diff --git a/.gitignore b/.gitignore index d454c6d1d61b..c6d3ab790179 100644 --- a/.gitignore +++ b/.gitignore @@ -2,7 +2,7 @@ *.slo *.lo *.o - +*.page # Compiled Dynamic libraries *.so *.dylib @@ -45,3 +45,11 @@ Debug *save *csv .Rproj.user +*.cpage.col +*.cpage +xgboost +xgboost.mpi +xgboost.mock +train* +rabit + diff --git a/Makefile b/Makefile index 3230661d4643..9716f8149146 100644 --- a/Makefile +++ b/Makefile @@ -1,8 +1,8 @@ export CC = gcc export CXX = g++ -export LDFLAGS= -pthread -lm - -export CFLAGS = -Wall -O3 -msse2 -Wno-unknown-pragmas -fPIC -pedantic +export MPICXX = mpicxx +export LDFLAGS= -Lrabit/lib -pthread -lm +export CFLAGS = -Wall -O3 -msse2 -Wno-unknown-pragmas -fPIC -Irabit/include ifeq ($(no_omp),1) CFLAGS += -DDISABLE_OPENMP @@ -10,33 +10,61 @@ else CFLAGS += -fopenmp endif +# by default use c++11 +ifeq ($(no_cxx11),1) +else + CFLAGS += +endif + # specify tensor path -BIN = xgboost -OBJ = updater.o gbm.o io.o +BIN = xgboost +MOCKBIN = xgboost.mock +OBJ = updater.o gbm.o io.o main.o +MPIBIN = xgboost.mpi SLIB = wrapper/libxgboostwrapper.so -.PHONY: clean all python Rpack +.PHONY: clean all mpi python Rpack librabit librabit_mpi -all: $(BIN) $(OBJ) $(SLIB) +all: $(BIN) $(OBJ) $(SLIB) $(MOCKBIN) +mpi: $(MPIBIN) + +# rules to get rabit library +librabit: + if [ ! -d rabit ]; then git clone https://github.com/tqchen/rabit.git; fi + cd rabit;make lib/librabit.a lib/librabit_mock.a; cd - +librabit_mpi: + if [ ! -d rabit ]; then git clone https://github.com/tqchen/rabit.git; fi + cd rabit;make lib/librabit_mpi.a; cd - python: wrapper/libxgboostwrapper.so # now the wrapper takes in two files. io and wrapper part -wrapper/libxgboostwrapper.so: wrapper/xgboost_wrapper.cpp $(OBJ) -updater.o: src/tree/updater.cpp src/tree/*.hpp src/*.h src/tree/*.h -gbm.o: src/gbm/gbm.cpp src/gbm/*.hpp src/gbm/*.h +updater.o: src/tree/updater.cpp src/tree/*.hpp src/*.h src/tree/*.h src/utils/*.h +gbm.o: src/gbm/gbm.cpp src/gbm/*.hpp src/gbm/*.h io.o: src/io/io.cpp src/io/*.hpp src/utils/*.h src/learner/dmatrix.h src/*.h -xgboost: src/xgboost_main.cpp src/utils/*.h src/*.h src/learner/*.hpp src/learner/*.h $(OBJ) -wrapper/libxgboostwrapper.so: wrapper/xgboost_wrapper.cpp src/utils/*.h src/*.h src/learner/*.hpp src/learner/*.h $(OBJ) +main.o: src/xgboost_main.cpp src/utils/*.h src/*.h src/learner/*.hpp src/learner/*.h +xgboost.mpi: updater.o gbm.o io.o main.o librabit_mpi +xgboost.mock: updater.o gbm.o io.o main.o librabit +xgboost: updater.o gbm.o io.o main.o librabit +wrapper/libxgboostwrapper.so: wrapper/xgboost_wrapper.cpp src/utils/*.h src/*.h src/learner/*.hpp src/learner/*.h updater.o gbm.o io.o librabit $(BIN) : - $(CXX) $(CFLAGS) $(LDFLAGS) -o $@ $(filter %.cpp %.o %.c, $^) + $(CXX) $(CFLAGS) -o $@ $(filter %.cpp %.o %.c, $^) $(LDFLAGS) -lrabit + +$(MOCKBIN) : + $(CXX) $(CFLAGS) -o $@ $(filter %.cpp %.o %.c, $^) $(LDFLAGS) -lrabit_mock $(SLIB) : - $(CXX) $(CFLAGS) -fPIC $(LDFLAGS) -shared -o $@ $(filter %.cpp %.o %.c, $^) + $(CXX) $(CFLAGS) -fPIC -shared -o $@ $(filter %.cpp %.o %.c, $^) $(LDFLAGS) -lrabit $(OBJ) : $(CXX) -c $(CFLAGS) -o $@ $(firstword $(filter %.cpp %.c, $^) ) +$(MPIOBJ) : + $(MPICXX) -c $(CFLAGS) -o $@ $(firstword $(filter %.cpp %.c, $^) ) + +$(MPIBIN) : + $(MPICXX) $(CFLAGS) -o $@ $(filter %.cpp %.o %.c, $^) $(LDFLAGS) -lrabit_mpi + install: cp -f -r $(BIN) $(INSTALL_PATH) @@ -62,4 +90,4 @@ Rpack: R CMD check --as-cran xgboost*.tar.gz clean: - $(RM) $(OBJ) $(BIN) $(SLIB) *.o */*.o */*/*.o *~ */*~ */*/*~ + $(RM) $(OBJ) $(BIN) $(MPIBIN) $(MPIOBJ) $(SLIB) *.o */*.o */*/*.o *~ */*~ */*/*~ diff --git a/R-package/src/Makevars b/R-package/src/Makevars index 44dce490e615..5cea1438c886 100644 --- a/R-package/src/Makevars +++ b/R-package/src/Makevars @@ -4,6 +4,4 @@ PKGROOT=../../ PKG_CPPFLAGS= -DXGBOOST_CUSTOMIZE_MSG_ -DXGBOOST_CUSTOMIZE_PRNG_ -DXGBOOST_STRICT_CXX98_ -I$(PKGROOT) PKG_CXXFLAGS= $(SHLIB_OPENMP_CFLAGS) PKG_LIBS = $(SHLIB_OPENMP_CFLAGS) -OBJECTS= xgboost_R.o xgboost_assert.o $(PKGROOT)/wrapper/xgboost_wrapper.o $(PKGROOT)/src/io/io.o $(PKGROOT)/src/gbm/gbm.o $(PKGROOT)/src/tree/updater.o - - +OBJECTS= xgboost_R.o xgboost_assert.o $(PKGROOT)/wrapper/xgboost_wrapper.o $(PKGROOT)/src/io/io.o $(PKGROOT)/src/gbm/gbm.o $(PKGROOT)/src/tree/updater.o $(PKGROOT)/src/sync/sync_empty.o diff --git a/R-package/src/Makevars.win b/R-package/src/Makevars.win index 289f1a15a435..5cea1438c886 100644 --- a/R-package/src/Makevars.win +++ b/R-package/src/Makevars.win @@ -4,4 +4,4 @@ PKGROOT=../../ PKG_CPPFLAGS= -DXGBOOST_CUSTOMIZE_MSG_ -DXGBOOST_CUSTOMIZE_PRNG_ -DXGBOOST_STRICT_CXX98_ -I$(PKGROOT) PKG_CXXFLAGS= $(SHLIB_OPENMP_CFLAGS) PKG_LIBS = $(SHLIB_OPENMP_CFLAGS) -OBJECTS= xgboost_R.o xgboost_assert.o $(PKGROOT)/wrapper/xgboost_wrapper.o $(PKGROOT)/src/io/io.o $(PKGROOT)/src/gbm/gbm.o $(PKGROOT)/src/tree/updater.o +OBJECTS= xgboost_R.o xgboost_assert.o $(PKGROOT)/wrapper/xgboost_wrapper.o $(PKGROOT)/src/io/io.o $(PKGROOT)/src/gbm/gbm.o $(PKGROOT)/src/tree/updater.o $(PKGROOT)/src/sync/sync_empty.o diff --git a/R-package/src/xgboost_R.cpp b/R-package/src/xgboost_R.cpp index 5a8ddbf52de7..b837d45fc6a9 100644 --- a/R-package/src/xgboost_R.cpp +++ b/R-package/src/xgboost_R.cpp @@ -4,10 +4,11 @@ #include #include #include -#include "xgboost_R.h" #include "wrapper/xgboost_wrapper.h" #include "src/utils/utils.h" #include "src/utils/omp.h" +#include "xgboost_R.h" + using namespace std; using namespace xgboost; @@ -290,4 +291,4 @@ extern "C" { UNPROTECT(1); return out; } -} \ No newline at end of file +} diff --git a/build.sh b/build.sh index 35a566cccf1c..f6fbcb74b808 100755 --- a/build.sh +++ b/build.sh @@ -3,6 +3,13 @@ # basically, it first try to make with OpenMP, if fails, disable OpenMP and make again # This will automatically make xgboost for MAC users who do not have openmp support # In most cases, type make will give what you want + +# download rabit +if [ ! -d rabit ]; then + git clone https://github.com/tqchen/rabit.git +else + cd rabit; git pull; cd .. +fi if make; then echo "Successfully build multi-thread xgboost" else diff --git a/demo/README.md b/demo/README.md index c1f9a210cd64..908cd062721e 100644 --- a/demo/README.md +++ b/demo/README.md @@ -32,6 +32,8 @@ This is a list of short codes introducing different functionalities of xgboost a [python](guide-python/cross_validation.py) [R](../R-package/demo/cross_validation.R) [Julia](https://github.com/antinucleon/XGBoost.jl/blob/master/demo/cross_validation.jl) +* Predicting leaf indices + [python](guide-python/predict_leaf_indices.py) Basic Examples by Tasks ==== diff --git a/demo/guide-python/README.md b/demo/guide-python/README.md index 3625c40f561f..bc1c219d0b1d 100644 --- a/demo/guide-python/README.md +++ b/demo/guide-python/README.md @@ -6,3 +6,4 @@ XGBoost Python Feature Walkthrough * [Predicting using first n trees](predict_first_ntree.py) * [Generalized Linear Model](generalized_linear_model.py) * [Cross validation](cross_validation.py) +* [Predicting leaf indices](predict_leaf_indices.py) diff --git a/demo/guide-python/predict_leaf_indices.py b/demo/guide-python/predict_leaf_indices.py new file mode 100755 index 000000000000..291ad1ee7831 --- /dev/null +++ b/demo/guide-python/predict_leaf_indices.py @@ -0,0 +1,22 @@ +#!