Fix using logloss for early stopping when task_type=GPU and eval_metric is not implemented on GPU.
Open
bug
GPU
python
- Dominant language
- C++
- Stars
- 9.1k
- Forks
- 1.3k
- PR merge metrics
- No merged PRs in 30d
Description
Problem:
It seems that currently when task_type='GPU' and eval_metric is not implemented on GPU (for example, AUC or BrierLoss), catboost silently uses logloss for early stopping (instead of user-supplied metric). In the screenshot you can see that in the widget UI best iteration is displayed correctly, but training is stopped too early and best_iteration_ is set incorrectly.
catboost version: 1.2.5
Operating System: Win
CPU: Intel. OK
GPU: Nvidia. **Problem**
Contributor guide
Assessment
This issue has not been assessed yet.