MPSCNNConvolutionTranspose(3)
NAME
MPSCNNConvolutionTranspose
SYNOPSIS
#import <MPSCNNConvolution.h>
Inherits MPSCNNKernel.
Instance Methods
(nonnull instancetype) - initWithDevice:weights:
(nonnull instancetype) - initWithDevice:
(nullable instancetype) - initWithCoder:device:
(MPSImage *__nonnull) -
encodeToCommandBuffer:sourceImage:convolutionState:
Properties
NSUInteger inputFeatureChannels
NSUInteger outputFeatureChannels
NSInteger kernelOffsetX
NSInteger kernelOffsetY
NSUInteger groups
Additional Inherited Members
Detailed Description
This depends on Metal.framework The MPSCNNConvolutionTranspose
specifies a transposed convolution. The MPSCNNConvolutionTranspose
convolves the input image with a set of filters, each producing one
feature map in the output image.
Some third-party frameworks may rotate the weights spatially by 180
degrees for Convolution Transpose. MPS uses the weights specified by
the developer as-is and does not perform any rotation. The developer
may need to rotate the weights appropriately in case this rotation is
needed before the convolution transpose is applied.
When the stride in any dimension is greater than 1, the convolution
transpose puts (stride - 1) zeroes in-between the source image pixels
to create an expanded image. Then a convolution is done over the
expanded image to generate the output of the convolution transpose.
Intermediate image size = (srcSize - 1) * Stride + 1
Examples:
So in case of sride == 2 (this behaves same in both dimensions)
Source image:
_______________
| | | | |
| 1 | 2 | 3 | 4 |
| | | | |
---------------
Intermediate Image:
___________________________
| | | | | | | |
| 1 | 0 | 2 | 0 | 3 | 0 | 4 |
| | | | | | | |
---------------------------
NOTE on Offset:
There are 2 types of offsets defined:
1) The Offset defined in MPSCNNKernel from which MPSCNNConvolutionTranspose inherits. This offset is applied to from where
the kernel will be applied on the source.
2) The kernelOffsetX and kernelOffsetY which is the offset applied to the kernel when it is finally applied on the intermediate
image.
So totalOffset = Offset * stride + kernelOffset
The offset defined by user refers to the coordinate frame of the expanded image
(we are showing only 1 dimension X it can be extended to Y dimension as well) :
X indicates where the convolution transpose begins:
Intermediate Image: Offset = 0, kernelOffset = 0
___________________________
| | | | | | | |
| 1 | 0 | 2 | 0 | 3 | 0 | 4 |
| X | | | | | | |
---------------------------
X indicates where the convolution transpose begins:
Intermediate Image: Offset = 0, kernelOffset = 1
___________________________
| | | | | | | |
| 1 | 0 | 2 | 0 | 3 | 0 | 4 |
| | X | | | | | |
---------------------------
X indicates where the convolution transpose begins:
Intermediate Image: Offset = 0, kernelOffset = -1
___________________________
| | | | | | | |
X | 1 | 0 | 2 | 0 | 3 | 0 | 4 |
| | | | | | | |
---------------------------
So if the user wanted to apply an offset of 2 on the source image of convolution transpose:
Source image:
_______________
| | | | |
| 1 | 2 | 3 | 4 |
| | | X | |
---------------
offset = 2, kernelOffset = 0
Intermediate Image:
___________________________
| | | | | | | |
| 1 | 0 | 2 | 0 | 3 | 0 | 4 |
| | | | | X | | |
---------------------------
Method Documentation
- (MPSImage * __nonnull) encodeToCommandBuffer: (nonnull id<
MTLCommandBuffer >) commandBuffer(MPSImage *__nonnull)
sourceImage(MPSCNNConvolutionState *__nullable) convolutionState
Encode a MPSCNNKernel into a command Buffer. Create a texture to hold
the result and return it. In the first iteration on this method,
encodeToCommandBuffer:sourceImage:destinationImage: some work was left
for the developer to do in the form of correctly setting the offset
property and sizing the result buffer. With the introduction of the
padding policy (see padding property) the filter can do this work
itself. If you would like to have some input into what sort of MPSImage
(e.g. temporary vs. regular) or what size it is or where it is
allocated, you may set the destinationImageAllocator to allocate the
image yourself.
This method uses the MPSNNPadding padding property to figure out how to
size the result image and to set the offset property. See discussion in
MPSNeuralNetworkTypes.h.
Note: the regular encodeToCommandBuffer:sourceImage: method may be used
when no state is needed, such as when the convolution transpose
operation is not balanced by a matching convolution object upstream.
Parameters:
commandBuffer The command buffer
sourceImage A MPSImage to use as the source images for the filter.
convolutionState A valid MPSCNNConvolutionState from the
MPSCNNConvoluton counterpart to this MPSCNNConvolutionTranspose. If
there is no forward convolution counterpart, pass NULL here. This
state affects the sizing the result.
Returns:
A MPSImage or MPSTemporaryImage allocated per the
destinationImageAllocator containing the output of the graph. The
offset property will be adjusted to reflect the offset used during
the encode. The returned image will be automatically released when
the command buffer completes. If you want to keep it around for
longer, retain the image. (ARC will do this for you if you use it
later.)
- (nullable instancetype) initWithCoder: (NSCoder *__nonnull)
aDecoder(nonnull id< MTLDevice >) device
<NSSecureCoding> support
Reimplemented from MPSCNNKernel.
- (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 MPSCNNKernel.
- (nonnull instancetype) initWithDevice: (nonnull id< MTLDevice >)
device(nonnull id< MPSCNNConvolutionDataSource >) weights
Initializes a convolution kernel
Parameters:
device The MTLDevice on which this MPSCNNConvolutionTranspose
filter will be used
weights A pointer to a object that conforms to the
MPSCNNConvolutionDataSource protocol. The
MPSCNNConvolutionDataSource protocol declares the methods that an
instance of MPSCNNConvolutionTranspose uses to obtain the weights
and bias terms for the CNN convolutionTranspose filter. Currently
we support only Float32 weights.
Returns:
A valid MPSCNNConvolution object or nil, if failure.
Property Documentation
- groups [read], [nonatomic], [assign]
Number of groups input and output channels are divided into.
- inputFeatureChannels [read], [nonatomic], [assign]
The number of feature channels per pixel in the input image.
- kernelOffsetX [read], [write], [nonatomic], [assign]
Offset in X from which the kernel starts sliding
- kernelOffsetY [read], [write], [nonatomic], [assign]
Offset in Y from which the kernel starts sliding
- outputFeatureChannels [read], [nonatomic], [assign]
The number of feature channels per pixel in the output image.
Author
Generated automatically by Doxygen for
MetalPerformanceShaders.framework from the source code.
Version MetalPerformanceShaders-Thu2Jul 13 2017 MPSCNNConvolutionTranspose(3)
Mac OS X 10.12.6 - Generated Sun Oct 29 15:37:01 CDT 2017
