| Safe Haskell | Safe-Inferred |
|---|---|
| Language | Haskell2010 |
Torch.GraduallyTyped.NN.Transformer.RoBERTa.Common
Synopsis
- type RoBERTaDType = 'Float
- robertaDType :: SDType RoBERTaDType
- type RoBERTaDataType = 'DataType RoBERTaDType
- robertaDataType :: SDataType RoBERTaDataType
- robertaDropoutP :: Double
- type RoBERTaPosEncDim = 'Dim ('Name "*") ('Size 514)
- robertaPosEncDim :: SDim RoBERTaPosEncDim
- robertaEps :: Double
- robertaMaxPositionEmbeddings :: Int
- robertaPadTokenId :: Int
- robertaBOSTokenId :: Int
- robertaEOSTokenId :: Int
- robertaAttentionMaskBias :: Double
- type family RoBERTaModelF (transformerHead :: TransformerHead) (numLayers :: Nat) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (headDim :: Dim (Name Symbol) (Size Nat)) (headEmbedDim :: Dim (Name Symbol) (Size Nat)) (embedDim :: Dim (Name Symbol) (Size Nat)) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) (ffnDim :: Dim (Name Symbol) (Size Nat)) (vocabDim :: Dim (Name Symbol) (Size Nat)) (typeVocabDim :: Dim (Name Symbol) (Size Nat)) (hasDropout :: HasDropout) :: Type where ...
- robertaModelSpec :: forall transformerHead numLayers gradient device headDim headEmbedDim embedDim inputEmbedDim ffnDim vocabDim typeVocabDim hasDropout. (SingI headDim, SingI headEmbedDim, SingI embedDim, SingI inputEmbedDim, SingI ffnDim, SingI vocabDim, SingI typeVocabDim) => STransformerHead transformerHead -> SNat numLayers -> SGradient gradient -> SDevice device -> SHasDropout hasDropout -> ModelSpec (RoBERTaModelF transformerHead numLayers gradient device headDim headEmbedDim embedDim inputEmbedDim ffnDim vocabDim typeVocabDim hasDropout)
- mkRoBERTaInput :: 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])) => SDim batchDim -> SDim seqDim -> SDevice device -> [[Int]] -> m output
Documentation
type RoBERTaDType = 'Float Source #
RoBERTa dType.
robertaDType :: SDType RoBERTaDType Source #
RoBERTa dType singleton.
type RoBERTaDataType = 'DataType RoBERTaDType Source #
RoBERTa data type.
robertaDataType :: SDataType RoBERTaDataType Source #
RoBERTa data type singleton.
robertaDropoutP :: Double Source #
RoBERTa dropout rate. 'dropout_rate = 0.1'
type RoBERTaPosEncDim = 'Dim ('Name "*") ('Size 514) Source #
RoBERTa positional encoding dimension.
Note the two extra dimensions.
robertaPosEncDim :: SDim RoBERTaPosEncDim Source #
RoBERTa positional encoding dimension singleton.
robertaEps :: Double Source #
RoBERTa layer-norm epsilon. 'layer_norm_epsilon = 1e-5'
robertaMaxPositionEmbeddings :: Int Source #
RoBERTa maximum number of position embeddings. 'max_position_embeddings = 514'
robertaPadTokenId :: Int Source #
RoBERTa padding token id. 'pad_token_id = 1'
robertaBOSTokenId :: Int Source #
RoBERTa begin-of-sentence token id. 'bos_token_id = 0'
robertaEOSTokenId :: Int Source #
RoBERTa end-of-sentence token id. 'eos_token_id = 0'
robertaAttentionMaskBias :: Double Source #
RoBERTa attention mask bias
type family RoBERTaModelF (transformerHead :: TransformerHead) (numLayers :: Nat) (gradient :: Gradient RequiresGradient) (device :: Device (DeviceType Nat)) (headDim :: Dim (Name Symbol) (Size Nat)) (headEmbedDim :: Dim (Name Symbol) (Size Nat)) (embedDim :: Dim (Name Symbol) (Size Nat)) (inputEmbedDim :: Dim (Name Symbol) (Size Nat)) (ffnDim :: Dim (Name Symbol) (Size Nat)) (vocabDim :: Dim (Name Symbol) (Size Nat)) (typeVocabDim :: Dim (Name Symbol) (Size Nat)) (hasDropout :: HasDropout) :: Type where ... Source #
Specifies the RoBERTa model.
Equations
| RoBERTaModelF transformerHead numLayers gradient device headDim headEmbedDim embedDim inputEmbedDim ffnDim vocabDim typeVocabDim hasDropout = GSimplifiedEncoderOnlyTransformer (GEncoderOnlyTransformerF 'RoBERTa transformerHead numLayers gradient device RoBERTaDataType headDim headEmbedDim embedDim inputEmbedDim ffnDim RoBERTaPosEncDim vocabDim typeVocabDim hasDropout) MkAbsPos MkTransformerPaddingMask (MkTransformerAttentionMask RoBERTaDataType) |
robertaModelSpec :: forall transformerHead numLayers gradient device headDim headEmbedDim embedDim inputEmbedDim ffnDim vocabDim typeVocabDim hasDropout. (SingI headDim, SingI headEmbedDim, SingI embedDim, SingI inputEmbedDim, SingI ffnDim, SingI vocabDim, SingI typeVocabDim) => STransformerHead transformerHead -> SNat numLayers -> SGradient gradient -> SDevice device -> SHasDropout hasDropout -> ModelSpec (RoBERTaModelF transformerHead numLayers gradient device headDim headEmbedDim embedDim inputEmbedDim ffnDim vocabDim typeVocabDim hasDropout) Source #
Specifies the parameters of a RoBERTa model.
transformerHead: the head of the RoBERTa model.numLayers: the number of layers in the RoBERTa model.gradient: whether to compute the gradient of the RoBERTa model.device: the computational device on which the RoBERTa model parameters are to be allocated.
mkRoBERTaInput :: 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])) => SDim batchDim -> SDim seqDim -> SDevice device -> [[Int]] -> m output Source #