MPSCNNGradientKernel(3) MetalPerformanceShaders.framework MPSCNNGradientKernel(3)
NAME
MPSCNNGradientKernel
SYNOPSIS
#import <MPSCNNKernel.h>
Inherits MPSCNNBinaryKernel.
Inherited by MPSCNNArithmeticGradient, MPSCNNBatchNormalizationGradient, MPSCNNBatchNormalizationStatisticsGradient,
MPSCNNConvolutionGradient, MPSCNNCrossChannelNormalizationGradient, MPSCNNDropoutGradient, MPSCNNInstanceNormalizationGradient,
MPSCNNLocalContrastNormalizationGradient, MPSCNNLogSoftMaxGradient, MPSCNNNeuronGradient, MPSCNNPoolingGradient, MPSCNNSoftMaxGradient,
MPSCNNSpatialNormalizationGradient, and MPSCNNUpsamplingGradient.
Instance Methods
(nonnull instancetype) - initWithDevice:
(nullable instancetype) - initWithCoder:device:
(MPSImage *__nonnull) - encodeToCommandBuffer:sourceGradient:sourceImage:gradientState:
(void) - encodeToCommandBuffer:sourceGradient:sourceImage:gradientState:destinationGradient:
(MPSImageBatch *__nonnull) - encodeBatchToCommandBuffer:sourceGradients:sourceImages:gradientStates:
(void) - encodeBatchToCommandBuffer:sourceGradients:sourceImages:gradientStates:destinationGradients:
Properties
NSInteger kernelOffsetX
NSInteger kernelOffsetY
Additional Inherited Members
Detailed Description
Gradient kernels are the backwards pass of a MPSCNNKernel used during training to calculate gradient back propagation. These take as
arguments the gradient result from the next filter and the source image for the forward version of the filter. There is also a
MPSNNGradientState passed from MPSCNNKernel to MPSCNNGradientKernel that contains information about the MPSCNNKernel parameters at the time
it encoded and possibly also additional MTLResources to enable it to do its job.
Training graph (partial):
---> input image ---------> MPSCNNKernel ------> resultImage ------>-->-->-->.
| |
'------. MPSNNGradientState loss estimation
| |
V V V
<--- result gradient <- MPSCNNGradientKernel <--- input gradient <--<--<--<---'
In general operation, starting with the input image, the sequence of events is:
1a) Invoke padding policy to find result size for MPSCNNKernel. This
also configures some MPSCNNKernel parameters such as offset.
1b) Use the MPSImageDescriptor from 1a to make resultImage.
1c) Call MPSCNNKernel -encode...
2) stages 1a-c are repeated for other forward passes in the inference portion of the graph
3) We estimate the loss resulting from the whole inference computation so far (see MPSCNNLoss.h>
4) stages 5a-c are repeated for corresponding backward gradient passes in the graph
5a) Invoke padding policy on the MPSCNNGradientKernel shown above. This sets the
MPSCNNGradientKernel parameters to correspond with those in the forward pass
5b) The result gradient for the MPSCNNGradientKernel is created from the MPSImageDescriptor from 5a
5c) Call MPSCNNGradientKernel -encode with the input image, input gradient, result gradient and MPSNNGradientState
6) pass the result gradient on to leftward gradient passes.
For MPSCNNKernels that are trained, there may be other accompanying training kernels that
need to be called in addition to the gradient kernel to update convolution weights or batch
normalization parameters, for example. Steps 1a-c and 5a-c can be combined in a single -encode
call. These return the result image or gradient out the left hand side.
For purposes of inheritance the gradient image is the MPSCNNBinaryKernel primary image
and the source image is the MPSCNNBinaryKernel secondary image. Various secondary properties
such as kernel size are copies of the forward inference pass parameters of similar name
are set automatically when -[MPSCNNGradientKernel destinationImageDescriptorForSourceImages:sourceStates:]
is called.
Method Documentation
- (MPSImageBatch*__nonnull) encodeBatchToCommandBuffer: (__nonnull id< MTLCommandBuffer >) commandBuffer(MPSImageBatch *__nonnull)
sourceGradients(MPSImageBatch *__nonnull) sourceImages(MPSStateBatch *__nonnull) gradientStates
Encode a gradient filter and return a gradient During training, gradient filters are used to calculate the gradient associated with the
loss for each feature channel in the forward pass source image. For those nodes that are trainable, these are then used to refine the value
used in the trainable parameter. They consume a source gradient image which contains the gradients corresponding with the forward pass
destination image, and calculate the gradients corresponding to the forward pass source image.
