Class prediction using the C API

[lnicola]

I have a model and an input dataset. I can run them fine using the command-line tool or the Python API. But when trying to use the C API, I get the wrong classes, while the raw values seem close to what the command-line tools produces. I'm probably using the API wrong, but I can't figure out at the moment.

See attached repro. Usage:

$ catboost calc -m model.cbm --input-path subset.csv --has-header --delimiter , --prediction-type Class
# notice how objects 361, 394, 399 have class 911

$ gcc test_catboost.c -o test_catboost -lcatboostmodel && ./test_catboost model.cbm
# notice how every class is 1

In addition, it seems that when trying to use a subset of the input data (the last 100 objects or so), it starts working.

repro.zip

[lnicola]

Looks like APT_PROBABILITY works fine. But I think I different results when running on a subset of the data, so there might still be something strange here.