def test_bin_vector_expr():
expr = ast.BinVectorExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3),
ast.NumVal(4)]),
ast.BinNumOpType.ADD)
expected_code = """
score <- function(input) {
return((c(1.0, 2.0)) + (c(3.0, 4.0)))
}
"""
interpreter = RInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_vector_expr():
expr = ast.BinVectorExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3),
ast.NumVal(4)]),
ast.BinNumOpType.MUL)
interpreter = interpreters.PythonInterpreter()
expected_code = """
import numpy as np
def score(input):
return (np.asarray([1, 2])) * (np.asarray([3, 4]))
"""
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_vector_expr():
expr = ast.BinVectorExpr(
ast.VectorVal([ast.NumVal(1), ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3), ast.NumVal(4)]),
ast.BinNumOpType.ADD)
expected_code = """
Module Model
Function AddVectors(ByRef v1() As Double, ByRef v2() As Double) As Double()
Dim resLength As Integer
resLength = UBound(v1) - LBound(v1)
Dim result() As Double
ReDim result(resLength)
Dim i As Integer
For i = LBound(v1) To UBound(v1)
result(i) = v1(i) + v2(i)
Next i
AddVectors = result
End Function
Function MulVectorNumber(ByRef v1() As Double, ByVal num As Double) As Double()
Dim resLength As Integer
resLength = UBound(v1) - LBound(v1)
Dim result() As Double
ReDim result(resLength)
Dim i As Integer
For i = LBound(v1) To UBound(v1)
result(i) = v1(i) * num
Next i
MulVectorNumber = result
End Function
Function Score(ByRef inputVector() As Double) As Double()
Dim var0(1) As Double
var0(0) = 1.0
var0(1) = 2.0
Dim var1(1) As Double
var1(0) = 3.0
var1(1) = 4.0
Score = AddVectors(var0, var1)
End Function
End Module
"""
interpreter = VisualBasicInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def _assemble_multi_class_output(self):
support_vectors = self.model.support_vectors_
coef = self.model.dual_coef_
intercept = self.model.intercept_
n_support = self.model.n_support_
n_support_len = len(n_support)
kernel_exprs = self._apply_kernel(support_vectors, to_reuse=True)
support_ranges = []
for i in range(n_support_len):
range_start = sum(n_support[:i])
range_end = range_start + n_support[i]
support_ranges.append((range_start, range_end))
# One-vs-one decisions.
decisions = []
for i in range(n_support_len):
for j in range(i + 1, n_support_len):
kernel_weight_mul_ops = [
utils.mul(kernel_exprs[k], ast.NumVal(coef[i][k]))
for k in range(*support_ranges[j])
]
kernel_weight_mul_ops.extend([
utils.mul(kernel_exprs[k], ast.NumVal(coef[j - 1][k]))
for k in range(*support_ranges[i])
])
decision = utils.apply_op_to_expressions(
ast.BinNumOpType.ADD,
ast.NumVal(intercept[len(decisions)]),
*kernel_weight_mul_ops)
decisions.append(decision)
return ast.VectorVal(decisions)
def test_bin_vector_expr():
expr = ast.BinVectorExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3),
ast.NumVal(4)]),
ast.BinNumOpType.ADD)
expected_code = """
let private addVectors v1 v2 = List.map2 (+) v1 v2
let private mulVectorNumber v1 num = List.map (fun i -> i * num) v1
let score (input : double list) =
addVectors ([1.0; 2.0]) ([3.0; 4.0])
"""
interpreter = FSharpInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_class_sigmoid_output_transform():
estimator = lightgbm.LGBMClassifier(n_estimators=1,
random_state=1,
max_depth=1,
sigmoid=0.5)
utils.get_binary_classification_model_trainer()(estimator)
assembler = assemblers.LightGBMModelAssembler(estimator)
actual = assembler.assemble()
sigmoid = ast.BinNumExpr(
ast.NumVal(1),
ast.BinNumExpr(
ast.NumVal(1),
ast.ExpExpr(
ast.BinNumExpr(
ast.NumVal(0),
ast.BinNumExpr(
ast.NumVal(0.5),
ast.IfExpr(
ast.CompExpr(ast.FeatureRef(20),
ast.NumVal(16.795),
ast.CompOpType.GT),
ast.NumVal(0.5500419366076967),