/usr/bin/python +import sys +import numpy as np +sys.path.append('../../wrapper') +import xgboost as xgb + +### load data in do training +dtrain = xgb.DMatrix('../data/agaricus.txt.train') +dtest = xgb.DMatrix('../data/agaricus.txt.test') +param = {'max_depth':2, 'eta':1, 'silent':1, 'objective':'binary:logistic' } +watchlist = [(dtest,'eval'), (dtrain,'train')] +num_round = 3 +bst = xgb.train(param, dtrain, num_round, watchlist) + +print ('start testing predict the leaf indices') +### predict using first 2 tree +leafindex = bst.predict(dtest, ntree_limit=2, pred_leaf = True) +print leafindex.shape +print leafindex +### predict all trees +leafindex = bst.predict(dtest, pred_leaf = True) +print leafindex.shape diff --git a/demo/guide-python/runall.sh b/demo/guide-python/runall.sh index 2dd2c20b04fe..8f4f9832a6d4 100755 --- a/demo/guide-python/runall.sh +++ b/demo/guide-python/runall.sh @@ -4,4 +4,5 @@ python custom_objective.py python boost_from_prediction.py python generalized_linear_model.py python cross_validation.py -rm -rf *~ *.model *.buffer \ No newline at end of file +python predict_leaf_indices.py +rm -rf *~ *.model *.buffer diff --git a/multi-node/README.md b/multi-node/README.md new file mode 100644 index 000000000000..ce37daeab15d --- /dev/null +++ b/multi-node/README.md @@ -0,0 +1,36 @@ +Distributed XGBoost +====== +This folder contains information of Distributed XGBoost. + +* The distributed version is built on Rabit:[Reliable Allreduce and Broadcast Library](https://github.com/tqchen/rabit) + - Rabit is a portable library that provides fault-tolerance for Allreduce calls for distributed machine learning + - This makes xgboost portable and fault-tolerant against node failures +* You can run Distributed XGBoost on platforms including Hadoop(see [hadoop folder](hadoop)) and MPI + - Rabit only replies a platform to start the programs, so it should be easy to port xgboost to most platforms + +Build +===== +* In the root folder, run ```./build.sh```, this will give you xgboost, which uses rabit allreduce + +Notes +==== +* Rabit handles all the fault tolerant and communications efficiently, we only use platform specific command to start programs + - The Hadoop version does not rely on Mapreduce to do iterations + - You can expect xgboost not suffering the drawbacks of iterative MapReduce program +* The design choice was made because Allreduce is very natural and efficient for distributed tree building + - In current version of xgboost, the distributed version is only adds several lines of Allreduce synchronization code +* The multi-threading nature of xgboost is inheritated in distributed mode + - This means xgboost efficiently use all the threads in one machine, and communicates only between machines + - Remember to run on xgboost process per machine and this will give you maximum speedup +* For more information about rabit and how it works, see the [tutorial](https://github.com/tqchen/rabit/tree/master/guide) + +Solvers +===== +There are two solvers in distributed xgboost. You can check for local demo of the two solvers, see [row-split](row-split) and [col-split](col-split) + * Column-based solver split data by column, each node work on subset of columns, + it uses exactly the same algorithm as single node version. + * Row-based solver split data by row, each node work on subset of rows, + it uses an approximate histogram count algorithm, and will only examine subset of + potential split points as opposed to all split points. + - This is the mode used by current hadoop version, since usually data was stored by rows in many industry system + diff --git a/multi-node/col-split/README.md b/multi-node/col-split/README.md new file mode 100644 index 000000000000..3ea0799fead0 --- /dev/null +++ b/multi-node/col-split/README.md @@ -0,0 +1,19 @@ +Distributed XGBoost: Column Split Version +==== +* run ```bash mushroom-col-rabit.sh ``` + - mushroom-col-rabit.sh starts xgboost job using rabit's allreduce +* run ```bash mushroom-col-rabit-mock.sh ``` + - mushroom-col-rabit-mock.sh starts xgboost job using rabit's allreduce, inserts suicide signal at certain point and test recovery + +How to Use +==== +* First split the data by column, +* In the config, specify data file as containing a wildcard %d, where %d is the rank of the node, each node will load their part of data +* Enable column split mode by ```dsplit=col``` + +Notes +==== +* The code is multi-threaded, so you want to run one process per node +* The code will work correctly as long as union of each column subset is all the columns we are interested in. + - The column subset can overlap with each other. +* It uses exactly the same algorithm as single node version, to examine all potential split points. diff --git a/multi-node/col-split/mushroom-col-rabit-mock.sh b/multi-node/col-split/mushroom-col-rabit-mock.sh new file mode 100755 index 000000000000..65e62309af38 --- /dev/null +++ b/multi-node/col-split/mushroom-col-rabit-mock.sh @@ -0,0 +1,25 @@ +#!/bin/bash +if [[ $# -ne 1 ]] +then + echo "Usage: nprocess" + exit -1 +fi + +# +# This script is same as mushroom-col except that we will be using xgboost instead of xgboost-mpi +# xgboost used built in tcp-based allreduce module, and can be run on more enviroment, so long as we know how to start job by modifying ../submit_job_tcp.py +# +rm -rf train.col* *.model +k=$1 + +# split the lib svm file into k subfiles +python splitsvm.py ../../demo/data/agaricus.txt.train train $k + +# run xgboost mpi +../../rabit/tracker/rabit_demo.py -n $k ../../xgboost.mock mushroom-col.conf dsplit=col mock=0,2,0,0 mock=1,2,0,0 mock=2,2,8,0 mock=2,3,0,0 + +# the model can be directly loaded by single machine xgboost solver, as usuall +#../../xgboost mushroom-col.conf task=dump model_in=0002.model fmap=../../demo/data/featmap.txt name_dump=dump.nice.$k.txt + + +#cat dump.nice.