This PR replaces the original RABIT implementation with a new one, which has already been partially merged into XGBoost. The new one features:
- Federated learning for both CPU and GPU.
- NCCL.
- More data types.
- A unified interface for all the underlying implementations.
- Improved timeout handling for both tracker and workers.
- Exhausted tests with metrics (fixed a couple of bugs along the way).
- A reusable tracker for Python and JVM packages.
- Add `numBoostedRound` to jvm packages
- Remove rabit checkpoint version.
- Change the starting version of training continuation in JVM [breaking].
- Redefine the checkpoint version policy in jvm package. [breaking]
- Rename the Python check point callback parameter. [breaking]
- Unifies the checkpoint policy between Python and JVM.
---------
Co-authored-by: Stephan T. Lavavej <stl@nuwen.net>
Co-authored-by: Jiaming Yuan <jm.yuan@outlook.com>
Co-authored-by: Joe <25804777+ByteSizedJoe@users.noreply.github.com>
* 1. Add parameters to set feature names and feature types
2. Save feature names and feature types to native json model
* Change serialization and deserialization format to ubj.
Previously, we use `libsvm` as default when format is not specified. However, the dmlc
data parser is not particularly robust against errors, and the most common type of error
is undefined format.
Along with which, we will recommend users to use other data loader instead. We will
continue the maintenance of the parsers as it's currently used for many internal tests
including federated learning.
* Use array interface for CSC matrix.
Use array interface for CSC matrix and align the interface with CSR and dense.
- Fix nthread issue in the R package DMatrix.
- Unify the behavior of handling `missing` with other inputs.
- Unify the behavior of handling `missing` around R, Python, Java, and Scala DMatrix.
- Expose `num_non_missing` to the JVM interface.
- Deprecate old CSR and CSC constructors.
This PR removes auto-detection of MUSL-based Linux systems in favor of system properties the user can set to configure a specific path for a native library.
* Add feature score support for linear model.
* Port R interface to the new implementation.
* Add linear model support in Python.
Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu>
* Add getNumFeature to the Java API
* Add getNumFeature to the Scala API
* Add unit tests for getNumFeature
Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu>
* fix type error
* Validate number of features.
* resolve comments
* add feature size for LabelPoint and DataBatch
* pass the feature size to native
* move feature size validating tests into a separate suite
* resolve comments
Co-authored-by: fis <jm.yuan@outlook.com>
* Add BigDenseMatrix
* ability to create DMatrix with bigger than Integer.MAX_VALUE size arrays
* uses sun.misc.Unsafe
* make DMatrix test work from a jar as well
* Add public group getter for java and scala
* Remove unnecessary param from javadoc
* Fix typo
* Fix another typo
* Add semicolon
* Fix javadoc return statement
* Fix missing return statement
* Add a unit test
* Make train in xgboost4j respect print params
Previously no setting in params argument of Booster::train would prevent
the Rabit.trackerPrint call. This can fill up a lot of screen space in
the case that many folds are being trained.
* Setting "silent" in this map to "true", "True", a non-zero integer, or
a string that can be parsed to such an int will prevent printing.
* Setting "verbose_eval" to "False" or "false" will prevent printing.
* Setting "verbose_eval" to an int (or a String parseable to an int) n
will result in printing every n steps, or no printing is n is zero.
This is to match the python behaviour described here:
https://www.kaggle.com/c/rossmann-store-sales/discussion/17499
* Fixed 'slient' typo in xgboost4j test
* private access on two methods
* Fix early stop with xgboost4j-spark
* Update XGBoost.java
* Update XGBoost.java
* Update XGBoost.java
To use -Float.MAX_VALUE as the lower bound, in case there is positive metric.
* Only update best score if the current score is better (no update when equal)
* Update xgboost-spark tutorial to fix early stopping docs.
* Updates to Booster to support other feature importances
* Add returns for Java methods
* Pass Scala style checks
* Pass Java style checks
* Fix indents
* Use class instead of enum
* Return map string double
* A no longer broken build, thanks to mvn package local build
* Add a unit test to increase code coverage back
* Address code review on main code
* Add more unit tests for different feature importance scores
* Address more CR
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* fix scalastyle error
* fix scalastyle error
* fix scalastyle error
* fix scalastyle error
* wrap iterators
* enable copartition training and validationset
* add parameters
* converge code path and have init unit test
* enable multi evals for ranking
* unit test and doc
* update example
* fix early stopping
* address the offline comments
* udpate doc
* test eval metrics
* fix compilation issue
* fix example
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* fix scalastyle error
* fix scalastyle error
* fix scalastyle error
* fix scalastyle error
* update version
* 0.82
* fix early stopping condition
* remove unused
* update comments
* udpate comments
* update test
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* add back train method but mark as deprecated
* fix scalastyle error
* fix scalastyle error
* fix scalastyle error
* fix scalastyle error
* temp
* add method for classifier and regressor
* update tutorial
* address the comments
* update
* [jvm-packages] Prevent dispose being called twice when finalize
* Convert SIGSEGV to XGBoostError
* Avoid creating a new SBooster with the same JBooster
* Address CR Comments
* [jvm-packages] Train Booster from an existing model
* Align Scala API with Java API
* Existing model should not load rabit checkpoint
* Address minor comments
* Implement saving temporary boosters and loading previous booster
* Add more unit tests for loadPrevBooster
* Add params to XGBoostEstimator
* (1) Move repartition out of the temp model saving loop (2) Address CR comments
* Catch a corner case of training next model with fewer rounds
* Address comments
* Refactor newly added methods into TmpBoosterManager
* Add two files which is missing in previous commit
* Rename TmpBooster to checkpoint
* Allowed subsampling test from the training data frame/RDD
The implementation requires storing 1 - trainTestRatio points in memory
to make the sampling work.
