Safe Haskell	Safe-Inferred
Language	Haskell2010

Torch.GraduallyTyped.NN.Transformer.GLMHead

Synopsis

data LMHeadHasScaling
- = LMHeadWithScaling
- | LMHeadWithoutScaling
data GLMHead (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) (dense :: Type) (activation :: Type) (layerNorm :: Type) (decoder :: Type) (bias :: Type) where
- GLMHead :: forall inputEmbedDim dense activation layerNorm decoder bias. {..} -> GLMHead inputEmbedDim dense activation layerNorm decoder bias
data GBias (bias :: Type) where
- GBias :: forall bias. bias -> GBias bias
type family GLMHeadF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) (vocabDim :: Dim (Name Symbol) (Size Nat)) :: Type where ...
type family LMHeadDenseF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) :: Type where ...
type family LMHeadActivationF (style :: TransformerStyle) :: Type where ...
type family LMHeadLayerNormF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) :: Type where ...
type family LMHeadDecoderF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) (vocabDim :: Dim (Name Symbol) (Size Nat)) :: Type where ...
type family LMHeadBiasF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (vocabDim :: Dim (Name Symbol) (Size Nat)) :: Type where ...
lmHeadSpec :: forall style gradient device dataType inputEmbedDim vocabDim. STransformerStyle style -> SGradient gradient -> SDevice device -> SDataType dataType -> SDim inputEmbedDim -> SDim vocabDim -> Double -> ModelSpec (GLMHeadF style gradient device dataType inputEmbedDim vocabDim)
type family LMHeadOutputF (style :: TransformerStyle) (decoderOutput :: Type) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (vocabDim :: Dim (Name Symbol) (Size Nat)) :: Type where ...

Documentation

data LMHeadHasScaling Source #

A data type that represents whether or not the language modelling head has a scaled decoder output.

Constructors

LMHeadWithScaling
LMHeadWithoutScaling

Instances

Instances details

Generic LMHeadHasScaling Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Associated Types type Rep LMHeadHasScaling :: Type -> Type Source # Methods from :: LMHeadHasScaling -> Rep LMHeadHasScaling x Source # to :: Rep LMHeadHasScaling x -> LMHeadHasScaling Source #
Show LMHeadHasScaling Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods showsPrec :: Int -> LMHeadHasScaling -> ShowS Source # show :: LMHeadHasScaling -> String Source # showList :: [LMHeadHasScaling] -> ShowS Source #
Eq LMHeadHasScaling Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods (==) :: LMHeadHasScaling -> LMHeadHasScaling -> Bool Source # (/=) :: LMHeadHasScaling -> LMHeadHasScaling -> Bool Source #
Ord LMHeadHasScaling Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods compare :: LMHeadHasScaling -> LMHeadHasScaling -> Ordering Source # (<) :: LMHeadHasScaling -> LMHeadHasScaling -> Bool Source # (<=) :: LMHeadHasScaling -> LMHeadHasScaling -> Bool Source # (>) :: LMHeadHasScaling -> LMHeadHasScaling -> Bool Source # (>=) :: LMHeadHasScaling -> LMHeadHasScaling -> Bool Source # max :: LMHeadHasScaling -> LMHeadHasScaling -> LMHeadHasScaling Source # min :: LMHeadHasScaling -> LMHeadHasScaling -> LMHeadHasScaling Source #
HasStateDict LMHeadHasScaling Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec LMHeadHasScaling -> StateDictKey -> m LMHeadHasScaling Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> LMHeadHasScaling -> m () Source #
HasInitialize LMHeadHasScaling generatorDevice LMHeadHasScaling generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods initialize :: MonadThrow m => ModelSpec LMHeadHasScaling -> Generator generatorDevice -> m (LMHeadHasScaling, Generator generatorDevice) Source #
type Rep LMHeadHasScaling Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead type Rep LMHeadHasScaling = D1 ('MetaData "LMHeadHasScaling" "Torch.GraduallyTyped.NN.Transformer.GLMHead" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "LMHeadWithScaling" 'PrefixI 'False) (U1 :: Type -> Type) :+: C1 ('MetaCons "LMHeadWithoutScaling" 'PrefixI 'False) (U1 :: Type -> Type))
type ModelSpec LMHeadHasScaling Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead type ModelSpec LMHeadHasScaling = LMHeadHasScaling

data GLMHead (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) (dense :: Type) (activation :: Type) (layerNorm :: Type) (decoder :: Type) (bias :: Type) where Source #

Generic language modelling head for transformer encoders and decoders.

inputEmbedDim is the dimension of the input embedding.
dense is a dense layer.
activation is an activation function.
layerNorm is a layer normalization layer.
decoder is a decoder layer.
bias is a bias layer.