$k.txt diff --git a/multi-node/col-split/mushroom-col-rabit.sh b/multi-node/col-split/mushroom-col-rabit.sh new file mode 100755 index 000000000000..f958305aa50d --- /dev/null +++ b/multi-node/col-split/mushroom-col-rabit.sh @@ -0,0 +1,28 @@ +#!/bin/bash +if [[ $# -ne 1 ]] +then + echo "Usage: nprocess" + exit -1 +fi + +# +# This script is same as mushroom-col except that we will be using xgboost instead of xgboost-mpi +# xgboost used built in tcp-based allreduce module, and can be run on more enviroment, so long as we know how to start job by modifying ../submit_job_tcp.py +# +rm -rf train.col* *.model +k=$1 + +# split the lib svm file into k subfiles +python splitsvm.py ../../demo/data/agaricus.txt.train train $k + +# run xgboost mpi +../../rabit/tracker/rabit_demo.py -n $k ../../xgboost mushroom-col.conf dsplit=col + +# the model can be directly loaded by single machine xgboost solver, as usuall +../../xgboost mushroom-col.conf task=dump model_in=0002.model fmap=../../demo/data/featmap.txt name_dump=dump.nice.$k.txt + +# run for one round, and continue training +../../rabit/tracker/rabit_demo.py -n $k ../../xgboost mushroom-col.conf dsplit=col num_round=1 +../../rabit/tracker/rabit_demo.py -n $k ../../xgboost mushroom-col.conf mushroom-col.conf dsplit=col model_in=0001.model + +cat dump.nice.$k.txt diff --git a/multi-node/col-split/mushroom-col.conf b/multi-node/col-split/mushroom-col.conf new file mode 100644 index 000000000000..2c779a44da53 --- /dev/null +++ b/multi-node/col-split/mushroom-col.conf @@ -0,0 +1,35 @@ +# General Parameters, see comment for each definition +# choose the booster, can be gbtree or gblinear +booster = gbtree +# choose logistic regression loss function for binary classification +objective = binary:logistic + +# Tree Booster Parameters +# step size shrinkage +eta = 1.0 +# minimum loss reduction required to make a further partition +gamma = 1.0 +# minimum sum of instance weight(hessian) needed in a child +min_child_weight = 1 +# maximum depth of a tree +max_depth = 3 + +# Task Parameters +# the number of round to do boosting +num_round = 2 +# 0 means do not save any model except the final round model +save_period = 0 +use_buffer = 0 + +# The path of training data %d is the wildcard for the rank of the data +# The idea is each process take a feature matrix with subset of columns +# +data = "train.col%d" + +# The path of validation data, used to monitor training process, here [test] sets name of the validation set +eval[test] = "../../demo/data/agaricus.txt.test" +# evaluate on training data as well each round +eval_train = 1 + +# The path of test data, need to use full data of test, try not use it, or keep an subsampled version +test:data = "../../demo/data/agaricus.txt.test" diff --git a/multi-node/col-split/splitsvm.py b/multi-node/col-split/splitsvm.py new file mode 100644 index 000000000000..365aef610c84 --- /dev/null +++ b/multi-node/col-split/splitsvm.py @@ -0,0 +1,32 @@ +#!/usr/bin/python +import sys +import random + +# split libsvm file into different subcolumns +if len(sys.argv) < 4: + print ('Usage: k') + exit(0) + +random.seed(10) +fmap = {} + +k = int(sys.argv[3]) +fi = open( sys.argv[1], 'r' ) +fos = [] + +for i in range(k): + fos.append(open( sys.argv[2]+'.col%d' % i, 'w' )) + +for l in open(sys.argv[1]): + arr = l.split() + for f in fos: + f.write(arr[0]) + for it in arr[1:]: + fid = int(it.split(':')[0]) + if fid not in fmap: + fmap[fid] = random.randint(0, k-1) + fos[fmap[fid]].write(' '+it) + for f in fos: + f.write('\n') +for f in fos: + f.close() diff --git a/multi-node/hadoop/README.md b/multi-node/hadoop/README.md new file mode 100644 index 000000000000..d1dde8ba3404 --- /dev/null +++ b/multi-node/hadoop/README.md @@ -0,0 +1,43 @@ +Distributed XGBoost: Hadoop Version +==== +* The script in this fold shows an example of how to run distributed xgboost on hadoop platform. +* It relies on [Rabit Library](https://github.com/tqchen/rabit) (Reliable Allreduce and Broadcast Interface) and Hadoop Streaming. Rabit provides an interface to aggregate gradient values and split statistics, that allow xgboost to run reliably on hadoop. You do not need to care how to update model in each iteration, just use the script ```rabit_hadoop.py```. For those who want to know how it exactly works, plz refer to the main page of [Rabit](https://github.com/tqchen/rabit). +* Quick start: run ```bash run_mushroom.sh ``` + - This is the hadoop version of binary classification example in the demo folder. + - More info of the usage of xgboost can be refered to [wiki page](https://github.com/tqchen/xgboost/wiki) + +Before you run the script +==== +* Make sure you have set up the hadoop environment. +* If you want to only use single machine multi-threading, try single machine examples in the [demo folder](../../demo). +* Build: run ```bash build.sh``` in the root folder, it will automatically download rabit and build xgboost. +* Check whether the environment variable $HADOOP_HOME exists (e.g. run ```echo $HADOOP_HOME```). If not, please set up hadoop-streaming.jar path in rabit_hadoop.py. + +How to Use +==== +* Input data format: LIBSVM format. The example here uses generated data in demo/data folder. +* Put the training data in HDFS (hadoop distributed file system). +* Use rabit ```rabit_hadoop.py``` to submit training task to hadoop, and save the final model file. +* Get the final model file from HDFS, and locally do prediction as well as visualization of model. + +Single machine vs Hadoop version +==== +If you have used xgboost (single machine version) before, this section will show you how to run xgboost on hadoop with a slight modification on conf file. +* Hadoop version needs to set up how many slave nodes/machines/workers you would like to use at first. +* IO: instead of reading and writing file locally, hadoop version use "stdin" to read training file and use "stdout" to store the final model file. Therefore, you should change the parameters "data" and "model_out" in conf file to ```data=stdin``` and ```model_out=stdout```. +* File cache: ```rabit_hadoop.py``` also provide several ways to cache necesary files, including binary file (xgboost), conf file, small size of dataset which used for eveluation during the training process, and so on. + - Any file used in config file, excluding stdin, should be cached in the script. ```rabit_hadoop.py``` will automatically cache files in the command line. For example, ```rabit_hadoop.py -n 3 -i $hdfsPath/agaricus.txt.train -o $hdfsPath/mushroom.final.model $localPath/xgboost mushroom.hadoop.conf``` will cache "xgboost" and "mushroom.hadoop.conf". + - You could also use "-f" to manually cache one or more files, like ```-f file1 -f file2``` or ```-f file1#file2``` (use "#" to spilt file names). + - The local path of cached files in command is "./". + - Since the cached files will be packaged and delivered to hadoop slave nodes, the cached file should not be large. For instance, trying to cache files of GB size may reduce the performance. +* Hadoop version also support evaluting each training round. You just need to modify parameters "eval_train". +* More details of submission can be referred to the usage of ```rabit_hadoop.py```. +* The model saved by hadoop version is compatible with single machine version. + +Notes +==== +* The code has been tested on MapReduce 1 (MRv1) and YARN. + - We recommend to run it on MapReduce 2 (MRv2, YARN) so that multi-threading can be enabled. +* The code is optimized with multi-threading, so you will want to run one xgboost per node/worker for best performance. + - You will want to set to be number of cores you have on each machine. + - You will need YARN to set specify number of cores of each worker diff --git a/multi-node/hadoop/mushroom.hadoop.conf b/multi-node/hadoop/mushroom.hadoop.conf new file mode 100644 index 000000000000..a4e885d54974 --- /dev/null +++ b/multi-node/hadoop/mushroom.hadoop.conf @@ -0,0 +1,39 @@ +# General Parameters, see comment for each definition +# choose the booster, can be gbtree or gblinear +booster = gbtree +# choose logistic regression loss function for binary classification +objective = binary:logistic + +# Tree Booster Parameters +# step size shrinkage +eta = 1.0 +# minimum loss reduction required to make a further partition +gamma = 1.0 +# minimum sum of instance weight(hessian) needed in a child +min_child_weight = 1 +# maximum depth of a tree +max_depth = 3 + +# Task Parameters +# the number of round to do boosting +num_round = 2 +# 0 means do not save any model except the final round model +save_period = 0 +# evaluate on training data as well each round +# eval_train = 1 +# The path of validation data, used to monitor training process, here [test] sets name of the validation set +# eval[test] = "agaricus.txt.test" + +# Plz donot modify the following parameters +# The path of training data +data = stdin +# The path of model file +model_out = stdout +# split pattern of xgboost +dsplit = row + +<<<<<<< HEAD +# evaluate on training data as well each round +eval_train = 1 +======= +>>>>>>> df3f87c182cc12ccc9ac1f9cafbe01ea7ebf0ac4 diff --git a/multi-node/hadoop/run_mushroom.sh b/multi-node/hadoop/run_mushroom.sh new file mode 100755 index 000000000000..5f133298e378 --- /dev/null +++ b/multi-node/hadoop/run_mushroom.sh @@ -0,0 +1,23 @@ +#!/bin/bash +if [ "$#" -lt 3 ]; +then + echo "Usage: " + exit -1 +fi + +# put the local training file to HDFS +hadoop fs -mkdir $3/data +hadoop fs -put ../../demo/data/agaricus.txt.train $3/data + +../../rabit/tracker/rabit_hadoop.py -n $1 -nt $2 -i $3/data/agaricus.txt.train -o $3/mushroom.final.model ../../xgboost mushroom.hadoop.conf nthread=$2 + +# get the final model file +hadoop fs -get $3/mushroom.final.model/part-00000 ./final.model + +# output prediction task=pred +../../xgboost mushroom.hadoop.conf task=pred model_in=final.model test:data=../../demo/data/agaricus.txt.test +# print the boosters of final.model in dump.raw.txt +../../xgboost mushroom.hadoop.conf task=dump model_in=final.model name_dump=dump.raw.txt +# use the feature map in printing for better visualization +../../xgboost mushroom.hadoop.conf task=dump model_in=final.model fmap=../../demo/data/featmap.txt name_dump=dump.nice.txt +cat dump.nice.txt diff --git a/multi-node/row-split/README.md b/multi-node/row-split/README.md new file mode 100644 index 000000000000..30e2528d33f0 --- /dev/null +++ b/multi-node/row-split/README.md @@ -0,0 +1,18 @@ +Distributed XGBoost: Row Split Version +==== +* You might be interested to checkout the [Hadoop example](../hadoop) +* Machine Rabit: run ```bash machine-row-rabit.sh ``` + - machine-col-rabit.sh starts xgboost job using rabit + +How to Use +==== +* First split the data by rows +* In the config, specify data file as containing a wildcard %d, where %d is the rank of the node, each node will load their part of data +* Enable ow split mode by ```dsplit=row``` + +Notes +==== +* The code is multi-threaded, so you want to run one xgboost-mpi per node +* Row-based solver split data by row, each node work on subset of rows, it uses an approximate histogram count algorithm, + and will only examine subset of potential split points as opposed to all split points. + diff --git a/multi-node/row-split/machine-row-rabit-mock.sh b/multi-node/row-split/machine-row-rabit-mock.sh new file mode 100755 index 000000000000..b8ef10b2d219 --- /dev/null +++ b/multi-node/row-split/machine-row-rabit-mock.sh @@ -0,0 +1,20 @@ +#!/bin/bash +if [[ $# -ne 1 ]] +then + echo "Usage: nprocess" + exit -1 +fi + +rm -rf train-machine.row* *.model +k=$1 +# make machine data +cd ../../demo/regression/ +python mapfeat.py +python mknfold.py machine.txt 1 +cd - + +# split the lib svm file into k subfiles +python splitrows.py ../../demo/regression/machine.txt.train train-machine $k + +# run xgboost mpi +../../rabit/tracker/rabit_demo.py -n $k ../../xgboost.mock machine-row.conf dsplit=row num_round=3 mock=1,1,1,0 mock=0,0,3,0 mock=2,2,3,0 diff --git a/multi-node/row-split/machine-row-rabit.sh b/multi-node/row-split/machine-row-rabit.sh new file mode 100755 index 000000000000..69f94b9d1442 --- /dev/null +++ b/multi-node/row-split/machine-row-rabit.sh @@ -0,0 +1,24 @@ +#!