[BUG] Treelite format needs to be updated to accommodate 32-bit and 64-bit floats

[hcho3]

Describe the bug
#837 added Treelite as a submodule. Currently, Treelite uses single-precision floating-point (float) to store test thresholds and leaf outputs in decision trees. This is good for XGBoost models, but however, some other packages such as LightGBM use double-precision floating-point (double).

Steps/Code to reproduce bug
See dmlc/treelite#98, dmlc/treelite#95, where Treelite is not able to match prediction results for LightGBM models. Due to semantics of floating-point numbers, there is always truncation
whenever we convert between float and double.

As part of 1.0 release of Treelite (dmlc/treelite#111), I'd like to propose an update to the exchange format to support multiple data types. This way, the Treelite exchange format will represent LightGBM models faithfully. I'll post an update when a concrete proposal is ready.

@canonizer

[canonizer]

What about also supporting float16 for models?

[hcho3]

@canonizer Good idea. I also have a client who wants to use integer types (floating-point math is slow on microcontrollers)

[hcho3]

@canonizer Can you point me to a reference implementation of float16? The standard C does not provide 16-bit floating-point type.

Update. this seems promising.

[hcho3]

@canonizer I'm preparing a pull request to Treelite to enable multiple types (float32, float64, float16, int) for thresholds. Stay tuned.

[hcho3]

@canonizer @teju85 dmlc/treelite#130 performs refactoring of Treelite codebase to support multiple data types (for now, int32, float32, float64). Finally, LightGBM and scikit-learn models can be represented exactly. Note: I'm breaking compatibility for the Protobuf spec (src/tree.proto).

[teju85]

Thanks @hcho3 for pointing us to this. I have filed an issue in our repo to make sure this is not forgotten.

[hcho3]

Consolidating to #2160.