Categories should be non-negative numbers ERROR

Hello everyone, I have been using Auto-Sklearn for about a month and am currently attempting to enter phase 2 of my research when I run into a problem. The error is that no problem will occur with a small amount of time left this task, but I know that after 30 minutes of running/searching in the spatial dimensions of possibilities, it crashed. I am also aware that some issues have been opened/closed regarding it, but none of them have actually aided me even though they presents the same kind of problem, as I have tried all of the suggested solutions without success. The following is the error:

raise ValueError('Categories should be non-negative numbers. ' ValueError: Categories should be non-negative numbers. NOTE: floats will be casted to integers.

To ensure the type of my columns in my dataset, before any run of the fit() method I am printing the dtype of each of my column and all of them are well “converted” in category / float64 or int64 which further in the Auto-Sklearn framework will be seen as “Categorical” and “Numerical”. However, the problem is still there and I do not know where to go deeper in the framework to find a solution to this problem. Thank you very much if you have anything else to suggest.

Steps to reproduce the behavior:

1. Using the version of Auto-Sklearn 0.12.6.
2. Using this parameter for the classifier:

time_left_for_this_task 180
per_run_time_limit 1
n_jobs 4
memory_limit 5000
seed 85
resampling_strategy holdout
ensemble_size 50

1. Using that dataset: https://archive.ics.uci.edu/ml/datasets/Estimation+of+obesity+levels+based+on+eating+habits+and+physical+condition+
2. Splitting (randomly or not it does not matter for the bug I guess) in 10 folds the dataset.
3. Looping over the folds and drop one which will be used as the test_set for later use (not relevant for the bug) and the remaining 9 folds are used for training.
4. 10 classifier are then outputted from the Auto-ML pipeline over the loop but one of them (at the end) crashed. Note: It is even more weirder than it does not crashed at the beginning.

About stracktrace and log file, I do not know where exactly to call

Please give details about your installation:

• OS: Mac OSX Big sur or Ubuntu 18 LTS.
• On my Mac OS installed or on an Ubuntu virtual machine.
• Python version 3.8.8 for both.
• Auto-sklearn version: 0.12.6.

[simonprovost]

Everything has been working very well since the update of the 0.13.0 as recommended. I tried over an extensive number of datasets plus hours of run (50 hours in total) and everything is being handled very well. Cheers all!

[eddiebergman]

@felidsche thank you for another fix for other OSX users!

Seeing as this was addressed in release 0.13.0 I will close this issue but please feel free to re-use this thread if the problem occurs on any version greater 0.13.0