Safe Haskell	None
Language	GHC2024

MnistCnnRanked2

Description

Ranked tensor-based implementation of Convolutional Neural Network for classification of MNIST digits. Sports 2 hidden layers.

With the current CPU backend it's slow enough that it's hard to see if it trains.

Synopsis

type ADCnnMnistParametersShaped (target :: Target) (h :: Natural) (w :: Natural) (kh :: Natural) (kw :: Natural) (c_out :: Natural) (n_hidden :: Natural) r = ((target (TKS '[c_out, 1, kh + 1, kw + 1] r), target (TKS '[c_out] r)), (target (TKS '[c_out, c_out, kh + 1, kw + 1] r), target (TKS '[c_out] r)), (target (TKS '[n_hidden, (c_out * Div h 4) * Div w 4] r), target (TKS '[n_hidden] r)), (target (TKS '[SizeMnistLabel, n_hidden] r), target (TKS '[SizeMnistLabel] r)))
type ADCnnMnistParameters (target :: Target) r = ((target (TKR 4 r), target (TKR 1 r)), (target (TKR 4 r), target (TKR 1 r)), (target (TKR 2 r), target (TKR 1 r)), (target (TKR 2 r), target (TKR 1 r)))
convMnistLayerR :: (ADReady target, GoodScalar r, Differentiable r) => target (TKR 4 r) -> target (TKR 4 r) -> target (TKR 1 r) -> target (TKR 4 r)
convMnistTwoR :: (ADReady target, GoodScalar r, Differentiable r) => Int -> Int -> Int -> PrimalOf target (TKR 4 r) -> ADCnnMnistParameters target r -> target (TKR 2 r)
convMnistLossFusedR :: (ADReady target, ADReady (PrimalOf target), GoodScalar r, Differentiable r) => Int -> (PrimalOf target (TKR 3 r), PrimalOf target (TKR 2 r)) -> ADCnnMnistParameters target r -> target ('TKScalar r)
convMnistTestR :: forall (target :: TK -> Type) r. (target ~ Concrete, GoodScalar r, Differentiable r) => Int -> MnistDataBatchR r -> ADCnnMnistParameters Concrete r -> r

Documentation

type ADCnnMnistParametersShaped (target :: Target) (h :: Natural) (w :: Natural) (kh :: Natural) (kw :: Natural) (c_out :: Natural) (n_hidden :: Natural) r = ((target (TKS '[c_out, 1, kh + 1, kw + 1] r), target (TKS '[c_out] r)), (target (TKS '[c_out, c_out, kh + 1, kw + 1] r), target (TKS '[c_out] r)), (target (TKS '[n_hidden, (c_out * Div h 4) * Div w 4] r), target (TKS '[n_hidden] r)), (target (TKS '[SizeMnistLabel, n_hidden] r), target (TKS '[SizeMnistLabel] r))) Source #

The differentiable type of all trainable parameters of this nn. Shaped version, statically checking all dimension widths.

Due to subtraction complicating posititive number type inference, kh denotes kernel height minus one and analogously kw is kernel width minus one.

type ADCnnMnistParameters (target :: Target) r = ((target (TKR 4 r), target (TKR 1 r)), (target (TKR 4 r), target (TKR 1 r)), (target (TKR 2 r), target (TKR 1 r)), (target (TKR 2 r), target (TKR 1 r))) Source #

The differentiable type of all trainable parameters of this nn. Ranked version.

convMnistLayerR Source #

Arguments

:: (ADReady target, GoodScalar r, Differentiable r)
=> target (TKR 4 r)	[c_out, c_in, kh + 1, kw + 1]
-> target (TKR 4 r)	[batch_size, c_in, h, w]
-> target (TKR 1 r)	[c_out]
-> target (TKR 4 r)	[batch_size, c_out, h `Div` 2, w `Div` 2]

A single convolutional layer with relu and maxPool.

convMnistTwoR Source #

Arguments

:: (ADReady target, GoodScalar r, Differentiable r)
=> Int
-> Int
-> Int
-> PrimalOf target (TKR 4 r)	input images `[batch_size, 1, SizeMnistHeight, SizeMnistWidth]`
-> ADCnnMnistParameters target r	parameters
-> target (TKR 2 r)	output classification `[SizeMnistLabel, batch_size]`

Composition of two convolutional layers.

convMnistLossFusedR Source #

Arguments

:: (ADReady target, ADReady (PrimalOf target), GoodScalar r, Differentiable r)
=> Int	batch_size
-> (PrimalOf target (TKR 3 r), PrimalOf target (TKR 2 r))	[batch_size, SizeMnistLabel]
-> ADCnnMnistParameters target r
-> target ('TKScalar r)

The neural network composed with the SoftMax-CrossEntropy loss function.

convMnistTestR :: forall (target :: TK -> Type) r. (target ~ Concrete, GoodScalar r, Differentiable r) => Int -> MnistDataBatchR r -> ADCnnMnistParameters Concrete r -> r Source #

A function testing the neural network given testing set of inputs and the trained parameters.