Developer Guide for Intel® Data Analytics Acceleration Library 2019 Update 4
The backward two-dimensional transposed convolution layer accepts the input described below. Pass the Input ID as a parameter to the methods that provide input for your algorithm. For more details, see Algorithms.
Input ID |
Input |
|
---|---|---|
inputGradient |
Pointer to tensor G of size n1 × nKernels × n3 × n4 that stores the input gradient computed on the preceding layer. This input can be an object of any class derived from Tensor. |
|
inputFromForward |
Collection of input data for the backward two-dimensional transposed convolution layer. This collection can contain objects of any class derived from Tensor. |
|
Element ID |
Element |
|
auxData |
Pointer to tensor X of size n1 × l2 × l3 × l4 that stores the input data for the forward two-dimensional transposed convolution layer. This input can be an object of any class derived from Tensor. |
|
auxWeights |
Pointer to the tensor K of size l2 × nKernels/nGroups × m3 × m4 that stores a set of kernel weights. This input can be an object of any class derived from Tensor. |
For common parameters of neural network layers, see Common Parameters.
In addition to the common parameters, the backward two-dimensional transposed convolution layer has the following parameters:
Parameter |
Default Value |
Description |
|
---|---|---|---|
algorithmFPType |
float |
The floating-point type that the algorithm uses for intermediate computations. Can be float or double. |
|
method |
defaultDense |
Performance-oriented computation method, the only method supported by the layer. |
|
kernelSizes |
KernelSizes(2, 2) |
Data structure representing the sizes mi, i ∈ {3, 4}, of the two-dimensional kernel subtensor. |
|
indices |
Indices(2,3) |
Data structure representing the dimensions for applying transposed convolution kernels. |
|
strides |
Strides(2, 2) |
Data structure representing the intervals si, i ∈ {3, 4}, on which the kernel should be applied to the input. |
|
paddings |
Paddings(0, 0) |
Data structure representing the number of data elements pi, i ∈ {3, 4}, to implicitly add to each side of the two-dimensional subtensor along which backward two-dimensional transposed convolution is performed. |
|
nKernels |
n/a |
Number of kernels applied to the layer input data. |
|
groupDimension |
1 |
Dimension to which grouping is applied. |
|
nGroups |
1 |
Number of groups into which the input data is split in dimension groupDimension. |
|
propagateGradient |
false |
Flag that specifies whether the forward layer propagates the gradient. |
The backward two-dimensional transposed convolution layer calculates the result described below. Pass the Result ID as a parameter to the methods that access the results of your algorithm. For more details, see Algorithms.
Result ID |
Result |
|
---|---|---|
gradient |
Pointer to tensor Z of size n1 × l2 × l3 × l4 that stores the computed gradient. This result can be an object of any class derived from Tensor. |
|
weightDerivatives |
Pointer to the tensor of size l2 × nKernels/nGroups × m3 × m4 that stores the derivatives ∂Ε / ∂krcuv, where r = {0, ..., l2}, c = {0, ..., nKernels/nGroups}, u = {0, ..., m3 - 1}, and v = {0, ..., m4 - 1}. This result can be an object of any class derived from Tensor. |
|
biasDerivatives |
Pointer to tensor of size nKernels that stores the derivatives ∂Ε / ∂br , where r = {0, ..., nKernels - 1}. This result can be an object of any class derived from Tensor. |