MPSCNNConvolutionDescriptor(3)
NAME
MPSCNNConvolutionDescriptor
SYNOPSIS
#import <MPSCNNConvolution.h>
Inherits NSObject, <NSSecureCoding>, and <NSCopying>.
Inherited by MPSCNNDepthWiseConvolutionDescriptor, and
MPSCNNSubPixelConvolutionDescriptor.
Instance Methods
(void) - encodeWithCoder:
(nullable instancetype) - initWithCoder:
(void) -
setBatchNormalizationParametersForInferenceWithMean:variance:gamma:beta:epsilon:
(void) - setNeuronType:parameterA:parameterB:
(MPSCNNNeuronType) - neuronType
(float) - neuronParameterA
(float) - neuronParameterB
(void) - setNeuronPReLUParametersA:
Class Methods
(nonnull instancetype) +
cnnConvolutionDescriptorWithKernelWidth:kernelHeight:inputFeatureChannels:outputFeatureChannels:neuronFilter:
(nonnull instancetype) +
cnnConvolutionDescriptorWithKernelWidth:kernelHeight:inputFeatureChannels:outputFeatureChannels:
Properties
NSUInteger kernelWidth
NSUInteger kernelHeight
NSUInteger inputFeatureChannels
NSUInteger outputFeatureChannels
NSUInteger strideInPixelsX
NSUInteger strideInPixelsY
NSUInteger groups
NSUInteger dilationRateX
NSUInteger dilationRateY
const MPSCNNNeuron *__nullable neuron
const MPSCNNNeuron *__nullable BOOL supportsSecureCoding
Detailed Description
This depends on Metal.framework The MPSCNNConvolutionDescriptor
specifies a convolution descriptor
Method Documentation
+ (nonnull instancetype) cnnConvolutionDescriptorWithKernelWidth:
(NSUInteger) kernelWidth(NSUInteger) kernelHeight(NSUInteger)
inputFeatureChannels(NSUInteger) outputFeatureChannels
Creates a convolution descriptor.
Parameters:
kernelWidth The width of the filter window. Must be > 0. Large
values will take a long time.
kernelHeight The height of the filter window. Must be > 0. Large
values will take a long time.
inputFeatureChannels The number of feature channels in the input
image. Must be >= 1.
outputFeatureChannels The number of feature channels in the output
image. Must be >= 1.
Returns:
A valid MPSCNNConvolutionDescriptor object or nil, if failure.
+ (nonnull instancetype) cnnConvolutionDescriptorWithKernelWidth:
(NSUInteger) kernelWidth(NSUInteger) kernelHeight(NSUInteger)
inputFeatureChannels(NSUInteger) outputFeatureChannels(const
MPSCNNNeuron *__nullable) neuronFilter
This method is deprecated. Please use neuronType, neuronParameterA and
neuronParameterB properites to fuse neuron with convolution.
Parameters:
kernelWidth The width of the filter window. Must be > 0. Large
values will take a long time.
kernelHeight The height of the filter window. Must be > 0. Large
values will take a long time.
inputFeatureChannels The number of feature channels in the input
image. Must be >= 1.
outputFeatureChannels The number of feature channels in the output
image. Must be >= 1.
neuronFilter An optional neuron filter that can be applied to the
output of convolution.
Returns:
A valid MPSCNNConvolutionDescriptor object or nil, if failure.
- (void) encodeWithCoder: (NSCoder *__nonnull) aCoder
<NSSecureCoding> support
- (nullable instancetype) initWithCoder: (NSCoder *__nonnull) aDecoder
<NSSecureCoding> support
- (float) neuronParameterA
Getter funtion for neuronType set using
setNeuronType:parameterA:parameterB method
- (float) neuronParameterB
Getter funtion for neuronType set using
setNeuronType:parameterA:parameterB method
- (MPSCNNNeuronType) neuronType
Getter funtion for neuronType set using
setNeuronType:parameterA:parameterB method
- (void) setBatchNormalizationParametersForInferenceWithMean: (const float
*__nonnull) mean(const float *__nonnull) variance(const float
*__nullable) gamma(const float *__nullable) beta(const float) epsilon
Adds batch normalization for inference, it copies all the float arrays
provided, expecting outputFeatureChannels elements in each.
This method will be used to pass in batch normalization parameters to
the convolution during the init call. For inference we modify weights
and bias going in convolution or Fully Connected layer to combine and
optimize the layers.
w: weights for a corresponding output feature channel
b: bias for a corresponding output feature channel
W: batch normalized weights for a corresponding output feature channel
B: batch normalized bias for a corresponding output feature channel
I = gamma / sqrt(variance + epsilon), J = beta - ( I * mean )
W = w * I
B = b * I + J
Every convolution has (OutputFeatureChannel * kernelWidth * kernelHeight * InputFeatureChannel) weights
I, J are calculated, for every output feature channel separately to get the corresponding weights and bias
Thus, I, J are calculated and then used for every (kernelWidth * kernelHeight * InputFeatureChannel)
weights, and this is done OutputFeatureChannel number of times for each output channel.
thus, internally, batch normalized weights are computed as:
W[no][i][j][ni] = w[no][i][j][ni] * I[no]
no: index into outputFeatureChannel
i : index into kernel Height
j : index into kernel Width
ni: index into inputFeatureChannel
One usually doesn't see a bias term and batch normalization together as batch normalization potentially cancels
out the bias term after training, but in MPS if the user provides it, batch normalization will use the above
formula to incorporate it, if user does not have bias terms then put a float array of zeroes in the convolution
init for bias terms of each output feature channel.
this comes from:
https://arxiv.org/pdf/1502.03167v3.pdf
Parameters:
mean Pointer to an array of floats of mean for each output feature
channel
variance Pointer to an array of floats of variance for each output
feature channel
gamma Pointer to an array of floats of gamma for each output
feature channel
beta Pointer to an array of floats of beta for each output feature
channel
epsilon A small float value used to have numerical stability in the
code
- (void) setNeuronToPReLUWithParametersA: (NSData *__nonnull) A
Add per-channel neuron parameters A for PReLu neuron activation
functions.
