error: ‘inf’ undeclared with export_lib

[travisbrady]

Hi, treelite looks excellent, but so far I can't seem to get my lightgbm model to export correctly.
Any pointers?

In [10]: tl.export_lib(toolchain='gcc', libpath='tree_test.so', verbose=False)
[19:31:42] /io/treelite/src/compiler/recursive.cc:188: Parallel compilation disabled; all member trees will be dump to a single source file. This may increase compilation time and memory usage.
[19:31:42] /home/ubuntu/anaconda3/envs/lgb207/lib/python3.6/site-packages/treelite/contrib/__init__.py:99: WARNING: some of the source files are long. Expect long compilation time. You may want to adjust the parameter parallel_comp.

---------------------------------------------------------------------------
TreeliteError                             Traceback (most recent call last)
<ipython-input-10-7ed5ccaceb4c> in <module>()
----> 1 tl.export_lib(toolchain='gcc', libpath='tree_test.so', verbose=False)

~/anaconda3/envs/lgb207/lib/python3.6/site-packages/treelite/frontend.py in export_lib(self, toolchain, libpath, params, compiler, verbose, nthread, options)
     90       self.compile(temp_dir, params, compiler, verbose)
     91       temp_libpath = create_shared(toolchain, temp_dir, nthread,
---> 92                                    verbose, options)
     93       shutil.move(temp_libpath, libpath)
     94

~/anaconda3/envs/lgb207/lib/python3.6/site-packages/treelite/contrib/__init__.py in create_shared(toolchain, dirpath, nthread, verbose, options)
    112   else:
    113     raise ValueError('toolchain {} not supported'.format(toolchain))
--> 114   libpath = _create_shared(dirpath, recipe, nthread, options, verbose)
    115   if verbose:
    116     log_info(__file__, lineno(),

~/anaconda3/envs/lgb207/lib/python3.6/site-packages/treelite/contrib/gcc.py in _create_shared(dirpath, recipe, nthread, options, verbose)
     90         f.write(result[id]['stdout'] + '\n')
     91       raise TreeliteError('Error occured in worker #{}: '.format(id) +\
---> 92                           '{}'.format(result[id]['stdout']))
     93
     94   # 2. Package objects into a dynamic shared library (.so)

TreeliteError: Error occured in worker #0: tmpl86ac6yh.c: In function ‘predict_margin’:
tmpl86ac6yh.c:10072:61: error: ‘inf’ undeclared (first use in this function)
           if ( (data[1].missing != -1) && data[1].fvalue <= inf) {
                                                             ^
tmpl86ac6yh.c:10072:61: note: each undeclared identifier is reported only once for each function it appears in

[hcho3]

@travisbrady This is interesting. Can you attach the LightGBM model file? I'm really not sure how that 'inf' constant got there.

[travisbrady]

Absolutely, apologies for the delay. Had to anonymize this.
Model is here: https://gist.github.com/travisbrady/0dce5b62c9c2de5ee9c2900bc930099e

[hcho3]

@travisbrady Thanks! A quick inspection does reveal the occurrence of constant 1e+300 in 8 trees (Trees 34, 100, 143, 160, 173, 199, 203, 248). Since 1e+300 is way too big to be stored in a 32-bit float, it gets converted into inf (infinity).

According to this page, LightGBM uses 1e+300 to indicate an infinite threshold. I will push a bug fix soon to accommodate infinite thresholds.

Thanks a lot for taking time to report this problem.

[travisbrady]

@hcho3 No thank you so much for looking into this so quickly.
I'm really excited about treelite and would love to contribute. Incidentally, do you have any benchmarks? Perhaps comparing against xgboost, lightgbm and maybe https://github.com/ajtulloch/sklearn-compiledtrees ?

[hcho3]

@travisbrady We're currently working on it. Our goal is to have 1) deployment tutorial and 2) benchmarks, before the official launch (some time next month?) Thanks for your interest.

[hcho3]

@travisbrady I've pushed a bug fix to address the issue. A new binary wheel (version 0.1a10) will be available on PyPI within an hour or two.

[hcho3]

As a side note, it may take a really long time to compile your model. To reduce compile time, you should enable both nthread option (uses multiple cores) and parallel_comp option (internally splits C program into multiple pieces).

Example:

model.export_lib('gcc', libpath='./libtest.dylib',
                 params={'parallel_comp':4}, nthread=4, verbose=True)

On my machine (specification), the above line takes about 2 minutes to complete.

[hcho3]

The new wheel is now available. Install it by writing

pip install treelite==0.1a10