An alternative approach would be to construct the full DMatrix and then
slice it deterministically into train/test. The peak memory consumption
of such scenario, however, is twice the dataset size.
* Removed duplication from 'XGBoost.train'
Scala callers can (and should) use names to supply a subset of
parameters. Method overloading is not required.
* Reuse XGBoost seed parameter to stabilize train/test splitting
* Added early stopping support to non-distributed XGBoost
Closes#1544
* Added early-stopping to distributed XGBoost
* Moved construction of 'watches' into a separate method
This commit also fixes the handling of 'baseMargin' which previously
was not added to the validation matrix.
* Addressed review comments
* Converted ml.dmlc.xgboost4j.LabeledPoint to Scala
This allows to easily integrate LabeledPoint with Spark DataFrame APIs,
which support encoding/decoding case classes out of the box. Alternative
solution would be to keep LabeledPoint in Java and make it a Bean by
generating boilerplate getters/setters. I have decided against that, even
thought the conversion in this PR implies a public API change.
I also had to remove the factory methods fromSparseVector and
fromDenseVector because a) they would need to be duplicated to support
overloaded calls with extra data (e.g. weight); and b) Scala would expose
them via mangled $.MODULE$ which looks ugly in Java.
Additionally, this commit makes it possible to switch to LabeledPoint in
all public APIs and effectively to pass initial margin/group as part of
the point. This seems to be the only reliable way of implementing distributed
learning with these data. Note that group size format used by single-node
XGBoost is not compatible with that scenario, since the partition split
could divide a group into two chunks.
* Switched to ml.dmlc.xgboost4j.LabeledPoint in RDD-based public APIs
Note that DataFrame-based and Flink APIs are not affected by this change.
* Removed baseMargin argument in favour of the LabeledPoint field
* Do a single pass over the partition in buildDistributedBoosters
Note that there is no formal guarantee that
val repartitioned = rdd.repartition(42)
repartitioned.zipPartitions(repartitioned.map(_ + 1)) { it1, it2, => ... }
would do a single shuffle, but in practice it seems to be always the case.
* Exposed baseMargin in DataFrame-based API
* Addressed review comments
* Pass baseMargin to XGBoost.trainWithDataFrame via params
* Reverted MLLabeledPoint in Spark APIs
As discussed, baseMargin would only be supported for DataFrame-based APIs.
* Cleaned up baseMargin tests
- Removed RDD-based test, since the option is no longer exposed via
public APIs
- Changed DataFrame-based one to check that adding a margin actually
affects the prediction
* Pleased Scalastyle
* Addressed more review comments
* Pleased scalastyle again
* Fixed XGBoost.fromBaseMarginsToArray
which always returned an array of NaNs even if base margin was not
specified. Surprisingly this only failed a few tests.
* add back train method but mark as deprecated
* fix scalastyle error
* first commit in scala binding for fast histo
* java test
* add missed scala tests
* spark training
* add back train method but mark as deprecated
* fix scalastyle error
* local change
* first commit in scala binding for fast histo
* local change
* fix df frame test
* [jvm-packages] Scala implementation of the Rabit tracker.
A Scala implementation of RabitTracker that is interface-interchangable with the
Java implementation, ported from `tracker.py` in the
[dmlc-core project](https://github.com/dmlc/dmlc-core).
* [jvm-packages] Updated Akka dependency in pom.xml.
* Refactored the RabitTracker directory structure.
* Fixed premature stopping of connection handler.
Added a new finite state "AwaitingPortNumber" to explicitly wait for the
worker to send the port, and close the connection. Stopping the actor
prematurely sends a TCP RST to the worker, causing the worker to crash
on AssertionError.
* Added interface IRabitTracker so that user can switch implementations.
* Default timeout duration changes.
* Dependency for Akka tests.
* Removed the main function of RabitTracker.
* A skeleton for testing Akka-based Rabit tracker.
* waitFor() in RabitTracker no longer throws exceptions.
* Completed unit test for the 'start' command of Rabit tracker.
* Preliminary support for Rabit Allreduce via JNI (no prepare function support yet.)
* Fixed the default timeout duration.
* Use Java container to avoid serialization issues due to intermediate wrappers.
* Added tests for Allreduce/model training using Scala Rabit tracker.
* Added spill-over unit test for the Scala Rabit tracker.
* Fixed a typo.
* Overhaul of RabitTracker interface per code review.
- Removed methods start() waitFor() (no arguments) from IRabitTracker.
