Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
4,231 Commits 32 Branches 59 Tags
cd3d14ad0e4e851cc855d045db1800855347cd76
Commit Graph
353 Commits
Author SHA1 Message Date
ShvetsKSSHVETS, KIRILLfis
 		cd3d14ad0e Add float32 histogram (#5624) 
* new single_precision_histogram param was added.

Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com>
Co-authored-by: fis <jm.yuan@outlook.com>
 2020-06-03 11:24:53 +08:00
Rory Mitchell f779980f7e gpu_hist performance tweaks (#5707) 
* Remove device vectors

* Remove allreduce synchronize

* Remove double buffer
 2020-05-29 16:48:53 +12:00
Andy Adinets 646def51e0 C++14 for xgboost (#5664) 2020-05-21 12:26:40 +12:00
ShvetsKSandSHVETS, KIRILL dd01e4ba8d Distributed optimizations for 'hist' method with CPUs (#5557) 
Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com>
 2020-05-20 06:03:03 +03:00
Jiaming Yuan 535479e69f Add JSON schema to model dump. (#5660) 2020-05-15 10:18:43 +08:00
Oleksandr Kuvshynov 4e64e2ef8e skip missing lookup if nothing is missing in CPU hist partition kernel. (#5644) 
* [xgboost] skip missing lookup if nothing is missing
 2020-05-12 05:50:08 +03:00
Rory Mitchell fcf57823b6 Reduce device synchronisation (#5631) 
* Reduce device synchronisation

* Initialise pinned memory
 2020-05-07 21:19:46 +12:00
Jiaming Yuan eaf2a00b5c Enhance nvtx support. (#5636) 2020-05-06 22:54:24 +08:00
Rory Mitchell b9649e7b8e Refactor gpu_hist split evaluation (#5610) 
* Refactor

* Rewrite evaluate splits

* Add more tests
 2020-04-30 08:58:12 +12:00
Jiaming Yuan c90457f489 Refactor the CLI. (#5574) 
* Enable parameter validation.
* Enable JSON.
* Catch `dmlc::Error`.
* Show help message.
 2020-04-26 10:56:33 +08:00
Andy Adinets 73142041b9 For histograms, opting into maximum shared memory available per block. (#5491) 2020-04-21 14:56:42 +12:00
Rory Mitchell b2827a80e1 Use non-synchronising scan (#5560) 2020-04-20 15:51:34 +12:00
Rory Mitchell d6d1035950 gpu_hist performance fixes (#5558) 
* Remove unnecessary cuda API calls

* Fix histogram memory growth
 2020-04-19 12:21:13 +12:00
Jiaming Yuan c245eb8755 Fix r interaction constraints (#5543) 
* Unify the parsing code.

* Cleanup.
 2020-04-18 06:53:51 +08:00
ShvetsKSandSHVETS, KIRILL a2d86b8e4b Optimizations for RNG in InitData kernel (#5522) 
* optimizations for subsampling in InitData

* optimizations for subsampling in InitData

Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com>
 2020-04-16 18:24:32 +03:00
Rory Mitchell e268fb0093 Use thrust functions instead of custom functions (#5544) 2020-04-16 21:41:16 +12:00
Rory Mitchell ca4e05660e Purge device_helpers.cuh (#5534) 
* Simplifications with caching_device_vector

* Purge device helpers
 2020-04-15 21:51:56 +12:00
Jiaming Yuan 866a477319 Unify max nodes. (#5497) 2020-04-10 19:26:35 +08:00
Jiaming Yuan bd653fad4c Remove distcol updater. (#5507) 
Closes #5498.
 2020-04-10 12:52:56 +08:00
Jiaming Yuan 7d52c0b8c2 Requires setting leaf stat when expanding tree. (#5501) 
* Fix GPU Hist feature importance.
 2020-04-10 12:27:03 +08:00
Jiaming YuanandHyunsu Cho 0012f2ef93 Upgrade clang-tidy on CI. (#5469) 
* Correct all clang-tidy errors.
* Upgrade clang-tidy to 10 on CI.

Co-authored-by: Hyunsu Cho <chohyu01@cs.washington.edu>
 2020-04-05 04:42:29 +08:00
Jiaming Yuan 939973630d Accept other gradient types for split entry. (#5467) 2020-04-03 10:38:44 +08:00
ShvetsKS 27a8e36fc3 Reducing memory consumption for 'hist' method on CPU (#5334) 2020-03-28 14:45:52 +13:00
Jiaming Yuan ab7a46a1a4 Check whether current updater can modify a tree. (#5406) 
* Check whether current updater can modify a tree.