A gradient filter consumes a MPSNNGradientState object which captured various forward pass properties such as offset and edgeMode at the
time the forward pass was encoded. These are transferred to the MPSCNNBinaryKernel secondary image properties automatically when this
method creates its destination image.
Parameters:
commandBuffer The MTLCommandBuffer on which to encode
sourceGradients The gradient images from the 'next' filter in the graph
sourceImages The images used as source image from the forward pass
gradientStates The MPSNNGradientState or MPSNNBinaryGradientState subclass produced by the forward pass
Reimplemented in MPSCNNBatchNormalizationStatisticsGradient.
- (void) encodeBatchToCommandBuffer: (__nonnull id< MTLCommandBuffer >) commandBuffer(MPSImageBatch *__nonnull) sourceGradients(MPSImageBatch
*__nonnull) sourceImages(MPSStateBatch *__nonnull) gradientStates(MPSImageBatch *__nonnull) destinationGradients
Encode a gradient filter and return a gradient During training, gradient filters are used to calculate the gradient associated with the
loss for each feature channel in the forward pass source image. For those nodes that are trainable, these are then used to refine the value
used in the trainable parameter. They consume a source gradient image which contains the gradients corresponding with the forward pass
destination image, and calculate the gradients corresponding to the forward pass source image.
A gradient filter consumes a MPSNNGradientState object which captured various forward pass properties such as offset and edgeMode at the
time the forward pass was encoded. These are transferred to the MPSCNNBinaryKernel secondary image properties automatically when you use
-[MPSCNNGradientKernel destinationImageDescriptorForSourceImages:sourceStates:]. If you do not call this method, then you are responsible
for configuring all of the primary and secondary image properties in MPSCNNBinaryKernel. Please see class description for expected ordering
of operations.
Parameters:
commandBuffer The MTLCommandBuffer on which to encode
sourceGradients The gradient images from the 'next' filter in the graph
sourceImages The image used as source images from the forward pass
gradientStates An array of the MPSNNGradientState or MPSNNBinaryGradientState subclass produced by the forward pass
destinationGradients The MPSImages into which to write the filter result
- (MPSImage*__nonnull) encodeToCommandBuffer: (__nonnull id< MTLCommandBuffer >) commandBuffer(MPSImage *__nonnull) sourceGradient(MPSImage
*__nonnull) sourceImage(MPSState *__nonnull) gradientState
Encode a gradient filter and return a gradient During training, gradient filters are used to calculate the gradient associated with the
loss for each feature channel in the forward pass source image. For those nodes that are trainable, these are then used to refine the value
used in the trainable parameter. They consume a source gradient image which contains the gradients corresponding with the forward pass
destination image, and calculate the gradients corresponding to the forward pass source image.
A gradient filter consumes a MPSNNGradientState object which captured various forward pass properties such as offset and edgeMode at the
time the forward pass was encoded. These are transferred to the MPSCNNBinaryKernel secondary image properties automatically when this
method creates its destination image.
Parameters:
commandBuffer The MTLCommandBuffer on which to encode
sourceGradient The gradient image from the 'next' filter in the graph (in the inference direction)
sourceImage The image used as source image by the forward inference pass
gradientState The MPSNNGradientState or MPSNNBinaryGradientState subclass produced by the forward inference pass
Returns:
The result gradient from the gradient filter
Reimplemented in MPSCNNBatchNormalizationStatisticsGradient.
- (void) encodeToCommandBuffer: (__nonnull id< MTLCommandBuffer >) commandBuffer(MPSImage *__nonnull) sourceGradient(MPSImage *__nonnull)
sourceImage(MPSState *__nonnull) gradientState(MPSImage *__nonnull) destinationGradient
Encode a gradient filter and return a gradient During training, gradient filters are used to calculate the gradient associated with the
loss for each feature channel in the forward pass source image. For those nodes that are trainable, these are then used to refine the value
used in the trainable parameter. They consume a source gradient image which contains the gradients corresponding with the forward pass
destination image, and calculate the gradients corresponding to the forward pass source image.
