secretflow.ml.boost.sgb_v.factory.components#

Classes:

`OrderMapManager`()
`GradientEncryptor`()	Manage all encryptions related to y, gradients, hessians
`Sampler`()
`LeafManager`()
`DataPreprocessor`()
`ModelBuilder`()	Functions related to build models including making predictions and add trees
`LossComputer`()	Compute loss, gradients and hessians
`TreeTrainer`()
`NodeSelector`()
`LevelWiseCache`()
`Shuffler`()
`BucketSumCalculator`()
`SplitFinder`()
`SplitTreeBuilder`()
`LevelWiseTreeTrainer`()

class secretflow.ml.boost.sgb_v.factory.components.OrderMapManager[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(params)
`get_params`(params)
`set_devices`(devices)
`build_order_map`(x)
`get_order_map`()
`get_feature_buckets`()
`get_bucket_lists`(col_choices_list)
`compute_left_child_selects`(actor_index, ...)
`batch_query_split_points_each_party`(queries_list)
`batch_compute_left_child_selects_each_party`(...)

__init__() → None[源代码]#

show_params()[源代码]#

set_params(params: Dict)[源代码]#

get_params(params: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

build_order_map(x: FedNdarray)[源代码]#

get_order_map() → FedNdarray[源代码]#

get_feature_buckets() → List[PYUObject][源代码]#

get_bucket_lists(col_choices_list: List[PYUObject]) → List[PYUObject][源代码]#

compute_left_child_selects(actor_index: int, feature: int, split_point_index: int, sampled_indices: Optional[List[int]] = None) → PYUObject[源代码]#

batch_query_split_points_each_party(queries_list: List[PYUObject]) → List[PYUObject][源代码]#

batch_compute_left_child_selects_each_party(split_feature_buckets_each_party: List[PYUObject], sampled_indices: Optional[List[int]] = None) → List[PYUObject][源代码]#

class secretflow.ml.boost.sgb_v.factory.components.GradientEncryptor[源代码]#

基类：Component

Manage all encryptions related to y, gradients, hessians

Methods:

`__init__`()
`show_params`()
`set_devices`(devices)
`get_params`(params)
`set_params`(params)
`pack`(g, h)
`encrypt`(gh, tree_index)
`cache_to_workers`(encrypted_gh, gh)
`get_move_config`(pyu)

__init__()[源代码]#

show_params()[源代码]#

set_devices(devices: Devices)[源代码]#

get_params(params: dict)[源代码]#

set_params(params: dict)[源代码]#

pack(g: PYUObject, h: PYUObject) → PYUObject[源代码]#

encrypt(gh: PYUObject, tree_index: int) → HEUObject[源代码]#

cache_to_workers(encrypted_gh: HEUObject, gh: PYUObject) → Dict[PYU, Union[HEUObject, PYUObject]][源代码]#

get_move_config(pyu)[源代码]#

class secretflow.ml.boost.sgb_v.factory.components.Sampler[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(params)
`get_params`(params)
`set_devices`(_)
`generate_col_choices`(feature_buckets)	Generate column sample choices.
`generate_row_choices`(row_num)
`apply_vector_sampling`(x, indices)	Sample x for a single partition.
`apply_v_fed_sampling`(X[, row_choices, ...])	Sample X based on row choices and col choices.

__init__()[源代码]#

show_params()[源代码]#

set_params(params: dict)[源代码]#

get_params(params: dict)[源代码]#

set_devices(_: Devices)[源代码]#

generate_col_choices(feature_buckets: List[PYUObject]) → Tuple[List[PYUObject], List[PYUObject]][源代码]#

Generate column sample choices.

参数:: feature_buckets (List[PYUObject]) – Behind PYUObject is List[int], bucket num for each feature.
返回:: first list is column choices, second is total number of buckets after sampling
返回类型:: Tuple[List[PYUObject], List[PYUObject]]

generate_row_choices(row_num) → Union[None, ndarray][源代码]#

apply_vector_sampling(x: PYUObject, indices: Union[PYUObject, ndarray])[源代码]#: Sample x for a single partition. Assuming we have a column vector. Assume the indices was generated from row sampling by sampler

apply_v_fed_sampling(X: FedNdarray, row_choices: Union[None, ndarray, PYUObject] = None, col_choices: List[Union[None, ndarray, PYUObject]] = []) → FedNdarray[源代码]#

Sample X based on row choices and col choices. Assume the choices were generated by sampler.

