scatterNDWithData(_:updates:indices:batchDimensions:mode:name:)

Declaration

func scatterNDWithData(_ dataTensor: MPSGraphTensor, updates updatesTensor: MPSGraphTensor, indices indicesTensor: MPSGraphTensor, batchDimensions: Int, mode: MPSGraphScatterMode, name: String?) -> MPSGraphTensor

Parameters

• dataTensor:

  Tensor containing inital values of same shape as result tensor

• updatesTensor:

  Tensor containing slices to be inserted into the result tensor.

• indicesTensor:

  Tensor containg the result indices to insert slices at

• batchDimensions:

  The number of batch dimensions

• mode:

  The type of update to use on the destination

• name:

  The name for the operation.

Return Value

A valid MPSGraphTensor object

Discussion

Scatters the slices in updatesTensor to the result tensor along the indices in indicesTensor, on top of dataTensor. The scatter is defined as

B = batchDims 
U = updates.rank - B 
P = res.rank - B 
Q = inds.rank - B 
K = inds.shape[-1] 
index_slice = indices[i_{b0},...,i_{bB},i_{0},..,i_{Q-1}] 
res[...] = data[...] 
res[i_{b0},...,i_{bB},index_slice[0],...,index_slice[K-1]] += updates[i_{b0},...,i_{bB},i_{0},...,i_{Q-1}] // Note += is used but this depends on mode 

Collisions will be updated according to mode, and slices not set by indices are set to 0. The tensors have the following shape requirements

K <= P 
U = (P-K) + Q-1 
data.shape = res.shape 
indices.shape[0:Q-1] = updates.shape[0:Q-1] 
updates.shape[Q:U] = res.shape[K:P]