Upgrading xgboost 0.90 to 1

XGBoost Sklearn estimator method get_params no longer returns internally configured values.

Removes the grow_local_histmaker updater.

Removes single precision histogram due to floating point accuracy errors.

Deprecated CUDA architectures are no longer supported in the release binaries.

Replaces the rabit module with the new collective module, a drop-in replacement with runtime backend selection.

Removes the timeoutRequestWorkers parameter, no longer needed with barrier mode support.

Removes the label encoder deprecated in 1.3.

Removes the old callback deprecated in 1.3.

Removes lz4 compression during development of the new external memory interface.

Speeds up GPU histogram building 2-3 times on large datasets and removes the deterministic_histogram parameter, making the GPU algorithm always deterministic.

Removes the n_gpus parameter, which was deprecated in the 1.0 release.

Removes feature grouping in the CPU hist tree method, which had been disabled long ago.

Removes the magrittr dependency from the R package.

Renames all data parameters in prediction functions to X for sklearn estimator interface compliance.

Removes generation of pseudo feature names with DMatrix when inputs like np.ndarray lack column names.

Drops saving the binary format for memory snapshot.

Changes the default evaluation metric for the binary:logitraw objective to logloss.

Removes cupy.array_equal because it is not compatible with cuPy 7.8.

Removes the silent parameter; use verbosity instead.

Removes the Makefile; CMake is now used exclusively to build XGBoost, except the R package which uses Autotools.

Removes the long-broken distcol updater.

Removes SimpleCSRSource.

Removes unnecessary DMatrix methods.

Removes the n_gpus parameter; multi-GPU training now requires a distributed framework such as Dask or Spark.

Removes the deprecated gpu_exact training method; use gpu_hist instead.

Removes the deprecated learning_rates parameter in Python; use the callback API instead.

Removes the num_roots parameter, since training code always uses a single root node.

Removes GPU-specific objectives such as gpu:reg:linear; GPU is now used automatically when available.

Moves all fit parameters not related to input data into the sklearn estimator constructor, settable via set_params.

Changes the C API for Quantile DMatrix to be consistent with the new external memory implementation.

Disallows importing non-dask estimators from xgboost.dask.

Upgrades cuDF and RMM to 0.18 nightlies and requires RMM 0.18+ for the RMM plugin.

Changes DaskXGBClassifier.predict() to produce class predictions; use predict_proba() for probability predictions.

Custom evaluation metric now receives raw untransformed predictions and must transform them itself.

Sets output_margin to True for custom objectives in R and Python, returning raw predictions.

The C API function XGBoosterPredict() now requires an extra training parameter.

XGBoost now builds exclusively with CMake; the Makefile is being sunset, except for the R package which uses Autotools.

De-duplicates parameters and docstrings in the constructors of Scikit-Learn models.

Adds a new security policy.

Removes jackson from the JVM dependencies, fixing CVE-2020-36518.

Python 3.8 is now the minimum required Python version.

Requires Python 3.7 as the lowest supported Python version.

Starting in the next release, the Python package will require Pip 19.3+ due to the use of manylinux2014 tag.

XGBoost Python package now requires Python 3.6 and later.

XGBoost Python package now requires Pip 19.0 and higher.

Requires Python 3.5 or newer for the XGBoost Python package.

Deprecates use_label_encoder in XGBClassifier, affecting only the indicator parameter.

Removes an old warning introduced in 1.3.

Deprecates JVM constructors with implicit missing value due to confusing behaviors.

This release will be the last release to support CUDA 10.0.

Users are strongly encouraged to use xgb.save() and xgb.save.raw() instead of saveRDS() for persisting models.

Python 3.5 is deprecated; this is the last release supporting it, and 1.2.0 will require Python 3.6.

Scala 2.11 support in XGBoost4J may be dropped if a future release adopts Spark 3, which requires Scala 2.12+.

Deprecates the XGDMatrixSetGroup C API in favor of XGDMatrixSetUIntInfo.