C# Class Seq2SeqSharp.Tools.BaseSeq2SeqFramework

This is a framework for neural network training. It includes many core parts, such as backward propagation, parameters updates, memory management, computing graph managment, corpus shuffle & batching, I/O for model, logging & monitoring, checkpoints. You need to create your network inherited from this class, implmenet forward part only and pass it to TrainOneEpoch method for training

Datei anzeigen Open project: SciSharp/Seq2SeqSharp

Public Methods

Method	Description
BaseSeq2SeqFramework ( Array deviceIds, ProcessorTypeEnums processorType, string modelFilePath, float memoryUsageRatio = 0.9f, Array compilerOptions = null ) : AdvUtils
CreateComputGraph ( int deviceIdIdx, bool needBack = true ) : IComputeGraph
LoadModel ( Func InitializeParameters ) : IModelMetaData	Load model from given file
RunNetwork ( Func ForwardOnSingleDevice, List sntPairBatchs, int batchSplitFactor ) : AdvUtils
SaveModel ( IModelMetaData modelMetaData ) : bool

Private Methods

Method	Description
CopyWeightsFromDefaultDeviceToAllOtherDevices ( ) : void	Copy weights from default device to all other devices
CreateCheckPoint ( IEnumerable validCorpus, List metrics, IModelMetaData modelMetaData, Func ForwardOnSingleDevice, double avgCostPerWordInTotal ) : void
GetParametersFromDefaultDevice ( ) : List
LoadParameters ( Stream stream ) : void
Register ( object childValue, string name ) : void
RegisterTrainableParameters ( object obj ) : void
RunTest ( List inputTokens, Func ForwardOnSingleDevice ) : List>
RunValid ( IEnumerable validCorpus, Func RunNetwork, List metrics, bool outputToFile = false ) : bool	Evaluate the quality of model on valid corpus.
RunValidParallel ( Func RunNetwork, List metrics, bool outputToFile, List srcSents, List refSents, List hypSents, List sntPairBatchs ) : void
SaveParameters ( Stream stream ) : void
SumGradientsToTensorsInDefaultDevice ( ) : void	Sum up gradients in all devices and keep them in the default device
TrainOneEpoch ( int ep, IEnumerable trainCorpus, IEnumerable validCorpus, ILearningRate learningRate, AdamOptimizer solver, List metrics, IModelMetaData modelMetaData, Func ForwardOnSingleDevice ) : void
ZeroGradientOnAllDevices ( ) : void

Method Details

BaseSeq2SeqFramework() public method

public BaseSeq2SeqFramework ( Array deviceIds, ProcessorTypeEnums processorType, string modelFilePath, float memoryUsageRatio = 0.9f, Array compilerOptions = null ) : AdvUtils
deviceIds	Array
processorType	ProcessorTypeEnums
modelFilePath	string
memoryUsageRatio	float
compilerOptions	Array
return	AdvUtils

CreateComputGraph() public method

public CreateComputGraph ( int deviceIdIdx, bool needBack = true ) : IComputeGraph
deviceIdIdx	int
needBack	bool
return	IComputeGraph

LoadModel() public method

Load model from given file

public LoadModel ( Func InitializeParameters ) : IModelMetaData
InitializeParameters	Func
return	IModelMetaData

RunNetwork() public method

public RunNetwork ( Func ForwardOnSingleDevice, List sntPairBatchs, int batchSplitFactor ) : AdvUtils
ForwardOnSingleDevice	Func
sntPairBatchs	List
batchSplitFactor	int
return	AdvUtils

SaveModel() public method

public SaveModel ( IModelMetaData modelMetaData ) : bool
modelMetaData	IModelMetaData
return	bool