/bin/bash +if [[ $# -ne 1 ]] +then + echo "Usage: nprocess" + exit -1 +fi + +rm -rf train-machine.row* *.model +k=$1 +# make machine data +cd ../../demo/regression/ +python mapfeat.py +python mknfold.py machine.txt 1 +cd - + +# split the lib svm file into k subfiles +python splitrows.py ../../demo/regression/machine.txt.train train-machine $k + +# run xgboost mpi +../../rabit/tracker/rabit_mpi.py -n $k ../../xgboost machine-row.conf dsplit=row num_round=3 eval_train=1 + +# run xgboost-mpi save model 0001, continue to run from existing model +../../rabit/tracker/rabit_mpi.py -n $k ../../xgboost machine-row.conf dsplit=row num_round=1 +../../rabit/tracker/rabit_mpi.py -n $k ../../xgboost machine-row.conf dsplit=row num_round=2 model_in=0001.model diff --git a/multi-node/row-split/machine-row.conf b/multi-node/row-split/machine-row.conf new file mode 100644 index 000000000000..c0cba3da8cf9 --- /dev/null +++ b/multi-node/row-split/machine-row.conf @@ -0,0 +1,30 @@ +# General Parameters, see comment for each definition +# choose the tree booster, can also change to gblinear +booster = gbtree +# this is the only difference with classification, use reg:linear to do linear classification +# when labels are in [0,1] we can also use reg:logistic +objective = reg:linear + +# Tree Booster Parameters +# step size shrinkage +eta = 1.0 +# minimum loss reduction required to make a further partition +gamma = 1.0 +# minimum sum of instance weight(hessian) needed in a child +min_child_weight = 1 +# maximum depth of a tree +max_depth = 3 +# Task parameters +# the number of round to do boosting +num_round = 2 +# 0 means do not save any model except the final round model +save_period = 0 +use_buffer = 0 + +# The path of training data +data = "train-machine.row%d" +# The path of validation data, used to monitor training process, here [test] sets name of the validation set +eval[test] = "../../demo/regression/machine.txt.test" +# The path of test data +test:data = "../../demo/regression/machine.txt.test" + diff --git a/multi-node/row-split/splitrows.py b/multi-node/row-split/splitrows.py new file mode 100644 index 000000000000..2e9d1184d44f --- /dev/null +++ b/multi-node/row-split/splitrows.py @@ -0,0 +1,24 @@ +#!/usr/bin/python +import sys +import random + +# split libsvm file into different rows +if len(sys.argv) < 4: + print ('Usage: k') + exit(0) + +random.seed(10) + +k = int(sys.argv[3]) +fi = open( sys.argv[1], 'r' ) +fos = [] + +for i in range(k): + fos.append(open( sys.argv[2]+'.row%d' % i, 'w' )) + +for l in open(sys.argv[1]): + i = random.randint(0, k-1) + fos[i].write(l) + +for f in fos: + f.close() diff --git a/src/data.h b/src/data.h index 2ea5f222a996..162a31bfe6c2 100644 --- a/src/data.h +++ b/src/data.h @@ -138,9 +138,10 @@ class IFMatrix { virtual utils::IIterator *ColIterator(const std::vector &fset) = 0; /*! * \brief check if column access is supported, if not, initialize column access + * \param enabled whether certain feature should be included in column access * \param subsample subsample ratio when generating column access */ - virtual void InitColAccess(float subsample) = 0; + virtual void InitColAccess(const std::vector &enabled, float subsample) = 0; // the following are column meta data, should be able to answer them fast /*! \return whether column access is enabled */ virtual bool HaveColAccess(void) const = 0; diff --git a/src/gbm/gblinear-inl.hpp b/src/gbm/gblinear-inl.hpp index 473914b6e683..8cbe8becf437 100644 --- a/src/gbm/gblinear-inl.hpp +++ b/src/gbm/gblinear-inl.hpp @@ -33,16 +33,17 @@ class GBLinear : public IGradBooster { model.param.SetParam(name, val); } } - virtual void LoadModel(utils::IStream &fi) { + virtual void LoadModel(utils::IStream &fi, bool with_pbuffer) { model.LoadModel(fi); } - virtual void SaveModel(utils::IStream &fo) const { + virtual void SaveModel(utils::IStream &fo, bool with_pbuffer) const { model.SaveModel(fo); } virtual void InitModel(void) { model.InitModel(); } virtual void DoBoost(IFMatrix *p_fmat, + int64_t buffer_offset, const BoosterInfo &info, std::vector *in_gpair) { std::vector &gpair = *in_gpair; @@ -135,8 +136,13 @@ class GBLinear : public IGradBooster { } } } - - virtual std::vector DumpModel(const utils::FeatMap& fmap, int option) { + virtual void PredictLeaf(IFMatrix *p_fmat, + const BoosterInfo &info, + std::vector *out_preds, + unsigned ntree_limit = 0) { + utils::Error("gblinear does not support predict leaf index"); + } + virtual std::vector DumpModel(const utils::FeatMap& fmap, int option) { std::stringstream fo(""); fo << "bias:\n"; for (int i = 0; i < model.param.num_output_group; ++i) { @@ -151,8 +157,8 @@ class GBLinear : public IGradBooster { std::vector v; v.push_back(fo.str()); return v; - } - + } + protected: inline void Pred(const RowBatch::Inst &inst, float *preds) { for (int gid = 0; gid < model.param.num_output_group; ++gid) { diff --git a/src/gbm/gbm.cpp b/src/gbm/gbm.cpp index e280fdd4a525..fe8d778e4be0 100644 --- a/src/gbm/gbm.cpp +++ b/src/gbm/gbm.cpp @@ -1,5 +1,6 @@ #define _CRT_SECURE_NO_WARNINGS #define _CRT_SECURE_NO_DEPRECATE +#define NOMINMAX #include #include "./gbm.h" #include "./gbtree-inl.hpp" diff --git a/src/gbm/gbm.h b/src/gbm/gbm.h index 07dade4aca7e..57b8c05732e1 100644 --- a/src/gbm/gbm.h +++ b/src/gbm/gbm.h @@ -27,25 +27,44 @@ class IGradBooster { /*! * \brief load model from stream * \param fi input stream + * \param with_pbuffer whether the incoming data contains pbuffer */ - virtual void LoadModel(utils::IStream &fi) = 0; + virtual void LoadModel(utils::IStream &fi, bool with_pbuffer) = 0; /*! * \brief save model to stream * \param fo output stream + * \param with_pbuffer whether save out pbuffer */ - virtual void SaveModel(utils::IStream &fo) const = 0; + virtual void SaveModel(utils::IStream &fo, bool with_pbuffer) const = 0; /*! * \brief initialize the model */ virtual void InitModel(void) = 0; + /*! + * \brief reset the predict buffer + * this will invalidate all the previous cached results + * and recalculate from scratch + */ + virtual void ResetPredBuffer(size_t num_pbuffer) {} + /*! + * \brief whether the model allow lazy checkpoint + * return true if model is only updated in DoBoost + * after all Allreduce calls + */ + virtual bool AllowLazyCheckPoint(void) const { + return false; + } /*! * \brief peform update to the model(boosting) * \param p_fmat feature matrix that provide access to features + * \param buffer_offset buffer index offset of these instances, if equals -1 + * this means we do not have buffer index allocated to the gbm * \param info meta information about training * \param in_gpair address of the gradient pair statistics of the data * the booster may change content of gpair */ virtual void DoBoost(IFMatrix *p_fmat, + int64_t buffer_offset, const BoosterInfo &info, std::vector *in_gpair) = 0; /*! @@ -65,6 +84,19 @@ class IGradBooster { const BoosterInfo &info, std::vector *out_preds, unsigned ntree_limit = 0) = 0; + /*! + * \brief predict the leaf index of each tree, the output will be nsample * ntree vector + * this is only valid in gbtree predictor + * \param p_fmat feature matrix + * \param info extra side information that may be needed for prediction + * \param out_preds output vector to hold the predictions + * \param ntree_limit limit the number of trees used in prediction, when it equals 0, this means + * we do not limit number of trees, this parameter is