Constructors

GLMHead

Fields

:: forall inputEmbedDim dense activation layerNorm decoder bias. { lmHeadInputEmbedDim :: SDim inputEmbedDim
the dimension of the input embedding.
, lmHeadDense :: dense
the dense layer.
, lmHeadActivation :: activation
the activation function.
, lmHeadLayerNorm :: layerNorm
the layer normalization layer.
, lmHeadDecoder :: decoder
the decoder layer.
, lmHeadBias :: bias
the bias layer.
, lmHeadHasScaling :: LMHeadHasScaling
whether or not the head has a scaled decoder output.
} -> GLMHead inputEmbedDim dense activation layerNorm decoder bias

Instances

Instances details

Generic (GLMHead inputEmbedDim dense activation layerNorm decoder bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Associated Types type Rep (GLMHead inputEmbedDim dense activation layerNorm decoder bias) :: Type -> Type Source # Methods from :: GLMHead inputEmbedDim dense activation layerNorm decoder bias -> Rep (GLMHead inputEmbedDim dense activation layerNorm decoder bias) x Source # to :: Rep (GLMHead inputEmbedDim dense activation layerNorm decoder bias) x -> GLMHead inputEmbedDim dense activation layerNorm decoder bias Source #
(Show dense, Show activation, Show layerNorm, Show decoder, Show bias) => Show (GLMHead inputEmbedDim dense activation layerNorm decoder bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods showsPrec :: Int -> GLMHead inputEmbedDim dense activation layerNorm decoder bias -> ShowS Source # show :: GLMHead inputEmbedDim dense activation layerNorm decoder bias -> String Source # showList :: [GLMHead inputEmbedDim dense activation layerNorm decoder bias] -> ShowS Source #
(HasStateDict dense, HasStateDict activation, HasStateDict layerNorm, HasStateDict decoder, HasStateDict bias) => HasStateDict (GLMHead inputEmbedDim dense activation layerNorm decoder bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec (GLMHead inputEmbedDim dense activation layerNorm decoder bias) -> StateDictKey -> m (GLMHead inputEmbedDim dense activation layerNorm decoder bias) Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> GLMHead inputEmbedDim dense activation layerNorm decoder bias -> m () Source #
(HasInitialize dense generatorDevice dense' generatorDevice0, HasInitialize activation generatorDevice0 activation' generatorDevice1, HasInitialize layerNorm generatorDevice1 layerNorm' generatorDevice2, HasInitialize decoder generatorDevice2 decoder' generatorDevice3, HasInitialize bias generatorDevice3 bias' generatorOutputDevice) => HasInitialize (GLMHead inputEmbedDim dense activation layerNorm decoder bias) generatorDevice (GLMHead inputEmbedDim dense' activation' layerNorm' decoder' bias') generatorOutputDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods initialize :: MonadThrow m => ModelSpec (GLMHead inputEmbedDim dense activation layerNorm decoder bias) -> Generator generatorDevice -> m (GLMHead inputEmbedDim dense' activation' layerNorm' decoder' bias', Generator generatorOutputDevice) Source #
(HasForward dense (Tensor gradient layout device dataType shape) generatorDevice tensor0 generatorDevice0, HasForward activation tensor0 generatorDevice0 tensor1 generatorDevice1, HasForward layerNorm tensor1 generatorDevice1 tensor2 generatorDevice2, HasForward decoder tensor2 generatorDevice2 (Tensor gradient3 layout3 device3 dataType3 shape3) generatorDevice3, HasForward bias (Tensor gradient3 layout3 device3 dataType3 shape3) generatorDevice3 output generatorOutputDevice) => HasForward (GLMHead inputEmbedDim dense activation layerNorm decoder bias) (Tensor gradient layout device dataType shape) generatorDevice output generatorOutputDevice Source #	`HasForward` instance for `LMHead`. ┌───────┐ │ input │ └───┬───┘ │ ▼ (lmHeadDense) ▼ (lmHeadActivation) ▼ (lmHeadLayerNorm) ▼ lmHeadDecoder ▼ (scaling) ▼ (lmHeadBias) │ ▼ ┌───────────────┐ │ decoderOutput │ └───────────────┘
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods forward :: MonadThrow m => GLMHead inputEmbedDim dense activation layerNorm decoder bias -> Tensor gradient layout device dataType shape -> Generator generatorDevice -> m (output, Generator generatorOutputDevice) Source #
type Rep (GLMHead inputEmbedDim dense activation layerNorm decoder bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead type Rep (GLMHead inputEmbedDim dense activation layerNorm decoder bias) = D1 ('MetaData "GLMHead" "Torch.GraduallyTyped.NN.Transformer.GLMHead" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "GLMHead" 'PrefixI 'True) ((S1 ('MetaSel ('Just "lmHeadInputEmbedDim") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 (SDim inputEmbedDim)) :: (S1 ('MetaSel ('Just "lmHeadDense") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 dense) :: S1 ('MetaSel ('Just "lmHeadActivation") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 activation))) :: ((S1 ('MetaSel ('Just "lmHeadLayerNorm") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 layerNorm) :: S1 ('MetaSel ('Just "lmHeadDecoder") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 decoder)) :: (S1 ('MetaSel ('Just "lmHeadBias") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 bias) :: S1 ('MetaSel ('Just "lmHeadHasScaling") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 LMHeadHasScaling)))))
type ModelSpec (GLMHead inputEmbedDim dense activation layerNorm decoder bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead type ModelSpec (GLMHead inputEmbedDim dense activation layerNorm decoder bias) = GLMHead inputEmbedDim (ModelSpec dense) (ModelSpec activation) (ModelSpec layerNorm) (ModelSpec decoder) (ModelSpec bias)

