Safe Haskell	Safe-Inferred
Language	Haskell2010

Torch.Typed.Optim

Synopsis

type LearningRate device dtype = Tensor device dtype '[]
type Loss device dtype = Tensor device dtype '[]
data ZerosLike = ZerosLike
class Optimizer optim gradients tensors dtype device where
- step :: LearningRate device dtype -> HList gradients -> HList tensors -> optim -> (HList tensors, optim)
runStep :: forall model optim parameters gradients tensors dtype device. (Parameterized model, parameters ~ Parameters model, HasGrad (HList parameters) (HList gradients), tensors ~ gradients, HMap' ToDependent parameters tensors, Castable (HList gradients) [ATenTensor], Optimizer optim gradients tensors dtype device, HMapM' IO MakeIndependent tensors parameters) => model -> optim -> Loss device dtype -> LearningRate device dtype -> IO (model, optim)
runStep' :: forall model optim parameters gradients tensors dtype device. (Parameterized model, parameters ~ Parameters model, tensors ~ gradients, HMap' ToDependent parameters tensors, Optimizer optim gradients tensors dtype device, HMapM' IO MakeIndependent tensors parameters) => model -> optim -> LearningRate device dtype -> HList gradients -> IO (model, optim)
data GD = GD
mkGD :: GD
newtype GDStep device dtype = GDStep (LearningRate device dtype)
gd :: forall gradients tensors dtype device. HZipWith (GDStep device dtype) tensors gradients tensors => LearningRate device dtype -> HList gradients -> HList tensors -> GD -> (HList tensors, GD)
data GDM (momenta :: [Type]) = GDM {
- beta :: Float
- momenta :: HList momenta
}
mkGDM :: forall parameters momenta. HMap' ZerosLike parameters momenta => Float -> HList parameters -> GDM momenta
data GDMStep device dtype = GDMStep Float (LearningRate device dtype)
gdm :: forall gradients tensors momenta gdmStep dtype device. (HZipWith3 (GDMStep device dtype) tensors gradients momenta gdmStep, HMap' AFst gdmStep tensors, HMap' ASnd gdmStep momenta) => LearningRate device dtype -> HList gradients -> HList tensors -> GDM momenta -> (HList tensors, GDM momenta)
type AdamIter = Tensor '('CPU, 0) 'Int64 '[]
data Adam (momenta :: [Type]) = Adam {
- iter :: AdamIter
- beta1 :: Float
- beta2 :: Float
- momenta1 :: HList momenta
- momenta2 :: HList momenta
}
mkAdam :: forall parameters momenta. HMap' ZerosLike parameters momenta => AdamIter -> Float -> Float -> HList parameters -> Adam momenta
newtype AdamMomentum1Update = AdamMomentum1Update Float
newtype AdamMomentum2Update = AdamMomentum2Update Float
data AdamBiasAdjustment = AdamBiasAdjustment AdamIter Float
data AdamParameterUpdate device dtype = AdamParameterUpdate Float (LearningRate device dtype)
adam :: forall gradients tensors momenta adamStep dtype device. (HZipWith AdamMomentum1Update momenta gradients momenta, HZipWith AdamMomentum2Update momenta gradients momenta, HMap' AdamBiasAdjustment momenta momenta, HZipWith3 (AdamParameterUpdate device dtype) tensors momenta momenta tensors) => LearningRate device dtype -> HList gradients -> HList tensors -> Adam momenta -> (HList tensors, Adam momenta)

Documentation

type LearningRate device dtype = Tensor device dtype '[] Source #

type Loss device dtype = Tensor device dtype '[] Source #

data ZerosLike Source #

Constructors

ZerosLike

Instances

Instances details

(parameter ~ Parameter device dtype shape, momentum ~ Tensor device dtype shape, TensorOptions shape dtype device) => Apply' ZerosLike parameter momentum Source #
Instance details Defined in Torch.Typed.Optim Methods apply' :: ZerosLike -> parameter -> momentum Source #

class Optimizer optim gradients tensors dtype device where Source #

Methods

step :: LearningRate device dtype -> HList gradients -> HList tensors -> optim -> (HList tensors, optim) Source #

Instances

Instances details

HZipWith (GDStep device dtype) tensors gradients tensors => Optimizer GD (gradients :: [k]) (tensors :: [k]) dtype device Source #
Instance details Defined in Torch.Typed.Optim Methods step :: LearningRate device dtype -> HList gradients -> HList tensors -> GD -> (HList tensors, GD) Source #
(HZipWith AdamMomentum1Update momenta gradients momenta, HZipWith AdamMomentum2Update momenta gradients momenta, HMap' AdamBiasAdjustment momenta momenta, HZipWith3 (AdamParameterUpdate device dtype) tensors momenta momenta tensors) => Optimizer (Adam momenta) (gradients :: [Type]) (tensors :: [Type]) dtype device Source #
Instance details Defined in Torch.Typed.Optim Methods step :: LearningRate device dtype -> HList gradients -> HList tensors -> Adam momenta -> (HList tensors, Adam momenta) Source #
(HZipWith3 (GDMStep device dtype) tensors gradients momenta gdmStep, HMap' AFst gdmStep tensors, HMap' ASnd gdmStep momenta) => Optimizer (GDM momenta) (gradients :: [Type]) (tensors :: [Type]) dtype device Source #
Instance details Defined in Torch.Typed.Optim Methods step :: LearningRate device dtype -> HList gradients -> HList tensors -> GDM momenta -> (HList tensors, GDM momenta) Source #

