| Safe Haskell | None |
|---|---|
| Language | Haskell2010 |
Backprop.Learn.Model.Neural
Contents
Synopsis
- type FCp = LRp
- fc :: (KnownNat i, KnownNat o) => Model (Just (FCp i o)) Nothing (R i) (R o)
- fca :: (KnownNat i, KnownNat o) => (forall z. Reifies z W => BVar z (R o) -> BVar z (R o)) -> Model (Just (FCp i o)) Nothing (R i) (R o)
- fcWeights :: Lens (FCp i o) (FCp i' o) (L o i) (L o i')
- fcBias :: forall i o. Lens' (FCp i o) (R o)
- fcr :: (KnownNat i, KnownNat o, KnownNat s) => (forall z. Reifies z W => BVar z (R o) -> BVar z (R s)) -> Model (Just (FCRp s i o)) (Just (R s)) (R i) (R o)
- fcra :: (KnownNat i, KnownNat o, KnownNat s) => (forall z. Reifies z W => BVar z (R o) -> BVar z (R o)) -> (forall z. Reifies z W => BVar z (R o) -> BVar z (R s)) -> Model (Just (FCRp s i o)) (Just (R s)) (R i) (R o)
- type FCRp s i o = FCp (i + s) o
- fcrBias :: forall s i o. Lens' (FCRp s i o) (R o)
- fcrInputWeights :: (KnownNat s, KnownNat i, KnownNat i', KnownNat o) => Lens (FCRp s i o) (FCRp s i' o) (L o i) (L o i')
- fcrStateWeights :: (KnownNat s, KnownNat s', KnownNat i, KnownNat o) => Lens (FCRp s i o) (FCRp s' i o) (L o s) (L o s')
Feed-forward
fc :: (KnownNat i, KnownNat o) => Model (Just (FCp i o)) Nothing (R i) (R o) Source #
Fully connected feed-forward layer with bias. Parameterized by its initialization distribution.
Note that this has no activation function; to use as a model with
activation function, chain it with an activation function using RMap,
:.~, etc.; see FCA for a convenient type synonym and constructor.
Without any activation function, this is essentially a multivariate linear regression.
With the logistic function as an activation function, this is
essentially multivariate logistic regression. (See logReg)
fca :: (KnownNat i, KnownNat o) => (forall z. Reifies z W => BVar z (R o) -> BVar z (R o)) -> Model (Just (FCp i o)) Nothing (R i) (R o) Source #
Convenient synonym for an fC post-composed with a simple
parameterless activation function.
Recurrent
Arguments
| :: (KnownNat i, KnownNat o, KnownNat s) | |
| => (forall z. Reifies z W => BVar z (R o) -> BVar z (R s)) | store |
| -> Model (Just (FCRp s i o)) (Just (R s)) (R i) (R o) |
Fully connected recurrent layer with bias.
Arguments
| :: (KnownNat i, KnownNat o, KnownNat s) | |
| => (forall z. Reifies z W => BVar z (R o) -> BVar z (R o)) | |
| -> (forall z. Reifies z W => BVar z (R o) -> BVar z (R s)) | store |
| -> Model (Just (FCRp s i o)) (Just (R s)) (R i) (R o) |
Convenient synonym for an fcr post-composed with a simple
parameterless activation function.