List of TODOs before first release about treelite HOT 13 CLOSED

@yupbank In fact, the link is quite relevant in that someone already suggested a Javascript backend.

from treelite.

Any progress on the R package yet? @hcho3

from treelite.

No R package planned?

from treelite.

@pommedeterresautee For now I'm pushing its addition to after the public release.

from treelite.

If there is native xgboost support, then likely we only need two functions, compile and predict, righ @hcho3 ?

from treelite.

@tqchen That's correct. The pred_transform function was referring to an internal interface. On the outside, you'd still have compile and predict. The latest commit adds the option to transform margins into probabilities.

I'm very close to finalizing Python interface (probably tonight).

from treelite.

this might be interesting for you ?
i can make the template for xgboost actually

https://github.com/nok/sklearn-porter/blob/master/sklearn_porter/classifier/RandomForestClassifier/__init__.py

from treelite.

@yupbank This is interesting indeed. Thanks for the link. I really like their use of templates to support multiple back-end languages.

One issue with XGBoost is that XGBoost models don't expose their internal representations on Python side. So you'd need to write C code to interact with XGBoost.

from treelite.

@hcho3: Can you share any sample prediction times for the treelite compiled models vs their reference implementations? I'm very interested in how the XGBoost Boston Housing samples from the documentation performs, for example. I realize this may vary wildly depending on the model, parameters, and architecture, but it would be very interesting to see a few data points. Thanks!

UPDATE: 7d8b88e

from treelite.

Was thinking to it recently too...
On R you can compile C++ on fly with Rcpp, does it make sense with this package?
In that case it would be much more easier to perform than on Python (because everything required is already managed by R devtools)

from treelite.

The big point about reticulate is that you are forcing R users to have a working Python installation. Am I right?

from treelite.

@pommedeterresautee Yea, I think it might not worth using it then given this project is primarily based on C unlike Keras, etc.

from treelite.

I don't have a definite time estimate yet, but an R package is definitely on my to-do list. (In fact, that is why treelite was designed to accommodate random forests.) I've been lately busy writing a workshop paper on treelite. I'm not exactly familiar with R, so it may take some time to add R support.

from treelite.