globalAveragePoolingGPU function

GPUTensor<Vector> globalAveragePoolingGPU(
  1. GPUTensor<Matrix> input,
  2. CommandBuffer tape, {
  3. GPUTensor<Vector>? outTensor,
})

Collapses a matrix input to a vector of their average values of their feature columns. The operations are appended to tape for execution.

Implementation

GPUTensor<Vector> globalAveragePoolingGPU(
    GPUTensor<Matrix> input, CommandBuffer tape,
    {GPUTensor<Vector>? outTensor}) {
  //int sequenceLength = input.shape[0];
  int dModel = input.shape[1];

  List<int> outShape = <int>[dModel];
  GPUTensor<Vector> out = outTensor ?? GPUTensor<Vector>.empty(outShape);

  tape.putInt(OP_GLOBAL_AVG_POOL_FORWARD);
  tape.putString(input.id);
  tape.putString(out.id);

  out.creator = GPUNode(
    <GPUTensor>[input],
    (CommandBuffer backwardTape) {
      backwardTape.putInt(OP_GLOBAL_AVG_POOL_BACKWARD);
      backwardTape.putString(out.gradId);
      backwardTape.putString(input.gradId);
    },
    opName: 'global_avg_pool_gpu',
  );

  return out;
}