data GBias (bias :: Type) where Source #

Generic data type for biasing the language model head.

Constructors

GBias :: forall bias. bias -> GBias bias

Instances

Instances details

Generic (GBias bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Associated Types type Rep (GBias bias) :: Type -> Type Source # Methods from :: GBias bias -> Rep (GBias bias) x Source # to :: Rep (GBias bias) x -> GBias bias Source #
Show bias => Show (GBias bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods showsPrec :: Int -> GBias bias -> ShowS Source # show :: GBias bias -> String Source # showList :: [GBias bias] -> ShowS Source #
Eq bias => Eq (GBias bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods (==) :: GBias bias -> GBias bias -> Bool Source # (/=) :: GBias bias -> GBias bias -> Bool Source #
Ord bias => Ord (GBias bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods compare :: GBias bias -> GBias bias -> Ordering Source # (<) :: GBias bias -> GBias bias -> Bool Source # (<=) :: GBias bias -> GBias bias -> Bool Source # (>) :: GBias bias -> GBias bias -> Bool Source # (>=) :: GBias bias -> GBias bias -> Bool Source # max :: GBias bias -> GBias bias -> GBias bias Source # min :: GBias bias -> GBias bias -> GBias bias Source #
HasStateDict bias => HasStateDict (GBias bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec (GBias bias) -> StateDictKey -> m (GBias bias) Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> GBias bias -> m () Source #
HasInitialize (GBias (NamedModel (Tensor biasGradient biasLayout biasDevice biasDataType biasShape))) generatorDevice (GBias (NamedModel (Tensor biasGradient biasLayout biasDevice biasDataType biasShape))) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods initialize :: MonadThrow m => ModelSpec (GBias (NamedModel (Tensor biasGradient biasLayout biasDevice biasDataType biasShape))) -> Generator generatorDevice -> m (GBias (NamedModel (Tensor biasGradient biasLayout biasDevice biasDataType biasShape)), Generator generatorDevice) Source #
HasInitialize (GBias (Tensor biasGradient biasLayout biasDevice biasDataType biasShape)) generatorDevice (GBias (Tensor biasGradient biasLayout biasDevice biasDataType biasShape)) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods initialize :: MonadThrow m => ModelSpec (GBias (Tensor biasGradient biasLayout biasDevice biasDataType biasShape)) -> Generator generatorDevice -> m (GBias (Tensor biasGradient biasLayout biasDevice biasDataType biasShape), Generator generatorDevice) Source #
HasInitialize (GBias ()) generatorDevice (GBias ()) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods initialize :: MonadThrow m => ModelSpec (GBias ()) -> Generator generatorDevice -> m (GBias (), Generator generatorDevice) Source #
(shape' ~ BroadcastShapesF shape biasShape, Catch shape', output ~ Tensor (gradient <\|> biasGradient) (layout <+> biasLayout) (device <+> biasDevice) (dataType <+> biasDataType) shape') => HasForward (GBias (NamedModel (Tensor biasGradient biasLayout biasDevice biasDataType biasShape))) (Tensor gradient layout device dataType shape) generatorDevice output generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods forward :: MonadThrow m => GBias (NamedModel (Tensor biasGradient biasLayout biasDevice biasDataType biasShape)) -> Tensor gradient layout device dataType shape -> Generator generatorDevice -> m (output, Generator generatorDevice) Source #
(shape' ~ BroadcastShapesF shape biasShape, Catch shape', output ~ Tensor (gradient <\|> biasGradient) (layout <+> biasLayout) (device <+> biasDevice) (dataType <+> biasDataType) shape') => HasForward (GBias (Tensor biasGradient biasLayout biasDevice biasDataType biasShape)) (Tensor gradient layout device dataType shape) generatorDevice output generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods forward :: MonadThrow m => GBias (Tensor biasGradient biasLayout biasDevice biasDataType biasShape) -> Tensor gradient layout device dataType shape -> Generator generatorDevice -> m (output, Generator generatorDevice) Source #
HasForward (GBias ()) (Tensor gradient layout device dataType shape) generatorDevice (Tensor gradient layout device dataType shape) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead Methods forward :: MonadThrow m => GBias () -> Tensor gradient layout device dataType shape -> Generator generatorDevice -> m (Tensor gradient layout device dataType shape, Generator generatorDevice) Source #
type Rep (GBias bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead type Rep (GBias bias) = D1 ('MetaData "GBias" "Torch.GraduallyTyped.NN.Transformer.GLMHead" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "GBias" 'PrefixI 'False) (S1 ('MetaSel ('Nothing :: Maybe Symbol) 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 bias)))
type ModelSpec (GBias bias) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.GLMHead type ModelSpec (GBias bias) = GBias (ModelSpec bias)

