convolution3DWeightsGradient(_:source:outputShape:forwardConvolutionDescriptor:name:)
Creates a 3D convolution gradient operation with respect to the weights tensor of the forward convolution.
Declaration
func convolution3DWeightsGradient(_ incomingGradient: MPSGraphTensor, source: MPSGraphTensor, outputShape: [NSNumber], forwardConvolutionDescriptor: MPSGraphConvolution3DOpDescriptor, name: String?) -> MPSGraphTensorParameters
- incomingGradient:
Incoming loss gradient tensor
- outputShape:
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, convolution3DWeightsGradientWithIncomingGradientTensor returns tensor dL/dW = dL/dR * dR/dW, where dL/dR is the incomingGradient parameter.