package yggdrasil_decision_forests.model.distributed_decision_tree.proto

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: distributed_gradient_boosted_trees.proto.DistributedGradientBoostedTreesTrainingConfig

• optional float max_unbalance_ratio = 1

  Maximum relative difference in resource usage of a worker compared to the "balanced state" (i.e. all the workers are received the same amount of work). In other words, the memory / cpu usage of worker can increase by up to "max_unbalance_ratio" * "initial / default / balanced usage". For example, suppose a 100MB dataset splitted among 10 workers. In the balanced state, each worker receive 1/10 of the dataset i.e. 10MB of data. If `max_unbalance_ratio = 0.2`, a fast worker can receive up to 10MB * 1.2 = 12MB of data.

• optional int32 estimation_window_length = 2

  Number of consecutive observations used to estimate the worker's speed.

• optional int32 dynamic_balancing_frequency_iteration = 3

  Number of iterations in between each optimization rounds. Set -1 to disable.

• optional int32 dynamic_balancing_frequency_seconds = 4

  Number of seconds in between each optimization rounds. Set -1 to disable.

• optional int32 max_balancing_changes_per_dynamic_balancing = 5

  Maximum number of individual changes (i.e. transferring one feature from one worker to another) that can be done in a single balancing optimization. Note that the most optimization are applied, the worst the quality of the cost (and therefore the less accurate the balancing).

• optional float median_margin_ratio = 6

  Workers with an average running time greater than the median running time times "median_margin_ratio" are candidate for optimization i.e. have some of their features taken.

• optional LoadBalancerOptions.Internal internal = 8

Used in: LoadBalancerOptions

• optional bool random_dynamic_balancing = 1

  If true, continuously and randomly change the load balancing. For unit test only.

Definition and statistics of a given split.

Used in: SplitPerOpenNode

• optional decision_tree.proto.NodeCondition condition = 1

• optional decision_tree.proto.LabelStatistics label_statistics_neg = 2

• optional decision_tree.proto.LabelStatistics label_statistics_pos = 3

List of splits. One for each currently open node.

Used in: distributed_gradient_boosted_trees.proto.WorkerRequest.EvaluateSplits, distributed_gradient_boosted_trees.proto.WorkerResult.FindSplits

• repeated Split splits = 1

A "plan" is a partially ordered sequence of instructions of the form "worker i request the evaluation of split j to worker k". A plan allows to synchronize efficiently the evaluation split values of a set of features.

• repeated SplitSharingPlan.Round rounds = 1

  Rounds define the ordering in between the orders. Requests within rounds are not ordered i.e. can be executed in any order / in parallel.

Used in: Round, distributed_gradient_boosted_trees.proto.WorkerRequest.ShareSplits

• repeated RequestItem items = 1

• optional bool last_request_of_plan = 2

  If true, this is the last request received by this worker in this plan.

Used in: Request

• optional int32 src_worker = 1

• repeated int32 features = 2

• repeated RequestItem.Split splits = 3

Weak models + split idx that use this feature.

Used in: RequestItem, distributed_gradient_boosted_trees.proto.WorkerRequest.GetSplitValue

• optional int32 weak_model_idx = 1

• optional int32 split_idx = 2

Used in: SplitSharingPlan

• map<int32, Request> requests = 1

  Requests of a given destination worker. dst_worker_idx -> {Requests}. In other words, worker "dst_worker_idx" will request data to the "src_worker" workers.