def _split_into_subroutines(self, trees_ast, trees_num_leaves):
result = []
subroutine_trees = []
subroutine_sum_leaves = 0
for tree, num_leaves in zip(trees_ast, trees_num_leaves):
next_sum = subroutine_sum_leaves + num_leaves
if subroutine_trees and next_sum > self._leaves_cutoff_threshold:
# Exceeded the max leaves in the current subroutine,
# finalize this one and start a new one.
partial_result = utils.apply_op_to_expressions(
ast.BinNumOpType.ADD, *subroutine_trees)
result.append(ast.SubroutineExpr(partial_result))
subroutine_trees = []
subroutine_sum_leaves = 0
subroutine_sum_leaves += num_leaves
subroutine_trees.append(tree)
if subroutine_trees:
partial_result = utils.apply_op_to_expressions(
ast.BinNumOpType.ADD, *subroutine_trees)
result.append(ast.SubroutineExpr(partial_result))
return result
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 _linear_to_ast(coef, intercept):
feature_weight_mul_ops = [
utils.mul(ast.FeatureRef(index), ast.NumVal(value))
for index, value in enumerate(coef)
]
return utils.apply_op_to_expressions(ast.BinNumOpType.ADD,
ast.NumVal(intercept),
*feature_weight_mul_ops)
def _assemble_single_output(self, trees, base_score=0):
if self._tree_limit:
trees = trees[:self._tree_limit]
trees_ast = [self._assemble_tree(t) for t in trees]
result_ast = utils.apply_op_to_expressions(ast.BinNumOpType.ADD,
ast.NumVal(base_score),
*trees_ast)
return ast.SubroutineExpr(result_ast)
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.BinNumExpr(e, exp_sum_expr, ast.BinNumOpType.DIV)
for e in exp_exprs
]
def _rbf_kernel(self, support_vector):
elem_wise = [
ast.PowExpr(
utils.sub(ast.NumVal(support_element), ast.FeatureRef(i)),
ast.NumVal(2))
for i, support_element in enumerate(support_vector)
]
kernel = utils.apply_op_to_expressions(ast.BinNumOpType.ADD,
*elem_wise)
kernel = utils.mul(self._neg_gamma_expr, kernel)
return ast.ExpExpr(kernel)
def _assemble_single_output(self, estimator_params,
base_score=0, split_idx=0):
estimators_ast = self._assemble_estimators(estimator_params, split_idx)
tmp_ast = utils.apply_op_to_expressions(
ast.BinNumOpType.ADD,
ast.NumVal(base_score),
*estimators_ast)
result_ast = self._final_transform(tmp_ast)
return result_ast
def assemble(self):
trees = self.model.estimators_
def assemble_tree_expr(t):
assembler = TreeModelAssembler(t)
return assembler.assemble()
assembled_trees = [assemble_tree_expr(t) for t in trees]
return utils.apply_bin_op(
utils.apply_op_to_expressions(ast.BinNumOpType.ADD,
*assembled_trees),
ast.NumVal(1 / self.model.n_estimators), ast.BinNumOpType.MUL)
def assemble(self):
coef = 1.0 / self.model.n_estimators
trees = self.model.estimators_
def assemble_tree_expr(t):
assembler = TreeModelAssembler(t)
return utils.apply_bin_op(
ast.SubroutineExpr(assembler.assemble()),
ast.NumVal(coef),
ast.BinNumOpType.MUL)
assembled_trees = [assemble_tree_expr(t) for t in trees]
return utils.apply_op_to_expressions(
ast.BinNumOpType.ADD, *assembled_trees)
def _assemble_single_output(self):
support_vectors = self.model.support_vectors_
coef = self.model.dual_coef_[0]
intercept = self.model.intercept_[0]
kernel_exprs = self._apply_kernel(support_vectors)
kernel_weight_mul_ops = []
for index, value in enumerate(coef):
kernel_weight_mul_ops.append(
utils.mul(kernel_exprs[index], ast.NumVal(value)))
return utils.apply_op_to_expressions(ast.BinNumOpType.ADD,
ast.NumVal(intercept),
*kernel_weight_mul_ops)
def _assemble_single_output(self, idx=0):
