[jvm-packages]Fix early stopping condition #3928
Conversation
| for (int shift = 0; shift < earlyStoppingRounds - 1; shift++) { | ||
| // the initial value of onTrack is false and if the metrics in any of `earlyStoppingRounds` | ||
| // iterations goes to the expected direction, we should consider these `earlyStoppingRounds` | ||
| // as `onTrack` | ||
| onTrack |= maximizeEvaluationMetrics ? |
There was a problem hiding this comment.
Does this |= mean that, if the metric is moving in the right direction any two consecutive steps within the earlyStoppingRounds from the current iteration, then this method will return true?
This may not be what people normally expect from setting early stopping. For example, I'm getting a real training progress below with earlyStoppingSteps set to 20: the training should stop around iterations 120 since the maximum PR AUC was observed around iteration 100. But the current logic seems to look for any upward pieces within earlyStoppingSteps and keep training.
In the python-package(see here), training stops if the current iteration is earlyStoppingSteps away from the best iteration. Should the spark version be consistent with the python implementation?
There was a problem hiding this comment.
good catch, are you interested in filing a PR or an issue?
There was a problem hiding this comment.
I don't really write in Scala. So created this issue instead.
No description provided.