secretflow.ml.boost.homo_boost.tree_core

secretflow.ml.boost.homo_boost.tree_core.criterion

Classes:

Criterion()	Base class for split criterion
XgboostCriterion([reg_lambda, reg_alpha, ...])	XgboostCriterion 分裂规则类 .

class secretflow.ml.boost.homo_boost.tree_core.criterion.Criterion

基类：ABC

Base class for split criterion

Methods:

split_gain(left_node_sum, right_node_sum)

abstract split_gain(left_node_sum, right_node_sum)

class secretflow.ml.boost.homo_boost.tree_core.criterion.XgboostCriterion(reg_lambda: float = 0.1, reg_alpha: float = 0, decimal: int = 10)

基类：Criterion

XgboostCriterion 分裂规则类 .. attribute:: reg_lambda

L2 regularization term on weight

reg_alpha

L1 regularization term on weight

decimal

truncate parms

Methods:

__init__([reg_lambda, reg_alpha, decimal])
split_gain(node_sum, left_node_sum, ...)	Calculate split gain :param node_sum: After the split, Grad and Hess at this node :param left_node_sum: After the split, Grad and Hess at the left split point :param right_node_sum: After the split, Grad and Hess at the right split point
truncate(f[, decimal])	Truncate control precision can reduce training time with early stop
node_gain(sum_grad, sum_hess)	Calculate node gain :param sum_grad: Sum of gradient :param sum_hess: Sum of hessian
node_weight(sum_grad, sum_hess)	Calculte node weight :param sum_grad: Sum of gradient :param sum_hess: Sum of hessian

__init__(reg_lambda: float = 0.1, reg_alpha: float = 0, decimal: int = 10)

split_gain(node_sum: Tuple[float, float], left_node_sum: Tuple[float, float], right_node_sum: Tuple[float, float]) → float

Calculate split gain :param node_sum: After the split, Grad and Hess at this node :param left_node_sum: After the split, Grad and Hess at the left split point :param right_node_sum: After the split, Grad and Hess at the right split point

返回:

Split gain of this split

返回类型:

gain

static truncate(f, decimal=10)

Truncate control precision can reduce training time with early stop

node_gain(sum_grad: float, sum_hess: float) → float

Calculate node gain :param sum_grad: Sum of gradient :param sum_hess: Sum of hessian

返回:

Gain of this node

node_weight(sum_grad: float, sum_hess: float) → float

Calculte node weight :param sum_grad: Sum of gradient :param sum_hess: Sum of hessian

返回:

Weight of this node

secretflow.ml.boost.homo_boost.tree_core.decision_tree

Classes:

DecisionTree([tree_param, data, ...])

Class for local version decision tree

class secretflow.ml.boost.homo_boost.tree_core.decision_tree.DecisionTree(tree_param: Optional[TreeParam] = None, data: Optional[DataFrame] = None, bin_split_points: Optional[ndarray] = None, tree_id: Optional[int] = None, group_id: Optional[int] = None, iter_round: Optional[int] = None, grad_key: str = 'grad', hess_key: str = 'hess', label_key: str = 'label')

基类：object

Class for local version decision tree

tree_param

params for tree build

data

training data, HdataFrame

bin_split_points

global binning infos

tree_id

tree id

group_id

group id indicates which class the tree classifies

iter_round

iteration round

hess_key

unique column name for hess value

grad_key

unique column name for grad value

Methods:

__init__([tree_param, data, ...])
feature_col_sample(all_features[, sample_rate])	Column sample for features :param all_features: A list of feature names for all columns :param sample_rate: subsample rate, a float-number in [0, 1]
get_feature_importance()
convert_bin_to_real()	convert bid to real value
columns_filter()
get_grad_hess_sum(data_frame)	calculate sum of grad and hess :param data_frame: data frame which contains hess and grad
update_feature_importance(split_info)	Calculate feature importance default split count :param split_info: Global optimal splitting information calculated from histogram
fit()	Entrance for local decision tree
update_tree(cur_to_split, split_info, ...)	Tree update function :param cur_to_split: List of nodes to be split :param split_info: Global optim split info :param cur_data_frames: List of dataframe in each node
init_xgboost_model(model_path)	Init standard xgboost model :param model_path: model path
save_xgboost_model(model_path, tree_nodes)	Transform tree info to standard xgboost model ref: https://xgboost.readthedocs.io/en/latest/dev/structxgboost_1_1TreeParam.html#aab8ff286e59f1bbab47bfa865da4a107 :param model_path: model path :param tree_nodes: federate decision tree internal model

