Contents

convolution2DWeightsGradient(_:source:outputShapeTensor:forwardConvolutionDescriptor:name:)

Creates a 2D convolution gradient operation with respect to weights tensor of forward convolution.

Declaration

func convolution2DWeightsGradient(_ gradient: MPSGraphTensor, source: MPSGraphTensor, outputShapeTensor: MPSGraphTensor, forwardConvolutionDescriptor: MPSGraphConvolution2DOpDescriptor, name: String?) -> MPSGraphTensor

Parameters

  • outputShapeTensor:

    4D int32 or Int64 Tensor. Shape of the forward pass source tensor

  • forwardConvolutionDescriptor:

    Forward convolution 2D op descriptor

  • name:

    The name for the operation.

Return Value

A valid MPSGraphTensor object

Discussion

If W is weights tensor to forward convolution, R is the result/returned tensor of forward convolution, and L is the loss function, convolution2DWeightsGradientWithIncomingGradientTensor returns tensor dL/dW = dL/dR * dR/dW, where dL/dR is the incomingGradient parameter.