Lightgbm model with ONNX backend is too slow in comparison to one with torch backend

The notebook LGBM-ONNX-example (URL) can be improved in several ways:

1. Add handy ‘pip install’ commands cell similar to other notebooks (for improving experience in Google Colab)
2. The example has high memory usage and also takes longer to run than the lightgbm’s pytorch backend example.

• The memory usage is higher than 16 GB and when run in Colab, etc, the kernel is killed.
• Similarly, the execution time also increases rapidly as we increase the number of samples.
• The above observations may be like that by design or may indicate a memory leak or inefficient implementations. Or is it some other factor? Any ideas welcome here. 😃

1. Fix typos in some of the cells (innitial_types vs input_types) so the users can execute the example.

Till we investigate and address the point 2, I propose that we reduce the number of samples in the example provided in the notebook so that it can be run in most systems that our users may use.

[interesaaat]

I did some profiling, and it looks that the problem is that the default opset (9) used to export into ONNX produces bad models. With opset 11 the performance are quite close to PyTorch (for GPU at least). Issue #211 should fix the problem.

[KranthiGV]

Yes, as you said, the problem seems to be in onnx converter. Thanks for the pointers here. I will do the comparison first and we can then file a bug appropriately.