MyCaffe.common.IXPersist< T > Interface Template Reference

The IXPersist interface is used by the CaffeControl to load and save weights. More...

Inheritance diagram for MyCaffe.common.IXPersist< T >:

Public Member Functions
byte[]	SaveWeights (BlobCollection< T > colBlobs, bool bSaveDiffs=false)
	Save the weights to a byte array. More...
BlobCollection< T >	LoadWeights (byte[] rgWeights, List< string > rgExpectedShapes, BlobCollection< T > colBlobs, bool bSizeToFit, out bool bLoadedDiffs, List< string > inputWtInfo=null, List< string > targetWtInfo=null, string strSkipBlobType=null)
	Loads new weights into a BlobCollection More...
byte[]	SaveSolverState (SolverState state, SolverParameter.SolverType type=SolverParameter.SolverType.SGD)
	Save the solver state to a byte array. More...
SolverState	LoadSolverState (byte[] rgState, SolverParameter.SolverType type=SolverParameter.SolverType.SGD)
	Load the solver state from a byte array. More...
WeightInfo< T >	LoadWeightInfo (byte[] rgWeights)
	Returns the weight information describing the weights containined within the weight bytes. More...
WeightInfo< T >	LoadWeightInfo (BlobCollection< T > colBlobs)
	Returns the weight information describing the weights containined within the Blob collection. More...

Detailed Description

The IXPersist interface is used by the CaffeControl to load and save weights.

Template Parameters

T	Specifies the base type float or double. Using float is recommended to conserve GPU memory.

Definition at line 186 of file Interfaces.cs.

Member Function Documentation

LoadSolverState()

SolverState MyCaffe.common.IXPersist< T >.LoadSolverState	(	byte[]	rgState,
		SolverParameter.SolverType	type = SolverParameter.SolverType.SGD
	)

Load the solver state from a byte array.

Parameters

rgState	Specifies the byte array containing the solver state.
type	Specifies the solver type.

Returns

The SolverState loaded is returned.

Implemented in MyCaffe.common.PersistCaffe< T >.

LoadWeightInfo() [1/2]

WeightInfo< T > MyCaffe.common.IXPersist< T >.LoadWeightInfo

(

BlobCollection< T >

colBlobs

)

Returns the weight information describing the weights containined within the Blob collection.

Parameters

colBlobs

Specifies the Blob collection containing the weights.

Returns

The weight information is returned.

Implemented in MyCaffe.common.PersistCaffe< T >.

LoadWeightInfo() [2/2]

WeightInfo< T > MyCaffe.common.IXPersist< T >.LoadWeightInfo

(

byte[]

rgWeights

)

Returns the weight information describing the weights containined within the weight bytes.

Parameters

rgWeights

Specifies the bytes containing the weights.

Returns

The weight information is returned.

Implemented in MyCaffe.common.PersistCaffe< T >.

LoadWeights()

BlobCollection< T > MyCaffe.common.IXPersist< T >.LoadWeights	(	byte[]	rgWeights,
		List< string >	rgExpectedShapes,
		BlobCollection< T >	colBlobs,
		bool	bSizeToFit,
		out bool	bLoadedDiffs,
		List< string >	inputWtInfo = null,
		List< string >	targetWtInfo = null,
		string	strSkipBlobType = null
	)

Loads new weights into a BlobCollection

NOTE: In order to maintain compatibility with the C++ Caffe, extra MyCaffe features may be added to the end of the weight file. After saving weights (see SaveWeights) in the format used by the C++ Caffe, MyCaffe writes the bytes "mycaffe.ai". All information after these bytes are specific to MyCaffe and allow for loading weights for models by Blob name and shape and loosen the C++ Caffe requirement that the 'number' of blobs match. Adding this functionality allows for training model, changing the model structure, and then re-using the trained weights in the new model.

Parameters

rgWeights	Specifies the weights themselves.
rgExpectedShapes	Specifies a list of expected shapes for each Blob where the weights are to be loaded.
colBlobs	Specifies the Blobs to load with the weights.
bSizeToFit	Optionally, specifies wether or not the weights should be re-sized. Note: resizing can render the weights useless, especially in deeper, layers.
bLoadedDiffs	Returns whether or not the diffs were loaded.
inputWtInfo	Optionally, specifies the weight info describing the input weight blobs to import by name. Note when used the number of blobs must match the number of targetWtInfo blobs. Otherwise, when null this parameter is ignored.
targetWtInfo	Optionally, specifies the weight info describing the target weight blobs to import by name. Note when used the number of blobs must match the number of inputWtInfo blobs. Otherwise, when null this parameter is ignored.
strSkipBlobType	Optionally, specifies a blob type where weights are NOT loaded. See Blob.BLOB_TYPE for the types of Blobs.

Returns

The collection of Blobs with newly loaded weights is returned.

Implemented in MyCaffe.common.PersistCaffe< T >.

SaveSolverState()

byte[] MyCaffe.common.IXPersist< T >.SaveSolverState	(	SolverState	state,
		SolverParameter.SolverType	type = SolverParameter.SolverType.SGD
	)

Save the solver state to a byte array.

Parameters

state	Specifies the solver state to save.
type	Specifies the solver type.

Returns

A byte array containing the solver state is returned.

Implemented in MyCaffe.common.PersistCaffe< T >.

SaveWeights()

byte[] MyCaffe.common.IXPersist< T >.SaveWeights	(	BlobCollection< T >	colBlobs,
		bool	bSaveDiffs = false
	)

Save the weights to a byte array.

NOTE: In order to maintain compatibility with the C++ Caffe, extra MyCaffe features may be added to the end of the weight file. After saving weights in the format used by the C++ Caffe, MyCaffe writes the bytes "mycaffe.ai". All information after these bytes are specific to MyCaffe and allow for loading weights for models by Blob name and shape and loosen the C++ Caffe requirement that the 'number' of blobs match. Adding this functionality allows for training model, changing the model structure, and then re-using the trained weights in the new model.

Parameters

colBlobs	Specifies the Blobs to save with the weights.
bSaveDiffs	Optionally, specifies to save the diff values.

Returns

The byte array containing the weights is returned.

Implemented in MyCaffe.common.PersistCaffe< T >.

---

The documentation for this interface was generated from the following file:

• C:/Data/Data/SS_Projects/Intelligence/GitHub/MyCaffe/MyCaffe/common/Interfaces.cs