Safe Haskell	Safe-Inferred
Language	Haskell2010

Synapse.NN.Training

Contents

Callbacks datatype and associated type aliases
Hyperparameters datatype
Training

Description

This module provides datatypes and functions that implement neural networks training.

Synopsis

type CallbackFnOnTrainBegin model optimizer a = IORef (model a) -> IORef [OptimizerParameters optimizer a] -> IO ()
type CallbackFnOnEpochBegin model optimizer a = IORef Int -> IORef (model a) -> IORef [OptimizerParameters optimizer a] -> IORef (BatchedDataset a) -> IO ()
type CallbackFnOnBatchBegin model optimizer a = IORef Int -> IORef Int -> IORef (model a) -> IORef [OptimizerParameters optimizer a] -> IORef (Sample (Mat a)) -> IORef a -> IO ()
type CallbackFnOnBatchEnd model optimizer a = IORef Int -> IORef Int -> IORef (model a) -> IORef [OptimizerParameters optimizer a] -> IORef (Vec a) -> IO ()
type CallbackFnOnEpochEnd model optimizer a = IORef Int -> IORef (model a) -> IORef [OptimizerParameters optimizer a] -> IO ()
type CallbackFnOnTrainEnd model optimizer a = IORef (model a) -> IORef [OptimizerParameters optimizer a] -> IORef (Vec (RecordedMetric a)) -> IO ()
data Callbacks model optimizer a = Callbacks {
- callbacksOnTrainBegin :: [CallbackFnOnTrainBegin model optimizer a]
- callbacksOnEpochBegin :: [CallbackFnOnEpochBegin model optimizer a]
- callbacksOnBatchBegin :: [CallbackFnOnBatchBegin model optimizer a]
- callbacksOnBatchEnd :: [CallbackFnOnBatchEnd model optimizer a]
- callbacksOnEpochEnd :: [CallbackFnOnEpochEnd model optimizer a]
- callbacksOnTrainEnd :: [CallbackFnOnTrainEnd model optimizer a]
}
emptyCallbacks :: Callbacks model optimizer a
data Hyperparameters a = Hyperparameters {
- hyperparametersEpochs :: Int
- hyperparametersBatchSize :: Int
- hyperparametersDataset :: VecDataset a
- hyperparametersLearningRate :: LearningRate a
- hyperparametersLoss :: Loss a
- hyperparametersMetrics :: Vec (Metric a)
}
newtype RecordedMetric a = RecordedMetric {
- unRecordedMetric :: Vec a
}
train :: (Symbolic a, Floating a, Ord a, Show a, RandomGen g, AbstractLayer model, Optimizer optimizer) => model a -> optimizer a -> Hyperparameters a -> Callbacks model optimizer a -> g -> IO (model a, [OptimizerParameters optimizer a], Vec (RecordedMetric a), g)

`Callbacks` datatype and associated type aliases

type CallbackFnOnTrainBegin model optimizer a Source #

Arguments

= IORef (model a)	Initial model state.
-> IORef [OptimizerParameters optimizer a]	Initial optimizer parameters.
-> IO ()

Type of callback that is called at the beginning of training.

type CallbackFnOnEpochBegin model optimizer a Source #

Arguments

= IORef Int	Current epoch.
-> IORef (model a)	Model state at the beginning of the epoch processing.
-> IORef [OptimizerParameters optimizer a]	Optimizer parameters at the beginning of the epoch processing.
-> IORef (BatchedDataset a)	Batched shuffled dataset.
-> IO ()

Type of callback that is called at the beginning of training epoch.

type CallbackFnOnBatchBegin model optimizer a Source #

Arguments

= IORef Int	Current epoch.
-> IORef Int	Current batch number.
-> IORef (model a)	Model state at the beginning of the batch processing.
-> IORef [OptimizerParameters optimizer a]	Optimizer parameters at the beginning of the batch processing.
-> IORef (Sample (Mat a))	Batch that is being processed.
-> IORef a	Learning rate value.
-> IO ()

Type of callback that is called at the beginning of training batch.

