singleGateRNNGradients(_:recurrentWeight:sourceGradient:zState:initState:descriptor:name:)
Creates a single-gate RNN gradient operation and returns the gradient tensor values.
Declaration
func singleGateRNNGradients(_ source: MPSGraphTensor, recurrentWeight: MPSGraphTensor, sourceGradient: MPSGraphTensor, zState: MPSGraphTensor, initState: MPSGraphTensor?, descriptor: MPSGraphSingleGateRNNDescriptor, name: String?) -> [MPSGraphTensor]Parameters
- source:
A tensor that contains the source data
x[t]with the data layout [T,N,I]. In caseinputWeight = nilandbidirectional = NOthen the layout is [T,N,H] and forinputWeight = nilandbidirectional = YESthe layout is [T,N,2H]. - recurrentWeight:
A tensor containing the recurrent weights
R. Forbidirectionalthe layout is [2,H,H] and otherwise it is [H,H]. Note: Forbidirectionalthis tensor must have a static shape. - sourceGradient:
The input gradient, that is the gradient of a tensor with respect to the first output of the forward pass.
- zState:
The second output of Singlegaternn(_:recurrentweight:inputweight:bias:initstate:mask:descriptor:name:) with
descriptor.training = YES. - initState:
The initial internal state of the RNN
h[-1]- optional, if missing the operation assumes zeroes. Forbidirectionalthe layout is [N,2H] and otherwise it is [N,H]. - descriptor:
A descriptor that defines the parameters for the RNN operation.
- name:
The name for the operation.
Return Value
A valid MPSGraphTensor array containing gradients for each input tensor, except for sourceGradient and mask. In case an input is nil, no gradient will be returned for it. The order of the gradients will be: for source, for recurrentWeight, for inputWeight, for bias and finally for initState.
Discussion
For details of this operation and parameters, refer to documentation of singleGateRNN(_:recurrentWeight:inputWeight:bias:initState:mask:descriptor:name:).