runStep :: forall model optim parameters gradients tensors dtype device. (Parameterized model, parameters ~ Parameters model, HasGrad (HList parameters) (HList gradients), tensors ~ gradients, HMap' ToDependent parameters tensors, Castable (HList gradients) [ATenTensor], Optimizer optim gradients tensors dtype device, HMapM' IO MakeIndependent tensors parameters) => model -> optim -> Loss device dtype -> LearningRate device dtype -> IO (model, optim) Source #

runStep' :: forall model optim parameters gradients tensors dtype device. (Parameterized model, parameters ~ Parameters model, tensors ~ gradients, HMap' ToDependent parameters tensors, Optimizer optim gradients tensors dtype device, HMapM' IO MakeIndependent tensors parameters) => model -> optim -> LearningRate device dtype -> HList gradients -> IO (model, optim) Source #

data GD Source #

Dummy state representation for GD Optimizer

Constructors

GD

Instances

Instances details

Parameterized GD Source #
Instance details Defined in Torch.Typed.Optim Associated Types type Parameters GD :: [Type] Source # Methods flattenParameters :: GD -> HList (Parameters GD) Source # replaceParameters :: GD -> HList (Parameters GD) -> GD Source #
HZipWith (GDStep device dtype) tensors gradients tensors => Optimizer GD (gradients :: [k]) (tensors :: [k]) dtype device Source #
Instance details Defined in Torch.Typed.Optim Methods step :: LearningRate device dtype -> HList gradients -> HList tensors -> GD -> (HList tensors, GD) Source #
type Parameters GD Source #
Instance details Defined in Torch.Typed.Optim type Parameters GD = '[] :: [Type]

mkGD :: GD Source #

newtype GDStep device dtype Source #

Constructors

GDStep (LearningRate device dtype)

Instances

Instances details

(parameter ~ Tensor device dtype shape, gradient ~ Tensor device dtype shape, shape ~ Broadcast ('[] :: [Nat]) shape, BasicArithmeticDTypeIsValid device dtype, KnownDevice device) => Apply' (GDStep device dtype) (parameter, gradient) parameter Source #
Instance details Defined in Torch.Typed.Optim Methods apply' :: GDStep device dtype -> (parameter, gradient) -> parameter Source #

gd :: forall gradients tensors dtype device. HZipWith (GDStep device dtype) tensors gradients tensors => LearningRate device dtype -> HList gradients -> HList tensors -> GD -> (HList tensors, GD) Source #

Gradient descent step with a dummy state variable

data GDM (momenta :: [Type]) Source #

State representation for GDM Optimizer

Constructors

GDM
Fields beta :: Float momenta :: HList momenta

Instances

Instances details

(HZipWith3 (GDMStep device dtype) tensors gradients momenta gdmStep, HMap' AFst gdmStep tensors, HMap' ASnd gdmStep momenta) => Optimizer (GDM momenta) (gradients :: [Type]) (tensors :: [Type]) dtype device Source #
Instance details Defined in Torch.Typed.Optim Methods step :: LearningRate device dtype -> HList gradients -> HList tensors -> GDM momenta -> (HList tensors, GDM momenta) Source #
Parameterized (GDM momenta) Source #
Instance details Defined in Torch.Typed.Optim Associated Types type Parameters (GDM momenta) :: [Type] Source # Methods flattenParameters :: GDM momenta -> HList (Parameters (GDM momenta)) Source # replaceParameters :: GDM momenta -> HList (Parameters (GDM momenta)) -> GDM momenta Source #
type Parameters (GDM momenta) Source #
Instance details Defined in Torch.Typed.Optim type Parameters (GDM momenta) = momenta

mkGDM :: forall parameters momenta. HMap' ZerosLike parameters momenta => Float -> HList parameters -> GDM momenta Source #

data GDMStep device dtype Source #

Constructors

GDMStep Float (LearningRate device dtype)

