mpscnnconvolutiondatasource-p(3) [mojave man page]
<MPSCNNConvolutionDataSource>(3) MetalPerformanceShaders.framework <MPSCNNConvolutionDataSource>(3) NAME
<MPSCNNConvolutionDataSource> SYNOPSIS
#include <MPSCNNConvolution.h> Detailed Description Provides convolution filter weights and bias terms The MPSCNNConvolutionDataSource protocol declares the methods that an instance of MPSCNNConvolution uses to obtain the weights and bias terms for the CNN convolution filter. Why? CNN weights can be large. If multiple copies of all the weights for all the convolutions are available unpacked in memory at the same time, some devices can run out of memory. The MPSCNNConvolutionDataSource is used to encapsulate a reference to the weights such as a file path, so that unpacking can be deferred until needed, then purged soon thereafter so that not all of the data must be in memory at the same time. MPS does not provide a class that conforms to this protocol. It is up to the developer to craft his own to encapsulate his data. Batch normalization and the neuron activation function are handled using the -descriptor method. Thread safety: The MPSCNNConvolutionDataSource object can be called by threads that are not the main thread. If you will be creating multiple MPSNNGraph objects concurrently in multiple threads and these share MPSCNNConvolutionDataSources, then the data source objects may be called reentrantly. Author Generated automatically by Doxygen for MetalPerformanceShaders.framework from the source code. Version MetalPerformanceShaders-100 Thu Feb 8 2018 <MPSCNNConvolutionDataSource>(3)
Check Out this Related Man Page
MPSCNNConvolutionGradientNode(3) MetalPerformanceShaders.framework MPSCNNConvolutionGradientNode(3) NAME
MPSCNNConvolutionGradientNode SYNOPSIS
#import <MPSNNGraphNodes.h> Inherits MPSNNGradientFilterNode, and <MPSNNTrainableNode>. Instance Methods (nonnull instancetype) - initWithSourceGradient:sourceImage:convolutionGradientState:weights: Class Methods (nonnull instancetype) + nodeWithSourceGradient:sourceImage:convolutionGradientState:weights: Additional Inherited Members Method Documentation - (nonnull instancetype) initWithSourceGradient: (MPSNNImageNode *__nonnull) sourceGradient(MPSNNImageNode *__nonnull) sourceImage(MPSCNNConvolutionGradientStateNode *__nonnull) gradientState(nullable id< MPSCNNConvolutionDataSource >) weights A node to represent the gradient calculation for convolution training. Parameters: sourceGradient The input gradient from the 'downstream' gradient filter. Often that is a neuron gradient filter node. sourceImage The input image from the forward convolution node gradientState The gradient state from the forward convolution weights The data source from the forward convolution. It may not contain an integrated neuron. Similary, any normalization should be broken out into a separate node. Pass nil to use the weights from the forward convolution pass. Returns: A MPSCNNConvolutionGradientNode + (nonnull instancetype) nodeWithSourceGradient: (MPSNNImageNode *__nonnull) sourceGradient(MPSNNImageNode *__nonnull) sourceImage(MPSCNNConvolutionGradientStateNode *__nonnull) gradientState(nullable id< MPSCNNConvolutionDataSource >) weights A node to represent the gradient calculation for convolution training. Parameters: sourceGradient The input gradient from the 'downstream' gradient filter. Often that is a neuron gradient filter node. sourceImage The input image from the forward convolution node gradientState The gradient state from the forward convolution weights The data source from the forward convolution. It may not contain an integrated neuron. Similary, any normalization should be broken out into a separate node. Pass nil to use the weights from the forward convolution pass. Returns: A MPSCNNConvolutionGradientNode Author Generated automatically by Doxygen for MetalPerformanceShaders.framework from the source code. Version MetalPerformanceShaders-100 Thu Feb 8 2018 MPSCNNConvolutionGradientNode(3)