ScatterNDLayer.md

CScatterNDLayer Class

• CScatterNDLayer Class
  • Settings
  • Trainable parameters
  • Inputs
  • Outputs

This class implements a layer that scatters updates over the given data blob.

for( int updateIndex = 0; updateIndex < UpdateCount; ++updateIndex )
    data[indices[updateIndex]] = updates[updateIndex];

where indices[...] is an integer vector of IndexDims elements which contains coordinates in the first IndexDims dimensions of the data blob.

Settings

There are no settings for this layer.

Trainable parameters

There are no trainable parameters for this layer.

Inputs

Layer has 3 inputs:

1. The set of objects of any data type. The product of first IndexDims dimensions is ObjectCount and the product of the rest of dimensions is ObjectSize.
2. The set of indices of integer data type of the following shape:

• BD_Channels of the blob must be equal to the IndexDims
• The product of the other dimensions must be equal to the UpdateCount.

3. The set of updates of the same data type as first input. It must contain UpdateCount * ObjectSize elements regardless of shape.

Outputs

The only output contains updated data of the same size and type as the first input.