Early Stopping Training: ClassificationScoreCalculator not working

[Aulig]

Im using a MultiLayerNetwork for classification. The early stopping trainer always stops the training after the number of epochs specified for epochs without improvement (new ScoreImprovementEpochTerminationCondition(maxNoImprovementEpochs)). The problem however is that the model considered the best by the early stopping trainer (epoch 0) is by no means better than the latest model. If i switch out the ClassificationScoreCalculator for a DataSetLossCalculator it acts as expected. Here is a single class model you can use to see it:

https://pastebin.com/Edvbg9Rz
(dataset is attached to this issue)
This is the stats it achieved for me using Metric.ACCURACY :

Termination reason: EpochTerminationCondition
Termination details: ScoreImprovementEpochTerminationCondition(maxEpochsWithNoImprovement=100, minImprovement=0.0)
Total epochs: 101
Best epoch number: 0
Score at best epoch: 0.2467700258397933

-----"Best" (According to earlystopping training) Model-----

---Testing Dataset---
Accuracy: 0,2468

---Training Dataset---
Accuracy: 0,2735

-----Latest Model-----

---Testing Dataset---
Accuracy: 0,4910

---Training Dataset---
Accuracy: 0,4688

minimalTraining.txt
minimalTesting.txt

[AlexDBlack]

Thanks for reporting: this bug has been fixed here: #5565

You can access the fix by using snapshots: https://deeplearning4j.org/snapshots

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.