In our experiment, origin lightgbm predict result is not equal treelite predict result about treelite HOT 12 CLOSED

@hcho3 well done!

from treelite.

when set the num_trees=1 the result is similar:
origin model result:

treelite retule:

when set num_trees=2 the result is begin different:
origin model result:

treelite result:

from treelite.

@henriezhang Is it possible to post your model?

from treelite.

my model is more than 2G, how post to you?

from treelite.

@henriezhang Dropbox or Google Drive link will work. I’ll make sure the bug is fixed.

from treelite.

please give me your email? I send you a simple model have the same problem

from treelite.

[email protected]

from treelite.

@hcho3 I have sanded you the model to your email.

from treelite.

@henriezhang #81 should fix the bug. Thanks so much for reporting!

from treelite.

On the same experiment parameters（num_tree=500, depth=20, leaves=255 ）, treelite cost more memory than before， It generate more source code than be before. And treelite write source code on disk after generate total source code. It may cause OOM .
Is there any way to solve this problem?
Thanks

from treelite.

@henuxhj This is because your model has high cardinality categorical features, and Treelite used to truncate high categorical values. So the larger code you get is the correct code (in the sense of giving correct prediction).

Please set parallel_comp option when compiling, to reduce memory consumption. This will break the 500 tree models into smaller pieces.

model.export_lib(toolchain='gcc', libpath='./a.so',
                verbose=True, params={'parallel_comp':500})

from treelite.

thank you so much， I'll try.

from treelite.