Safe Haskell	Safe-Inferred
Language	Haskell2010

Torch.GraduallyTyped.NN.Transformer.Type

Synopsis

data TransformerStyle
- = T5
- | ByT5
- | BART
- | MBART
- | Pegasus
- | BERT
- | RoBERTa
- | GPT2
type family T5Sym0 :: TransformerStyle where ...
type family ByT5Sym0 :: TransformerStyle where ...
type family BARTSym0 :: TransformerStyle where ...
type family MBARTSym0 :: TransformerStyle where ...
type family PegasusSym0 :: TransformerStyle where ...
type family BERTSym0 :: TransformerStyle where ...
type family RoBERTaSym0 :: TransformerStyle where ...
type family GPT2Sym0 :: TransformerStyle where ...
data STransformerStyle :: TransformerStyle -> Type where
- ST5 :: STransformerStyle ('T5 :: TransformerStyle)
- SByT5 :: STransformerStyle ('ByT5 :: TransformerStyle)
- SBART :: STransformerStyle ('BART :: TransformerStyle)
- SMBART :: STransformerStyle ('MBART :: TransformerStyle)
- SPegasus :: STransformerStyle ('Pegasus :: TransformerStyle)
- SBERT :: STransformerStyle ('BERT :: TransformerStyle)
- SRoBERTa :: STransformerStyle ('RoBERTa :: TransformerStyle)
- SGPT2 :: STransformerStyle ('GPT2 :: TransformerStyle)
data TransformerHead
- = WithoutHead
- | WithLMHead
type family WithoutHeadSym0 :: TransformerHead where ...
type family WithLMHeadSym0 :: TransformerHead where ...
data STransformerHead :: TransformerHead -> Type where
- SWithoutHead :: STransformerHead ('WithoutHead :: TransformerHead)
- SWithLMHead :: STransformerHead ('WithLMHead :: TransformerHead)
padded :: Integral n => n -> a -> [a] -> [a]
mkTransformerInput :: forall batchDim seqDim device m output. (MonadThrow m, SGetDim batchDim, SGetDim seqDim, Catch ('Shape '['Dim ('Name "*") 'UncheckedSize, 'Dim ('Name "*") 'UncheckedSize] <+> 'Shape '[batchDim, seqDim]), output ~ Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '[batchDim, seqDim])) => Int -> SDim batchDim -> SDim seqDim -> SDevice device -> [[Int]] -> m output
type MkPosC device shape seqDim seqName seqSize output = (SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), seqDim ~ 'Dim seqName seqSize, output ~ Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '['Dim ('Name "*") seqSize]))
mkPos :: forall m gradient layout device dataType shape seqDim seqName seqSize output. (MonadThrow m, MkPosC device shape seqDim seqName seqSize output) => Tensor gradient layout device dataType shape -> m output
data MkAbsPos
- = MkAbsPos
- | MkAbsPosWithOffset {
  - absPosOffset :: Int
  }
mkRelPos' :: Int -> Int -> Int -> Int -> [[Int]]
type MkRelPosC device shape seqDim seqName seqSize output = (SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), seqDim ~ 'Dim seqName seqSize, Catch ('['Dim ('Name "*") 'UncheckedSize, 'Dim ('Name "*") 'UncheckedSize] <+> '['Dim ('Name "*") seqSize, 'Dim ('Name "*") seqSize]), output ~ Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '['Dim ('Name "*") ('Size 1), 'Dim ('Name "*") seqSize, 'Dim ('Name "*") seqSize]))
mkRelPos :: forall m gradient layout device dataType shape relPosEncBucketDim seqDim seqName seqSize output. (MonadThrow m, MkRelPosC device shape seqDim seqName seqSize output) => SDim relPosEncBucketDim -> Int -> Tensor gradient layout device dataType shape -> m output
mkDecoderRelPos' :: Int -> Int -> Int -> Int -> [[Int]]
mkDecoderRelPos :: forall m gradient layout device dataType shape relPosEncBucketDim seqDim seqName seqSize output. (MonadThrow m, MkRelPosC device shape seqDim seqName seqSize output) => SDim relPosEncBucketDim -> Int -> Tensor gradient layout device dataType shape -> m output
data MkRelPos (relPosEncBucketDim :: Dim (Name Symbol) (Size Nat)) where
- MkRelPos :: forall relPosEncBucketDim. {..} -> MkRelPos relPosEncBucketDim
- MkDecoderRelPos :: forall relPosEncBucketDim. {..} -> MkRelPos relPosEncBucketDim
type MkTransformerPaddingMaskC layout device dataType shape output = (SGetDevice device, Catch (dataType <+> 'DataType 'Int64), Catch (BroadcastShapesF shape ('Shape '[])), output ~ Tensor ('Gradient 'WithoutGradient) (layout <+> 'Layout 'Dense) device ('DataType 'Bool) (BroadcastShapesF shape ('Shape '[])))
mkTransformerPaddingMask :: forall m gradient layout device dataType shape output. (MonadThrow m, MkTransformerPaddingMaskC layout device dataType shape output) => Int -> Tensor gradient layout device dataType shape -> m output
newtype MkTransformerPaddingMask = MkTransformerPaddingMask {
- padTokenId :: Int
}
type MkTransformerAttentionMaskC transformerDataType gradient layout device dataType shape seqDim output = (SGetLayout layout, SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), Catch (gradient <+> 'Gradient 'WithoutGradient), Catch (dataType <+> 'DataType 'Bool), Catch (UnsqueezeF ('SelectDim ('ByIndex 1)) shape), Catch (BroadcastShapesF (UnsqueezeF ('SelectDim ('ByIndex 1)) shape) ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim])), output ~ Tensor ('Gradient 'WithoutGradient) (layout <+> 'Layout 'Dense) device transformerDataType (BroadcastShapesF (UnsqueezeF ('SelectDim ('ByIndex 1)) shape) ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim])))
mkTransformerAttentionMask :: forall m transformerDataType gradient layout device dataType shape seqDim output. (MonadThrow m, MkTransformerAttentionMaskC transformerDataType gradient layout device dataType shape seqDim output) => SDataType transformerDataType -> Double -> Tensor gradient layout device dataType shape -> m output
data MkTransformerAttentionMask (dataType :: DataType DType) where
- MkTransformerAttentionMask :: forall dataType. {..} -> MkTransformerAttentionMask dataType
type MkTransformerDecoderAttentionMaskC transformerDataType layout device shape seqDim output = (SGetLayout layout, SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), Catch seqDim, Catch (UnsqueezeF ('SelectDim ('ByIndex 1)) shape), Catch (BroadcastShapesF ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim]) (UnsqueezeF ('SelectDim ('ByIndex 1)) shape)), Catch (BroadcastShapesF (BroadcastShapesF ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim]) (UnsqueezeF ('SelectDim ('ByIndex 1)) shape)) ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim])), output ~ Tensor ('Gradient 'WithoutGradient) (layout <+> 'Layout 'Dense) device transformerDataType (BroadcastShapesF (BroadcastShapesF ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim]) (UnsqueezeF ('SelectDim ('ByIndex 1)) shape)) ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim])))
mkTransformerDecoderAttentionMask :: forall m transformerDataType gradient layout device dataType shape seqDim output. (MonadThrow m, MkTransformerDecoderAttentionMaskC transformerDataType layout device shape seqDim output) => SDataType transformerDataType -> Double -> Tensor gradient layout device dataType shape -> m output
data MkTransformerDecoderAttentionMask (dataType :: DataType DType) where
- MkTransformerDecoderAttentionMask :: forall dataType. {..} -> MkTransformerDecoderAttentionMask dataType
type MkTransformerCrossAttentionMaskC transformerDataType decoderInputShape decoderInputSeqDim gradient layout device dataType shape seqDim output = (SGetLayout layout, SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), SGetShape decoderInputShape, decoderInputSeqDim ~ (decoderInputShape ! 1), Catch (gradient <+> 'Gradient 'WithoutGradient), Catch (dataType <+> 'DataType 'Bool), Catch (UnsqueezeF ('SelectDim ('ByIndex 1)) shape), Catch (BroadcastShapesF (UnsqueezeF ('SelectDim ('ByIndex 1)) shape) ('Shape '['Dim ('Name "*") ('Size 1), decoderInputSeqDim, seqDim])), output ~ Tensor ('Gradient 'WithoutGradient) (layout <+> 'Layout 'Dense) device transformerDataType (BroadcastShapesF (UnsqueezeF ('SelectDim ('ByIndex 1)) shape) ('Shape '['Dim ('Name "*") ('Size 1), decoderInputSeqDim, seqDim])))
mkTransformerCrossAttentionMask :: forall m transformerDataType decoderInputShape decoderInputSeqDim gradient layout device dataType shape seqDim output. (MonadThrow m, MkTransformerCrossAttentionMaskC transformerDataType decoderInputShape decoderInputSeqDim gradient layout device dataType shape seqDim output) => SDataType transformerDataType -> SShape decoderInputShape -> Double -> Tensor gradient layout device dataType shape -> m output
data MkTransformerCrossAttentionMask (dataType :: DataType DType) where
- MkTransformerCrossAttentionMask :: forall dataType. {..} -> MkTransformerCrossAttentionMask dataType
data ShiftRight fillValue where
- ShiftRight :: forall fillValue. fillValue -> ShiftRight fillValue

