Eli Osherovich

Catboost reliably segfaults with a certain set of parameters. Sometimes it's a silent segfault, sometimes it is #1436 (a hairy error message). It depends on the random_state i set in...

My very wild guess is that the segfaults are related to (low values of) the l2 regularization weight. When I remove it from the optuna search space I do not...

Hi @Evgueni-Petrov-aka-espetrov , unfortunately it does not help. Here is the set of parameters that segfaults: ```json { "task_type": "CPU", "thread_count": 4, "random_state": 42, "learning_rate": 0.025, "l2_leaf_reg": 2.444914148737743e-06, "bootstrap_type": "Bernoulli",...

@Evgueni-Petrov-aka-espetrov sorry, cannot release company's data. Do you use fuzzy testing for catboost?

@Evgueni-Petrov-aka-espetrov That would require some non-negligible amount of work on my side.... Will do it only if absolutely necessary. Meanwhile I downgraded catboost to the version `1.0.2` and my segfaults...

Unfortunately this is true `1.0.2` is buggy too. Increased the number of trials in optuna and got a failure again.... ``` FAIL (2021-11-29T20:59:57.073412+0200): This should be unreachable catboost/private/libs/algo/tensor_search_helpers.cpp:99 GetSplit() failed...

After trying the recent versions `1.0.x`, which all had segmentation faults and/or the error above (are there two different problems?) I downgraded catboost to `0.26.1` and so far had no...

I will check if I can do something. The difficulty lies in the complex flow that predates training. The data undergoes a bunch of transformations before it enters catboost. Namely,...

@rs can you please comment. Thanks.

I can do this if there is a real need from the keras side.