a Lens', set a value inside a BVar. Meant to evoke parallels to "set" from lens.

See documentation for .~~ for more information.

sequenceVar :: forall t a s. (Backprop a, Reifies s W, Traversable t) => BVar s (t a) -> t (BVar s a)

Extract all of the BVars out of a Traversable container of BVars.

Note that this associates gradients in order of occurrence in the original data structure; the second item in the gradient is assumed to correspond with the second item in the input, etc.; this can cause unexpected behavior in Foldable instances that don't have a fixed number of items.

collectVar :: forall t a s. (Backprop a, Backprop (t a), Reifies s W, Foldable t, Functor t) => t (BVar s a) -> BVar s (t a)

Collect all of the BVars in a container into a BVar of that container's contents.

Note that this associates gradients in order of occurrence in the original data structure; the second item in the total derivative and gradient is assumed to correspond with the second item in the input, etc.; this can cause unexpected behavior in Foldable instances that don't have a fixed number of items.

previewVar :: forall b a s. (Reifies s W, Backprop a) => Traversal' b a -> BVar s b -> Maybe (BVar s a)

Using a Traversal', extract a single value inside a BVar, if it exists. If more than one traversal target exists, returns te first. Meant to evoke parallels to preview from lens. Really only intended to be used wth Prism's, or up-to-one target traversals.

See documentation for ^^? for more information.

toListOfVar :: forall b a s. (Backprop a, Reifies s W) => Traversal' b a -> BVar s b -> [BVar s a]

Using a Traversal', extract all targeted values inside a BVar. Meant to evoke parallels to toListOf from lens.

See documentation for ^^.. for more information.

pattern T2 :: (Backprop a, Backprop b, Reifies s W) => BVar s a -> BVar s b -> BVar s (a, b)

Useful pattern for constructing and deconstructing BVars of two-tuples.

Since: 0.2.1.0

pattern T3 :: (Backprop a, Backprop b, Backprop c, Reifies s W) => BVar s a -> BVar s b -> BVar s c -> BVar s (a, b, c)

Useful pattern for constructing and deconstructing BVars three-tuples.

Since: 0.2.1.0

With Isomorphisms

isoVar :: (Backprop a, Backprop b, Reifies s W) => (a -> b) -> (b -> a) -> BVar s a -> BVar s b

Convert the value inside a BVar using a given isomorphism. Useful for things like constructors.

Warning: This is unsafe! It assumes that the isomorphisms themselves have derivative 1, so will break for things like exp & log. Basically, don't use this for any "numeric" isomorphisms.

Since: 0.1.4.0

isoVar2 :: (Backprop a, Backprop b, Backprop c, Reifies s W) => (a -> b -> c) -> (c -> (a, b)) -> BVar s a -> BVar s b -> BVar s c

Convert the values inside two BVars using a given isomorphism. Useful for things like constructors. See isoVar for caveats.

Since: 0.1.4.0

isoVar3 :: (Backprop a, Backprop b, Backprop c, Backprop d, Reifies s W) => (a -> b -> c -> d) -> (d -> (a, b, c)) -> BVar s a -> BVar s b -> BVar s c -> BVar s d

Convert the values inside three BVars using a given isomorphism. Useful for things like constructors. See isoVar for caveats.

Since: 0.1.4.0

isoVarN :: (Every Backprop as, Known Length as, Backprop b, Reifies s W) => (Tuple as -> b) -> (b -> Tuple as) -> Prod (BVar s) as -> BVar s b

Convert the values inside a tuple of BVars using a given isomorphism. Useful for things like constructors. See isoVar for caveats.

Since: 0.1.4.0

With Ops

This library provides a few primitive actions for manipulating BVars and the values inside them, including its Num, Fractional, and Floating instances, and lens-based operations like ^^., .~~ ^^?, and ^^...

However, the power of this library comes from manipulating many different types from libraries, like matrices and vectors. Libraries can provide their own BVar s a -> BVar s b functions, alongside (or in lieu of) a -> b functions for their types.

The easiest way to create a BVar function is to use liftOp with an Op constructor. For example, imagine a vector library providing a dot product function. We can write this using liftOp2 and op2:

dot :: BVar s Vec -> BVar s Vec -> BVar s Double
dot = liftOp2 . op2 $ \xs ys ->
        ( sum (zipWith (*) xs ys)
        , \g -> (map (*g) ys, map (*g) xs)
        )

We provide a function that, given the two inputs, returns:

1. The result of the function on those two inputs
2. A function taking the "total derivative", and returning the gradient with respect to each of the inputs.