only valid for gbtree, but not for gblinear + */ + virtual void PredictLeaf(IFMatrix *p_fmat, + const BoosterInfo &info, + std::vector *out_preds, + unsigned ntree_limit = 0) = 0; /*! * \brief dump the model in text format * \param fmap feature map that may help give interpretations of feature diff --git a/src/gbm/gbtree-inl.hpp b/src/gbm/gbtree-inl.hpp index ed52afa7df90..c08d15dd72c7 100644 --- a/src/gbm/gbtree-inl.hpp +++ b/src/gbm/gbtree-inl.hpp @@ -19,6 +19,8 @@ namespace gbm { */ class GBTree : public IGradBooster { public: + GBTree(void) { + } virtual ~GBTree(void) { this->Clear(); } @@ -37,7 +39,7 @@ class GBTree : public IGradBooster { tparam.SetParam(name, val); if (trees.size() == 0) mparam.SetParam(name, val); } - virtual void LoadModel(utils::IStream &fi) { + virtual void LoadModel(utils::IStream &fi, bool with_pbuffer) { this->Clear(); utils::Check(fi.Read(&mparam, sizeof(ModelParam)) != 0, "GBTree: invalid model file"); @@ -51,7 +53,7 @@ class GBTree : public IGradBooster { utils::Check(fi.Read(&tree_info[0], sizeof(int) * mparam.num_trees) != 0, "GBTree: invalid model file"); } - if (mparam.num_pbuffer != 0) { + if (mparam.num_pbuffer != 0 && with_pbuffer) { pred_buffer.resize(mparam.PredBufferSize()); pred_counter.resize(mparam.PredBufferSize()); utils::Check(fi.Read(&pred_buffer[0], pred_buffer.size() * sizeof(float)) != 0, @@ -60,7 +62,7 @@ class GBTree : public IGradBooster { "GBTree: invalid model file"); } } - virtual void SaveModel(utils::IStream &fo) const { + virtual void SaveModel(utils::IStream &fo, bool with_pbuffer) const { utils::Assert(mparam.num_trees == static_cast(trees.size()), "GBTree"); fo.Write(&mparam, sizeof(ModelParam)); for (size_t i = 0; i < trees.size(); ++i) { @@ -69,7 +71,7 @@ class GBTree : public IGradBooster { if (tree_info.size() != 0) { fo.Write(&tree_info[0], sizeof(int) * tree_info.size()); } - if (mparam.num_pbuffer != 0) { + if (mparam.num_pbuffer != 0 && with_pbuffer) { fo.Write(&pred_buffer[0], pred_buffer.size() * sizeof(float)); fo.Write(&pred_counter[0], pred_counter.size() * sizeof(unsigned)); } @@ -82,12 +84,23 @@ class GBTree : public IGradBooster { utils::Assert(mparam.num_trees == 0, "GBTree: model already initialized"); utils::Assert(trees.size() == 0, "GBTree: model already initialized"); } + virtual void ResetPredBuffer(size_t num_pbuffer) { + mparam.num_pbuffer = static_cast(num_pbuffer); + pred_buffer.clear(); pred_counter.clear(); + pred_buffer.resize(mparam.PredBufferSize(), 0.0f); + pred_counter.resize(mparam.PredBufferSize(), 0); + } + virtual bool AllowLazyCheckPoint(void) const { + return !(tparam.distcol_mode != 0 && mparam.num_output_group != 1); + } virtual void DoBoost(IFMatrix *p_fmat, + int64_t buffer_offset, const BoosterInfo &info, std::vector *in_gpair) { const std::vector &gpair = *in_gpair; - if (mparam.num_output_group == 1) { - this->BoostNewTrees(gpair, p_fmat, info, 0); + std::vector > new_trees; + if (mparam.num_output_group == 1) { + new_trees.push_back(BoostNewTrees(gpair, p_fmat, buffer_offset, info, 0)); } else { const int ngroup = mparam.num_output_group; utils::Check(gpair.size() % ngroup == 0, @@ -99,15 +112,18 @@ class GBTree : public IGradBooster { for (bst_omp_uint i = 0; i < nsize; ++i) { tmp[i] = gpair[i * ngroup + gid]; } - this->BoostNewTrees(tmp, p_fmat, info, gid); + new_trees.push_back(BoostNewTrees(tmp, p_fmat, buffer_offset, info, gid)); } } + for (int gid = 0; gid < mparam.num_output_group; ++gid) { + this->CommitModel(new_trees[gid], gid); + } } virtual void Predict(IFMatrix *p_fmat, int64_t buffer_offset, const BoosterInfo &info, std::vector *out_preds, - unsigned ntree_limit = 0) { + unsigned ntree_limit = 0) { int nthread; #pragma omp parallel { @@ -117,7 +133,6 @@ class GBTree : public IGradBooster { for (int i = 0; i < nthread; ++i) { thread_temp[i].Init(mparam.num_feature); } - std::vector &preds = *out_preds; const size_t stride = info.num_row * mparam.num_output_group; preds.resize(stride * (mparam.size_leaf_vector+1)); @@ -144,6 +159,22 @@ class GBTree : public IGradBooster { } } } + } + virtual void PredictLeaf(IFMatrix *p_fmat, + const BoosterInfo &info, + std::vector *out_preds, + unsigned ntree_limit) { + int nthread; + #pragma omp parallel + { + nthread = omp_get_num_threads(); + } + thread_temp.resize(nthread, tree::RegTree::FVec()); + for (int i = 0; i < nthread; ++i) { + thread_temp[i].Init(mparam.num_feature); + } + this->PredPath(p_fmat, info, out_preds, ntree_limit); + } virtual std::vector DumpModel(const utils::FeatMap& fmap, int option) { std::vector dump; @@ -184,13 +215,15 @@ class GBTree : public IGradBooster { tparam.updater_initialized = 1; } // do group specific group - inline void BoostNewTrees(const std::vector &gpair, - IFMatrix *p_fmat, - const BoosterInfo &info, - int bst_group) { + inline std::vector + BoostNewTrees(const std::vector &gpair, + IFMatrix *p_fmat, + int64_t buffer_offset, + const BoosterInfo &info, + int bst_group) { + std::vector new_trees; this->InitUpdater(); // create the trees - std::vector new_trees; for (int i = 0; i < tparam.num_parallel_tree; ++i) { new_trees.push_back(new tree::RegTree()); for (size_t j = 0; j < cfg.size(); ++j) { @@ -201,13 +234,52 @@ class GBTree : public IGradBooster { // update the trees for (size_t i = 0; i < updaters.size(); ++i) { updaters[i]->Update(gpair, p_fmat, info, new_trees); + } + // optimization, update buffer, if possible + // this is only under distributed column mode + // for safety check of lazy checkpoint + if ( + buffer_offset >= 0 && + new_trees.size() == 1 && updaters.size() > 0 && + updaters.back()->GetLeafPosition() != NULL) { + utils::Check(info.num_row == p_fmat->buffered_rowset().size(), + "distributed mode is not compatible with prob_buffer_row"); + this->UpdateBufferByPosition(p_fmat, + buffer_offset, bst_group, + *new_trees[0], + updaters.back()->GetLeafPosition()); } - // push back to model + return new_trees; + } + // commit new trees all at once + inline void CommitModel(const std::vector &new_trees, int bst_group) { for (size_t i = 0; i < new_trees.size(); ++i) { trees.push_back(new_trees[i]); tree_info.push_back(bst_group); } - mparam.num_trees += tparam.num_parallel_tree; + mparam.num_trees += static_cast(new_trees.size()); + } + // update buffer by pre-cached position + inline void UpdateBufferByPosition(IFMatrix *p_fmat, + int64_t buffer_offset, + int bst_group, + const tree::RegTree &new_tree, + const int* leaf_position) { + const std::vector &rowset = p_fmat->buffered_rowset(); + const bst_omp_uint