support_vectors = self.model.support_vectors_
coef = self._get_single_coef(idx)
intercept = self._get_single_intercept(idx)
kernel_exprs = self._apply_kernel(support_vectors)
kernel_weight_mul_ops = [
utils.mul(kernel_exprs[index], ast.NumVal(value))
for index, value in enumerate(coef)
]
return utils.apply_op_to_expressions(ast.BinNumOpType.ADD,
ast.NumVal(intercept),
*kernel_weight_mul_ops)
def _cosine_kernel(self, support_vector):
support_vector_norm = np.linalg.norm(support_vector)
if support_vector_norm == 0.0:
support_vector_norm = 1.0
feature_norm = ast.SqrtExpr(utils.apply_op_to_expressions(
ast.BinNumOpType.ADD, *[
utils.mul(ast.FeatureRef(i), ast.FeatureRef(i))
for i in range(len(support_vector))
]),
to_reuse=True)
safe_feature_norm = ast.IfExpr(utils.eq(feature_norm, ast.NumVal(0.0)),
ast.NumVal(1.0), feature_norm)
kernel = self._linear_kernel(support_vector / support_vector_norm)
kernel = utils.div(kernel, safe_feature_norm)
return kernel
def test_linear_model():
# Default updater ("shotgun") is nondeterministic
estimator = xgb.XGBRegressor(n_estimators=2,
random_state=1,
updater="coord_descent",
feature_selector="shuffle",
booster="gblinear")
utils.get_regression_model_trainer()(estimator)
assembler = XGBoostModelAssemblerSelector(estimator)
actual = assembler.assemble()
feature_weight_mul = [
ast.BinNumExpr(ast.FeatureRef(0), ast.NumVal(-0.154567),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(1), ast.NumVal(0.0815865),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(2), ast.NumVal(-0.0979713),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(3), ast.NumVal(4.80472),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(4), ast.NumVal(1.35478),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(5), ast.NumVal(0.327222),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(6), ast.NumVal(0.0610654),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(7), ast.NumVal(0.46989),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(8), ast.NumVal(-0.0674318),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(9), ast.NumVal(-0.000506212),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(10), ast.NumVal(0.0732867),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(11), ast.NumVal(0.0108842),
ast.BinNumOpType.MUL),
ast.BinNumExpr(ast.FeatureRef(12), ast.NumVal(-0.140096),
ast.BinNumOpType.MUL),
]
expected = ast.BinNumExpr(
ast.NumVal(0.5),
apply_op_to_expressions(ast.BinNumOpType.ADD, ast.NumVal(11.138),
*feature_weight_mul), ast.BinNumOpType.ADD)
assert utils.cmp_exprs(actual, expected)
def _assemble_single_output(self, trees, base_score=0):
if self._tree_limit:
trees = trees[:self._tree_limit]
trees_ast = [ast.SubroutineExpr(self._assemble_tree(t)) for t in trees]
to_sum = trees_ast
# In a large tree we need to generate multiple subroutines to avoid
# java limitations https://github.com/BayesWitnesses/m2cgen/issues/103.
trees_num_leaves = [self._count_leaves(t) for t in trees]
if sum(trees_num_leaves) > self._leaves_cutoff_threshold:
to_sum = self._split_into_subroutines(trees_ast, trees_num_leaves)
tmp_ast = utils.apply_op_to_expressions(ast.BinNumOpType.ADD,
ast.NumVal(base_score),
*to_sum)
result_ast = self._final_transform(tmp_ast)
return ast.SubroutineExpr(result_ast)
def _linear_kernel(self, support_vector):
elem_wise = [
utils.mul(ast.NumVal(support_element), ast.FeatureRef(i))
for i, support_element in enumerate(support_vector)
]
return utils.apply_op_to_expressions(ast.BinNumOpType.ADD, *elem_wise)