Documentation

data TransformerStyle Source #

A data type representing the style of a transformer. Every supported transformer has a constructor of this type.

Constructors

T5	`T5` transformer style, see https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html
ByT5	`ByT5` transformer style, see https://arxiv.org/abs/2105.13626
BART	`BART` transformer style, see https://arxiv.org/abs/1910.13461
MBART	`MBART` transformer style, see https://arxiv.org/abs/2001.08210
Pegasus	`Pegasus` transformer style, see https://ai.googleblog.com/2020/06/pegasus-state-of-art-model-for.html
BERT	`BERT` transformer style, see https://arxiv.org/abs/1810.04805
RoBERTa	`RoBERTa` transformer style, see https://arxiv.org/abs/1907.11692
GPT2	`GPT2` transformer style, see https://openai.com/blog/better-language-models/

Instances

Instances details

Show TransformerStyle Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods showsPrec :: Int -> TransformerStyle -> ShowS Source # show :: TransformerStyle -> String Source # showList :: [TransformerStyle] -> ShowS Source #
Eq TransformerStyle Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods (==) :: TransformerStyle -> TransformerStyle -> Bool Source # (/=) :: TransformerStyle -> TransformerStyle -> Bool Source #
SingKind TransformerStyle Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Demote TransformerStyle = (r :: Type) Source # Methods fromSing :: forall (a :: TransformerStyle). Sing a -> Demote TransformerStyle Source # toSing :: Demote TransformerStyle -> SomeSing TransformerStyle Source #
SingI 'BART Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'BART Source #
SingI 'BERT Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'BERT Source #
SingI 'ByT5 Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'ByT5 Source #
SingI 'GPT2 Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'GPT2 Source #
SingI 'MBART Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'MBART Source #
SingI 'Pegasus Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'Pegasus Source #
SingI 'RoBERTa Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'RoBERTa Source #
SingI 'T5 Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'T5 Source #
type Demote TransformerStyle Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Demote TransformerStyle = TransformerStyle
type Sing Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Sing = STransformerStyle