* Fix tree model JSON IO for pruned trees.
 2020-03-14 09:24:08 +08:00
Rory Mitchell b745b7acce Fix memory usage of device sketching (#5407) 2020-03-14 13:43:24 +13:00
Rory Mitchell 3ad4333b0e Partial rewrite EllpackPage (#5352) 2020-03-11 10:15:53 +13:00
Jiaming Yuan 8d06878bf9 Deterministic GPU histogram. (#5361) 
* Use pre-rounding based method to obtain reproducible floating point
  summation.
* GPU Hist for regression and classification are bit-by-bit reproducible.
* Add doc.
* Switch to thrust reduce for `node_sum_gradient`.
 2020-03-04 15:13:28 +08:00
Egor Smirnov 1b97eaf7a7 Optimized ApplySplit, BuildHist and UpdatePredictCache functions on CPU (#5244) 
* Split up sparse and dense build hist kernels.
* Add `PartitionBuilder`.
 2020-02-29 16:11:42 +08:00
Jiaming Yuan e0509b3307 Fix pruner. (#5335) 
* Honor the tree depth.
* Prevent pruning pruned node.
 2020-02-25 08:32:46 +08:00
Rory Mitchell b0ed3f0a66 Remove unnecessary DMatrix methods (#5324) 2020-02-25 12:40:39 +13:00
Jiaming Yuan 655cf17b60 Predict on Ellpack. (#5327) 
* Unify GPU prediction node.
* Add `PageExists`.
* Dispatch prediction on input data for GPU Predictor.
 2020-02-23 06:27:03 +08:00
Rong Ou e4b74c4d22 Gradient based sampling for GPU Hist (#5093) 
* Implement gradient based sampling for GPU Hist tree method.
* Add samplers and handle compacted page in GPU Hist.
 2020-02-04 10:31:27 +08:00
Egor Smirnov c67163250e Optimized BuildHist function (#5156) 2020-01-29 23:32:57 -08:00
Jiaming Yuan 7b65698187 Enforce correct data shape. (#5191) 
* Fix syncing DMatrix columns.
* notes for tree method.
* Enable feature validation for all interfaces except for jvm.
* Better tests for boosting from predictions.
* Disable validation on JVM.
 2020-01-13 15:48:17 +08:00
Egor Smirnov 7b17e76c5b Optimized EvaluateSplut function (#5138) 
* Add block based threading utilities.
 2019-12-31 18:18:42 +08:00
Jiaming Yuan 04db125699 Quick fix for memory leak in CPU Hist. (#5153) 
Closes https://github.com/dmlc/xgboost/issues/3579 .

* Don't use map.
 2019-12-31 14:05:53 +08:00
Jiaming Yuan 139ccc9902 Fix num_roots to be 1. (#5165) 2019-12-30 02:18:45 +08:00
Jiaming Yuan f3d7877802 Parameter validation (#5157) 
* Unused code.

* Split up old colmaker parameters from train param.

* Fix dart.

* Better name.
 2019-12-26 11:59:05 +08:00
Jiaming Yuan 3136185bc5 JSON configuration IO. (#5111) 
* Add saving/loading JSON configuration.
* Implement Python pickle interface with new IO routines.
* Basic tests for training continuation.
 2019-12-15 17:31:53 +08:00
Jiaming Yuan 208ab3b1ff Model IO in JSON. (#5110) 2019-12-11 11:20:40 +08:00
Jiaming Yuan e089e16e3d Pass pointer to model parameters. (#5101) 
* Pass pointer to model parameters.

This PR de-duplicates most of the model parameters except the one in
`tree_model.h`.  One difficulty is `base_score` is a model property but can be
changed at runtime by objective function.  Hence when performing model IO, we
need to save the one provided by users, instead of the one transformed by
objective.  Here we created an immutable version of `LearnerModelParam` that
represents the value of model parameter after configuration.
 2019-12-10 12:11:22 +08:00
Jiaming Yuan 1cb6bcc382 Remove dead code in colmaker. (#5105) 2019-12-10 09:32:37 +08:00
Egor Smirnov b1789b0346 added tracking execution time for UpdatePredictionCache function (#5107) 2019-12-10 01:32:56 +08:00
Jiaming Yuan 7ef5b78003 Implement JSON IO for updaters (#5094) 
* Implement JSON IO for updaters.

* Remove parameters in split evaluator.
 2019-12-07 00:24:00 +08:00
Jiaming Yuan 64af1ecf86 [Breaking] Remove num roots. (#5059) 2019-12-05 21:58:43 +08:00
Jiaming Yuan f3d8536702 Don't use 0 for "fresh leaf". (#5084) 
* Allow using right child as marker for Exact tree_method.
 2019-12-05 11:50:51 +08:00
Jiaming Yuan df9bdbbcb9 Fix parsing empty vector in parameter. (#5087) 2019-12-05 11:42:01 +08:00
Rong Ou 0afcc55d98 Support multiple batches in gpu_hist (#5014) 
* Initial external memory training support for GPU Hist tree method.
 2019-11-16 14:50:20 +08:00
Jiaming Yuan 97abcc7ee2 Extract interaction constraint from split evaluator. (#5034) 
*  Extract interaction constraints from split evaluator.

The reason for doing so is mostly for model IO, where num_feature and interaction_constraints are copied in split evaluator. Also interaction constraint by itself is a feature selector, acting like column sampler and it's inefficient to bury it deep in the evaluator chain. Lastly removing one another copied parameter is a win.

*  Enable inc for approx tree method.

As now the implementation is spited up from evaluator class, it's also enabled for approx method.

*  Removing obsoleted code in colmaker.

They are never documented nor actually used in real world. Also there isn't a single test for those code blocks.

*  Unifying the types used for row and column.

As the size of input dataset is marching to billion, incorrect use of int is subject to overflow, also singed integer overflow is undefined behaviour. This PR starts the procedure for unifying used index type to unsigned integers. There's optimization that can utilize this undefined behaviour, but after some testings I don't see the optimization is beneficial to XGBoost.
 2019-11-14 20:11:41 +08:00
Philip Hyunsu Cho f4e7b707c9 Revert #4529 (#5008) 
* Revert " Optimize ‘hist’ for multi-core CPU (#4529)"

This reverts commit 4d6590be3c.

* Fix build
 2019-11-12 09:35:03 -08:00