ast.NumVal(1.2782342253678096)),
ast.BinNumOpType.MUL), ast.BinNumOpType.SUB)),
ast.BinNumOpType.ADD),
ast.BinNumOpType.DIV,
to_reuse=True)
expected = ast.VectorVal([
ast.BinNumExpr(ast.NumVal(1), sigmoid, ast.BinNumOpType.SUB), sigmoid
])
assert utils.cmp_exprs(actual, expected)
def test_multi_class():
estimator = linear_model.LogisticRegression()
estimator.coef_ = np.array([[1, 2], [3, 4], [5, 6]])
estimator.intercept_ = np.array([7, 8, 9])
assembler = assemblers.SklearnLinearModelAssembler(estimator)
actual = assembler.assemble()
expected = ast.VectorVal([
ast.BinNumExpr(
ast.BinNumExpr(
ast.NumVal(7),
ast.BinNumExpr(ast.FeatureRef(0), ast.NumVal(1),
ast.BinNumOpType.MUL), ast.BinNumOpType.ADD),
ast.BinNumExpr(ast.FeatureRef(1), ast.NumVal(2),
ast.BinNumOpType.MUL), ast.BinNumOpType.ADD),
ast.BinNumExpr(
ast.BinNumExpr(
ast.NumVal(8),
ast.BinNumExpr(ast.FeatureRef(0), ast.NumVal(3),
ast.BinNumOpType.MUL), ast.BinNumOpType.ADD),
ast.BinNumExpr(ast.FeatureRef(1), ast.NumVal(4),
ast.BinNumOpType.MUL), ast.BinNumOpType.ADD),
ast.BinNumExpr(
ast.BinNumExpr(
ast.NumVal(9),
ast.BinNumExpr(ast.FeatureRef(0), ast.NumVal(5),
ast.BinNumOpType.MUL), ast.BinNumOpType.ADD),
ast.BinNumExpr(ast.FeatureRef(1), ast.NumVal(6),
ast.BinNumOpType.MUL), ast.BinNumOpType.ADD)
])
assert utils.cmp_exprs(actual, expected)
def test_multi_class():
estimator = lightgbm.LGBMClassifier(n_estimators=1,
random_state=1,
max_depth=1)
estimator.fit(np.array([[1], [2], [3]]), np.array([1, 2, 3]))
assembler = assemblers.LightGBMModelAssembler(estimator)
actual = assembler.assemble()
exponent = ast.ExpExpr(ast.BinNumExpr(ast.NumVal(0.0),
ast.NumVal(-1.0986122886681098),
ast.BinNumOpType.ADD),
to_reuse=True)
exponent_sum = ast.BinNumExpr(ast.BinNumExpr(exponent, exponent,
ast.BinNumOpType.ADD),
exponent,
ast.BinNumOpType.ADD,
to_reuse=True)
softmax = ast.BinNumExpr(exponent, exponent_sum, ast.BinNumOpType.DIV)
expected = ast.VectorVal([softmax] * 3)
assert utils.cmp_exprs(actual, expected)
def test_bin_vector_num_expr():
expr = ast.BinVectorNumExpr(
ast.VectorVal([ast.NumVal(1), ast.NumVal(2)]),
ast.NumVal(1),
ast.BinNumOpType.MUL)
expected_code = """
<?php
function addVectors(array $v1, array $v2) {
$result = array();
for ($i = 0; $i < count($v1); ++$i) {
$result[] = $v1[$i] + $v2[$i];
}
return $result;
}
function mulVectorNumber(array $v1, $num) {
$result = array();
for ($i = 0; $i < count($v1); ++$i) {
$result[] = $v1[$i] * $num;
}
return $result;
}
function score(array $input) {
return mulVectorNumber(array(1, 2), 1);
}
"""
interpreter = PhpInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_vector_num_expr():
expr = ast.BinVectorNumExpr(
ast.VectorVal([ast.NumVal(1), ast.NumVal(2)]),
ast.NumVal(1),
ast.BinNumOpType.MUL)
expected_code = """
function Add-Vectors([double[]] $v1, [double[]] $v2) {
[int] $length = $v1.Length
[double[]] $result = @(0) * $length
for ([int] $i = 0; $i -lt $length; ++$i) {
$result[$i] = $v1[$i] + $v2[$i]
}
return $result
}
function Mul-Vector-Number([double[]] $v1, [double] $num) {
[int] $length = $v1.Length
[double[]] $result = @(0) * $length
for ([int] $i = 0; $i -lt $length; ++$i) {
$result[$i] = $v1[$i] * $num
}
return $result
}
function Score([double[]] $InputVector) {
return Mul-Vector-Number $(@($(1), $(2))) $(1)
}
"""
interpreter = PowershellInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_vector_num_expr():
expr = ast.BinVectorNumExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]), ast.NumVal(1),
ast.BinNumOpType.MUL)
interpreter = GoInterpreter()
expected_code = """
func addVectors(v1, v2 []float64) []float64 {
result := make([]float64, len(v1))
for i := 0; i < len(v1); i++ {
result[i] = v1[i] + v2[i]
