Function SumGradHelper

tensorflow/cc/gradients/math_grad.cc:694–729

SumGradHelper returns the gradient for the Sum operator, and is used by SumGrad and MeanGrad.

// SumGradHelper returns the gradient for the Sum operator, and is used
// by SumGrad and MeanGrad.
Output SumGradHelper(const Scope& scope, const Operation& op,
                     const std::vector<Output>& grad_inputs) {
  // The partial derivative for any input along a "reduced" dimension
  // is just 1, so we only need replicate the output gradient on such a
  // dimension to its "expanded" shape.
  // Running example:
  // input is
  // [[a, b, c],
  //  [d, e, f]]
  // reduction_indices = [1]
  // Sum = [a + b + c, d + e + f]
  // if the gradient is [g1, g2]
  // We want the propagated gradient to be
  // [[g1, g1, g1],
  //  [g2, g2, g2]]

  // input_shape = [2, 3]
  auto input_shape = Shape(scope, op.input(0));

  // output_shape_kept_dims = [2, 1]
  auto output_shape_kept_dims =
      ReducedShapeHelper(scope, input_shape, op.input(1));

  // This step "flips" any 1s with values from the input_shape, and
  // replaces remaining entries with 1. This creates a shape that
  // shows how much each dimension in the incoming gradient should be
  // replicated.
  // tile_scaling = [1, 3]
  auto tile_scaling = SafeDivHelper(scope, input_shape, output_shape_kept_dims);

  // grad = [[g1], [g2]]
  auto grad = Reshape(scope, grad_inputs[0], output_shape_kept_dims);

  // tile(grad, tile_scaling) = [[g1, g1, g1], [g2, g2, g2]]
  return Tile(scope, grad, tile_scaling);
}

Status SumGrad(const Scope& scope, const Operation& op,
               const std::vector<Output>& grad_inputs,