Tracking deprecated features.

[trivialfis]

Feature	PR(or time)	Replacement	Removed
GPU Objectives.	#3643	Now all objectives having GPU port will use GPU by default.	#4690
grow_fast_histmaker	#3836	Renamed to grow_quantile_histmaker.	✔
Some of R parameters documented in R/xgb.importance.R.	#1964	??	??
num_pbuffer_deprecated	??	??	✔
python/training.py: train::learning_rates	#1797	Now use use callback API.	#5155
python/sklearn.py: seed	??	Now use random_state.	#4929
python/sklearn.py: nthread	??	Now use n_jobs.	#4929
XGDMatrixCreateFromCSC	#1600	XGDMatrixCreateFromCSCEx	✔
XGDMatrixCreateFromCSR	#1600	XGDMatrixCreateFromCSR_omp	✔
silent	#3982	Now use verbosity	#5476
gpu_exact	#4527	Users are encouraged to use hist or approx on GPU.	#4742
n_gpus	#4749	Use distributed frameworks like dask or spark	#6821
reg:linear	#4267	Renamed to reg:squarederror	#4267
XGDMatrixSetGroup	#4864	Use XGDMatrixSetUIntInfo instead.	✘
TreeParam.max_depth	#5101	This variable is never used (TrainParam.max_depth is still supported)	#5101
Old Python callback functions	#6199	A new callback interface is designed.	#7280
Label encoder in XGBClassifier (use_label_encoder=True)	#6269	User should perform label encoding manually.	#7357
ntree_limit in Python	#6668	Use iteration_range or model slicing instead.	#8345 (2.0.0)
positional arguments in Python package	#6365	Use keyword arguments instead.	✘
use_gpu in PySpark	#9390 (2.0.0)	Use device instead.	✘
gpu_hist tree method	#9385 (2.0.0)	Use device instead.	✘
gpu_coord_descent updater	#9507 (2.0.0)	Use device instead.	✘
predictor (cpu_predictor, gpu_predictor)	#9129 (2.0.0)	Use device instead.	#9129 (2.0.0)
gpu_id tree method	#9385 (2.0.0)	Use device instead.	✘
fit parameters in the sklearn interface	#6751 (1.6.0)	Use set_params instead.	#9986 (2.1.0
Command line interface	#9485 (2.1.0)	Use other language bindings instead	✘
Reading files from URLs like S3	#9504 (2.1.0)	Use third-party libraries instead.	#9504 (2.1.0)
MPI support in RABIT	#9525 (2.1.0)	Use RABIT or NCCL instead (along with grpc for federated)	#9525 (2.1.0)
XGDMatrixSetDenseInfo	#10139 (2.1.0)	Use XGDMatrixSetInfoFromInterface	✘
manylinux2014 Python wheels	#10475 (2.1.0)	Use a recent Linux distribution and manylinux2_28 wheels instead	✘

[trivialfis]

Inspired by @thvasilo in #3929 (Thanks). Let's keep track of deprecated features here. Great chance that I missed something, anyone who notice please help adding it back. :) @hcho3 Please help filling the ?? if it's still possible (please don't spend too much time on it, it's deprecated), otherwise I will summit a PR to remove everything that loss track.

[hcho3]

Let's discuss whether num_roots should be kept or not. I think it is not documented and experimental and should be deprecated. Already, I decided not to support it when I first wrote hist updater.

[trivialfis]

@hcho3 I tried to find out what it does. It turns out num_roots is not documented. Is there any reference for it?

[RAMitchell]

Its always 1 in code. According to @tqchen it was intended to be used for multiclass trees (he can maybe confirm this). There is also something in metainfo to do with root index that is not documented or used - suggest also deprecating this.

[tqchen]

I think it is fine dropping them, but let us also keep the backward compatibility in mind. It might be useful to first introduce json exchange format, and then allow upgrade via json

[hcho3]

@trivialfis Why is XGDMatrixCreateFromCSC deprecated? It is used by the R wrapper.

@tqchen Got it

[hcho3]

Just found out that TreeParam.max_depth is never set and never used:

xgboost/include/xgboost/tree_model.h

Line 34 in 3d81c48

int max_depth;

This field is different from the TrainParam.max_depth, which is given at

xgboost/src/tree/param.h

Line 29 in 42bf90e

int max_depth;

Every occurrence of max_depth refers to TrainParam.max_depth, not TreeParam.max_depth. (Just verified with CLion)

It is not possible to remove max_depth from TreeParam right away, since TreeParam is serialized byte-wise and we cannot change the byte count. However, we can certainly remove TreeParam.max_depth from the new JSON serialization schema.

[trivialfis]

@tqchen

It might be useful to first introduce json exchange format

Yes, JSON is the priority. This issue is intended to be long term tracking.

Why is XGDMatrixCreateFromCSC deprecated? It is used by the R wrapper.

No idea. The document said so:

xgboost/include/xgboost/c_api.h

Line 187 in 42bf90e

* \deprecated

[trivialfis]

@hcho3 I'm only tracking the features that's already decided to be deprecated here, so that we can have a time frame for removing those code, as you can see some of them are really old and still lingering.

Let's move whether some other features needed to be deprecated into new issues. :)

[hcho3]

@trivialfis

I'm only tracking the features that's already decided to be deprecated here, so that we can have a time frame for removing those code

Got it. I'll file a new issue to deprecate TreeParam.max_depth.

No idea. The document said so

Turns out XGDMatrixCreateFromCSC is indeed deprecated, in favor of XGDMatrixCreateFromCSCEx.

[thvasilo]

There's an issue on the board about num_feature which I guess was used to subsample features? This is taken care of by colsample* now so I guess this parameter should be removed. @trivialfis should we open an issue to track this?

[trivialfis]

@hcho3 Actually I deprecated a lots of things in #5101, but there is no user visible effect, as all deprecated parameters are duplicated in some way. I just can't remove them as binary IO is still the default IO method.