__init__(tree_param: Optional[TreeParam] = None, data: Optional[DataFrame] = None, bin_split_points: Optional[ndarray] = None, tree_id: Optional[int] = None, group_id: Optional[int] = None, iter_round: Optional[int] = None, grad_key: str = 'grad', hess_key: str = 'hess', label_key: str = 'label')

feature_col_sample(all_features: List[str], sample_rate: float = 1.0)

Column sample for features :param all_features: A list of feature names for all columns :param sample_rate: subsample rate, a float-number in [0, 1]

返回:

A dict of valid features, which will be use in this round built

返回类型:

valid_features

get_feature_importance()

convert_bin_to_real()

convert bid to real value

columns_filter()

get_grad_hess_sum(data_frame)

calculate sum of grad and hess :param data_frame: data frame which contains hess and grad

返回:

sum of grad hess: sum of hess

返回类型:

grad

update_feature_importance(split_info)

Calculate feature importance default split count :param split_info: Global optimal splitting information calculated from histogram

fit()

Entrance for local decision tree

update_tree(cur_to_split: List[Node], split_info: List[SplitInfo], cur_data_frames: List[DataFrame])

Tree update function :param cur_to_split: List of nodes to be split :param split_info: Global optim split info :param cur_data_frames: List of dataframe in each node

返回:

List of nodes to be evaluated in the next iteration next_layer_data: List of data to be evaluated in the next iteration

返回类型:

next_layer_node

init_xgboost_model(model_path: str)

Init standard xgboost model :param model_path: model path

save_xgboost_model(model_path: str, tree_nodes: List[Node])

Transform tree info to standard xgboost model ref: https://xgboost.readthedocs.io/en/latest/dev/structxgboost_1_1TreeParam.html#aab8ff286e59f1bbab47bfa865da4a107 :param model_path: model path :param tree_nodes: federate decision tree internal model

返回:

update standard xgboost model on the model path

secretflow.ml.boost.homo_boost.tree_core.feature_histogram

Classes:

HistogramBag([histogram, hid, p_hid])	Histogram container
FeatureHistogram()	Feature Histogram

class secretflow.ml.boost.homo_boost.tree_core.feature_histogram.HistogramBag(histogram: Optional[List] = None, hid: int = -1, p_hid: int = -1)

基类：object

Histogram container

histogram

Histogram list calculated by calculate_histogram

Type:

List

hid

histogram id

Type:

int

p_hid

parent histogram id

Type:

int

Attributes:

histogram
hid
p_hid

Methods:

binary_op(other, func[, inplace])
__init__([histogram, hid, p_hid])

histogram: List = None

hid: int = -1

p_hid: int = -1

binary_op(other, func: callable, inplace: bool = False)

__init__(histogram: Optional[List] = None, hid: int = -1, p_hid: int = -1) → None

class secretflow.ml.boost.homo_boost.tree_core.feature_histogram.FeatureHistogram

基类：object

Feature Histogram

Methods:

calculate_histogram(data_frame_list, ...[, ...])	Calculate histogram according to G and H histogram: [cols,[buckets,[sum_g,sum_h,count]]
calculate_single_histogram(data, bin_split_point)

static calculate_histogram(data_frame_list: List[DataFrame], bin_split_points: ndarray, valid_features: Optional[Dict] = None, use_missing: bool = False, grad_key: str = 'grad', hess_key: str = 'hess', thread_pool: Optional[ThreadPoolExecutor] = None)

Calculate histogram according to G and H histogram: [cols,[buckets,[sum_g,sum_h,count]]

参数:

• data_frame_list – A list of data frame, which contain grad and hess

• bin_split_points – global split point dicts

• valid_features – valid feature names Dict[id:bool]

• use_missing – whether missing value participate in train

• grad_key – unique column name for grad value

• hess_key – unique column name for hess value

返回:

一个List[histogram1, histogram2, …]

返回类型:

node_histograms

static calculate_single_histogram(data: ndarray, bin_split_point: ndarray)

secretflow.ml.boost.homo_boost.tree_core.feature_importance

Classes:

FeatureImportance([main_importance, ...])