type family T5Sym0 :: TransformerStyle where ... Source #

Equations

T5Sym0 = 'T5

type family ByT5Sym0 :: TransformerStyle where ... Source #

Equations

ByT5Sym0 = 'ByT5

type family BARTSym0 :: TransformerStyle where ... Source #

Equations

BARTSym0 = 'BART

type family MBARTSym0 :: TransformerStyle where ... Source #

Equations

MBARTSym0 = 'MBART

type family PegasusSym0 :: TransformerStyle where ... Source #

Equations

PegasusSym0 = 'Pegasus

type family BERTSym0 :: TransformerStyle where ... Source #

Equations

BERTSym0 = 'BERT

type family RoBERTaSym0 :: TransformerStyle where ... Source #

Equations

RoBERTaSym0 = 'RoBERTa

type family GPT2Sym0 :: TransformerStyle where ... Source #

Equations

GPT2Sym0 = 'GPT2

data STransformerStyle :: TransformerStyle -> Type where Source #

Constructors

ST5 :: STransformerStyle ('T5 :: TransformerStyle)
SByT5 :: STransformerStyle ('ByT5 :: TransformerStyle)
SBART :: STransformerStyle ('BART :: TransformerStyle)
SMBART :: STransformerStyle ('MBART :: TransformerStyle)
SPegasus :: STransformerStyle ('Pegasus :: TransformerStyle)
SBERT :: STransformerStyle ('BERT :: TransformerStyle)
SRoBERTa :: STransformerStyle ('RoBERTa :: TransformerStyle)
SGPT2 :: STransformerStyle ('GPT2 :: TransformerStyle)

data TransformerHead Source #

A data type representing the type of head used in a transformer.

Constructors

WithoutHead
WithLMHead

Instances

Instances details

SingKind TransformerHead Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Demote TransformerHead = (r :: Type) Source # Methods fromSing :: forall (a :: TransformerHead). Sing a -> Demote TransformerHead Source # toSing :: Demote TransformerHead -> SomeSing TransformerHead Source #
SingI 'WithLMHead Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'WithLMHead Source #
SingI 'WithoutHead Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods sing :: Sing 'WithoutHead Source #
type Demote TransformerHead Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Demote TransformerHead = TransformerHead
type Sing Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Sing = STransformerHead

type family WithoutHeadSym0 :: TransformerHead where ... Source #

Equations

WithoutHeadSym0 = 'WithoutHead

type family WithLMHeadSym0 :: TransformerHead where ... Source #

Equations

WithLMHeadSym0 = 'WithLMHead

data STransformerHead :: TransformerHead -> Type where Source #

Constructors

SWithoutHead :: STransformerHead ('WithoutHead :: TransformerHead)
SWithLMHead :: STransformerHead ('WithLMHead :: TransformerHead)

padded :: Integral n => n -> a -> [a] -> [a] Source #

mkTransformerInput Source #

Arguments

:: forall batchDim seqDim device m output. (MonadThrow m, SGetDim batchDim, SGetDim seqDim, Catch ('Shape '['Dim ('Name "") 'UncheckedSize, 'Dim ('Name "") 'UncheckedSize] <+> 'Shape '[batchDim, seqDim]), output ~ Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '[batchDim, seqDim]))
=> Int	padding token id
-> SDim batchDim	batch dimension singleton
-> SDim seqDim	sequence dimension singleton
-> SDevice device	device for the tensor
-> [[Int]]	batch of input ids
-> m output	input tensor

Converts a doubly-nested list of input ids to a batched input tensor. The outer list is over batches, the inner list over sequences. The batch size is inferred from the length of the outer list. The sequence length is inferred from the length of the inner list. The input ids are padded to the maximum sequence length. The output tensor is truncated to the maximum sequence length.

type MkPosC device shape seqDim seqName seqSize output = (SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), seqDim ~ 'Dim seqName seqSize, output ~ Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '['Dim ('Name "*") seqSize])) Source #

mkPos Source #

Arguments

:: forall m gradient layout device dataType shape seqDim seqName seqSize output. (MonadThrow m, MkPosC device shape seqDim seqName seqSize output)
=> Tensor gradient layout device dataType shape	input tensor
-> m output	positions of the input tokens

Computes absolute positions of the input tokens. Given an input tensor of shape [batchDim, Dim seqName seqSize], returns a tensor of shape [Dim "*" seqSize].

data MkAbsPos Source #

Constructors

MkAbsPos
MkAbsPosWithOffset
Fields absPosOffset :: Int

Instances

Instances details

Generic MkAbsPos Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Rep MkAbsPos :: Type -> Type Source # Methods from :: MkAbsPos -> Rep MkAbsPos x Source # to :: Rep MkAbsPos x -> MkAbsPos Source #
Show MkAbsPos Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods showsPrec :: Int -> MkAbsPos -> ShowS Source # show :: MkAbsPos -> String Source # showList :: [MkAbsPos] -> ShowS Source #
Eq MkAbsPos Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods (==) :: MkAbsPos -> MkAbsPos -> Bool Source # (/=) :: MkAbsPos -> MkAbsPos -> Bool Source #
Ord MkAbsPos Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods compare :: MkAbsPos -> MkAbsPos -> Ordering Source # (<) :: MkAbsPos -> MkAbsPos -> Bool Source # (<=) :: MkAbsPos -> MkAbsPos -> Bool Source # (>) :: MkAbsPos -> MkAbsPos -> Bool Source # (>=) :: MkAbsPos -> MkAbsPos -> Bool Source # max :: MkAbsPos -> MkAbsPos -> MkAbsPos Source # min :: MkAbsPos -> MkAbsPos -> MkAbsPos Source #
HasStateDict MkAbsPos Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec MkAbsPos -> StateDictKey -> m MkAbsPos Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> MkAbsPos -> m () Source #
HasInitialize MkAbsPos generatorDevice MkAbsPos generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods initialize :: MonadThrow m => ModelSpec MkAbsPos -> Generator generatorDevice -> m (MkAbsPos, Generator generatorDevice) Source #
MkPosC device shape seqDim seqName seqSize output => HasForward MkAbsPos (Tensor gradient layout device dataType shape) generatorDevice (Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '['Dim ('Name "*") seqSize])) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods forward :: MonadThrow m => MkAbsPos -> Tensor gradient layout device dataType shape -> Generator generatorDevice -> m (Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '['Dim ('Name "*") seqSize]), Generator generatorDevice) Source #
type Rep MkAbsPos Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Rep MkAbsPos = D1 ('MetaData "MkAbsPos" "Torch.GraduallyTyped.NN.Transformer.Type" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "MkAbsPos" 'PrefixI 'False) (U1 :: Type -> Type) :+: C1 ('MetaCons "MkAbsPosWithOffset" 'PrefixI 'True) (S1 ('MetaSel ('Just "absPosOffset") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Int)))
type ModelSpec MkAbsPos Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type ModelSpec MkAbsPos = MkAbsPos

mkRelPos' :: Int -> Int -> Int -> Int -> [[Int]] Source #

Computes relative positions of the input tokens to the encoder.

