Grenade.Recurrent.Layers.LSTM

Documentation

data LSTM :: Nat -> Nat -> * where

Long Short Term Memory Recurrent unit

This is a Peephole formulation, so the recurrent shape is just the cell state, the previous output is not held or used at all.

Constructors

LSTM :: (KnownNat input, KnownNat output) => !(LSTMWeights input output) -> !(LSTMWeights input output) -> LSTM input output

Instances

Show (LSTM i o)

Methods

showsPrec :: Int -> LSTM i o -> ShowS

show :: LSTM i o -> String

showList :: [LSTM i o] -> ShowS

(KnownNat i, KnownNat o) => Serialize (LSTM i o)

Methods

put :: Putter (LSTM i o)

get :: Get (LSTM i o)

(KnownNat i, KnownNat o) => UpdateLayer (LSTM i o)

Associated Types

type Gradient (LSTM i o) :: *

Methods

runUpdate :: LearningParameters -> LSTM i o -> Gradient (LSTM i o) -> LSTM i o

createRandom :: MonadRandom m => m (LSTM i o)

runUpdates :: LearningParameters -> LSTM i o -> [Gradient (LSTM i o)] -> LSTM i o

(KnownNat i, KnownNat o) => RecurrentUpdateLayer (LSTM i o)

Associated Types

type RecurrentShape (LSTM i o) :: Shape

(KnownNat i, KnownNat o) => RecurrentLayer (LSTM i o) (D1 i) (D1 o)

Associated Types

type RecTape (LSTM i o) (D1 i :: Shape) (D1 o :: Shape) :: *

Methods

runRecurrentForwards :: LSTM i o -> S (RecurrentShape (LSTM i o)) -> S (D1 i) -> (RecTape (LSTM i o) (D1 i) (D1 o), S (RecurrentShape (LSTM i o)), S (D1 o))

runRecurrentBackwards :: LSTM i o -> RecTape (LSTM i o) (D1 i) (D1 o) -> S (RecurrentShape (LSTM i o)) -> S (D1 o) -> (Gradient (LSTM i o), S (RecurrentShape (LSTM i o)), S (D1 i))

type Gradient (LSTM i o)

type Gradient (LSTM i o) = LSTMWeights i o

type RecurrentShape (LSTM i o)

type RecurrentShape (LSTM i o) = D1 o

type RecTape (LSTM i o) (D1 i) (D1 o)

type RecTape (LSTM i o) (D1 i) (D1 o) = (S (D1 o), S (D1 i))

data LSTMWeights :: Nat -> Nat -> * where

Constructors

LSTMWeights :: (KnownNat input, KnownNat output) => {..} -> LSTMWeights input output

Fields

• lstmWf :: !(L output input)
• lstmUf :: !(L output output)
• lstmBf :: !(R output)
• lstmWi :: !(L output input)
• lstmUi :: !(L output output)
• lstmBi :: !(R output)
• lstmWo :: !(L output input)
• lstmUo :: !(L output output)
• lstmBo :: !(R output)
• lstmWc :: !(L output input)
• lstmBc :: !(R output)

randomLSTM :: forall m i o. (MonadRandom m, KnownNat i, KnownNat o) => m (LSTM i o)

Generate an LSTM layer with random Weights one can also just call createRandom from UpdateLayer

Has forget gate biases set to 1 to encourage early learning.

https://github.com/karpathy/char-rnn/commit/0dfeaa454e687dd0278f036552ea1e48a0a408c9