}
return result
}
func mulVectorNumber(v1 []float64, num float64) []float64 {
result := make([]float64, len(v1))
for i := 0; i < len(v1); i++ {
result[i] = v1[i] * num
}
return result
}
func score(input []float64) []float64 {
return mulVectorNumber([]float64{1.0, 2.0}, 1.0)
}"""
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_vector_num_expr():
expr = ast.BinVectorNumExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]), ast.NumVal(1),
ast.BinNumOpType.MUL)
interpreter = interpreters.CInterpreter()
expected_code = """
void assign_array(double source[], double *target, int size) {
for(int i = 0; i < size; ++i)
target[i] = source[i];
}
void add_vectors(double *v1, double *v2, int size, double *result) {
for(int i = 0; i < size; ++i)
result[i] = v1[i] + v2[i];
}
void mul_vector_number(double *v1, double num, int size, double *result) {
for(int i = 0; i < size; ++i)
result[i] = v1[i] * num;
}
void score(double * input, double * output) {
double var0[2];
mul_vector_number((double[]){1, 2}, 1, 2, var0);
assign_array(var0, output, 2);
}"""
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_vector_num_expr():
expr = ast.BinVectorNumExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]), ast.NumVal(1),
ast.BinNumOpType.MUL)
expected_code = """
namespace ML {
public static class Model {
public static double[] Score(double[] input) {
return MulVectorNumber(new double[2] {1.0, 2.0}, 1.0);
}
private static double[] AddVectors(double[] v1, double[] v2) {
double[] result = new double[v1.Length];
for (int i = 0; i < v1.Length; ++i) {
result[i] = v1[i] + v2[i];
}
return result;
}
private static double[] MulVectorNumber(double[] v1, double num) {
double[] result = new double[v1.Length];
for (int i = 0; i < v1.Length; ++i) {
result[i] = v1[i] * num;
}
return result;
}
}
}
"""
interpreter = CSharpInterpreter()
assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_vector_num_expr():
expr = ast.BinVectorNumExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]), ast.NumVal(1),
ast.BinNumOpType.MUL)
expected_code = """
List<double> score(List<double> input) {
return mulVectorNumber([1.0, 2.0], 1.0);
}
List<double> addVectors(List<double> v1, List<double> v2) {
List<double> result = new List<double>(v1.length);
for (int i = 0; i < v1.length; i++) {
result[i] = v1[i] + v2[i];
}
return result;
}
List<double> mulVectorNumber(List<double> v1, double num) {
List<double> result = new List<double>(v1.length);
for (int i = 0; i < v1.length; i++) {
result[i] = v1[i] * num;
}
return result;
}
"""
interpreter = DartInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_multi_class():
estimator = xgboost.XGBClassifier(n_estimators=1, random_state=1,
max_depth=1)
estimator.fit(np.array([[1], [2], [3]]), np.array([1, 2, 3]))
assembler = assemblers.XGBoostModelAssembler(estimator)
actual = assembler.assemble()
exponent = ast.ExpExpr(
ast.SubroutineExpr(
ast.BinNumExpr(
ast.NumVal(0.5),
ast.NumVal(0.0),
ast.BinNumOpType.ADD)),
to_reuse=True)
exponent_sum = ast.BinNumExpr(
ast.BinNumExpr(exponent, exponent, ast.BinNumOpType.ADD),
exponent,
ast.BinNumOpType.ADD,
to_reuse=True)
softmax = ast.BinNumExpr(exponent, exponent_sum, ast.BinNumOpType.DIV)
expected = ast.VectorVal([softmax] * 3)
assert utils.cmp_exprs(actual, expected)
def test_bin_vector_num_expr():
expr = ast.BinVectorNumExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]), ast.NumVal(1),
ast.BinNumOpType.MUL)
interpreter = interpreters.JavascriptInterpreter()
expected_code = """
function score(input) {
return mulVectorNumber([1, 2], 1);
}
function addVectors(v1, v2) {
let result = new Array(v1.length);
for (let i = 0; i < v1.length; i++) {
result[i] = v1[i] + v2[i];
}
return result;
}