type family GLMHeadF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) (vocabDim :: Dim (Name Symbol) (Size Nat)) :: Type where ... Source #

Equations

GLMHeadF style gradient device dataType inputEmbedDim vocabDim = GLMHead inputEmbedDim (LMHeadDenseF style gradient device dataType inputEmbedDim) (LMHeadActivationF style) (LMHeadLayerNormF style gradient device dataType inputEmbedDim) (LMHeadDecoderF style gradient device dataType inputEmbedDim vocabDim) (LMHeadBiasF style gradient device dataType vocabDim)

type family LMHeadDenseF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) :: Type where ... Source #

Specifies the dense layer of the language model head.

Equations

LMHeadDenseF 'T5 _ _ _ _ = ()
LMHeadDenseF 'ByT5 gradient device dataType inputEmbedDim = LMHeadDenseF 'T5 gradient device dataType inputEmbedDim
LMHeadDenseF 'BART _ _ _ _ = ()
LMHeadDenseF 'MBART gradient device dataType inputEmbedDim = LMHeadDenseF 'BART gradient device dataType inputEmbedDim
LMHeadDenseF 'Pegasus gradient device dataType inputEmbedDim = LMHeadDenseF 'BART gradient device dataType inputEmbedDim
LMHeadDenseF 'BERT gradient device dataType inputEmbedDim = NamedModel (GLinearF 'WithBias gradient device dataType inputEmbedDim inputEmbedDim)
LMHeadDenseF 'RoBERTa gradient device dataType inputEmbedDim = LMHeadDenseF 'BERT gradient device dataType inputEmbedDim

type family LMHeadActivationF (style :: TransformerStyle) :: Type where ... Source #

Specifies the activation function of the language model head.

Equations

LMHeadActivationF 'T5 = ()
LMHeadActivationF 'ByT5 = LMHeadActivationF 'T5
LMHeadActivationF 'BART = ()
LMHeadActivationF 'MBART = LMHeadActivationF 'BART
LMHeadActivationF 'Pegasus = LMHeadActivationF 'BART
LMHeadActivationF 'BERT = Gelu
LMHeadActivationF 'RoBERTa = LMHeadActivationF 'BERT

type family LMHeadLayerNormF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) :: Type where ... Source #

Specifies the layer normalization layer of the language model head.

Equations

LMHeadLayerNormF 'T5 _ _ _ _ = ()
LMHeadLayerNormF 'ByT5 gradient device dataType inputEmbedDim = LMHeadLayerNormF 'T5 gradient device dataType inputEmbedDim
LMHeadLayerNormF 'BART _ _ _ _ = ()
LMHeadLayerNormF 'MBART gradient device dataType inputEmbedDim = LMHeadLayerNormF 'BART gradient device dataType inputEmbedDim
LMHeadLayerNormF 'Pegasus gradient device dataType inputEmbedDim = LMHeadLayerNormF 'BART gradient device dataType inputEmbedDim
LMHeadLayerNormF 'BERT gradient device dataType inputEmbedDim = NamedModel (LayerNorm 'WithBias gradient device dataType ('Shape '[inputEmbedDim]))
LMHeadLayerNormF 'RoBERTa gradient device dataType inputEmbedDim = LMHeadLayerNormF 'BERT gradient device dataType inputEmbedDim

type family LMHeadDecoderF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) (vocabDim :: Dim (Name Symbol) (Size Nat)) :: Type where ... Source #

Specifies the decoder layer of the language model head.

