FastTreeRankingTrainer.Options 类

定义

public sealed class FastTreeRankingTrainer.Options : Microsoft.ML.Trainers.FastTree.BoostedTreeOptions, Microsoft.ML.Runtime.IComponentFactory<Microsoft.ML.ITrainer>
type FastTreeRankingTrainer.Options = class
    inherit BoostedTreeOptions
    interface IComponentFactory<ITrainer>
    interface IComponentFactory
Public NotInheritable Class FastTreeRankingTrainer.Options
Inherits BoostedTreeOptions
Implements IComponentFactory(Of ITrainer)
继承
实现
IComponentFactory IComponentFactory<Microsoft.ML.ITrainer>

构造函数

FastTreeRankingTrainer.Options()

创建一个具有默认值的新 FastTreeRankingTrainer.Options 对象。

字段

AllowEmptyTrees

如果无法进行根拆分,则允许继续训练。

(继承自 TreeOptions)
BaggingExampleFraction

每个包中使用的训练示例的百分比。 默认值为 0.7 (70%) 。

(继承自 TreeOptions)
BaggingSize

每个袋子中的树数 (0,用于禁用装袋) 。

(继承自 TreeOptions)
BestStepRankingRegressionTrees

使用最佳回归步骤树的选项。

(继承自 BoostedTreeOptions)
Bias

用于计算分类特征的每个特征箱的梯度的偏差。

(继承自 TreeOptions)
Bundling

捆绑低填充箱。Bundle.None (0) : no bundleling, Bundle.AggregateLowPopulation (1) : Bundle low population, Bundle.Neighbort (2) : Neighbor low population bundle.