>>> mkRelPos' 32 128 21 17
[[0,17,18,19,20,21,22,23,24,24,24,24,25,25,25,25,26],[1,0,17,18,19,20,21,22,23,24,24,24,24,25,25,25,25],[2,1,0,17,18,19,20,21,22,23,24,24,24,24,25,25,25],[3,2,1,0,17,18,19,20,21,22,23,24,24,24,24,25,25],[4,3,2,1,0,17,18,19,20,21,22,23,24,24,24,24,25],[5,4,3,2,1,0,17,18,19,20,21,22,23,24,24,24,24],[6,5,4,3,2,1,0,17,18,19,20,21,22,23,24,24,24],[7,6,5,4,3,2,1,0,17,18,19,20,21,22,23,24,24],[8,7,6,5,4,3,2,1,0,17,18,19,20,21,22,23,24],[8,8,7,6,5,4,3,2,1,0,17,18,19,20,21,22,23],[8,8,8,7,6,5,4,3,2,1,0,17,18,19,20,21,22],[8,8,8,8,7,6,5,4,3,2,1,0,17,18,19,20,21],[9,8,8,8,8,7,6,5,4,3,2,1,0,17,18,19,20],[9,9,8,8,8,8,7,6,5,4,3,2,1,0,17,18,19],[9,9,9,8,8,8,8,7,6,5,4,3,2,1,0,17,18],[9,9,9,9,8,8,8,8,7,6,5,4,3,2,1,0,17],[10,9,9,9,9,8,8,8,8,7,6,5,4,3,2,1,0],[10,10,9,9,9,9,8,8,8,8,7,6,5,4,3,2,1],[10,10,10,9,9,9,9,8,8,8,8,7,6,5,4,3,2],[10,10,10,10,9,9,9,9,8,8,8,8,7,6,5,4,3],[10,10,10,10,10,9,9,9,9,8,8,8,8,7,6,5,4]]

type MkRelPosC device shape seqDim seqName seqSize output = (SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), seqDim ~ 'Dim seqName seqSize, Catch ('['Dim ('Name "*") 'UncheckedSize, 'Dim ('Name "*") 'UncheckedSize] <+> '['Dim ('Name "*") seqSize, 'Dim ('Name "*") seqSize]), output ~ Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '['Dim ('Name "*") ('Size 1), 'Dim ('Name "*") seqSize, 'Dim ('Name "*") seqSize])) Source #

mkRelPos Source #

Arguments

:: forall m gradient layout device dataType shape relPosEncBucketDim seqDim seqName seqSize output. (MonadThrow m, MkRelPosC device shape seqDim seqName seqSize output)
=> SDim relPosEncBucketDim	bucket dimension
-> Int	maximum distance
-> Tensor gradient layout device dataType shape	input tensor
-> m output	relative positions of the input tokens

Computes relative positions of the input tokens to the encoder. Given an input tensor of shape [batchDim, Dim seqName seqSize], returns a tensor of shape [1, Dim "*" seqSize, Dim "*" seqSize].

mkDecoderRelPos' :: Int -> Int -> Int -> Int -> [[Int]] Source #

Computes relative positions of the input tokens to the decoder.

>>> mkDecoderRelPos' 32 128 21 17
[[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],[2,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],[3,2,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0],[4,3,2,1,0,0,0,0,0,0,0,0,0,0,0,0,0],[5,4,3,2,1,0,0,0,0,0,0,0,0,0,0,0,0],[6,5,4,3,2,1,0,0,0,0,0,0,0,0,0,0,0],[7,6,5,4,3,2,1,0,0,0,0,0,0,0,0,0,0],[8,7,6,5,4,3,2,1,0,0,0,0,0,0,0,0,0],[9,8,7,6,5,4,3,2,1,0,0,0,0,0,0,0,0],[10,9,8,7,6,5,4,3,2,1,0,0,0,0,0,0,0],[11,10,9,8,7,6,5,4,3,2,1,0,0,0,0,0,0],[12,11,10,9,8,7,6,5,4,3,2,1,0,0,0,0,0],[13,12,11,10,9,8,7,6,5,4,3,2,1,0,0,0,0],[14,13,12,11,10,9,8,7,6,5,4,3,2,1,0,0,0],[15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0,0],[16,15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0],[16,16,15,14,13,12,11,10,9,8,7,6,5,4,3,2,1],[16,16,16,15,14,13,12,11,10,9,8,7,6,5,4,3,2],[17,16,16,16,15,14,13,12,11,10,9,8,7,6,5,4,3],[17,17,16,16,16,15,14,13,12,11,10,9,8,7,6,5,4]]

mkDecoderRelPos Source #

Arguments

:: forall m gradient layout device dataType shape relPosEncBucketDim seqDim seqName seqSize output. (MonadThrow m, MkRelPosC device shape seqDim seqName seqSize output)
=> SDim relPosEncBucketDim	bucket dimension
-> Int	maximum distance
-> Tensor gradient layout device dataType shape	decoder input tensor
-> m output	relative positions of the input tokens

Computes relative positions of the input tokens to the decoder. Given an input tensor of shape [batchDim, Dim seqName seqSize], returns a tensor of shape [1, Dim "*" seqSize, Dim "*" seqSize].

data MkRelPos (relPosEncBucketDim :: Dim (Name Symbol) (Size Nat)) where Source #