type CallbackFnOnBatchEnd model optimizer a Source #

Arguments

= IORef Int	Current epoch.
-> IORef Int	Current batch number.
-> IORef (model a)	Model state at the end of the batch processing.
-> IORef [OptimizerParameters optimizer a]	Optimizer parameters at the end of the batch processing.
-> IORef (Vec a)	Metrics that were recorded on this batch.
-> IO ()

Type of callback that is called at the end of training batch.

type CallbackFnOnEpochEnd model optimizer a Source #

Arguments

= IORef Int	Current epoch.
-> IORef (model a)	Model state at the end of the epoch processing.
-> IORef [OptimizerParameters optimizer a]	Optimizer parameters at the end of the epoch processing.
-> IO ()

Type of callback that is called at the end of training epoch.

type CallbackFnOnTrainEnd model optimizer a Source #

Arguments

= IORef (model a)	Model state at the end of the training.
-> IORef [OptimizerParameters optimizer a]	Optimizer parameters at the end of the training.
-> IORef (Vec (RecordedMetric a))	Recorded metrics.
-> IO ()

Type of callback that is called at the end of training.

data Callbacks model optimizer a Source #

Callbacks record datatype holds all callbacks for the training.

All callbacks take IORefs to various training parameters, which allows to affect training in any way possible.

This interface should be used with caution, because some changes might break the training completely.

Constructors

Callbacks

Fields

callbacksOnTrainBegin :: [CallbackFnOnTrainBegin model optimizer a]
Callbacks that will be called at the beginning of training.
callbacksOnEpochBegin :: [CallbackFnOnEpochBegin model optimizer a]
Callbacks that will be called at the beginning of training epoch processing.
callbacksOnBatchBegin :: [CallbackFnOnBatchBegin model optimizer a]
Callbacks that will be called at the beginning of training batch processing.
callbacksOnBatchEnd :: [CallbackFnOnBatchEnd model optimizer a]
Callbacks that will be called at the end of training batch processing.
callbacksOnEpochEnd :: [CallbackFnOnEpochEnd model optimizer a]
Callbacks that will be called at the end of training epoch processing.
callbacksOnTrainEnd :: [CallbackFnOnTrainEnd model optimizer a]
Callbacks that will be called at the end of training.

emptyCallbacks :: Callbacks model optimizer a Source #

Returns empty Callbacks record. It could also be used to build your own callbacks upon.

`Hyperparameters` datatype

data Hyperparameters a Source #

Hyperparameters datatype represents configuration of a training.

Constructors

Hyperparameters

Fields

hyperparametersEpochs :: Int
Number of epochs in the training.
hyperparametersBatchSize :: Int
Size of batches that will be used in the training.
hyperparametersDataset :: VecDataset a
Dataset with samples of vector functions.
hyperparametersLearningRate :: LearningRate a
LearningRate that will be used in the training.
hyperparametersLoss :: Loss a
Loss that will be used in the training.
hyperparametersMetrics :: Vec (Metric a)
Metrics that will be recorded during training.

newtype RecordedMetric a Source #

RecordedMetric newtype wraps vector of results of metrics.

Constructors

RecordedMetric
Fields unRecordedMetric :: Vec a Results of metric recording.

Training

train Source #

Arguments

:: (Symbolic a, Floating a, Ord a, Show a, RandomGen g, AbstractLayer model, Optimizer optimizer)
=> model a	Trained model.
-> optimizer a	Optimizer that will be during training.
-> Hyperparameters a	Hyperparameters of training.
-> Callbacks model optimizer a	Callbacks that will be used during training.
-> g	Generator of random values that will be used to shuffle dataset.
-> IO (model a, [OptimizerParameters optimizer a], Vec (RecordedMetric a), g)	Updated model, optimizer parameters at the end of training, vector of recorded metrics (loss is also recorded and is the first in vector), updated generator of random values.

train function allows training neural networks on datasets with specified parameters.

Callbacks datatype and associated type aliases

Hyperparameters datatype

Training

`Callbacks` datatype and associated type aliases

`Hyperparameters` datatype