(继承自 TreeOptions)
CategoricalSplit

是否根据多个分类特征值执行拆分。

(继承自 TreeOptions)
CompressEnsemble

压缩树系综。

(继承自 TreeOptions)
CustomGains

与每个相关性标签关联的收益的逗号分隔列表。

DiskTranspose

在执行转置时,是利用磁盘还是数据的本机转置设施 () 。

(继承自 TreeOptions)
DropoutRate

树正则化的删除率。

(继承自 BoostedTreeOptions)
EnablePruning

启用训练后树修剪以避免过度拟合。 它需要验证集。

(继承自 BoostedTreeOptions)
EntropyCoefficient

熵 (正则化) 系数介于 0 和 1 之间。

(继承自 TreeOptions)
ExampleWeightColumnName

用于示例权重的列。

(继承自 TrainerInputBaseWithWeight)
ExecutionTime

打印执行时间细分为 ML.NET 通道。

(继承自 TreeOptions)
FeatureColumnName

要用于功能的列。

(继承自 TrainerInputBase)
FeatureFirstUsePenalty

该特征首先使用惩罚系数。

(继承自 TreeOptions)
FeatureFlocks

是否在数据集准备期间对特征进行集合以加快训练速度。

(继承自 TreeOptions)
FeatureFraction

(随机选择) 用于每次迭代的特征部分。 如果只需要 90% 的功能,请使用 0.9。 较低的数字有助于减少过度拟合。

(继承自 TreeOptions)
FeatureFractionPerSplit

特征的分数 (随机选择) 用于每次拆分。 如果值为 0.9,则会降低 90% 的所有功能。

(继承自 TreeOptions)
FeatureReusePenalty

特征重新使用正则化) 系数 (惩罚。

(继承自 TreeOptions)
FeatureSelectionSeed

活动特征选择的种子。

(继承自 TreeOptions)
FilterZeroLambdas

在训练期间筛选零 lambda。

(继承自 BoostedTreeOptions)
GainConfidenceLevel

树拟合获取置信度要求。 仅当其可能性与随机选择增益高于此值时,才考虑增益。

(继承自 TreeOptions)
GetDerivativesSampleRate

在 GetDerivatives 函数中对每个查询采样 1(以 k 为单位)。

(继承自 BoostedTreeOptions)
HistogramPoolSize

池中的直方图数 (介于 2 和 numLeaves) 之间。

(继承自 TreeOptions)
LabelColumnName

用于标签的列。

(继承自 TrainerInputBaseWithLabel)
LearningRate

学习速率。

(继承自 BoostedTreeOptions)
MaximumBinCountPerFeature

每个特征的最大不同值(箱)数。

(继承自 TreeOptions)
MaximumCategoricalGroupCountPerNode

对分类功能进行拆分时要考虑的最大分类拆分组数。 拆分组是拆分点的集合。 这用于在存在许多分类特征时减少过度拟合。

(继承自 TreeOptions)
MaximumCategoricalSplitPointCount

在分类特征上进行拆分时要考虑的最大分类拆分点。

(继承自 TreeOptions)
MaximumNumberOfLineSearchSteps

括号后行搜索步骤数。

(继承自 BoostedTreeOptions)
MaximumTreeOutput

单个树输出的绝对值的上限。

(继承自 BoostedTreeOptions)
MemoryStatistics

将内存统计信息打印到 ML.NET 通道。

(继承自 TreeOptions)
MinimumExampleCountPerLeaf

形成新树叶所需的最小数据点数。

(继承自 TreeOptions)
MinimumExampleFractionForCategoricalSplit

要考虑拆分的箱中的最小分类示例百分比。 默认值为所有训练示例的 0.1%。

(继承自 TreeOptions)
MinimumExamplesForCategoricalSplit

要考虑拆分的箱中的最小分类示例计数。

(继承自 TreeOptions)
MinimumStepSize

最小行搜索步骤大小。

(继承自 BoostedTreeOptions)
NdcgTruncationLevel

LambdaMAR 算法中使用的最大 NDCG 截断。

NumberOfLeaves

每个回归树中的最大叶数。

(继承自 TreeOptions)
NumberOfThreads

要使用的线程数。

(继承自 TreeOptions)
NumberOfTrees

在集成中创建的决策树总数。

(继承自 TreeOptions)
OptimizationAlgorithm

要使用的优化算法。

(继承自 BoostedTreeOptions)
PruningThreshold

修剪的容差阈值。

(继承自 BoostedTreeOptions)
PruningWindowSize

用于修剪的移动窗口大小。

(继承自 BoostedTreeOptions)
RandomStart

训练从 /r1) 确定的随机排序 (开始。

(继承自 BoostedTreeOptions)
RowGroupColumnName

用于示例 groupId 的列。

(继承自 TrainerInputBaseWithGroupId)
Seed

随机数生成器的种子。

(继承自 TreeOptions)
Shrinkage

收缩。

(继承自 BoostedTreeOptions)
Smoothing

用于树正则化的平滑参数。

(继承自 TreeOptions)
SoftmaxTemperature

用于选择特征的随机 softmax 分布的温度。

(继承自 TreeOptions)
SparsifyThreshold

使用稀疏特征表示形式所需的稀疏级别。

(继承自 TreeOptions)
TestFrequency

每 k 轮计算训练/有效/测试的指标值。

(继承自 TreeOptions)
UseDcg

是否使用折扣累积增益 (DCG) 而不是规范化 DCG (NDCG) 进行训练。

UseLineSearch

确定是否对步长使用行搜索。

(继承自 BoostedTreeOptions)
UseTolerantPruning

使用窗口和容错进行修剪。

(继承自 BoostedTreeOptions)
WriteLastEnsemble

编写最后一个合奏,而不是通过提前停止确定的合奏。

(继承自 BoostedTreeOptions)

属性

EarlyStoppingMetric

提前停止指标。

EarlyStoppingRule

满足指定条件后,用于终止训练过程的提前停止规则。 可能的选择是 EarlyStoppingRuleBase的实现,例如 TolerantEarlyStoppingRuleGeneralityLossRule

(继承自 BoostedTreeOptions)

显式接口实现

IComponentFactory<ITrainer>.CreateComponent(IHostEnvironment)

FastTreeRankingTrainerFastTree 中使用的 选项 (选项)

适用于