Constructors

MkRelPos
Fields :: forall relPosEncBucketDim. { relPosEncBucketDim :: SDim relPosEncBucketDim , relPosMaxDistance :: Int } -> MkRelPos relPosEncBucketDim
MkDecoderRelPos
Fields :: forall relPosEncBucketDim. { decoderRelPosEncBucketDim :: SDim relPosEncBucketDim , decoderRelPosMaxDistance :: Int } -> MkRelPos relPosEncBucketDim

Instances

Instances details

Generic (MkRelPos relPosEncBucketDim) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Rep (MkRelPos relPosEncBucketDim) :: Type -> Type Source # Methods from :: MkRelPos relPosEncBucketDim -> Rep (MkRelPos relPosEncBucketDim) x Source # to :: Rep (MkRelPos relPosEncBucketDim) x -> MkRelPos relPosEncBucketDim Source #
Show (MkRelPos relPosEncBucketDim) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods showsPrec :: Int -> MkRelPos relPosEncBucketDim -> ShowS Source # show :: MkRelPos relPosEncBucketDim -> String Source # showList :: [MkRelPos relPosEncBucketDim] -> ShowS Source #
HasStateDict (MkRelPos relPosEncBucketDim) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec (MkRelPos relPosEncBucketDim) -> StateDictKey -> m (MkRelPos relPosEncBucketDim) Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> MkRelPos relPosEncBucketDim -> m () Source #
HasInitialize (MkRelPos relPosEncBucketDim) generatorDevice (MkRelPos relPosEncBucketDim) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods initialize :: MonadThrow m => ModelSpec (MkRelPos relPosEncBucketDim) -> Generator generatorDevice -> m (MkRelPos relPosEncBucketDim, Generator generatorDevice) Source #
MkRelPosC device shape seqDim seqName seqSize output => HasForward (MkRelPos relPosEncBucketDim) (Tensor gradient layout device dataType shape) generatorDevice (Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '['Dim ('Name "") ('Size 1), 'Dim ('Name "") seqSize, 'Dim ('Name "*") seqSize])) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods forward :: MonadThrow m => MkRelPos relPosEncBucketDim -> Tensor gradient layout device dataType shape -> Generator generatorDevice -> m (Tensor ('Gradient 'WithoutGradient) ('Layout 'Dense) device ('DataType 'Int64) ('Shape '['Dim ('Name "") ('Size 1), 'Dim ('Name "") seqSize, 'Dim ('Name "*") seqSize]), Generator generatorDevice) Source #
type Rep (MkRelPos relPosEncBucketDim) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Rep (MkRelPos relPosEncBucketDim) = D1 ('MetaData "MkRelPos" "Torch.GraduallyTyped.NN.Transformer.Type" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "MkRelPos" 'PrefixI 'True) (S1 ('MetaSel ('Just "relPosEncBucketDim") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 (SDim relPosEncBucketDim)) :: S1 ('MetaSel ('Just "relPosMaxDistance") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Int)) :+: C1 ('MetaCons "MkDecoderRelPos" 'PrefixI 'True) (S1 ('MetaSel ('Just "decoderRelPosEncBucketDim") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 (SDim relPosEncBucketDim)) :: S1 ('MetaSel ('Just "decoderRelPosMaxDistance") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Int)))
type ModelSpec (MkRelPos relPosEncBucketDim) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type ModelSpec (MkRelPos relPosEncBucketDim) = MkRelPos relPosEncBucketDim

type MkTransformerPaddingMaskC layout device dataType shape output = (SGetDevice device, Catch (dataType <+> 'DataType 'Int64), Catch (BroadcastShapesF shape ('Shape '[])), output ~ Tensor ('Gradient 'WithoutGradient) (layout <+> 'Layout 'Dense) device ('DataType 'Bool) (BroadcastShapesF shape ('Shape '[]))) Source #

mkTransformerPaddingMask Source #

Arguments

:: forall m gradient layout device dataType shape output. (MonadThrow m, MkTransformerPaddingMaskC layout device dataType shape output)
=> Int	padding token id
-> Tensor gradient layout device dataType shape	input tensor
-> m output	padding mask

Computes the padding mask for a transformer. Given an input tensor of shape [batchDim, Dim seqName seqSize], returns a tensor of shape [batchDim, Dim "*" seqSize].

newtype MkTransformerPaddingMask Source #

Constructors

MkTransformerPaddingMask
Fields padTokenId :: Int

Instances

Instances details

Generic MkTransformerPaddingMask Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Rep MkTransformerPaddingMask :: Type -> Type Source # Methods from :: MkTransformerPaddingMask -> Rep MkTransformerPaddingMask x Source # to :: Rep MkTransformerPaddingMask x -> MkTransformerPaddingMask Source #
Show MkTransformerPaddingMask Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods showsPrec :: Int -> MkTransformerPaddingMask -> ShowS Source # show :: MkTransformerPaddingMask -> String Source # showList :: [MkTransformerPaddingMask] -> ShowS Source #
HasStateDict MkTransformerPaddingMask Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec MkTransformerPaddingMask -> StateDictKey -> m MkTransformerPaddingMask Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> MkTransformerPaddingMask -> m () Source #
HasInitialize MkTransformerPaddingMask generatorDevice MkTransformerPaddingMask generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods initialize :: MonadThrow m => ModelSpec MkTransformerPaddingMask -> Generator generatorDevice -> m (MkTransformerPaddingMask, Generator generatorDevice) Source #
MkTransformerPaddingMaskC layout device dataType shape output => HasForward MkTransformerPaddingMask (Tensor gradient layout device dataType shape) generatorDevice output generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods forward :: MonadThrow m => MkTransformerPaddingMask -> Tensor gradient layout device dataType shape -> Generator generatorDevice -> m (output, Generator generatorDevice) Source #
type Rep MkTransformerPaddingMask Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Rep MkTransformerPaddingMask = D1 ('MetaData "MkTransformerPaddingMask" "Torch.GraduallyTyped.NN.Transformer.Type" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'True) (C1 ('MetaCons "MkTransformerPaddingMask" 'PrefixI 'True) (S1 ('MetaSel ('Just "padTokenId") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Int)))
type ModelSpec MkTransformerPaddingMask Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type ModelSpec MkTransformerPaddingMask = MkTransformerPaddingMask