Feature importance class

class secretflow.ml.boost.homo_boost.tree_core.feature_importance.FeatureImportance(main_importance: float = 0, other_importance: float = 0, main_type: str = 'split')

基类：object

Feature importance class

main_importance

main importance value, ref main_type

other_importance

other importance value, ref opposite to main_type

main_type

type of importance, eg:gain

Methods:

__init__([main_importance, ...])
add_gain(val)
add_split(val)

__init__(main_importance: float = 0, other_importance: float = 0, main_type: str = 'split')

add_gain(val: float)

add_split(val: float)

secretflow.ml.boost.homo_boost.tree_core.loss_function

Classes:

LossFunction(obj_name)

Inner define for loss functions

class secretflow.ml.boost.homo_boost.tree_core.loss_function.LossFunction(obj_name: str)

基类：object

Inner define for loss functions

obj_name

Name of loss function in [“binary:logistic”,# logistic regression “reg:logistic”, # logistic regression for binary classification, output probability “multi:softmax”, # logistic regression for binary classification, output score before logistic transformation “multi:softprob”, # logistic regression for binary classification, output probability “reg:squarederror” # for multi label classification ]

Methods:

__init__(obj_name)
obj_function()

__init__(obj_name: str)

obj_function()

secretflow.ml.boost.homo_boost.tree_core.node

Classes:

Node([id, fid, bid, weight, is_leaf, ...])

Tree Node

class secretflow.ml.boost.homo_boost.tree_core.node.Node(id: Optional[int] = None, fid: Optional[int] = None, bid: Optional[int] = None, weight: float = 0.0, is_leaf: bool = False, sum_grad: Optional[float] = None, sum_hess: Optional[float] = None, left_nodeid: int = -1, right_nodeid: int = -1, missing_dir: int = 1, sample_num: int = 0, parent_nodeid: Optional[int] = None, is_left_node: bool = False, sibling_nodeid: Optional[int] = None, loss_change: float = 0.0)

基类：object

Tree Node

id

node id

Type:

int

fid

feature id

Type:

int

bid

bucket id

Type:

int

weight

node weight

Type:

float

is_leaf

whether this node is leaf

Type:

bool

sum_grad

sum of grad

Type:

float

sum_hess

sum of hess

Type:

float

left_nodeid

left node id

Type:

int

right_nodeid

right node id

Type:

int

missing_dir

which branch to go when encounting missing value default 1->right

Type:

int

sample_num

num of data sample

Type:

int

parent_nodeid

parent nodeid

Type:

int

is_left_node

is this node if left child of the parent

Type:

bool

sibling_nodeid

sibling node id

Type:

int

loss_change

the loss change.

Type:

float

Attributes:

id
fid
bid
weight
is_leaf
sum_grad
sum_hess
left_nodeid
right_nodeid
missing_dir
sample_num
parent_nodeid
is_left_node
sibling_nodeid
loss_change

Methods:

__init__([id, fid, bid, weight, is_leaf, ...])

id: int = None

fid: int = None

bid: int = None

weight: float = 0.0

is_leaf: bool = False

sum_grad: float = None

sum_hess: float = None

left_nodeid: int = -1

right_nodeid: int = -1

missing_dir: int = 1

sample_num: int = 0

parent_nodeid: int = None

is_left_node: bool = False

sibling_nodeid: int = None

loss_change: float = 0.0

__init__(id: Optional[int] = None, fid: Optional[int] = None, bid: Optional[int] = None, weight: float = 0.0, is_leaf: bool = False, sum_grad: Optional[float] = None, sum_hess: Optional[float] = None, left_nodeid: int = -1, right_nodeid: int = -1, missing_dir: int = 1, sample_num: int = 0, parent_nodeid: Optional[int] = None, is_left_node: bool = False, sibling_nodeid: Optional[int] = None, loss_change: float = 0.0) → None

secretflow.ml.boost.homo_boost.tree_core.splitter

Classes:

