- thrust::copy() called from dvec::copy() for gpairs invoked a GPU kernel instead of
cudaMemcpy()
- this resulted in illegal memory access if the GPU running the kernel could not access
the data being copied
- new version of dvec::copy() for thrust::device_ptr iterators calls cudaMemcpy(),
avoiding the problem.
* Added GPU objective function and no-copy interface.
- xgboost::HostDeviceVector<T> syncs automatically between host and device
- no-copy interfaces have been added
- default implementations just sync the data to host
and call the implementations with std::vector
- GPU objective function, predictor, histogram updater process data
directly on GPU
* Fix#2905
* Fix gpu_exact test failures
* Fix bug in GPU prediction where multiple calls to batch prediction can produce incorrect results
* Fix GPU documentation formatting
- Implement colsampling, subsampling for gpu_hist_experimental
- Optimised multi-GPU implementation for gpu_hist_experimental
- Make nccl optional
- Add Volta architecture flag
- Optimise RegLossObj
- Add timing utilities for debug verbose mode
- Bump required cuda version to 8.0
* Fatal error if GPU algorithm selected without GPU support compiled
* Resolve type conversion warnings
* Fix gpu unit test failure
* Fix compressed iterator edge case
* Fix python unit test failures due to flake8 update on pip
Problem:
Fast histogram updater crashes whenever subsampling picks zero rows
Diagnosis:
Row set data structure uses "nullptr" internally to indicate a non-existent
row set. Since you cannot take the address of the first element of an empty
vector, a valid row set ends up getting "nullptr" as well.
Fix:
Use an arbitrary value (not equal to "nullptr") to bypass nullptr check.
* SHAP values for feature contributions
* Fix commenting error
* New polynomial time SHAP value estimation algorithm
* Update API to support SHAP values
* Fix merge conflicts with updates in master
* Correct submodule hashes
* Fix variable sized stack allocation
* Make lint happy
* Add docs
* Fix typo
* Adjust tolerances
* Remove unneeded def
* Fixed cpp test setup
* Updated R API and cleaned up
* Fixed test typo
* [R] MSVC compatibility
* [GPU] allow seed in BernoulliRng up to size_t and scale to uint32_t
* R package build with cmake and CUDA
* R package CUDA build fixes and cleanups
* always export the R package native initialization routine on windows
* update the install instructions doc
* fix lint
* use static_cast directly to set BernoulliRng seed
* [R] demo for GPU accelerated algorithm
* tidy up the R package cmake stuff
* R pack cmake: installs main dependency packages if needed
* [R] version bump in DESCRIPTION
* update NEWS
* added short missing/sparse values explanations to FAQ
* Removal of redundant code/files.
* Removal of exact namespace in GPU plugin
* Revert double precision histograms to single precision for performance on Maxwell/Kepler
* repared serialization after update process; fixes#2545
* non-stratified folds in python could omit some data instances
* Makefile: fixes for older makes on windows; clean R-package too
* make cub to be a shallow submodule
* improve $(MAKE) recovery
* for MinGW, drop the 'lib' prefix from shared library name
* fix defines for 'g++ 4.8 or higher' to include g++ >= 5
* fix compile warnings
* [Appveyor] add MinGW with python; remove redundant jobs
* [Appveyor] also do python build for one of msvc jobs
It has been reported that new parallel algorithm (#2493) results in excessive
message usage (see issue #2326). Until issues are resolved, XGBoost should use
the old parallel algorithm by default. The user would have to specify
`enable_feature_grouping=1` manually to enable the new algorithm.
* Patch to improve multithreaded performance scaling
Change parallel strategy for histogram construction.
Instead of partitioning data rows among multiple threads, partition feature
columns instead. Useful heuristics for assigning partitions have been adopted
from LightGBM project.
* Add missing header to satisfy MSVC
* Restore max_bin and related parameters to TrainParam
* Fix lint error
* inline functions do not require static keyword
* Feature grouping algorithm accepting FastHistParam
Feature grouping algorithm accepts many parameters (3+), and it gets annoying to
pass them one by one. Instead, simply pass the reference to FastHistParam. The
definition of FastHistParam has been moved to a separate header file to
accomodate this change.
Use int32_t explicitly when serializing version field of dmatrix in binary
format. On ILP64 architectures, although very little, size of int is 64 bits.
* [gblinear] add features contribution prediction; fix DumpModel bug
* [gbtree] minor changes to PredContrib
* [R] add feature contribution prediction to R
* [R] bump up version; update NEWS
* [gblinear] fix the base_margin issue; fixes#1969
* [R] list of matrices as output of multiclass feature contributions
* [gblinear] make order of DumpModel coefficients consistent: group index changes the fastest
* Fix compilation on OS X with GCC 7
Compilation failed with
In file included from src/tree/tree_updater.cc:6:0:
include/xgboost/tree_updater.h:75:46: error: 'function' is not a member of 'std'
std::function<TreeUpdater* ()> > {
caused by a missing <functional> include.
* Fixed another occurence of that issue spotted by @ClimberPG
* Add prediction of feature contributions
This implements the idea described at http://blog.datadive.net/interpreting-random-forests/
which tries to give insight in how a prediction is composed of its feature contributions
and a bias.
* Support multi-class models
* Calculate learning_rate per-tree instead of using the one from the first tree
* Do not rely on node.base_weight * learning_rate having the same value as the node mean value (aka leaf value, if it were a leaf); instead calculate them (lazily) on-the-fly
* Add simple test for contributions feature
* Check against param.num_nodes instead of checking for non-zero length
* Loop over all roots instead of only the first
Reported in issue #2165. Dynamic scheduling of OpenMP loops involve
implicit synchronization. To implement synchronization, libgomp uses futex
(fast userspace mutex), whereas MinGW uses kernel-space mutex, which is more
costly. With chunk size of 1, synchronization overhead may become prohibitive
on Windows machines.
Solution: use 'guided' schedule to minimize the number of syncs
Storing and then loading a model loses any eval_metric that was
provided. This causes implementations that always store/load, like
xgboost4j-spark, to be unable to eval with the desired metric.
This log appears to fire every time I ask the python package to make a prediction. It's the only log that fires from XGBoost. When we're getting predictions on millions of items a day in production, this log seems out of place.
* Bugfix 1: Fix segfault in multithreaded ApplySplitSparseData()
When there are more threads than rows in rowset, some threads end up
with empty ranges, causing them to crash. (iend - 1 needs to be
accessible as part of algorithm)
Fix: run only those threads with nonempty ranges.
* Add regression test for Bugfix 1
* Moving python_omp_test to existing python test group
Turns out you don't need to set "OMP_NUM_THREADS" to enable
multithreading. Just add nthread parameter.
* Bugfix 2: Fix corner case of ApplySplitSparseData() for categorical feature
When split value is less than all cut points, split_cond is set
incorrectly.
Fix: set split_cond = -1 to indicate this scenario
* Bugfix 3: Initialize data layout indicator before using it
data_layout_ is accessed before being set; this variable determines
whether feature 0 is included in feat_set.
Fix: re-order code in InitData() to initialize data_layout_ first
* Adding regression test for Bugfix 2
Unfortunately, no regression test for Bugfix 3, as there is no
way to deterministically assign value to an uninitialized variable.