参数:

X (FedNdarray) – Array to sample from
row_choices (Union[None, np.ndarray, PYUObject]) – row sampling choices. devices are assumed to be ordered as X.
col_choices (List[Union[None, np.ndarray,PYUObject]) – col sampling choices. devices are assumed to be ordered as X.

返回:

subsampled X shape (Tuple[int, int]): shape of X_sub

返回类型:

X_sub (FedNdarray)

class secretflow.ml.boost.sgb_v.factory.components.LeafManager[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(params)
`get_params`(params)
`set_devices`(devices)
`clear_leaves`()
`extend_leaves`(pruned_node_selects, ...)
`get_leaf_selects`()
`get_leaf_indices`()
`compute_leaf_weights`(g, h)

__init__() → None[源代码]#

show_params()[源代码]#

set_params(params: dict)[源代码]#

get_params(params: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

clear_leaves()[源代码]#

extend_leaves(pruned_node_selects: List[ndarray], pruned_node_indices: List[int])[源代码]#

get_leaf_selects()[源代码]#

get_leaf_indices()[源代码]#

compute_leaf_weights(g, h)[源代码]#

class secretflow.ml.boost.sgb_v.factory.components.DataPreprocessor[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(_)
`get_params`(_)
`set_devices`(_)
`validate`(dataset, label)

__init__() → None[源代码]#

show_params()[源代码]#

set_params(_)[源代码]#

get_params(_)[源代码]#

set_devices(_)[源代码]#

validate(dataset, label) → Tuple[FedNdarray, Tuple[int, int], PYUObject, Tuple[int, int]][源代码]#

class secretflow.ml.boost.sgb_v.factory.components.ModelBuilder[源代码]#

基类：Component

Functions related to build models including making predictions and add trees

Methods:

`__init__`()
`show_params`()
`set_params`(params)
`get_params`(params)
`set_devices`(devices)
`init_pred`(sample_num)
`init_model`()
`insert_tree`(tree)
`get_tree_num`()
`finish`()

__init__()[源代码]#

show_params()[源代码]#

set_params(params: dict)[源代码]#

get_params(params: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

init_pred(sample_num: int) → PYUObject[源代码]#

init_model()[源代码]#

insert_tree(tree: DistributedTree)[源代码]#

get_tree_num() → int[源代码]#

finish() → SgbModel[源代码]#

class secretflow.ml.boost.sgb_v.factory.components.LossComputer[源代码]#

基类：Component

Compute loss, gradients and hessians

Methods:

`__init__`()
`show_params`()
`set_params`(params)
`get_params`(params)
`set_devices`(devices)
`compute_gh`(y, pred)

__init__() → None[源代码]#

show_params()[源代码]#

set_params(params: dict)[源代码]#

get_params(params: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

compute_gh(y: Union[PYUObject, ndarray], pred: Union[PYUObject, ndarray]) → Tuple[PYUObject, PYUObject][源代码]#

class secretflow.ml.boost.sgb_v.factory.components.TreeTrainer[源代码]#

基类：Composite

Methods:

`show_params`()
`set_params`(params)
`set_devices`(devices)
`train_tree`(cur_tree_num, order_map_manager, ...)	train on training data

show_params()[源代码]#

set_params(params: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

abstract train_tree(cur_tree_num, order_map_manager, y, pred, x_shape) → DistributedTree[源代码]#: train on training data

class secretflow.ml.boost.sgb_v.factory.components.NodeSelector[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(_)
`get_params`(_)
`set_devices`(devices)
`root_select`(sample_num)
`is_list_empty`(any_list)
`pick_children_node_ss`(node_select_list)
`get_child_select`(nodes_s, lchild_ss, ...)	compute the next level's sample select indices.

