Implementation:Rapidsai Cuml FIL Postproc Ops

Knowledge Sources	Rapidsai_Cuml
Domains	Machine_Learning, Forest_Inference
Last Updated	2026-02-08 12:00 GMT

Overview

Defines enumerations for element-wise and row-wise post-processing operations applied to Forest Inference Library (FIL) model outputs.

Description

This header provides two enum types within the ML::fil namespace that control how raw tree ensemble outputs are post-processed before returning predictions:

• row_op: Row-wise operations applied across all output columns of a single sample. Encoded as bit flags:
  • disable (0b00100000): No row-wise post-processing.
  • softmax (0b01000000): Apply softmax normalization across the row (used for multi-class classification).
  • max_index (0b10000000): Return the index of the maximum value in the row (argmax for class prediction).

• element_op: Element-wise operations applied to each individual output value. Encoded as bit flags:
  • disable (0b00000000): No element-wise post-processing.
  • signed_square (0b00000001): Compute sign-preserving square (sign(x) * x^2).
  • hinge (0b00000010): Apply hinge function (max(0, x) or similar threshold).
  • sigmoid (0b00000100): Apply sigmoid (logistic) transformation.
  • exponential (0b00001000): Apply exponential transformation.
  • logarithm_one_plus_exp (0b00010000): Apply log(1 + exp(x)) (softplus).

The bit-flag encoding allows these operations to be combined and efficiently dispatched in GPU kernels.

Usage

Use these enums when configuring FIL inference to specify how raw tree ensemble scores should be transformed into final predictions. For example, use element_op::sigmoid for binary classification probabilities and row_op::softmax with row_op::max_index for multi-class classification.

Code Reference

Source Location

• Repository: Rapidsai_Cuml
• File: cpp/include/cuml/fil/postproc_ops.hpp

Signature

namespace ML {
namespace fil {

enum struct row_op : unsigned char {
  disable   = 0b00100000,
  softmax   = 0b01000000,
  max_index = 0b10000000
};

enum struct element_op : unsigned char {
  disable                = 0b00000000,
  signed_square          = 0b00000001,
  hinge                  = 0b00000010,
  sigmoid                = 0b00000100,
  exponential            = 0b00001000,
  logarithm_one_plus_exp = 0b00010000
};

}  // namespace fil
}  // namespace ML

Import

#include <cuml/fil/postproc_ops.hpp>

I/O Contract

Inputs

Name	Type	Required	Description
(N/A -- these are enum definitions)

Outputs

Name	Type	Description
row_op	enum struct (unsigned char)	Enumerator value selecting the row-wise post-processing operation
element_op	enum struct (unsigned char)	Enumerator value selecting the element-wise post-processing operation

Enum Values

row_op

Enumerator	Bit Pattern	Description
disable	0b00100000	No row-wise operation
softmax	0b01000000	Softmax normalization across the row
max_index	0b10000000	Return index of the maximum value (argmax)

element_op

Enumerator	Bit Pattern	Description
disable	0b00000000	No element-wise operation
signed_square	0b00000001	Sign-preserving square transformation
hinge	0b00000010	Hinge (thresholding) transformation
sigmoid	0b00000100	Sigmoid (logistic) transformation
exponential	0b00001000	Exponential transformation
logarithm_one_plus_exp	0b00010000	Softplus: log(1 + exp(x))

Usage Examples

#include <cuml/fil/postproc_ops.hpp>

void configure_fil_postprocessing() {
    // For binary classification: apply sigmoid to raw scores
    auto elem_op = ML::fil::element_op::sigmoid;
    auto r_op    = ML::fil::row_op::disable;

    // For multi-class classification: apply softmax then argmax
    auto elem_op_multi = ML::fil::element_op::disable;
    auto r_op_multi    = ML::fil::row_op::softmax;
    // Or to get the predicted class index directly:
    auto r_op_argmax   = ML::fil::row_op::max_index;

    // Pass these enum values to FIL inference configuration...
}

Related Pages

• Environment:Rapidsai_Cuml_CUDA_GPU