Data.NeuralNetwork

This module defines an abstract interface for neural network and a protocol for its backends to follow.

Documentation

class Component a where

Abstraction of a neural network component

Minimal complete definition

forwardT, output, backward

Associated Types

type Run a :: * -> *

execution environment

type Inp a

the type of input and in-error

type Out a

the type of output and out-error

data Trace a

the trace of a forward propagation

Methods

forwardT :: a -> Inp a -> Run a (Trace a)

Forward propagation

forward :: Applicative (Run a) => a -> Inp a -> Run a (Out a)

Forward propagation

output :: Trace a -> Out a

extract the output value from the trace

backward :: a -> Trace a -> Out a -> Float -> Run a (a, Inp a)

Backward propagation

learn

Arguments

:: (Component n, Monad (Run n))

=> (Out n -> Out n -> Run n (Out n))

derivative of the error function

-> Float

learning rate

-> n

neuron network

-> (Inp n, Out n)

input and expect output

-> Run n n

updated network

By giving a way to measure the error, learn can update the neural network component.

relu :: (Num a, Ord a) => a -> a

default RELU and derivative of RELU

relu' :: (Num a, Ord a) => a -> a

default RELU and derivative of RELU

cost' :: (Num a, Ord a) => a -> a -> a

default derivative of error measurement

class Backend b s where

Abstraction of backend to carry out the specification

Minimal complete definition

witness, compile

Associated Types

type Env b :: * -> *

environment to compile the specification

type ConvertFromSpec s :: *

result type of compile

Methods

witness :: b -> s -> Dict (Monad (Env b), Monad (Run (ConvertFromSpec s)), Component (ConvertFromSpec s), RunInEnv (Run (ConvertFromSpec s)) (Env b))

necessary constraints of the resulting type

compile :: b -> s -> Env b (ConvertFromSpec s)

compile the specification to runnable component.

class (Monad r, Monad e) => RunInEnv r e where

Lifting from one monad to another. It is not necessary that the Env and Run maps to the same execution environment, but the Run one should be able to be lifted to Env one.

Minimal complete definition

run

Methods

run :: r a -> e a

data a :++ b infixr 0

Specification: stacking layer

Constructors

a :++ b infixr 0

data SpecIn1D

Specification: 1D input

Constructors

In1D Int

dimension of input

data SpecIn2D

Specification: 2D input

Constructors

In2D Int Int

dimension of input

data SpecReshape2DAs1D

Specification: reshaping layer

Constructors

Reshape2DAs1D

data SpecFullConnect

Specification: full connection layer

Constructors

FullConnect Int

number of neurals

data SpecConvolution

Specification: convolution layer

Constructors

Convolution Int Int Int

number of output channels, size of kernel, size of padding

data SpecMaxPooling

Specification: max pooling layer

Constructors

MaxPooling Int