RF/GBT SurrogateModel triggers unnecessary recalculation

[pfleck]

When the surrogate-storage mechanism is used for RFs and GBTs, the full model is still created after a run and passed to the surrogate model for immediate use. It currently is stored as a Lazy object: actualModel = new Lazy<IRandomForestModel>(() => model);

When a surrogate model is cloned, it checks whether the Lazy object is initialized and clones the full model if it is available:

IRandomForestModel clonedModel = null;
if (original.actualModel.IsValueCreated)
  clonedModel = cloner.Clone(original.ActualModel);
actualModel = new Lazy<IRandomForestModel>(() => clonedModel ?? RecalculateModel());

In case the cloning constructor is called immediately after creation of the surrogate Model, IsValueCreated is actually returning false since the Lazy object itself has not yet accessed its Value property. Thus the full model gets lost, and an unnecessary recalculation of the model is required later.

[mkommend]

r17272: Fixed cloning of RF & GBT surrogate models with already created models.

[pfleck]

Works as expected.

Code:

• Alternatively to the Lazy<T>, we could also have a look at System.Threading.LazyInitializer. This could help avoiding to store the full model twice (in the Lazy<T> and the fullModel field). This way we could also get rid of the CreateLazyInitFunc, which I don't really like.
• RandomForestModelSurrogate.cs
  • actualModel = CreateLazyInitFunc(); on line 86 is not necessary since it is already initialized in the this-constructor call.
• GradientBoostedTreesModelSurrogate.cs
  • actualModel = CreateLazyInitFunc(); on line 120 is not necessary since it is already initialized in the this-constructor call.

[pfleck]

r17278 Removed duplicated LazyInitFunc initialization.

[mkommend]

r17494: Merged r17272 and r17278 into stable.