MPSCNNLossLabels(3) MetalPerformanceShaders.framework MPSCNNLossLabels(3)
NAME
MPSCNNLossLabels
SYNOPSIS
#import <MPSCNNLoss.h>
Inherits MPSState.
Instance Methods
(nonnull instancetype) - init
(nonnull instancetype) - initWithDevice:labelsDescriptor:
(nonnull instancetype) - initWithDevice:lossImageSize:labelsDescriptor:weightsDescriptor:
(nonnull MPSImage *) - lossImage
Additional Inherited Members
Detailed Description
This depends on Metal.framework. The MPSCNNLossLabels is used to hold the per-element weights buffer used by both MPSCNNLoss forward
filter and MPSCNNLossGradient backward filter. The MPSCNNLoss forward filter populates the MPSCNNLossLabels object and the
MPSCNNLossGradient backward filter consumes the state object.
Method Documentation
- (nonnull instancetype) init
Use one of the interfaces below instead.
Reimplemented from MPSState.
- (nonnull instancetype) initWithDevice: (nonnull id< MTLDevice >) device(MPSCNNLossDataDescriptor *_Nonnull) labelsDescriptor
Set labels (aka targets, ground truth) for the MPSCNNLossLabels object. The labels and weights data are copied into internal storage. The
computed loss can either be a scalar value (in batch mode, a single value per image in a batch) or it can be one value per feature channel.
Thus, the size of the loss image must either match the size of the input source image or be {1, 1, 1}, which results in a scalar value. In
this convinience initializer, the assumed size of the loss image is {1, 1, 1}.
Parameters:
device Device the state resources will be created on.
labelsDescriptor Describes the labels data. This includes:
o The per-element labels data. The data must be in floating point format.
o Data layout of labels data. See MPSImage.h for more information.
o Size of labels data: (width, height, feature channels}.
o Optionally, row bytes of labels data.
o Optionally, slice bytes of labels data.
- (nonnull instancetype) initWithDevice: (nonnull id< MTLDevice >) device(MTLSize) lossImageSize(MPSCNNLossDataDescriptor *_Nonnull)
labelsDescriptor(MPSCNNLossDataDescriptor *_Nullable) weightsDescriptor
Set labels (aka targets, ground truth) and weights for the MPSCNNLossLabels object. Weights are optional. The labels and weights data are
copied into internal storage.
Parameters:
device Device the state resources will be created on.
lossImageSize The size of the resulting loss image: { width, height, featureChannels }. The computed loss can either be a scalar value
(in batch mode, a single value per image in a batch) or it can be one value per feature channel. Thus, the size of the loss image must
either match the size of the input source image or be {1, 1, 1}, which results in a scalar value.
labelsDescriptor Describes the labels data. This includes:
o The per-element labels data. The data must be in floating point format.
o Data layout of labels data. See MPSImage.h for more information.
o Size of labels data: (width, height, feature channels}.
o Optionally, row bytes of labels data.
o Optionally, slice bytes of labels data.
weightsDescriptor Describes the weights data. This includes:
o The per-element weights data. The data must be in floating point format.
o Data layout of weights data. See MPSImage.h for more information.
o Size of weights data: (width, height, feature channels}.
o Optionally, row bytes of weights data.
o Optionally, slice bytes of weights data. This parameter is optional. If you are using a single weight, please use the weight property
of the MPSCNNLossDescriptor object.
- (nonnull MPSImage*) lossImage
Loss image accessor method.
Returns:
An autoreleased MPSImage object, containing the loss data. The loss data is populated in the -encode call, thus the contents are
undefined until you -encode the filter.
In order to gaurantee that the image is correctly synchronized for CPU side access, it is the application's responsibility to call the
[gradientState synchronizeOnCommandBuffer:] method before accessing the data in the image.
Author
Generated automatically by Doxygen for MetalPerformanceShaders.framework from the source code.
Version MetalPerformanceShaders-100 Thu Feb 8 2018 MPSCNNLossLabels(3)