A gradient filter consumes a MPSNNGradientState object which captured various forward pass properties such as offset and edgeMode at the
time the forward pass was encoded. These are transferred to the MPSCNNBinaryKernel secondary image properties automatically when you use
-[MPSCNNGradientKernel destinationImageDescriptorForSourceImages:sourceStates:]. If you do not call this method, then you are responsible
for configuring all of the primary and secondary image properties in MPSCNNBinaryKernel. Please see class description for expected ordering
of operations.
Parameters:
commandBuffer The MTLCommandBuffer on which to encode
sourceGradient The gradient image from the 'next' filter in the graph
sourceImage The image used as source image from the forward pass
gradientState The MPSNNGradientState and MPSNNBinaryGradientState subclass produced by the forward pass
destinationGradient The MPSImage into which to write the filter result
Reimplemented in MPSCNNBatchNormalizationStatisticsGradient.
- (nullable instancetype) initWithCoder: (NSCoder *__nonnull) aDecoder(nonnull id< MTLDevice >) device
NSSecureCoding compatability While the standard NSSecureCoding/NSCoding method -initWithCoder: should work, since the file can't know
which device your data is allocated on, we have to guess and may guess incorrectly. To avoid that problem, use initWithCoder:device
instead.
Parameters:
aDecoder The NSCoder subclass with your serialized MPSKernel
device The MTLDevice on which to make the MPSKernel
Returns:
A new MPSKernel object, or nil if failure.
Reimplemented from MPSCNNBinaryKernel.
Reimplemented in MPSCNNConvolutionGradient, MPSCNNFullyConnectedGradient, MPSCNNPoolingAverageGradient, MPSCNNPoolingMaxGradient,
MPSCNNPoolingL2NormGradient, MPSCNNDilatedPoolingMaxGradient, MPSCNNSoftMaxGradient, MPSCNNLogSoftMaxGradient,
MPSCNNCrossChannelNormalizationGradient, MPSCNNPoolingGradient, MPSCNNLocalContrastNormalizationGradient, MPSCNNNeuronGradient,
MPSCNNDropoutGradient, and MPSCNNSpatialNormalizationGradient.
- (nonnull instancetype) initWithDevice: (nonnull id< MTLDevice >) device
Standard init with default properties per filter type
Parameters:
device The device that the filter will be used on. May not be NULL.
Returns:
A pointer to the newly initialized object. This will fail, returning nil if the device is not supported. Devices must be
MTLFeatureSet_iOS_GPUFamily2_v1 or later.
Reimplemented from MPSCNNBinaryKernel.
Reimplemented in MPSCNNConvolutionGradient, MPSCNNFullyConnectedGradient, MPSCNNSoftMaxGradient, MPSCNNLogSoftMaxGradient,
MPSCNNPoolingGradient, MPSCNNArithmeticGradient, MPSCNNAddGradient, MPSCNNSubtractGradient, MPSCNNMultiplyGradient, MPSCNNNeuronGradient,
MPSCNNDropoutGradient, and MPSCNNUpsamplingGradient.
Property Documentation
- kernelOffsetX [read], [write], [nonatomic], [assign]
Offset in the kernel reference frame to position the kernel in the X dimension In some cases, the input gradient must be upsampled with
zero insertion to account for things like strides in the forward MPSCNNKernel pass. As such, the offset, which describes a X,Y offset in
the source coordinate space is insufficient to fully describe the offset applied to a kernel. The kernel offset is the offset after
upsampling. Both the source offset and kernel offset are additive: effective offset = source offset * stride + kernel offset. The offset is
applied to the (upsampled) source gradient
- kernelOffsetY [read], [write], [nonatomic], [assign]
Offset in the kernel reference frame to position the kernel in the Y dimension In some cases, the input gradient must be upsampled with
zero insertion to account for things like strides in the forward MPSCNNKernel pass. As such, the offset, which describes a X,Y offset in
the source coordinate space is insufficient to fully describe the offset applied to a kernel. The kernel offset is the offset after
upsampling. Both the source offset and kernel offset are additive: effective offset = source offset * stride + kernel offset. The offset is
applied to the (upsampled) source gradient
Author
Generated automatically by Doxygen for MetalPerformanceShaders.framework from the source code.
Version MetalPerformanceShaders-100 Thu Feb 8 2018 MPSCNNGradientKernel(3)