See documentation in Numeric.Backprop.Op for more information on the second part (the gradient).

Nice Ops are how backprop links together BVars and tracks them to determine their gradient. Ideally, users would never have to deal with these when backpropagating their own functions, and library authors providing their matrix and vector operations, etc. would provide BVar variants of their normal operations.

In fact, BVar operations could even be defined instead of normal operations, since it is easy to go from BVar s a -> BVar s b to a -> b, using evalBP, and this carries virtually zero overhead, so some libraries might even provide BVar versions by default.

liftOp :: forall as b s. (Every Backprop as, Known Length as, Backprop b, Reifies s W) => Op as b -> Prod (BVar s) as -> BVar s b

Lift an Op with an arbitrary number of inputs to a function on the appropriate number of BVars.

Should preferably be used only by libraries to provide primitive BVar functions for their types for users.

See Numeric.Backprop and documentation for liftOp for more information, and Numeric.Backprop.Op for a mini-tutorial on using Prod and Tuple.

liftOp1 :: forall a b s. (Backprop a, Backprop b, Reifies s W) => Op '[a] b -> BVar s a -> BVar s b

Lift an Op with a single input to be a function on a single BVar.

Should preferably be used only by libraries to provide primitive BVar functions for their types for users.

See Numeric.Backprop and documentation for liftOp for more information.

liftOp2 :: forall a b c s. (Backprop a, Backprop b, Backprop c, Reifies s W) => Op '[a, b] c -> BVar s a -> BVar s b -> BVar s c

Lift an Op with two inputs to be a function on a two BVars.

Should preferably be used only by libraries to provide primitive BVar functions for their types for users.

See Numeric.Backprop and documentation for liftOp for more information.

liftOp3 :: forall a b c d s. (Backprop a, Backprop b, Backprop c, Backprop d, Reifies s W) => Op '[a, b, c] d -> BVar s a -> BVar s b -> BVar s c -> BVar s d

Lift an Op with three inputs to be a function on a three BVars.

Should preferably be used only by libraries to provide primitive BVar functions for their types for users.

See Numeric.Backprop and documentation for liftOp for more information.

Op

newtype Op as a

An Op as a describes a differentiable function from as to a.

For example, a value of type

Op '[Int, Bool] Double

is a function from an Int and a Bool, returning a Double. It can be differentiated to give a gradient of an Int and a Bool if given a total derivative for the Double. If we call Bool \(2\), then, mathematically, it is akin to a:

\[ f : \mathbb{Z} \times 2 \rightarrow \mathbb{R} \]

See runOp, gradOp, and gradOpWith for examples on how to run it, and Op for instructions on creating it.

It is simpler to not use this type constructor directly, and instead use the op2, op1, op2, and op3 helper smart constructors.

See Numeric.Backprop.Op for a mini-tutorial on using Prod and Tuple.

To use an Op with the backprop library, see liftOp, liftOp1, liftOp2, and liftOp3.

Constructors

Op

Construct an Op by giving a function creating the result, and also a continuation on how to create the gradient, given the total derivative of a.

See the module documentation for Numeric.Backprop.Op for more details on the function that this constructor and Op expect.

Fields

• runOpWith :: Tuple as -> (a, a -> Tuple as)

  Run the function that the Op encodes, returning a continuation to compute the gradient, given the total derivative of a. See documentation for Numeric.Backprop.Op for more information.

Instances

(Known [*] (Length *) as, Every * Floating as, Every * Fractional as, Every * Num as, Floating a) => Floating (Op as a)

Methods

pi :: Op as a

exp :: Op as a -> Op as a

log :: Op as a -> Op as a

sqrt :: Op as a -> Op as a

(**) :: Op as a -> Op as a -> Op as a

logBase :: Op as a -> Op as a -> Op as a

sin :: Op as a -> Op as a

cos :: Op as a -> Op as a

tan :: Op as a -> Op as a

asin :: Op as a -> Op as a

acos :: Op as a -> Op as a

atan :: Op as a -> Op as a

sinh :: Op as a -> Op as a

cosh :: Op as a -> Op as a

tanh :: Op as a -> Op as a

asinh :: Op as a -> Op as a

acosh :: Op as a -> Op as a

atanh :: Op as a -> Op as a

log1p :: Op as a -> Op as a

expm1 :: Op as a -> Op as a

log1pexp :: Op as a -> Op as a

log1mexp :: Op as a -> Op as a

(Known [*] (Length *) as, Every * Fractional as, Every * Num as, Fractional a) => Fractional (Op as a)

