MPSRNNDescriptor(3) MetalPerformanceShaders.framework MPSRNNDescriptor(3)

MPSRNNDescriptor

#import <MPSRNNLayer.h>

Inherits NSObject.

Inherited by MPSGRUDescriptor, MPSLSTMDescriptor, and MPSRNNSingleGateDescriptor.


NSUInteger inputFeatureChannels
NSUInteger outputFeatureChannels
BOOL useLayerInputUnitTransformMode
BOOL useFloat32Weights
MPSRNNSequenceDirection layerSequenceDirection

This depends on Metal.framework The MPSRNNDescriptor specifies a Recursive neural network block/layer descriptor.

- inputFeatureChannels [read], [write], [nonatomic], [assign]

The number of feature channels per pixel in the input image or number of rows in the input matrix.

- layerSequenceDirection [read], [write], [nonatomic], [assign]

When the layer specified with this descriptor is used to process a sequence of inputs by calling

See also:

encodeBidirectionalSequenceToCommandBuffer then this parameter defines in which direction the sequence is processed. The operation of the layer is: (yt, ht, ct) = f(xt,ht-1,ct-1) for MPSRNNSequenceDirectionForward and (yt, ht, ct) = f(xt,ht+1,ct+1) for MPSRNNSequenceDirectionBackward, where xt is the output of the previous layer that encodes in the same direction as this layer, (or the input image or matrix if this is the first layer in stack with this direction).

MPSRNNImageInferenceLayer and

MPSRNNMatrixInferenceLayer.

- outputFeatureChannels [read], [write], [nonatomic], [assign]

The number of feature channels per pixel in the destination image or number of rows in the destination matrix.

- useFloat32Weights [read], [write], [nonatomic], [assign]

If YES, then MPSRNNMatrixInferenceLayer uses 32-bit floating point numbers internally for weights when computing matrix transformations. If NO, then 16-bit, half precision floating point numbers are used. Currently MPSRNNImageInferenceLayer ignores this property and the convolution operations always convert FP32 weights into FP16 for better performance. Defaults to NO.

- useLayerInputUnitTransformMode [read], [write], [nonatomic], [assign]

if YES then use identity transformation for all weights (W, Wr, Wi, Wf, Wo, Wc) affecting input x_j in this layer, even if said weights are specified as nil. For example 'W_ij * x_j' is replaced by 'x_j' in formulae defined in MPSRNNSingleGateDescriptor. Defaults to NO.

Generated automatically by Doxygen for MetalPerformanceShaders.framework from the source code.

Mon Jul 9 2018 Version MetalPerformanceShaders-119.3