type MkTransformerAttentionMaskC transformerDataType gradient layout device dataType shape seqDim output = (SGetLayout layout, SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), Catch (gradient <+> 'Gradient 'WithoutGradient), Catch (dataType <+> 'DataType 'Bool), Catch (UnsqueezeF ('SelectDim ('ByIndex 1)) shape), Catch (BroadcastShapesF (UnsqueezeF ('SelectDim ('ByIndex 1)) shape) ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim])), output ~ Tensor ('Gradient 'WithoutGradient) (layout <+> 'Layout 'Dense) device transformerDataType (BroadcastShapesF (UnsqueezeF ('SelectDim ('ByIndex 1)) shape) ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim]))) Source #

mkTransformerAttentionMask Source #

Arguments

:: forall m transformerDataType gradient layout device dataType shape seqDim output. (MonadThrow m, MkTransformerAttentionMaskC transformerDataType gradient layout device dataType shape seqDim output)
=> SDataType transformerDataType	data type singleton of the transformer
-> Double	attention mask bias (typically a large negative number)
-> Tensor gradient layout device dataType shape	encoder padding mask
-> m output

Creates a bidirectional attention mask for a transformer. Given a padding mask of shape [batchDim, seqDim], returns a tensor of shape [batchDim, seqDim, seqDim].

data MkTransformerAttentionMask (dataType :: DataType DType) where Source #

Constructors

MkTransformerAttentionMask
Fields :: forall dataType. { attentionMaskDataType :: SDataType dataType , attentionMaskBias :: Double } -> MkTransformerAttentionMask dataType

Instances

Instances details

Generic (MkTransformerAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Rep (MkTransformerAttentionMask dataType) :: Type -> Type Source # Methods from :: MkTransformerAttentionMask dataType -> Rep (MkTransformerAttentionMask dataType) x Source # to :: Rep (MkTransformerAttentionMask dataType) x -> MkTransformerAttentionMask dataType Source #
Show (MkTransformerAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods showsPrec :: Int -> MkTransformerAttentionMask dataType -> ShowS Source # show :: MkTransformerAttentionMask dataType -> String Source # showList :: [MkTransformerAttentionMask dataType] -> ShowS Source #
HasStateDict (MkTransformerAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec (MkTransformerAttentionMask dataType) -> StateDictKey -> m (MkTransformerAttentionMask dataType) Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> MkTransformerAttentionMask dataType -> m () Source #
HasInitialize (MkTransformerAttentionMask dataType) generatorDevice (MkTransformerAttentionMask dataType) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods initialize :: MonadThrow m => ModelSpec (MkTransformerAttentionMask dataType) -> Generator generatorDevice -> m (MkTransformerAttentionMask dataType, Generator generatorDevice) Source #
MkTransformerAttentionMaskC dataType inputGradient inputLayout inputDevice inputDataType inputShape seqDim output => HasForward (MkTransformerAttentionMask dataType) (Tensor inputGradient inputLayout inputDevice inputDataType inputShape) generatorDevice output generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods forward :: MonadThrow m => MkTransformerAttentionMask dataType -> Tensor inputGradient inputLayout inputDevice inputDataType inputShape -> Generator generatorDevice -> m (output, Generator generatorDevice) Source #
type Rep (MkTransformerAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Rep (MkTransformerAttentionMask dataType) = D1 ('MetaData "MkTransformerAttentionMask" "Torch.GraduallyTyped.NN.Transformer.Type" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "MkTransformerAttentionMask" 'PrefixI 'True) (S1 ('MetaSel ('Just "attentionMaskDataType") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 (SDataType dataType)) :*: S1 ('MetaSel ('Just "attentionMaskBias") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Double)))
type ModelSpec (MkTransformerAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type ModelSpec (MkTransformerAttentionMask dataType) = MkTransformerAttentionMask dataType

type MkTransformerDecoderAttentionMaskC transformerDataType layout device shape seqDim output = (SGetLayout layout, SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), Catch seqDim, Catch (UnsqueezeF ('SelectDim ('ByIndex 1)) shape), Catch (BroadcastShapesF ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim]) (UnsqueezeF ('SelectDim ('ByIndex 1)) shape)), Catch (BroadcastShapesF (BroadcastShapesF ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim]) (UnsqueezeF ('SelectDim ('ByIndex 1)) shape)) ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim])), output ~ Tensor ('Gradient 'WithoutGradient) (layout <+> 'Layout 'Dense) device transformerDataType (BroadcastShapesF (BroadcastShapesF ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim]) (UnsqueezeF ('SelectDim ('ByIndex 1)) shape)) ('Shape '['Dim ('Name "*") ('Size 1), seqDim, seqDim]))) Source #

mkTransformerDecoderAttentionMask Source #

Arguments

:: forall m transformerDataType gradient layout device dataType shape seqDim output. (MonadThrow m, MkTransformerDecoderAttentionMaskC transformerDataType layout device shape seqDim output)
=> SDataType transformerDataType	data type singleton of the transformer
-> Double	attention mask bias (typically a large negative number)
-> Tensor gradient layout device dataType shape	decoder padding mask
-> m output

Creates a causal attention mask for a transformer decoder. Given a padding mask of shape [batchDim, seqDim], returns a tensor of shape [batchDim, seqDim, seqDim].

data MkTransformerDecoderAttentionMask (dataType :: DataType DType) where Source #

Constructors

MkTransformerDecoderAttentionMask
Fields :: forall dataType. { decoderAttentionMaskDataType :: SDataType dataType , decoderAttentionMaskBias :: Double } -> MkTransformerDecoderAttentionMask dataType