__init__() → None[源代码]#

show_params()[源代码]#

set_params(_: dict)[源代码]#

get_params(_: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

root_select(sample_num)[源代码]#

is_list_empty(any_list: Union[PYUObject, List]) → PYUObject[源代码]#

pick_children_node_ss(node_select_list: PYUObject) → Tuple[List[PYUObject], List[bool], int][源代码]#

get_child_select(nodes_s: List[ndarray], lchild_ss: List[ndarray], gain_is_cost_effective: List[bool], split_node_indices: List[int]) → Tuple[List[ndarray], List[int], List[ndarray], List[int]][源代码]#

compute the next level’s sample select indices.

参数:

nodes_s – List[np.ndarray]. sample select indices of each node from current level’s nodes.
lchilds_ss – List[np.ndarray]. left children’s sample selects idx for current level’s nodes. A non-empty single sample select is a np.ndarray with the shape n_samples * 1 and with entries being 0 and 1s. 1 indicates the sample remains in node.
gain_is_cost_effective – List[bool]. indicate whether node should be split.
split_node_indices – List[int]. node indices at the current level.

返回:

sample select indices for nodes in next tree level. node indices for the next level sample_selects for pruned nodes node_indices for pruned nodes

class secretflow.ml.boost.sgb_v.factory.components.LevelWiseCache[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(_)
`get_params`(_)
`set_devices`(devices)
`reset_level_caches`()
`collect_level_node_GH`(worker, bucket_sums, ...)
`get_level_nodes_GH`(worker)
`update_level_cache`(is_last_level, ...)

__init__()[源代码]#

show_params()[源代码]#

set_params(_: dict)[源代码]#

get_params(_: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

reset_level_caches()[源代码]#

collect_level_node_GH(worker, bucket_sums, is_lefts)[源代码]#

get_level_nodes_GH(worker) → List[源代码]#

update_level_cache(is_last_level, gain_is_cost_effective)[源代码]#

class secretflow.ml.boost.sgb_v.factory.components.Shuffler[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(params)
`get_params`(params)
`set_devices`(devices)
`reset_shuffle_masks`()
`create_shuffle_mask`(worker_index, key, ...)
`unshuffle_split_buckets`(...)	unshuffle split buckets viewed by each parition

__init__()[源代码]#

show_params()[源代码]#

set_params(params: dict)[源代码]#

get_params(params: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

reset_shuffle_masks()[源代码]#

create_shuffle_mask(worker_index: int, key: int, bucket_list: List[PYUObject]) → List[int][源代码]#

unshuffle_split_buckets(split_buckets_parition_wise: List[PYUObject]) → List[PYUObject][源代码]#

unshuffle split buckets viewed by each parition

参数:: split_buckets_parition_wise (List[PYUObject]) – PYUObject is List[int], split buckets viewed from this partition
返回:: unshuffled split buckets
返回类型:: List[List[PYUObject]]

class secretflow.ml.boost.sgb_v.factory.components.BucketSumCalculator[源代码]#

基类：Composite

Methods:

`__init__`()
`show_params`()
`set_params`(_)
`get_params`(_)
`set_devices`(devices)
`calculate_bucket_sum_level_wise`(shuffler, ...)
`update_level_cache`(is_last_level, ...)

__init__()[源代码]#

show_params()[源代码]#

set_params(_: dict)[源代码]#

get_params(_: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

calculate_bucket_sum_level_wise(shuffler: Shuffler, encrypted_gh_dict: Dict[PYU, HEUObject], children_split_node_selects: List[PYUObject], is_lefts: List[bool], order_map_sub: FedNdarray, bucket_num: int, bucket_lists: List[PYUObject], gradient_encryptor: GradientEncryptor, node_num: int) → Tuple[PYUObject, PYUObject][源代码]#

update_level_cache(is_last_level, gain_is_cost_effective)[源代码]#

class secretflow.ml.boost.sgb_v.factory.components.SplitFinder[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(params)
`get_params`(params)
`set_devices`(devices)
`find_best_splits`(G, H, tree_num, level)