SplitInfo([best_fid, best_bid, sum_grad, ...])	Split Info .
Splitter(criterion_method[, ...])	Split Calculate Class .

class secretflow.ml.boost.homo_boost.tree_core.splitter.SplitInfo(best_fid: Optional[int] = None, best_bid: Optional[int] = None, sum_grad: float = 0, sum_hess: float = 0, gain: Optional[float] = None, missing_dir: int = 1, sample_count: int = -1)

基类：object

Split Info .. attribute:: best_fid

best split on feature id

type:

int

best_bid

best split on bucket id

Type:

int

sum_grad

sum of grad

Type:

float

sum_hess

sum of hess

Type:

float

gain

split gain

Type:

float

missing_dir

which branch to go when encounting missing value default 1->right

Type:

int

sample_count

num of sample after split

Type:

int

Attributes:

best_fid
best_bid
sum_grad
sum_hess
gain
missing_dir
sample_count

Methods:

__init__([best_fid, best_bid, sum_grad, ...])

best_fid: int = None

best_bid: int = None

sum_grad: float = 0

sum_hess: float = 0

gain: float = None

missing_dir: int = 1

sample_count: int = -1

__init__(best_fid: Optional[int] = None, best_bid: Optional[int] = None, sum_grad: float = 0, sum_hess: float = 0, gain: Optional[float] = None, missing_dir: int = 1, sample_count: int = -1) → None

class secretflow.ml.boost.homo_boost.tree_core.splitter.Splitter(criterion_method: str, criterion_params: List = [0, 0, 10], min_impurity_split: float = 0.01, min_sample_split: int = 2, min_leaf_node: int = 1, min_child_weight: int = 1)

基类：object

Split Calculate Class .. attribute:: criterion_method

criterion method

criterion_params

criterion parms, eg[l1: 0.1, l2: 0.2]

min_impurity_split

minimum gain threshold of splitting

min_sample_split

minimum sample split of splitting, default to 2

min_leaf_node

minimum samples on node to split

min_child_weight

minimum sum of hess after split

Methods:

__init__(criterion_method[, ...])
find_split_once(histogram, valid_features, ...)	Find best split info from histogram
find_split(histograms, valid_features[, ...])	查找最优分裂点 :param histograms: a list of histogram :param valid_features: valid feature names Dict[id:bool] :param use_missing: whether missing value participate in train
node_gain(grad, hess)
node_weight(grad, hess)
split_gain(sum_grad, sum_hess, sum_grad_l, ...)

__init__(criterion_method: str, criterion_params: List = [0, 0, 10], min_impurity_split: float = 0.01, min_sample_split: int = 2, min_leaf_node: int = 1, min_child_weight: int = 1)

find_split_once(histogram: List, valid_features: Dict, use_missing: bool) → SplitInfo

Find best split info from histogram

参数:

• histogram – a three-dimensional matrix store G,H,Count

• valid_features – valid feature names Dict[id:bool]

• use_missing – whether missing value participate in train

返回:

best split point info

返回类型:

SplitInfo

find_split(histograms: List, valid_features: Dict, use_missing: bool = False) → List[SplitInfo]

查找最优分裂点 :param histograms: a list of histogram :param valid_features: valid feature names Dict[id:bool] :param use_missing: whether missing value participate in train

返回:

best split info on each node

返回类型:

tree_node_splitinfo

node_gain(grad: float, hess: float) → float

node_weight(grad: float, hess: float) → float

split_gain(sum_grad: float, sum_hess: float, sum_grad_l: float, sum_hess_l: float, sum_grad_r: float, sum_hess_r: float) → float