Safe Haskell	None
Language	GHC2024

Net

Contents

Parameter bundle
Layer primitives
Model
Loss
Boundary type

Description

Forward-only neural-network layer definitions built on the circuits ecosystem.

NetParams stores weights as Matrix and biases as Array vectors.
Linear layers use matVec.
The full model is a Para composition, so inference reuses the same parameter threading as the rest of circuits-learn.

Synopsis

data NetParams a = NetParams {
- w1 :: Matrix a
- b1 :: Array a
- w2 :: Matrix a
- b2 :: Array a
}
netParamsFromArrays :: Array a -> Array a -> Array a -> Array a -> NetParams a
linear1 :: (Additive a, Multiplicative a) => Para (NetParams a) (Array a) (Array a)
bias1 :: Num a => Para (NetParams a) (Array a) (Array a)
relu1 :: (Ord a, Additive a) => Para (NetParams a) (Array a) (Array a)
linear2 :: (Additive a, Multiplicative a) => Para (NetParams a) (Array a) (Array a)
bias2 :: Num a => Para (NetParams a) (Array a) (Array a)
model :: (Ord a, Num a, Additive a, Multiplicative a) => Para (NetParams a) (Array a) (Array a)
forward :: (Ord a, Num a, Additive a, Multiplicative a) => NetParams a -> Array a -> Array a
mseLoss :: (Fractional a, Additive a) => Array a -> Array a -> (a, Array a)
type Boundary a = These a a

Parameter bundle

data NetParams a Source #

Network parameters. Weights are stored as dense matrices so that Dense can be used for the linear maps; biases and activations remain as Array vectors.

Constructors

NetParams
Fields w1 :: Matrix a b1 :: Array a w2 :: Matrix a b2 :: Array a

Instances

Instances details

Eq a => Eq (NetParams a) Source #
Instance details Defined in Net Methods (==) :: NetParams a -> NetParams a -> Bool # (/=) :: NetParams a -> NetParams a -> Bool #
Show a => Show (NetParams a) Source #
Instance details Defined in Net Methods showsPrec :: Int -> NetParams a -> ShowS # show :: NetParams a -> String # showList :: [NetParams a] -> ShowS #

netParamsFromArrays :: Array a -> Array a -> Array a -> Array a -> NetParams a Source #

Build NetParams from plain Array weights and biases.

Layer primitives

linear1 :: (Additive a, Multiplicative a) => Para (NetParams a) (Array a) (Array a) Source #

First linear layer.

bias1 :: Num a => Para (NetParams a) (Array a) (Array a) Source #

First bias layer.

relu1 :: (Ord a, Additive a) => Para (NetParams a) (Array a) (Array a) Source #

ReLU activation.

linear2 :: (Additive a, Multiplicative a) => Para (NetParams a) (Array a) (Array a) Source #

Second linear layer.

bias2 :: Num a => Para (NetParams a) (Array a) (Array a) Source #

Second bias layer.

Model

model :: (Ord a, Num a, Additive a, Multiplicative a) => Para (NetParams a) (Array a) (Array a) Source #

Full 2-layer MLP: linear1 → bias1 → relu → linear2 → bias2.

forward :: (Ord a, Num a, Additive a, Multiplicative a) => NetParams a -> Array a -> Array a Source #

Run the model with explicit parameters.

Loss

mseLoss :: (Fractional a, Additive a) => Array a -> Array a -> (a, Array a) Source #

Mean-squared-error loss and its gradient w.r.t. the prediction.

L = (1/n) Σ (y - target)², dLdy = (2n)(y - target).

Boundary type

type Boundary a = These a a Source #

A These-based boundary distinguishes inference-only, gradient-only, and combined inference-with-gradient traffic at a layer boundary.

This prediction — inference only, no gradient.
That gradient — training signal only, no prediction.
These prediction gradient — training with prediction (common supervised case).