Instances

Instances details

Generic (MkTransformerDecoderAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Rep (MkTransformerDecoderAttentionMask dataType) :: Type -> Type Source # Methods from :: MkTransformerDecoderAttentionMask dataType -> Rep (MkTransformerDecoderAttentionMask dataType) x Source # to :: Rep (MkTransformerDecoderAttentionMask dataType) x -> MkTransformerDecoderAttentionMask dataType Source #
Show (MkTransformerDecoderAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods showsPrec :: Int -> MkTransformerDecoderAttentionMask dataType -> ShowS Source # show :: MkTransformerDecoderAttentionMask dataType -> String Source # showList :: [MkTransformerDecoderAttentionMask dataType] -> ShowS Source #
HasStateDict (MkTransformerDecoderAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec (MkTransformerDecoderAttentionMask dataType) -> StateDictKey -> m (MkTransformerDecoderAttentionMask dataType) Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> MkTransformerDecoderAttentionMask dataType -> m () Source #
HasInitialize (MkTransformerDecoderAttentionMask dataType) generatorDevice (MkTransformerDecoderAttentionMask dataType) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods initialize :: MonadThrow m => ModelSpec (MkTransformerDecoderAttentionMask dataType) -> Generator generatorDevice -> m (MkTransformerDecoderAttentionMask dataType, Generator generatorDevice) Source #
MkTransformerDecoderAttentionMaskC dataType decoderInputLayout decoderInputDevice decoderInputShape seqDim output => HasForward (MkTransformerDecoderAttentionMask dataType) (Tensor decoderInputGradient decoderInputLayout decoderInputDevice decoderInputDataType decoderInputShape) generatorDevice output generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods forward :: MonadThrow m => MkTransformerDecoderAttentionMask dataType -> Tensor decoderInputGradient decoderInputLayout decoderInputDevice decoderInputDataType decoderInputShape -> Generator generatorDevice -> m (output, Generator generatorDevice) Source #
type Rep (MkTransformerDecoderAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Rep (MkTransformerDecoderAttentionMask dataType) = D1 ('MetaData "MkTransformerDecoderAttentionMask" "Torch.GraduallyTyped.NN.Transformer.Type" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "MkTransformerDecoderAttentionMask" 'PrefixI 'True) (S1 ('MetaSel ('Just "decoderAttentionMaskDataType") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 (SDataType dataType)) :*: S1 ('MetaSel ('Just "decoderAttentionMaskBias") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Double)))
type ModelSpec (MkTransformerDecoderAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type ModelSpec (MkTransformerDecoderAttentionMask dataType) = MkTransformerDecoderAttentionMask dataType

type MkTransformerCrossAttentionMaskC transformerDataType decoderInputShape decoderInputSeqDim gradient layout device dataType shape seqDim output = (SGetLayout layout, SGetDevice device, SGetShape shape, seqDim ~ (shape ! 1), SGetShape decoderInputShape, decoderInputSeqDim ~ (decoderInputShape ! 1), Catch (gradient <+> 'Gradient 'WithoutGradient), Catch (dataType <+> 'DataType 'Bool), Catch (UnsqueezeF ('SelectDim ('ByIndex 1)) shape), Catch (BroadcastShapesF (UnsqueezeF ('SelectDim ('ByIndex 1)) shape) ('Shape '['Dim ('Name "*") ('Size 1), decoderInputSeqDim, seqDim])), output ~ Tensor ('Gradient 'WithoutGradient) (layout <+> 'Layout 'Dense) device transformerDataType (BroadcastShapesF (UnsqueezeF ('SelectDim ('ByIndex 1)) shape) ('Shape '['Dim ('Name "*") ('Size 1), decoderInputSeqDim, seqDim]))) Source #

mkTransformerCrossAttentionMask Source #

Arguments

:: forall m transformerDataType decoderInputShape decoderInputSeqDim gradient layout device dataType shape seqDim output. (MonadThrow m, MkTransformerCrossAttentionMaskC transformerDataType decoderInputShape decoderInputSeqDim gradient layout device dataType shape seqDim output)
=> SDataType transformerDataType	data type singleton of the transformer
-> SShape decoderInputShape	decoder input shape
-> Double	attention mask bias (typically a large negative number)
-> Tensor gradient layout device dataType shape	encoder padding mask
-> m output

Creates a cross-attention mask for an encoder-decoder transformer. Given an encoder padding mask of shape [batchDim, seqDim], and the shape [batchDim, decoderSeqDim] of the decoder's input, returns a tensor of shape [batchDim, decoderSeqDim, seqDim].

data MkTransformerCrossAttentionMask (dataType :: DataType DType) where Source #

Constructors

MkTransformerCrossAttentionMask
Fields :: forall dataType. { crossAttentionMaskDataType :: SDataType dataType , crossAttentionMaskBias :: Double } -> MkTransformerCrossAttentionMask dataType