function mulVectorNumber(v1, num) {
let result = new Array(v1.length);
for (let i = 0; i < v1.length; i++) {
result[i] = v1[i] * num;
}
return result;
}
"""
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_multi_class_sigmoid_output_transform():
estimator = lightgbm.LGBMClassifier(n_estimators=1,
random_state=1,
max_depth=1,
sigmoid=0.5,
objective="ovr")
estimator.fit(np.array([[1], [2], [3]]), np.array([1, 2, 3]))
assembler = assemblers.LightGBMModelAssembler(estimator)
actual = assembler.assemble()
sigmoid = ast.BinNumExpr(
ast.NumVal(1),
ast.BinNumExpr(
ast.NumVal(1),
ast.ExpExpr(
ast.BinNumExpr(
ast.NumVal(0),
ast.BinNumExpr(ast.NumVal(0.5), ast.NumVal(-1.3862943611),
ast.BinNumOpType.MUL),
ast.BinNumOpType.SUB)), ast.BinNumOpType.ADD),
ast.BinNumOpType.DIV)
expected = ast.VectorVal([sigmoid] * 3)
assert utils.cmp_exprs(actual, expected)
def test_bin_vector_num_expr():
expr = ast.BinVectorNumExpr(ast.VectorVal([ast.NumVal(1),
ast.NumVal(2)]), ast.NumVal(1),
ast.BinNumOpType.MUL)
interpreter = JavaInterpreter()
expected_code = """
public class Model {
public static double[] score(double[] input) {
return mulVectorNumber(new double[] {1.0, 2.0}, 1.0);
}
private static double[] addVectors(double[] v1, double[] v2) {
double[] result = new double[v1.length];
for (int i = 0; i < v1.length; i++) {
result[i] = v1[i] + v2[i];
}
return result;
}
private static double[] mulVectorNumber(double[] v1, double num) {
double[] result = new double[v1.length];
for (int i = 0; i < v1.length; i++) {
result[i] = v1[i] * num;
}
return result;
}
}"""
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_count_exprs():
assert ast.count_exprs(
ast.BinNumExpr(ast.NumVal(1), ast.NumVal(2), ast.BinNumOpType.ADD)
) == 3
assert ast.count_exprs(
ast.ExpExpr(ast.NumVal(2))
) == 2
assert ast.count_exprs(
ast.VectorVal([
ast.NumVal(2),
ast.TanhExpr(ast.NumVal(3))
])
) == 4
assert ast.count_exprs(
ast.IfExpr(
ast.CompExpr(ast.NumVal(2), ast.NumVal(0), ast.CompOpType.GT),
ast.NumVal(3),
ast.NumVal(4),
)
) == 6
assert ast.count_exprs(ast.NumVal(1)) == 1
def test_multi_class_best_ntree_limit():
base_score = 0.5
estimator = xgboost.XGBClassifier(n_estimators=100,
random_state=1,
max_depth=1,
base_score=base_score)
estimator.best_ntree_limit = 1
utils.train_model_classification(estimator)
assembler = assemblers.XGBoostModelAssembler(estimator)
actual = assembler.assemble()
estimator_exp_class1 = ast.ExpExpr(ast.SubroutineExpr(
ast.BinNumExpr(
ast.NumVal(0.5),
ast.SubroutineExpr(
ast.IfExpr(
ast.CompExpr(ast.FeatureRef(2),
ast.NumVal(2.45000005), ast.CompOpType.GTE),
ast.NumVal(-0.0733167157), ast.NumVal(0.143414631))),
ast.BinNumOpType.ADD)),
to_reuse=True)
estimator_exp_class2 = ast.ExpExpr(ast.SubroutineExpr(
ast.BinNumExpr(
ast.NumVal(0.5),
ast.SubroutineExpr(
ast.IfExpr(
ast.CompExpr(ast.FeatureRef(2), ast.NumVal(2.45000005),
ast.CompOpType.GTE), ast.NumVal(0.0344139598),
ast.NumVal(-0.0717073306))), ast.BinNumOpType.ADD)),
to_reuse=True)
estimator_exp_class3 = ast.ExpExpr(ast.SubroutineExpr(
ast.BinNumExpr(
ast.NumVal(0.5),
ast.SubroutineExpr(
ast.IfExpr(
ast.CompExpr(ast.FeatureRef(3), ast.NumVal(1.6500001),
ast.CompOpType.GTE), ast.NumVal(0.13432835),
ast.NumVal(-0.0644444525))), ast.BinNumOpType.ADD)),
to_reuse=True)
exp_sum = ast.BinNumExpr(ast.BinNumExpr(estimator_exp_class1,
estimator_exp_class2,
ast.BinNumOpType.ADD),
estimator_exp_class3,
ast.BinNumOpType.ADD,
to_reuse=True)
expected = ast.VectorVal([
ast.BinNumExpr(estimator_exp_class1, exp_sum, ast.BinNumOpType.DIV),
ast.BinNumExpr(estimator_exp_class2, exp_sum, ast.BinNumOpType.DIV),
ast.BinNumExpr(estimator_exp_class3, exp_sum, ast.BinNumOpType.DIV)
])
assert utils.cmp_exprs(actual, expected)