Equations

LMHeadDecoderF 'T5 gradient device dataType inputEmbedDim vocabDim = NamedModel (GLinearF 'WithoutBias gradient device dataType inputEmbedDim vocabDim)
LMHeadDecoderF 'ByT5 gradient device dataType inputEmbedDim vocabDim = LMHeadDecoderF 'T5 gradient device dataType inputEmbedDim vocabDim
LMHeadDecoderF 'BART gradient device dataType inputEmbedDim vocabDim = NamedModel (GLinearF 'WithoutBias gradient device dataType inputEmbedDim vocabDim)
LMHeadDecoderF 'MBART gradient device dataType inputEmbedDim vocabDim = LMHeadDecoderF 'BART gradient device dataType inputEmbedDim vocabDim
LMHeadDecoderF 'Pegasus gradient device dataType inputEmbedDim vocabDim = LMHeadDecoderF 'BART gradient device dataType inputEmbedDim vocabDim
LMHeadDecoderF 'BERT gradient device dataType inputEmbedDim vocabDim = NamedModel (GLinearF 'WithBias gradient device dataType inputEmbedDim vocabDim)
LMHeadDecoderF 'RoBERTa gradient device dataType inputEmbedDim vocabDim = LMHeadDecoderF 'BERT gradient device dataType inputEmbedDim vocabDim

type family LMHeadBiasF (style :: TransformerStyle) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (vocabDim :: Dim (Name Symbol) (Size Nat)) :: Type where ... Source #

Specifies the bias layer of the language model head.

Equations

LMHeadBiasF 'T5 _ _ _ _ = GBias ()
LMHeadBiasF 'ByT5 gradient device dataType vocabDim = LMHeadBiasF 'T5 gradient device dataType vocabDim
LMHeadBiasF 'BART gradient device dataType vocabDim = GBias (NamedModel (Tensor gradient ('Layout 'Dense) device dataType ('Shape '['Dim ('Name "*") ('Size 1), vocabDim])))
LMHeadBiasF 'MBART gradient device dataType vocabDim = LMHeadBiasF 'BART gradient device dataType vocabDim
LMHeadBiasF 'Pegasus gradient device dataType vocabDim = LMHeadBiasF 'BART gradient device dataType vocabDim
LMHeadBiasF 'BERT _ _ _ _ = GBias ()
LMHeadBiasF 'RoBERTa gradient device dataType vocabDim = LMHeadBiasF 'BERT gradient device dataType vocabDim

lmHeadSpec :: forall style gradient device dataType inputEmbedDim vocabDim. STransformerStyle style -> SGradient gradient -> SDevice device -> SDataType dataType -> SDim inputEmbedDim -> SDim vocabDim -> Double -> ModelSpec (GLMHeadF style gradient device dataType inputEmbedDim vocabDim) Source #

Specifies the parameters of the language model head.

type family LMHeadOutputF (style :: TransformerStyle) (decoderOutput :: Type) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (dataType :: DataType DType) (vocabDim :: Dim (Name Symbol) (Size Nat)) :: Type where ... Source #

Equations

LMHeadOutputF 'T5 decoderOutput _ _ _ _ = decoderOutput
LMHeadOutputF 'ByT5 decoderOutput gradient device dataType vocabDim = LMHeadOutputF 'T5 decoderOutput gradient device dataType vocabDim
LMHeadOutputF 'BART (Tensor gradient' layout' device' dataType' shape') gradient device dataType vocabDim = Tensor (gradient' <\|> gradient) (layout' <+> 'Layout 'Dense) (device' <+> device) (dataType' <+> dataType) (BroadcastShapesF shape' ('Shape '['Dim ('Name "*") ('Size 1), vocabDim]))
LMHeadOutputF 'MBART decoderOutput gradient device dataType vocabDim = LMHeadOutputF 'BART decoderOutput gradient device dataType vocabDim
LMHeadOutputF 'Pegasus decoderOutput gradient device dataType vocabDim = LMHeadOutputF 'BART decoderOutput gradient device dataType vocabDim
LMHeadOutputF 'RoBERTa decoderOutput _ _ _ _ = decoderOutput
LMHeadOutputF 'BERT decoderOutput _ _ _ _ = decoderOutput