Numeric.Neural.Pipes

This modules provides a "pipes"-based API for working with models.

Documentation

data TS f g a b c

The training state of a model.

Constructors

TS

Fields

tsModel :: Model f g a b c

updated model

tsGeneration :: Int

generation

tsEta :: Double

learning rate

tsBatchError :: Double

last training error

descentP

Arguments

:: (Foldable h, Monad m)

=> Model f g a b c

initial model

-> Int

first generation

-> (Int -> Double)

computes the learning rate from the generation

-> Pipe (h a) (TS f g a b c) m r

A Pipe for training a model: It consumes mini-batches of samples from upstream and pushes the updated training state downstream.

simpleBatchP

Arguments

:: MonadRandom m

=> [a]

all available samples

-> Int

the mini-batch size

-> Producer [a] m r

A simple Producer of mini-batches.

reportTSP

Arguments

:: Monad m

=> Int

report interval

-> (TS f g a b c -> m ())

report action

-> Pipe (TS f g a b c) (TS f g a b c) m r

A Pipe for progress reporting of model training.

consumeTSP

Arguments

:: Monad m

=> (TS f g a b c -> m (Maybe x))

check whether training is finished and what to return in that case

-> Consumer (TS f g a b c) m x

A Consumer of training states that decides when training is finished and then returns a value.