Methods

(/) :: Op as a -> Op as a -> Op as a

recip :: Op as a -> Op as a

fromRational :: Rational -> Op as a

(Known [*] (Length *) as, Every * Num as, Num a) => Num (Op as a)

Methods

(+) :: Op as a -> Op as a -> Op as a

(-) :: Op as a -> Op as a -> Op as a

(*) :: Op as a -> Op as a -> Op as a

negate :: Op as a -> Op as a

abs :: Op as a -> Op as a

signum :: Op as a -> Op as a

fromInteger :: Integer -> Op as a

Creation

op0 :: a -> Op '[] a

Create an Op that takes no inputs and always returns the given value.

There is no gradient, of course (using gradOp will give you an empty tuple), because there is no input to have a gradient of.

>>> runOp (op0 10) Ø
(10, Ø)

For a constant Op that takes input and ignores it, see opConst and opConst'.

opConst :: (Every Num as, Known Length as) => a -> Op as a

An Op that ignores all of its inputs and returns a given constant value.

>>> gradOp' (opConst 10) (1 ::< 2 ::< 3 ::< Ø)
(10, 0 ::< 0 ::< 0 ::< Ø)

idOp :: Op '[a] a

An Op that just returns whatever it receives. The identity function.

idOp = opIso id id

opConst' :: Every Num as => Length as -> a -> Op as a

A version of opConst taking explicit Length, indicating the number of inputs and their types.

Requiring an explicit Length is mostly useful for rare "extremely polymorphic" situations, where GHC can't infer the type and length of the the expected input tuple. If you ever actually explicitly write down as as a list of types, you should be able to just use opConst.

Giving gradients directly

op1 :: (a -> (b, b -> a)) -> Op '[a] b

Create an Op of a function taking one input, by giving its explicit derivative. The function should return a tuple containing the result of the function, and also a function taking the derivative of the result and return the derivative of the input.

If we have

\[ \eqalign{ f &: \mathbb{R} \rightarrow \mathbb{R}\cr y &= f(x)\cr z &= g(y) } \]

Then the derivative \( \frac{dz}{dx} \), it would be:

\[ \frac{dz}{dx} = \frac{dz}{dy} \frac{dy}{dx} \]

If our Op represents \(f\), then the second item in the resulting tuple should be a function that takes \(\frac{dz}{dy}\) and returns \(\frac{dz}{dx}\).

As an example, here is an Op that squares its input:

square :: Num a => Op '[a] a
square = op1 $ \x -> (x*x, \d -> 2 * d * x
                     )

Remember that, generally, end users shouldn't directly construct Ops; they should be provided by libraries or generated automatically.

op2 :: (a -> b -> (c, c -> (a, b))) -> Op '[a, b] c

Create an Op of a function taking two inputs, by giving its explicit gradient. The function should return a tuple containing the result of the function, and also a function taking the derivative of the result and return the derivative of the input.

If we have

\[ \eqalign{ f &: \mathbb{R}^2 \rightarrow \mathbb{R}\cr z &= f(x, y)\cr k &= g(z) } \]

Then the gradient \( \left< \frac{\partial k}{\partial x}, \frac{\partial k}{\partial y} \right> \) would be:

\[ \left< \frac{\partial k}{\partial x}, \frac{\partial k}{\partial y} \right> = \left< \frac{dk}{dz} \frac{\partial z}{dx}, \frac{dk}{dz} \frac{\partial z}{dy} \right> \]

If our Op represents \(f\), then the second item in the resulting tuple should be a function that takes \(\frac{dk}{dz}\) and returns \( \left< \frac{\partial k}{dx}, \frac{\partial k}{dx} \right> \).