Instances

Instances details

Generic (MkTransformerCrossAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Rep (MkTransformerCrossAttentionMask dataType) :: Type -> Type Source # Methods from :: MkTransformerCrossAttentionMask dataType -> Rep (MkTransformerCrossAttentionMask dataType) x Source # to :: Rep (MkTransformerCrossAttentionMask dataType) x -> MkTransformerCrossAttentionMask dataType Source #
Show (MkTransformerCrossAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods showsPrec :: Int -> MkTransformerCrossAttentionMask dataType -> ShowS Source # show :: MkTransformerCrossAttentionMask dataType -> String Source # showList :: [MkTransformerCrossAttentionMask dataType] -> ShowS Source #
HasStateDict (MkTransformerCrossAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec (MkTransformerCrossAttentionMask dataType) -> StateDictKey -> m (MkTransformerCrossAttentionMask dataType) Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> MkTransformerCrossAttentionMask dataType -> m () Source #
HasInitialize (MkTransformerCrossAttentionMask dataType) generatorDevice (MkTransformerCrossAttentionMask dataType) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods initialize :: MonadThrow m => ModelSpec (MkTransformerCrossAttentionMask dataType) -> Generator generatorDevice -> m (MkTransformerCrossAttentionMask dataType, Generator generatorDevice) Source #
MkTransformerCrossAttentionMaskC dataType decoderInputShape decoderInputSeqDim inputPaddingMaskGradient inputPaddingMaskLayout inputPaddingMaskDevice inputPaddingMaksDataType inputPaddingMaskShape seqDim output => HasForward (MkTransformerCrossAttentionMask dataType) (Tensor decoderInputGradient decoderInputLayout decoderInputDevice decoderInputDataType decoderInputShape, Tensor inputPaddingMaskGradient inputPaddingMaskLayout inputPaddingMaskDevice inputPaddingMaksDataType inputPaddingMaskShape) generatorDevice output generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods forward :: MonadThrow m => MkTransformerCrossAttentionMask dataType -> (Tensor decoderInputGradient decoderInputLayout decoderInputDevice decoderInputDataType decoderInputShape, Tensor inputPaddingMaskGradient inputPaddingMaskLayout inputPaddingMaskDevice inputPaddingMaksDataType inputPaddingMaskShape) -> Generator generatorDevice -> m (output, Generator generatorDevice) Source #
type Rep (MkTransformerCrossAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Rep (MkTransformerCrossAttentionMask dataType) = D1 ('MetaData "MkTransformerCrossAttentionMask" "Torch.GraduallyTyped.NN.Transformer.Type" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "MkTransformerCrossAttentionMask" 'PrefixI 'True) (S1 ('MetaSel ('Just "crossAttentionMaskDataType") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 (SDataType dataType)) :*: S1 ('MetaSel ('Just "crossAttentionMaskBias") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Double)))
type ModelSpec (MkTransformerCrossAttentionMask dataType) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type ModelSpec (MkTransformerCrossAttentionMask dataType) = MkTransformerCrossAttentionMask dataType

data ShiftRight fillValue where Source #

Constructors

ShiftRight
Fields :: forall fillValue. fillValue fill value for shift right -> ShiftRight fillValue

Instances

Instances details

Generic (ShiftRight fillValue) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Associated Types type Rep (ShiftRight fillValue) :: Type -> Type Source # Methods from :: ShiftRight fillValue -> Rep (ShiftRight fillValue) x Source # to :: Rep (ShiftRight fillValue) x -> ShiftRight fillValue Source #
Show fillValue => Show (ShiftRight fillValue) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods showsPrec :: Int -> ShiftRight fillValue -> ShowS Source # show :: ShiftRight fillValue -> String Source # showList :: [ShiftRight fillValue] -> ShowS Source #
Eq fillValue => Eq (ShiftRight fillValue) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods (==) :: ShiftRight fillValue -> ShiftRight fillValue -> Bool Source # (/=) :: ShiftRight fillValue -> ShiftRight fillValue -> Bool Source #
Ord fillValue => Ord (ShiftRight fillValue) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods compare :: ShiftRight fillValue -> ShiftRight fillValue -> Ordering Source # (<) :: ShiftRight fillValue -> ShiftRight fillValue -> Bool Source # (<=) :: ShiftRight fillValue -> ShiftRight fillValue -> Bool Source # (>) :: ShiftRight fillValue -> ShiftRight fillValue -> Bool Source # (>=) :: ShiftRight fillValue -> ShiftRight fillValue -> Bool Source # max :: ShiftRight fillValue -> ShiftRight fillValue -> ShiftRight fillValue Source # min :: ShiftRight fillValue -> ShiftRight fillValue -> ShiftRight fillValue Source #
HasStateDict (ShiftRight fillValue) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods fromStateDict :: (MonadIO m, MonadThrow m, MonadState StateDict m) => ModelSpec (ShiftRight fillValue) -> StateDictKey -> m (ShiftRight fillValue) Source # toStateDict :: (MonadThrow m, MonadState StateDict m) => StateDictKey -> ShiftRight fillValue -> m () Source #
(input ~ Tensor inputGradient inputLayout inputDevice inputDataType inputShape, SGetLayout inputLayout, SGetDevice inputDevice, SGetDataType inputDataType, SGetShape inputShape, inputBatchDim ~ (inputShape ! 0), inputSeqDim ~ (inputShape ! 1), Scalar fillValue, rightShiftedInput ~ Tensor (inputGradient <\|> 'Gradient 'WithoutGradient) inputLayout inputDevice inputDataType (ReplaceDimF ('SelectDim ('ByIndex 1 :: By Symbol Natural)) (inputShape <+> 'Shape '[inputBatchDim, inputSeqDim]) (AddDimF inputSeqDim ('Dim ('Name "*") ('Size 1))))) => HasForward (ShiftRight fillValue) input generator rightShiftedInput generator Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods forward :: MonadThrow m => ShiftRight fillValue -> input -> Generator generator -> m (rightShiftedInput, Generator generator) Source #
HasInitialize (ShiftRight fillValue) generatorDevice (ShiftRight fillValue) generatorDevice Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type Methods initialize :: MonadThrow m => ModelSpec (ShiftRight fillValue) -> Generator generatorDevice -> m (ShiftRight fillValue, Generator generatorDevice) Source #
type Rep (ShiftRight fillValue) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type Rep (ShiftRight fillValue) = D1 ('MetaData "ShiftRight" "Torch.GraduallyTyped.NN.Transformer.Type" "hasktorch-gradually-typed-0.2.0.0-1KV1aIPzzbp6JpSr37tC1K" 'False) (C1 ('MetaCons "ShiftRight" 'PrefixI 'False) (S1 ('MetaSel ('Nothing :: Maybe Symbol) 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 fillValue)))
type ModelSpec (ShiftRight fillValue) Source #
Instance details Defined in Torch.GraduallyTyped.NN.Transformer.Type type ModelSpec (ShiftRight fillValue) = ShiftRight fillValue