tensorflow-ops-0.3.0.0: Friendly layer around TensorFlow bindings.

Safe Haskell	None
Language	Haskell2010

TensorFlow.Minimize

Synopsis

type Minimizer a = forall m. MonadBuild m => [Variable a] -> [Tensor Value a] -> m ControlNode
minimizeWith :: (MonadBuild m, GradientCompatible a) => Minimizer a -> Tensor v a -> [Variable a] -> m ControlNode
gradientDescent :: GradientCompatible a => a -> Minimizer a
type OneOfAdamDataTypes t = OneOf '[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word32, Word64, Word8, Double, Float] t
data AdamConfig t = AdamConfig {
- adamLearningRate :: t
- adamBeta1 :: t
- adamBeta2 :: t
- adamEpsilon :: t
}
adam :: (OneOfAdamDataTypes t, Fractional t) => Minimizer t
adam' :: OneOfAdamDataTypes t => AdamConfig t -> Minimizer t

Documentation

type Minimizer a = forall m. MonadBuild m => [Variable a] -> [Tensor Value a] -> m ControlNode Source #

Functions that minimize a loss w.r.t. a set of Variables.

Generally only performs one step of an iterative algorithm.

Minimizers are defined as a function of the gradients instead of the loss so that users can apply transformations to the gradients.

minimizeWith Source #

Arguments

:: (MonadBuild m, GradientCompatible a)
=> Minimizer a
-> Tensor v a	Loss.
-> [Variable a]	Parameters of the loss function.
-> m ControlNode

Convenience wrapper around gradients and a Minimizer.

gradientDescent Source #

Arguments

:: GradientCompatible a
=> a	Learning rate.
-> Minimizer a

Perform one step of the gradient descent algorithm.

type OneOfAdamDataTypes t = OneOf '[Complex Double, Complex Float, Int16, Int32, Int64, Int8, Word16, Word32, Word64, Word8, Double, Float] t Source #

data AdamConfig t Source #

Constructors

AdamConfig
Fields adamLearningRate :: t adamBeta1 :: t adamBeta2 :: t adamEpsilon :: t

Instances

Instances details

Fractional t => Default (AdamConfig t) Source #
Instance details Defined in TensorFlow.Minimize Methods def :: AdamConfig t

adam :: (OneOfAdamDataTypes t, Fractional t) => Minimizer t Source #

Perform one step of the adam algorithm.

See https://arxiv.org/abs/1412.6980.

NOTE: Currently requires all Variables to have an initializedValue.

adam' :: OneOfAdamDataTypes t => AdamConfig t -> Minimizer t Source #