As an example, here is an Op that multiplies its inputs:

mul :: Num a => Op '[a, a] a
mul = op2' $ \x y -> (x*y, \d -> (d*y, x*d)
                     )

Remember that, generally, end users shouldn't directly construct Ops; they should be provided by libraries or generated automatically.

op3 :: (a -> b -> c -> (d, d -> (a, b, c))) -> Op '[a, b, c] d

Create an Op of a function taking three inputs, by giving its explicit gradient. See documentation for op2 for more details.

From Isomorphisms

opCoerce :: Coercible a b => Op '[a] b

An Op that coerces an item into another item whose type has the same runtime representation.

>>> gradOp' opCoerce (Identity 5) :: (Int, Identity Int)
(5, Identity 1)

opCoerce = opIso coerced coerce

opTup :: Op as (Tuple as)

An Op that takes as and returns exactly the input tuple.

>>> gradOp' opTup (1 ::< 2 ::< 3 ::< Ø)
(1 ::< 2 ::< 3 ::< Ø, 1 ::< 1 ::< 1 ::< Ø)

opIso :: (a -> b) -> (b -> a) -> Op '[a] b

An Op that runs the input value through an isomorphism.

Warning: This is unsafe! It assumes that the isomorphisms themselves have derivative 1, so will break for things like exp & log. Basically, don't use this for any "numeric" isomorphisms.

opIsoN :: (Tuple as -> b) -> (b -> Tuple as) -> Op as b

An Op that runs the input value through an isomorphism between a tuple of values and a value. See opIso for caveats.

In Numeric.Backprop.Op since version 0.1.2.0, but only exported from Numeric.Backprop since version 0.1.3.0.

Since: 0.1.2.0

opLens :: Num a => Lens' a b -> Op '[a] b

An Op that extracts a value from an input value using a Lens'.

Warning: This is unsafe! It assumes that it extracts a specific value unchanged, with derivative 1, so will break for things that numerically manipulate things before returning them.

No gradients

noGrad1 :: (a -> b) -> Op '[a] b

Create an Op with no gradient. Can be evaluated with evalOp, but will throw a runtime exception when asked for the gradient.

Can be used with BVar with liftOp1, and evalBP will work fine. gradBP and backprop will also work fine if the result is never used in the final answer, but will throw a runtime exception if the final answer depends on the result of this operation.

Useful if your only API is exposed through backprop. Just be sure to tell your users that this will explode when finding the gradient if the result is used in the final result.

Since: 0.1.3.0

noGrad :: (Tuple as -> b) -> Op as b

Create an Op with no gradient. Can be evaluated with evalOp, but will throw a runtime exception when asked for the gradient.

Can be used with BVar with liftOp, and evalBP will work fine. gradBP and backprop will also work fine if the result is never used in the final answer, but will throw a runtime exception if the final answer depends on the result of this operation.

Useful if your only API is exposed through backprop. Just be sure to tell your users that this will explode when finding the gradient if the result is used in the final result.

Since: 0.1.3.0

Utility

Inductive tuples/heterogeneous lists

data Prod k (f :: k -> *) (a :: [k]) :: forall k. (k -> *) -> [k] -> * where

Constructors

Ø :: Prod k f ([] k)

(:<) :: Prod k f ((:) k a1 as) infixr 5

Instances

Witness ØC ØC (Prod k f (Ø k))

Associated Types

type WitnessC ØC ØC (Prod k f (Ø k)) :: Constraint

Methods

(\\) :: ØC => (ØC -> r) -> Prod k f (Ø k) -> r

Functor1 k [k] (Prod k)

Methods

map1 :: (forall (a :: Prod k). f a -> g a) -> t f b -> t g b

Foldable1 k [k] (Prod k)

Methods

foldMap1 :: Monoid m => (forall (a :: Prod k). f a -> m) -> t f b -> m

Traversable1 k [k] (Prod k)

Methods

traverse1 :: Applicative h => (forall (a :: Prod k). f a -> h (g a)) -> t f b -> h (t g b)

IxFunctor1 k [k] (Index k) (Prod k)

Methods

imap1 :: (forall (a :: Index k). i b a -> f a -> g a) -> t f b -> t g b

IxFoldable1 k [k] (Index k) (Prod k)

Methods

ifoldMap1 :: Monoid m => (forall (a :: Index k). i b a -> f a -> m) -> t f b -> m

IxTraversable1 k [k] (Index k) (Prod k)