__init__() → None[源代码]#

show_params()[源代码]#

set_params(params: dict)[源代码]#

get_params(params: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

find_best_splits(G: ndarray, H: ndarray, tree_num: int, level: int) → Tuple[ndarray, ndarray][源代码]#

class secretflow.ml.boost.sgb_v.factory.components.SplitTreeBuilder[源代码]#

基类：Component

Methods:

`__init__`()
`show_params`()
`set_params`(_)
`get_params`(_)
`set_devices`(devices)
`reset`()
`set_col_choices_and_buckets`(col_choices, ...)
`split_bucket_to_partition`(split_buckets)	map split bucket to position in the partition or -1 if not in partition
`get_split_feature_list_wise_each_party`(...)	map the unmasked split buckets to feature and split point
`do_split_list_wise_each_party`(...)	insert split points to split trees
`do_split`(split_buckets, sampled_rows, ...)
`insert_split_trees_into_distributed_tree`(...)

__init__() → None[源代码]#

show_params()[源代码]#

set_params(_)[源代码]#

get_params(_)[源代码]#

set_devices(devices: Devices)[源代码]#

reset()[源代码]#

set_col_choices_and_buckets(col_choices: List[PYUObject], total_buckets: List[PYUObject], feature_buckets: List[PYUObject])[源代码]#

split_bucket_to_partition(split_buckets: PYUObject) → List[PYUObject][源代码]#

map split bucket to position in the partition or -1 if not in partition

参数:: split_buckets (PYUObject) – PYUObject is in fact a List[int].
返回:: each PYUObject is in fact a List[int]. split buckets viewed by each party
返回类型:: List[PYUObject]

get_split_feature_list_wise_each_party(un_shuffled_split_buckets_each_party: List[PYUObject]) → List[PYUObject][源代码]#

map the unmasked split buckets to feature and split point

参数:: split_buckets_each_party (List[PYUObject]) – split buckets viewed by each party. PYUOBject is a list of int. -1 if not here.
返回:: PYUObject is in fact a List[Union[None, Tuple[int, int]]], None if -1 else (feature_index, bucket_index) for split.
返回类型:: List[PYUObject]

do_split_list_wise_each_party(split_features: List[PYUObject], split_points: List[PYUObject], left_child_selects: List[PYUObject], gain_is_cost_effective: List[bool], node_indices: Union[List[int], PYUObject]) → List[List[int]][源代码]#

insert split points to split trees

参数:

split_features (List[PYUObject]) – party wise. each PYUObject is List[Tuple[int, int]]. len = node indices length.
split_points (List[PYUObject]) – : party wise. each PYUObject is List[float]. len = node indices length.
left_child_selects (List[PYUObject]) – party wise. each PYUObject is List[np.ndarray]
gain_is_cost_effective (List[bool]) – if gain is cost effective
node_indices (Union[List[int], PYUObject]) – node indices.

返回:

left child selects for the new split nodes.

返回类型:

left_child_selects

do_split(split_buckets: List[int], sampled_rows: List[int], gain_is_cost_effective: List[bool], node_indices: Union[List[int], PYUObject], shuffler: Shuffler, order_map_manager: OrderMapManager) → List[PYUObject][源代码]#

insert_split_trees_into_distributed_tree(distributed_tree: DistributedTree, leaf_node_indices: PYUObject)[源代码]#

class secretflow.ml.boost.sgb_v.factory.components.LevelWiseTreeTrainer[源代码]#

基类：TreeTrainer

Methods:

`__init__`()
`show_params`()
`set_params`(params)
`get_params`(params)
`set_devices`(devices)
`train_tree`(cur_tree_num, order_map_manager, ...)	train on training data

__init__() → None[源代码]#

show_params()[源代码]#

set_params(params: dict)[源代码]#

get_params(params: dict)[源代码]#

set_devices(devices: Devices)[源代码]#

train_tree(cur_tree_num, order_map_manager: OrderMapManager, y: PYUObject, pred: Union[PYUObject, ndarray], x_shape: Tuple[int, int]) → DistributedTree[源代码]#: train on training data