def test_multi_output():
expr = ast.IfExpr(
ast.CompExpr(ast.NumVal(1), ast.NumVal(1), ast.CompOpType.EQ),
ast.VectorVal([ast.NumVal(1), ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3), ast.NumVal(4)]))
expected_code = """
def score(input):
if (1.0) == (1.0):
var0 = [1.0, 2.0]
else:
var0 = [3.0, 4.0]
return var0
"""
interpreter = interpreters.PythonInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_multi_output():
expr = ast.IfExpr(
ast.CompExpr(ast.NumVal(1), ast.NumVal(1), ast.CompOpType.EQ),
ast.VectorVal([ast.NumVal(1), ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3), ast.NumVal(4)]))
expected_code = """
func score(input []float64) []float64 {
var var0 []float64
if (1.0) == (1.0) {
var0 = []float64{1.0, 2.0}
} else {
var0 = []float64{3.0, 4.0}
}
return var0
}"""
interpreter = GoInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_bin_vector_expr():
expr = ast.BinVectorExpr(
ast.VectorVal([ast.NumVal(1), ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3), ast.NumVal(4)]),
ast.BinNumOpType.ADD)
interpreter = PythonInterpreter()
expected_code = """
def add_vectors(v1, v2):
return [sum(i) for i in zip(v1, v2)]
def mul_vector_number(v1, num):
return [i * num for i in v1]
def score(input):
return add_vectors([1.0, 2.0], [3.0, 4.0])
"""
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_multi_output():
expr = ast.IfExpr(
ast.CompExpr(ast.NumVal(1), ast.NumVal(1), ast.CompOpType.EQ),
ast.VectorVal([ast.NumVal(1), ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3), ast.NumVal(4)]))
expected_code = """
let score (input : double list) =
let func0 =
if ((1.0) = (1.0)) then
[1.0; 2.0]
else
[3.0; 4.0]
func0
"""
interpreter = FSharpInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def test_raw_array():
expr = ast.VectorVal([ast.NumVal(3), ast.NumVal(4)])
expected_code = """
func score(input []float64) []float64 {
return []float64{3.0, 4.0}
}"""
interpreter = GoInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def softmax(exprs):
exp_exprs = [ast.ExpExpr(e, to_reuse=True) for e in exprs]
exp_sum_expr = utils.apply_op_to_expressions(ast.BinNumOpType.ADD,
*exp_exprs,
to_reuse=True)
return ast.VectorVal([
ast.BinNumExpr(e, exp_sum_expr, ast.BinNumOpType.DIV)
for e in exp_exprs
])
def _assemble_leaf(self, node_id):
scores = self._tree.value[node_id][0]
if self._is_vector_output:
score_sum = scores.sum() or 1.0
outputs = [ast.NumVal(s / score_sum) for s in scores]
return ast.VectorVal(outputs)
else:
assert len(scores) == 1, "Unexpected number of outputs"
return ast.NumVal(scores[0])
def test_multi_output():
expr = ast.IfExpr(
ast.CompExpr(ast.NumVal(1), ast.NumVal(1), ast.CompOpType.EQ),
ast.VectorVal([ast.NumVal(1), ast.NumVal(2)]),
ast.VectorVal([ast.NumVal(3), ast.NumVal(4)]))
expected_code = """
def score(input)
if (1.0) == (1.0)
var0 = [1.0, 2.0]
else
var0 = [3.0, 4.0]
end
var0
end
"""
interpreter = RubyInterpreter()
assert_code_equal(interpreter.interpret(expr), expected_code)
def test_raw_array():
expr = ast.VectorVal([ast.NumVal(3), ast.NumVal(4)])
expected_code = """
#include <string.h>
void score(double * input, double * output) {
memcpy(output, (double[]){3.0, 4.0}, 2 * sizeof(double));
}"""
interpreter = interpreters.CInterpreter()
utils.assert_code_equal(interpreter.interpret(expr), expected_code)
def _assemble_multi_class_output(self, trees):
# Multi-class output is calculated based on discussion in
# https://github.com/dmlc/xgboost/issues/1746#issuecomment-295962863
splits = _split_trees_by_classes(trees, self._output_size)
base_score = self._base_score
exprs = [self._assemble_single_output(t, base_score) for t in splits]
proba_exprs = utils.softmax_exprs(exprs)
return ast.VectorVal(proba_exprs)