Methods

itraverse1 :: Applicative h => (forall (a :: Index k). i b a -> f a -> h (g a)) -> t f b -> h (t g b)

TestEquality k f => TestEquality [k] (Prod k f)

Methods

testEquality :: f a -> f b -> Maybe ((Prod k f :~: a) b)

BoolEquality k f => BoolEquality [k] (Prod k f)

Methods

boolEquality :: f a -> f b -> Boolean ((Prod k f == a) b)

Eq1 k f => Eq1 [k] (Prod k f)

Methods

eq1 :: f a -> f a -> Bool

neq1 :: f a -> f a -> Bool

Ord1 k f => Ord1 [k] (Prod k f)

Methods

compare1 :: f a -> f a -> Ordering

(<#) :: f a -> f a -> Bool

(>#) :: f a -> f a -> Bool

(<=#) :: f a -> f a -> Bool

(>=#) :: f a -> f a -> Bool

Show1 k f => Show1 [k] (Prod k f)

Methods

showsPrec1 :: Int -> f a -> ShowS

show1 :: f a -> String

Read1 k f => Read1 [k] (Prod k f)

Methods

readsPrec1 :: Int -> ReadS (Some (Prod k f) f)

(Known [k] (Length k) as, Every k (Known k f) as) => Known [k] (Prod k f) as

Associated Types

type KnownC (Prod k f) (as :: Prod k f -> *) (a :: Prod k f) :: Constraint

Methods

known :: as a

(Witness p q (f a2), Witness s t (Prod a1 f as)) => Witness (p, s) (q, t) (Prod a1 f ((:<) a1 a2 as))

Associated Types

type WitnessC (p, s) (q, t) (Prod a1 f ((a1 :< a2) as)) :: Constraint

Methods

(\\) :: (p, s) => ((q, t) -> r) -> Prod a1 f ((a1 :< a2) as) -> r

ListC ((<$>) * Constraint Eq ((<$>) k * f as)) => Eq (Prod k f as)

Methods

(==) :: Prod k f as -> Prod k f as -> Bool

(/=) :: Prod k f as -> Prod k f as -> Bool

(ListC ((<$>) * Constraint Eq ((<$>) k * f as)), ListC ((<$>) * Constraint Ord ((<$>) k * f as))) => Ord (Prod k f as)

Methods

compare :: Prod k f as -> Prod k f as -> Ordering

(<) :: Prod k f as -> Prod k f as -> Bool

(<=) :: Prod k f as -> Prod k f as -> Bool

(>) :: Prod k f as -> Prod k f as -> Bool

(>=) :: Prod k f as -> Prod k f as -> Bool

max :: Prod k f as -> Prod k f as -> Prod k f as

min :: Prod k f as -> Prod k f as -> Prod k f as

ListC ((<$>) * Constraint Show ((<$>) k * f as)) => Show (Prod k f as)

Methods

showsPrec :: Int -> Prod k f as -> ShowS

show :: Prod k f as -> String

showList :: [Prod k f as] -> ShowS

ListC ((<$>) * Constraint Backprop ((<$>) * * f as)) => Backprop (Prod * f as)

Methods

zero :: Prod * f as -> Prod * f as

add :: Prod * f as -> Prod * f as -> Prod * f as

one :: Prod * f as -> Prod * f as

type WitnessC ØC ØC (Prod k f (Ø k))

type WitnessC ØC ØC (Prod k f (Ø k)) = ØC

type KnownC [k] (Prod k f) as

type KnownC [k] (Prod k f) as = (Known [k] (Length k) as, Every k (Known k f) as)

type WitnessC (p, s) (q, t) (Prod a1 f ((:<) a1 a2 as))

type WitnessC (p, s) (q, t) (Prod a1 f ((:<) a1 a2 as)) = (Witness p q (f a2), Witness s t (Prod a1 f as))

pattern (:>) :: forall k (f :: k -> *) (a :: k) (b :: k). f a -> f b -> Prod k f ((:) k a ((:) k b ([] k))) infix 6

Construct a two element Prod. Since the precedence of (:>) is higher than (:<), we can conveniently write lists like:

>>> a :< b :> c

Which is identical to:

>>> a :< b :< c :< Ø

only :: f a -> Prod k f ((:) k a ([] k))

Build a singleton Prod.