This method can be used to set per-channel neuron parameters A for
PReLU neuron functions that dictate unique value of this parameter for
each output feature channel If convolution preceeds this kind of neuron
/ activation function, setting these parameters here has the
performance advantage of merging the neuron with convolution,
eliminating a pass. If the neuron function is f(v,a,b), it will apply
OutputImage(x,y,i) = f( ConvolutionResult(x,y,i), A[i], B[i] ) where i in [0,outputFeatureChannels-1]
See https://arxiv.org/pdf/1502.01852.pdf for details.
All other neuron types, where parameter A and parameter B are shared
across channels must be set using
setNeuronOfType:parameterA:parameterB. Its an error to call this
function on any neuronType other than MPSCNNNeuronTypePReLU.
If batch normalization parameters are set, batch normalization will
preceed neuron application i.e. output of convolution is first batch
normalized followed by neuron activation. This function automatically
sets neuronType to MPSCNNNeuronTypePReLU.
Parameters:
A Array containing per-channel float values for neuron parameter A.
Number of entries must be equal to outputFeatureChannels.
- (void) setNeuronType: (MPSCNNNeuronType) neuronType(float)
parameterA(float) parameterB
Adds a neuron activation function to convolution descriptor.
This mathod can be used to add a neuron activation funtion of given
type with associated scalar parameters A and B that are shared across
all output channels. Neuron activation fucntion is applied to output of
convolution. This is a per-pixel operation that is fused with
convolution kernel itself for best performance. Note that this method
can only be used to fuse neuron of kind for which parameters A and B
are shared across all channels of convoution output. It is an error to
call this method for neuron activation functions like
MPSCNNNeuronTypePReLU, which require per-channel parameter values. For
those kind of neuron activation functions, use appropriate setter
functions.
Parameters:
neuronType type of neuron activation function. For full list see
MPSCNNNeuronType.h
parameterA parameterA of neuron activation that is shared across
all channels of convolution output.
parameterB parameterB of neuron activation that is shared across
all channels of convolution output.
Property Documentation
- dilationRateX [read], [write], [nonatomic], [assign]
dilationRateX property can be used to implement dilated convolution as
described in https://arxiv.org/pdf/1511.07122v3.pdf to aggregate global
information in dense prediction problems. Default value is 1. When set
to value > 1, original kernel width, kW is dilated to
kW_Dilated = (kW-1)*dilationRateX + 1
by inserting d-1 zeros between consecutive entries in each row of the
original kernel. The kernel is centered based on kW_Dilated.
- dilationRateY [read], [write], [nonatomic], [assign]
dilationRateY property can be used to implement dilated convolution as
described in https://arxiv.org/pdf/1511.07122v3.pdf to aggregate global
information in dense prediction problems. Default value is 1. When set
to value > 1, original kernel height, kH is dilated to
kH_Dilated = (kH-1)*dilationRateY + 1
by inserting d-1 rows of zeros between consecutive row of the original
kernel. The kernel is centered based on kH_Dilated.
- groups [read], [write], [nonatomic], [assign]
Number of groups input and output channels are divided into. The
default value is 1. Groups lets you reduce the parameterization. If
groups is set to n, input is divided into n groups with
inputFeatureChannels/n channels in each group. Similarly output is
divided into n groups with outputFeatureChannels/n channels in each
group. ith group in input is only connected to ith group in output so
number of weights (parameters) needed is reduced by factor of n. Both
inputFeatureChannels and outputFeatureChannels must be divisible by n
and number of channels in each group must be multiple of 4.
- inputFeatureChannels [read], [write], [nonatomic], [assign]
The number of feature channels per pixel in the input image.
- kernelHeight [read], [write], [nonatomic], [assign]
The height of the filter window. The default value is 3. Any positive
non-zero value is valid, including even values. The position of the top
edge of the filter window is given by offset.y - (kernelHeight>>1)
- kernelWidth [read], [write], [nonatomic], [assign]
The width of the filter window. The default value is 3. Any positive
non-zero value is valid, including even values. The position of the
left edge of the filter window is given by offset.x - (kernelWidth>>1)
- neuron [read], [write], [nonatomic], [retain]
MPSCNNNeuron filter to be applied as part of convolution. This is
applied after BatchNormalization in the end. Default is nil. This is
deprecated. You dont need to create MPSCNNNeuron object to fuse with
convolution. Use neuron properties in this descriptor.
- outputFeatureChannels [read], [write], [nonatomic], [assign]
The number of feature channels per pixel in the output image.
- strideInPixelsX [read], [write], [nonatomic], [assign]
The output stride (downsampling factor) in the x dimension. The default
value is 1.
- strideInPixelsY [read], [write], [nonatomic], [assign]
The output stride (downsampling factor) in the y dimension. The default
value is 1.
- (const MPSCNNNeuron* __nullable BOOL) supportsSecureCoding [read],
[nonatomic], [assign]
<NSSecureCoding> support
Author
Generated automatically by Doxygen for
MetalPerformanceShaders.framework from the source code.
Version MetalPerformanceShaders-Thu2Jul 13 2017 MPSCNNConvolutionDescriptor(3)
Mac OS X 10.13.1 - Generated Mon Nov 6 16:23:49 CST 2017