Instances

Instances details

(parameter ~ Tensor device dtype shape, gradient ~ Tensor device dtype shape, momentum ~ Tensor device dtype shape, shape ~ Broadcast ('[] :: [Nat]) shape, KnownDevice device, BasicArithmeticDTypeIsValid device dtype) => Apply' (GDMStep device dtype) (parameter, gradient, momentum) (parameter, momentum) Source #
Instance details Defined in Torch.Typed.Optim Methods apply' :: GDMStep device dtype -> (parameter, gradient, momentum) -> (parameter, momentum) Source #

gdm Source #

Arguments

:: forall gradients tensors momenta gdmStep dtype device. (HZipWith3 (GDMStep device dtype) tensors gradients momenta gdmStep, HMap' AFst gdmStep tensors, HMap' ASnd gdmStep momenta)
=> LearningRate device dtype	learning rate
-> HList gradients	model parameter gradient tensors
-> HList tensors	model parameter tensors
-> GDM momenta	beta and model parameter momentum tensors
-> (HList tensors, GDM momenta)	returns updated parameters and momenta

gradient descent with momentum step

type AdamIter = Tensor '('CPU, 0) 'Int64 '[] Source #

data Adam (momenta :: [Type]) Source #

State representation for Adam Optimizer

Constructors

Adam
Fields iter :: AdamIter beta1 :: Float beta2 :: Float momenta1 :: HList momenta momenta2 :: HList momenta

Instances

Instances details

(HZipWith AdamMomentum1Update momenta gradients momenta, HZipWith AdamMomentum2Update momenta gradients momenta, HMap' AdamBiasAdjustment momenta momenta, HZipWith3 (AdamParameterUpdate device dtype) tensors momenta momenta tensors) => Optimizer (Adam momenta) (gradients :: [Type]) (tensors :: [Type]) dtype device Source #
Instance details Defined in Torch.Typed.Optim Methods step :: LearningRate device dtype -> HList gradients -> HList tensors -> Adam momenta -> (HList tensors, Adam momenta) Source #
HAppendFD momenta momenta (momenta ++ momenta) => Parameterized (Adam momenta) Source #
Instance details Defined in Torch.Typed.Optim Associated Types type Parameters (Adam momenta) :: [Type] Source # Methods flattenParameters :: Adam momenta -> HList (Parameters (Adam momenta)) Source # replaceParameters :: Adam momenta -> HList (Parameters (Adam momenta)) -> Adam momenta Source #
type Parameters (Adam momenta) Source #
Instance details Defined in Torch.Typed.Optim type Parameters (Adam momenta) = AdamIter ': (momenta ++ momenta)

mkAdam :: forall parameters momenta. HMap' ZerosLike parameters momenta => AdamIter -> Float -> Float -> HList parameters -> Adam momenta Source #

newtype AdamMomentum1Update Source #

Constructors

AdamMomentum1Update Float

Instances

Instances details

(gradient ~ Tensor device dtype shape, momentum1 ~ Tensor device dtype shape, KnownDevice device) => Apply' AdamMomentum1Update (momentum1, gradient) momentum1 Source #	decaying average of the first momenta
Instance details Defined in Torch.Typed.Optim Methods apply' :: AdamMomentum1Update -> (momentum1, gradient) -> momentum1 Source #

newtype AdamMomentum2Update Source #

Constructors

AdamMomentum2Update Float

Instances

Instances details

(gradient ~ Tensor device dtype shape, momentum2 ~ Tensor device dtype shape, shape ~ Broadcast shape shape, KnownDevice device, BasicArithmeticDTypeIsValid device dtype) => Apply' AdamMomentum2Update (momentum2, gradient) momentum2 Source #	decaying average of the second momenta
Instance details Defined in Torch.Typed.Optim Methods apply' :: AdamMomentum2Update -> (momentum2, gradient) -> momentum2 Source #

data AdamBiasAdjustment Source #

Constructors

AdamBiasAdjustment AdamIter Float

Instances

Instances details

(momentum ~ Tensor device dtype shape, KnownDevice device, KnownDType dtype, shape ~ Reverse (Reverse shape), BasicArithmeticDTypeIsValid device dtype) => Apply' AdamBiasAdjustment momentum momentum Source #	bias adjustment
Instance details Defined in Torch.Typed.Optim Methods apply' :: AdamBiasAdjustment -> momentum -> momentum Source #

data AdamParameterUpdate device dtype Source #

Constructors

AdamParameterUpdate Float (LearningRate device dtype)

Instances

Instances details

(parameter ~ Tensor device dtype shape, momentum ~ Tensor device dtype shape, shape ~ Broadcast ('[] :: [Nat]) shape, KnownDevice device, BasicArithmeticDTypeIsValid device dtype, StandardFloatingPointDTypeValidation device dtype) => Apply' (AdamParameterUpdate device dtype) (parameter, momentum, momentum) parameter Source #	parameter update
Instance details Defined in Torch.Typed.Optim Methods apply' :: AdamParameterUpdate device dtype -> (parameter, momentum, momentum) -> parameter Source #

adam Source #

Arguments

:: forall gradients tensors momenta adamStep dtype device. (HZipWith AdamMomentum1Update momenta gradients momenta, HZipWith AdamMomentum2Update momenta gradients momenta, HMap' AdamBiasAdjustment momenta momenta, HZipWith3 (AdamParameterUpdate device dtype) tensors momenta momenta tensors)
=> LearningRate device dtype	learning rate
-> HList gradients	model parameter gradient tensors
-> HList tensors	model parameter tensors
-> Adam momenta	adam parameters - beta1, beta2, momenta1, momenta2, iteration
-> (HList tensors, Adam momenta)	returns new parameters + updated adam parameters

Adam step