head' :: Prod k f ((:<) k a as) -> f a

type Tuple = Prod * I

A Prod of simple Haskell types.

pattern (::<) :: forall a (as :: [*]). a -> Tuple as -> Tuple ((:<) * a as) infixr 5

Cons onto a Tuple.

only_ :: a -> Tuple ((:) * a ([] *))

Singleton Tuple.

newtype I a :: * -> *

Constructors

I

Fields

• getI :: a

Instances

Monad I

Methods

(>>=) :: I a -> (a -> I b) -> I b

(>>) :: I a -> I b -> I b

return :: a -> I a

fail :: String -> I a

Functor I

Methods

fmap :: (a -> b) -> I a -> I b

(<$) :: a -> I b -> I a

Applicative I

Methods

pure :: a -> I a

(<*>) :: I (a -> b) -> I a -> I b

liftA2 :: (a -> b -> c) -> I a -> I b -> I c

(*>) :: I a -> I b -> I b

(<*) :: I a -> I b -> I a

Foldable I

Methods

fold :: Monoid m => I m -> m

foldMap :: Monoid m => (a -> m) -> I a -> m

foldr :: (a -> b -> b) -> b -> I a -> b

foldr' :: (a -> b -> b) -> b -> I a -> b

foldl :: (b -> a -> b) -> b -> I a -> b

foldl' :: (b -> a -> b) -> b -> I a -> b

foldr1 :: (a -> a -> a) -> I a -> a

foldl1 :: (a -> a -> a) -> I a -> a

toList :: I a -> [a]

null :: I a -> Bool

length :: I a -> Int

elem :: Eq a => a -> I a -> Bool

maximum :: Ord a => I a -> a

minimum :: Ord a => I a -> a

sum :: Num a => I a -> a

product :: Num a => I a -> a

Traversable I

Methods

traverse :: Applicative f => (a -> f b) -> I a -> f (I b)

sequenceA :: Applicative f => I (f a) -> f (I a)

mapM :: Monad m => (a -> m b) -> I a -> m (I b)

sequence :: Monad m => I (m a) -> m (I a)

Witness p q a => Witness p q (I a)

Associated Types

type WitnessC p q (I a) :: Constraint

Methods

(\\) :: p => (q -> r) -> I a -> r

Eq a => Eq (I a)

Methods

(==) :: I a -> I a -> Bool

(/=) :: I a -> I a -> Bool

Num a => Num (I a)

Methods

(+) :: I a -> I a -> I a

(-) :: I a -> I a -> I a

(*) :: I a -> I a -> I a

negate :: I a -> I a

abs :: I a -> I a

signum :: I a -> I a

fromInteger :: Integer -> I a

Ord a => Ord (I a)

Methods

compare :: I a -> I a -> Ordering

(<) :: I a -> I a -> Bool

(<=) :: I a -> I a -> Bool

(>) :: I a -> I a -> Bool

(>=) :: I a -> I a -> Bool

max :: I a -> I a -> I a

min :: I a -> I a -> I a

Show a => Show (I a)

Methods

showsPrec :: Int -> I a -> ShowS

show :: I a -> String

showList :: [I a] -> ShowS

Backprop a => Backprop (I a)

Methods

zero :: I a -> I a

add :: I a -> I a -> I a

one :: I a -> I a

type WitnessC p q (I a)

type WitnessC p q (I a) = Witness p q a

Misc

class Reifies k (s :: k) a | s -> a

Minimal complete definition

reflect

Instances

KnownNat n => Reifies Nat n Integer

Methods

reflect :: proxy Integer -> a

KnownSymbol n => Reifies Symbol n String

Methods

reflect :: proxy String -> a

Reifies * Z Int

Methods

reflect :: proxy Int -> a

Reifies * n Int => Reifies * (D n) Int

Methods

reflect :: proxy Int -> a

Reifies * n Int => Reifies * (SD n) Int

Methods

reflect :: proxy Int -> a

Reifies * n Int => Reifies * (PD n) Int

Methods

reflect :: proxy Int -> a

(B * b0, B * b1, B * b2, B * b3, B * b4, B * b5, B * b6, B * b7, (~) * w0 (W b0 b1 b2 b3), (~) * w1 (W b4 b5 b6 b7)) => Reifies * (Stable w0 w1 a) a

Methods

reflect :: proxy a -> a