Repository navigation
FastTree LearningRate not settable thru arguments object #1983
Description
Issue
- What did you do?
Action<FastTreeBinaryClassificationTrainer.Arguments> argsFunc = (args) =>
{
args.LearningRates = 0.1;
};
var trainer = mlContext.BinaryClassification.Trainers.FastTree(advancedSettings: argsFunc);
-
What happened?
trainer.Args.LearningRate = the default learning rate of 0.2 -
What did you expect?
trainer.Args.LearningRate = 0.1
I think this happens because
if (Args.LearningRates != learningRate)
{
using (var ch = Host.Start($"Setting learning rate to: {learningRate} as supplied in the direct arguments."))
Args.LearningRates = learningRate;
}
in BoostingFastTree.cs
Activity
[-]FastTree LearningRate not settable thru arguments[/-][+]FastTree LearningRate not settable thru arguments object[/+]yaeldekel commentedon Jan 2, 2019
Hi @daholste , thank you for your question! I looked at the code, and it seems that in all the other arguments that can be specified directly (minLeaves, numTrees and minDataPointsInLeaves), the advanced settings Action overwrites the directly specified values. @sfilipi , do you know if this is intentional or not?
@daholste , until this issue is resolved, you can specify the learning rate directly and that should work.
sfilipi commentedon Jan 2, 2019
We had the directly specified values override the advanced settings, than decided to swap and to have the advancedSettings override the directly specified values. If this is not happening for this trainer, we should fix it.
daholste commentedon Jan 11, 2019
Also, for Fast Forest, the trainer constructor takes 'learning rate', but the arguments object does not have a property for learning rate. This differs from Fast Tree, LightGbm, and all other algos that accept learning rate, where learning rate is a field on the arguments object if it's a parameter that can be set. It would be awesome & really useful if learning rate could be added to the Fast Forest arguments object. Thank you!
singlis commentedon Jan 24, 2019
This issue was resolved with the PR above, so I am closing this issue.
At the same time @daholste mentioned not being able to set the learning rate on the Arguments for Fast Forest but can set the learning rate as part of the constructor. After investigating (thanks @yaeldekel ), there should not be a learningRate argument for FastForest -- this was added by mistake. I created a bug to remove learningRate from the FastForest constructor.
Here is the issue for reference: #2237
comment out fast forest trainer, per discussion on ML.NET open issue d…
code migration from machinelearning-automl to machinelearning repo fe…
comment out fast forest trainer, per discussion on ML.NET open issue d…