- The timeout in start(timeout) is now worker connection timeout, as tcp
socket binding timeout is less intuitive.
- Dropped time unit from start(...) and waitFor(...) methods; the default
time unit is millisecond.
- Moved random port number generation into the RabitTrackerHandler.
- Moved all Rabit-related classes to package ml.dmlc.xgboost4j.scala.rabit.
* More code refactoring and comments.
* Unified timeout constants. Readable tracker status code.
* Add comments to indicate that allReduce is for tests only. Removed all other variants.
* Removed unused imports.
* Simplified signatures of training methods.
- Moved TrackerConf into parameter map.
- Changed GeneralParams so that TrackerConf becomes a standalone parameter.
- Updated test cases accordingly.
* Changed monitoring strategies.
* Reverted monitoring changes.
* Update test case for Rabit AllReduce.
* Mix in UncaughtExceptionHandler into IRabitTracker to prevent tracker from hanging due to exceptions thrown by workers.
* More comprehensive test cases for exception handling and worker connection timeout.
* Handle executor loss due to unknown cause: the newly spawned executor will attempt to connect to the tracker. Interrupt tracker in such case.
* Per code-review, removed training timeout from TrackerConf. Timeout logic must be implemented explicitly and externally in the driver code.
* Reverted scalastyle-config changes.
* Visibility scope change. Interface tweaks.
* Use match pattern to handle tracker_conf parameter.
* Minor clarification in JNI code.
* Clearer intent in match pattern to suppress warnings.
* Removed Future from constructor. Block in start() and waitFor() instead.
* Revert inadvertent comment changes.
* Removed debugging information.
* Updated test cases that are a bit finicky.
* Added comments on the reasoning behind the unit tests for testing Rabit tracker robustness.
* Fixed BufferUnderFlow bug in decoding tracker 'print' command.
* Merge conflicts resolution.
* [jvm-packages] Scala implementation of the Rabit tracker.
A Scala implementation of RabitTracker that is interface-interchangable with the
Java implementation, ported from `tracker.py` in the
[dmlc-core project](https://github.com/dmlc/dmlc-core).
* [jvm-packages] Updated Akka dependency in pom.xml.
* Refactored the RabitTracker directory structure.
* Fixed premature stopping of connection handler.
Added a new finite state "AwaitingPortNumber" to explicitly wait for the
worker to send the port, and close the connection. Stopping the actor
prematurely sends a TCP RST to the worker, causing the worker to crash
on AssertionError.
* Added interface IRabitTracker so that user can switch implementations.
* Default timeout duration changes.
* Dependency for Akka tests.
* Removed the main function of RabitTracker.
* A skeleton for testing Akka-based Rabit tracker.
* waitFor() in RabitTracker no longer throws exceptions.
* Completed unit test for the 'start' command of Rabit tracker.
* Preliminary support for Rabit Allreduce via JNI (no prepare function support yet.)
* Fixed the default timeout duration.
* Use Java container to avoid serialization issues due to intermediate wrappers.
* Added tests for Allreduce/model training using Scala Rabit tracker.
* Added spill-over unit test for the Scala Rabit tracker.
* Fixed a typo.
* Overhaul of RabitTracker interface per code review.
- Removed methods start() waitFor() (no arguments) from IRabitTracker.
- The timeout in start(timeout) is now worker connection timeout, as tcp
socket binding timeout is less intuitive.
- Dropped time unit from start(...) and waitFor(...) methods; the default
time unit is millisecond.
- Moved random port number generation into the RabitTrackerHandler.
- Moved all Rabit-related classes to package ml.dmlc.xgboost4j.scala.rabit.
* More code refactoring and comments.
* Unified timeout constants. Readable tracker status code.
* Add comments to indicate that allReduce is for tests only. Removed all other variants.
* Removed unused imports.
* Simplified signatures of training methods.
- Moved TrackerConf into parameter map.
- Changed GeneralParams so that TrackerConf becomes a standalone parameter.
- Updated test cases accordingly.
* Changed monitoring strategies.
* Reverted monitoring changes.
* Update test case for Rabit AllReduce.
* Mix in UncaughtExceptionHandler into IRabitTracker to prevent tracker from hanging due to exceptions thrown by workers.
* More comprehensive test cases for exception handling and worker connection timeout.
* Handle executor loss due to unknown cause: the newly spawned executor will attempt to connect to the tracker. Interrupt tracker in such case.
* Per code-review, removed training timeout from TrackerConf. Timeout logic must be implemented explicitly and externally in the driver code.
* Reverted scalastyle-config changes.
* Visibility scope change. Interface tweaks.
* Use match pattern to handle tracker_conf parameter.
* Minor clarification in JNI code.
* Clearer intent in match pattern to suppress warnings.
* Removed Future from constructor. Block in start() and waitFor() instead.
* Revert inadvertent comment changes.
* Removed debugging information.
* Updated test cases that are a bit finicky.
* Added comments on the reasoning behind the unit tests for testing Rabit tracker robustness.
* create dmatrix with specified missing value
* update dmlc-core
* support for predict method in spark package
repartitioning
work around
* add more elements to work around training set empty partition issue
