Get partial derivative wrt params for batchnorm (subroutine version)
| Type | Intent | Optional | Attributes | Name | ||
|---|---|---|---|---|---|---|
| class(array_type), | intent(in) | :: | this | |||
| real(kind=real32), | intent(in), | dimension(:,:) | :: | upstream_grad | ||
| real(kind=real32), | intent(out), | dimension(:,:) | :: | output |
pure subroutine get_partial_batchnorm_right_val(this, upstream_grad, output) !! Get partial derivative wrt params for batchnorm (subroutine version) implicit none class(array_type), intent(in) :: this real(real32), dimension(:,:), intent(in) :: upstream_grad real(real32), dimension(:,:), intent(out) :: output integer :: c, num_dims, num_elements real(real32), allocatable :: x_hat(:,:) real(real32) :: mu, var, eps integer, dimension(size(this%shape)) :: input_shape input_shape = this%left_operand%shape select type(this) type is (batchnorm_array_type) eps = this%epsilon num_dims = size(this%shape) num_elements = product(this%shape(1:num_dims - 1)) output = 0._real32 allocate(x_hat(num_elements, size(upstream_grad,2))) do c = 1, input_shape(num_dims) mu = this%mean(c) var = this%variance(c) ! Normalised input x_hat(:,:) = ( & this%left_operand%val((c-1)*num_elements+1:c*num_elements,:) - mu & ) / sqrt(var + eps) output(c,1) = & sum(upstream_grad((c-1)*num_elements+1:c*num_elements,:) * x_hat) output(c + input_shape(num_dims),1) = & sum(upstream_grad((c-1)*num_elements+1:c*num_elements,:)) end do end select end subroutine get_partial_batchnorm_right_val