ndata = static_cast(rowset.size()); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < ndata; ++i) { + const bst_uint ridx = rowset[i]; + const int64_t bid = mparam.BufferOffset(buffer_offset + ridx, bst_group); + const int tid = leaf_position[ridx]; + utils::Assert(pred_counter[bid] == trees.size(), "cached buffer not up to date"); + utils::Assert(tid >= 0, "invalid leaf position"); + pred_buffer[bid] += new_tree[tid].leaf_value(); + for (int i = 0; i < mparam.size_leaf_vector; ++i) { + pred_buffer[bid + i + 1] += new_tree.leafvec(tid)[i]; + } + pred_counter[bid] += tparam.num_parallel_tree; + } } // make a prediction for a single instance inline void Pred(const RowBatch::Inst &inst, @@ -215,7 +287,8 @@ class GBTree : public IGradBooster { int bst_group, unsigned root_index, tree::RegTree::FVec *p_feats, - float *out_pred, size_t stride, unsigned ntree_limit) { + float *out_pred, size_t stride, + unsigned ntree_limit) { size_t itop = 0; float psum = 0.0f; // sum of leaf vector @@ -258,6 +331,39 @@ class GBTree : public IGradBooster { out_pred[stride * (i + 1)] = vec_psum[i]; } } + // predict independent leaf index + inline void PredPath(IFMatrix *p_fmat, + const BoosterInfo &info, + std::vector *out_preds, + unsigned ntree_limit) { + // number of valid trees + if (ntree_limit == 0 || ntree_limit > trees.size()) { + ntree_limit = static_cast(trees.size()); + } + std::vector &preds = *out_preds; + preds.resize(info.num_row * ntree_limit); + // start collecting the prediction + utils::IIterator *iter = p_fmat->RowIterator(); + iter->BeforeFirst(); + while (iter->Next()) { + const RowBatch &batch = iter->Value(); + // parallel over local batch + const bst_omp_uint nsize = static_cast(batch.size); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < nsize; ++i) { + const int tid = omp_get_thread_num(); + int64_t ridx = static_cast(batch.base_rowid + i); + tree::RegTree::FVec &feats = thread_temp[tid]; + feats.Fill(batch[i]); + for (unsigned j = 0; j < ntree_limit; ++j) { + int tid = trees[j]->GetLeafIndex(feats, info.GetRoot(ridx)); + preds[ridx * ntree_limit + j] = static_cast(tid); + } + feats.Drop(batch[i]); + } + } + } + // --- data structure --- /*! \brief training parameters */ struct TrainParam { @@ -270,6 +376,8 @@ class GBTree : public IGradBooster { int num_parallel_tree; /*! \brief whether updater is already initialized */ int updater_initialized; + /*! \brief distributed column mode */ + int distcol_mode; /*! \brief tree updater sequence */ std::string updater_seq; // construction @@ -278,6 +386,7 @@ class GBTree : public IGradBooster { updater_seq = "grow_colmaker,prune"; num_parallel_tree = 1; updater_initialized = 0; + distcol_mode = 0; } inline void SetParam(const char *name, const char *val){ using namespace std; @@ -286,6 +395,9 @@ class GBTree : public IGradBooster { updater_seq = val; updater_initialized = 0; } + if (!strcmp(name, "dsplit") && !strcmp(val, "col")) { + distcol_mode = 1; + } if (!strcmp(name, "nthread")) { omp_set_num_threads(nthread = atoi(val)); } diff --git a/src/io/io.cpp b/src/io/io.cpp index d251d7a96c03..0072618c65c6 100644 --- a/src/io/io.cpp +++ b/src/io/io.cpp @@ -1,15 +1,32 @@ #define _CRT_SECURE_NO_WARNINGS #define _CRT_SECURE_NO_DEPRECATE +#define NOMINMAX #include #include "./io.h" #include "../utils/io.h" #include "../utils/utils.h" #include "simple_dmatrix-inl.hpp" +#include "page_dmatrix-inl.hpp" +#include "page_fmatrix-inl.hpp" + // implements data loads using dmatrix simple for now namespace xgboost { namespace io { DataMatrix* LoadDataMatrix(const char *fname, bool silent, bool savebuffer) { + if (!strcmp(fname, "stdin")) { + DMatrixSimple *dmat = new DMatrixSimple(); + dmat->LoadText(fname, silent); + return dmat; + } + std::string tmp_fname; + const char *fname_ext = NULL; + if (strchr(fname, ';') != NULL) { + tmp_fname = fname; + char *ptr = strchr(&tmp_fname[0], ';'); + ptr[0] = '\0'; fname = &tmp_fname[0]; + fname_ext = ptr + 1; + } int magic; utils::FileStream fs(utils::FopenCheck(fname, "rb")); utils::Check(fs.Read(&magic, sizeof(magic)) != 0, "invalid input file format"); @@ -20,7 +37,27 @@ DataMatrix* LoadDataMatrix(const char *fname, bool silent, bool savebuffer) { dmat->LoadBinary(fs, silent, fname); fs.Close(); return dmat; - } + } + if (magic == DMatrixPage::kMagic) { + if (fname_ext == NULL) { + DMatrixPage *dmat = new DMatrixPage(); + dmat->Load(fs, silent, fname); + return dmat; + } else { + DMatrixColPage *dmat = new DMatrixColPage(fname_ext); + dmat->Load(fs, silent, fname, true); + return dmat; + } + } + if (magic == DMatrixColPage::kMagic) { + std::string sfname = fname; + if (fname_ext == NULL) { + sfname += ".col"; fname_ext = sfname.c_str(); + } + DMatrixColPage *dmat = new DMatrixColPage(fname_ext); + dmat->Load(fs, silent, fname); + return dmat; + } fs.Close(); DMatrixSimple *dmat = new DMatrixSimple(); @@ -29,11 +66,21 @@ DataMatrix* LoadDataMatrix(const char *fname, bool silent, bool savebuffer) { } void SaveDataMatrix(const DataMatrix &dmat, const char *fname, bool silent) { + if (!strcmp(fname + strlen(fname) - 5, ".page")) { + DMatrixPage::Save(fname, dmat, silent); + return; + } + if (!strcmp(fname + strlen(fname) - 6, ".cpage")) { + DMatrixColPage::Save(fname, dmat, silent); + return; + } if (dmat.magic == DMatrixSimple::kMagic) { const DMatrixSimple *p_dmat = static_cast(&dmat); p_dmat->SaveBinary(fname, silent); } else { - utils::Error("not implemented"); + DMatrixSimple smat; + smat.CopyFrom(dmat); + smat.SaveBinary(fname, silent); } } diff --git a/src/io/page_dmatrix-inl.hpp b/src/io/page_dmatrix-inl.hpp new file mode 100644 index 000000000000..4f70ff2e9fb3 --- /dev/null +++ b/src/io/page_dmatrix-inl.hpp @@ -0,0 +1,278 @@ +#ifndef XGBOOST_IO_PAGE_ROW_ITER_INL_HPP_ +#define XGBOOST_IO_PAGE_ROW_ITER_INL_HPP_ +/*! + * \file page_row_iter-inl.hpp + * row iterator based on sparse page + * \author Tianqi Chen + */ +#include +#include "../data.h" +#include "../utils/iterator.h" +#include "../utils/thread_buffer.h" +#include "./simple_fmatrix-inl.hpp" + +namespace xgboost { +namespace io { +/*! \brief page structure that can be used to store a rowbatch */ +struct RowBatchPage { + public: + explicit RowBatchPage(size_t page_size) : kPageSize(page_size) { + data_ = new int[kPageSize]; + utils::Assert(data_ != NULL, "fail to allocate row batch page"); + this->Clear(); + } + ~RowBatchPage(void) { + if (data_ != NULL) delete [] data_; + } + /*! + * \brief Push one row into page + * \param row an instance row + * \return false or true to push into + */ + inline bool PushRow(const RowBatch::Inst &row) { + const size_t dsize = row.length * sizeof(RowBatch::Entry); + if (FreeBytes() < dsize+ sizeof(int)) return false; + row_ptr(Size() + 1) = row_ptr(Size()) + row.length; + memcpy(data_ptr(row_ptr(Size())) , row.data, dsize); + ++data_[0]; + return true; + } + /*! + * \brief get a row batch representation from the page + * \param p_rptr a temporal space that can be used to provide + * ind_ptr storage for RowBatch + * \return a new RowBatch object + */ + inline RowBatch GetRowBatch(std::vector *p_rptr, size_t base_rowid) { + RowBatch batch; + batch.base_rowid = base_rowid; + batch.data_ptr = this->data_ptr(0); + batch.size = static_cast(this->Size()); + std::vector &rptr = *p_rptr; + rptr.resize(this->Size() + 1); + for (size_t i = 0; i < rptr.size(); ++i) { + rptr[i] = static_cast(this->row_ptr(static_cast(i))); + } + batch.ind_ptr = &rptr[0]; + return batch; + } + /*! \brief get i-th row from the batch */ + inline RowBatch::Inst operator[](int i) { + return RowBatch::Inst(data_ptr(0) + row_ptr(i), + static_cast(row_ptr(i+1) - row_ptr(i))); + } + /*! + * \brief clear the page, cleanup the content + */ + inline void Clear(void) { + memset(&data_[0], 0, sizeof(int) * kPageSize); + } + /*! + * \brief load one page form instream + * \return true if loading is successful + */ + inline bool Load(utils::IStream &fi) { + return fi.Read(&data_[0], sizeof(int) * kPageSize) != 0; + } + /*! \brief save one page into outstream */ + inline void Save(utils::IStream &fo) { + fo.Write(&data_[0], sizeof(int) * kPageSize); + } + /*! \return number of elements */ + inline int Size(void) const { + return data_[0]; + } + + protected: + /*! \return number of elements */ + inline size_t FreeBytes(void) { + return (kPageSize - (Size() + 2)) * sizeof(int) - + row_ptr(Size()) * sizeof(RowBatch::Entry); + } + /*! \brief equivalent row pointer at i */ + inline int& row_ptr(int i) { + return data_[kPageSize - i - 1]; + } + inline RowBatch::Entry* data_ptr(int i) { + return (RowBatch::Entry*)(&data_[1]) + i; + } + // content of data + int *data_; + // page size + const size_t kPageSize; +}; +/*! \brief thread buffer iterator */ +class ThreadRowPageIterator: public utils::IIterator { + public: + ThreadRowPageIterator(void) { + itr.SetParam("buffer_size", "2"); + page_ = NULL; + base_rowid_ = 0; + } + virtual ~ThreadRowPageIterator(void) {} + virtual void Init(void) { + } + virtual void BeforeFirst(void) { + itr.BeforeFirst(); + base_rowid_ = 0; + } + virtual bool Next(void) { + if (!itr.Next(page_)) return false; + out_ = page_->GetRowBatch(&tmp_ptr_, base_rowid_); + base_rowid_ += out_.size; + return true; + } + virtual const RowBatch &Value(void) const { + return out_; + } + /*! \brief load and initialize the iterator with fi */ + inline void Load(const utils::FileStream &fi) { + itr.get_factory().SetFile(fi); + itr.Init(); + this->BeforeFirst(); + } + /*! + * \brief save a row iterator to output stream, in row iterator format + */ + inline static void Save(utils::IIterator *iter, + utils::IStream &fo) { + RowBatchPage page(kPageSize); + iter->BeforeFirst(); + while (iter->Next()) { + const RowBatch &batch = iter->Value(); + for (size_t i = 0; i < batch.size; ++i) { + if (!page.PushRow(batch[i])) { + page.Save(fo); + page.Clear(); + utils::Check(page.PushRow(batch[i]), "row is too big"); + } + } + } + if (page.Size() != 0) page.Save(fo); + } + /*! \brief page size 64 MB */ + static const size_t kPageSize = 64 << 18; + + private: + // base row id + size_t base_rowid_; + // temporal ptr + std::vector tmp_ptr_; + // output data + RowBatch out_; + // page pointer type + typedef RowBatchPage* PagePtr; + // loader factory for page + struct Factory { + public: + size_t file_begin_; + utils::FileStream fi; + Factory(void) {} + inline void SetFile(const utils::FileStream &fi) { + this->fi = fi; + file_begin_ = this->fi.Tell(); + } + inline bool Init(void) { + return true; + } + inline void SetParam(const char *name, const char *val) {} + inline bool LoadNext(PagePtr &val) { + return val->Load(fi); + } + inline PagePtr Create(void) { + PagePtr a = new RowBatchPage(kPageSize); + return a; + } + inline void FreeSpace(PagePtr &a) { + delete a; + } + inline void Destroy(void) { + fi.Close(); + } + inline void BeforeFirst(void) { + fi.Seek(file_begin_); + } + }; + + protected: + PagePtr page_; + utils::ThreadBuffer itr; +}; + +/*! \brief data matrix using page */ +template +class DMatrixPageBase : public DataMatrix { + public: + DMatrixPageBase(void) : DataMatrix(kMagic) { + iter_ = new ThreadRowPageIterator(); + } + // virtual destructor + virtual ~DMatrixPageBase(void) { + // do not delete row iterator, since it is owned by fmat + // to be cleaned up in a more clear way + } + /*! \brief load and initialize the iterator with fi */ + inline void Load(utils::FileStream &fi, + bool silent = false, + const char *fname = NULL, + bool skip_magic_check = false) { + int tmagic; + utils::Check(fi.Read(&tmagic, sizeof(tmagic)) != 0, "invalid input file format"); + if (!skip_magic_check) { + utils::Check(tmagic == magic, "invalid format,magic number mismatch"); + } + this->info.LoadBinary(fi); + iter_->Load(fi); + if (!silent) { + utils::Printf("DMatrixPage: %lux%lu matrix is loaded", + static_cast(info.num_row()), + static_cast(info.num_col())); + if (fname != NULL) { + utils::Printf(" from %s\n", fname); + } else { + utils::Printf("\n"); + } + if (info.group_ptr.size() != 0) { + utils::Printf("data contains %u groups\n", (unsigned)info.group_ptr.size() - 1); + } + } + } + /*! \brief save a DataMatrix as DMatrixPage*/ + inline static void Save(const char* fname, const DataMatrix &mat, bool silent) { + utils::FileStream fs(utils::FopenCheck(fname, "wb")); + int magic = kMagic; + fs.Write(&magic, sizeof(magic)); + mat.info.SaveBinary(fs); + ThreadRowPageIterator::Save(mat.fmat()->RowIterator(), fs); + fs.Close(); + if (!silent) { + utils::Printf("DMatrixPage: %lux%lu is saved to %s\n", + static_cast(mat.info.num_row()), + static_cast(mat.info.num_col()), fname); + } + } + /*! \brief magic number used to identify DMatrix */ + static const int kMagic = TKMagic; + + protected: + /*! \brief row iterator */ + ThreadRowPageIterator *iter_; +}; + +class DMatrixPage : public DMatrixPageBase<0xffffab02> { + public: + DMatrixPage(void) { + fmat_ = new FMatrixS(iter_); + } + virtual ~DMatrixPage(void) { + delete fmat_; + } + virtual IFMatrix *fmat(void) const { + return fmat_; + } + /*! \brief the real fmatrix */ + IFMatrix *fmat_; +}; +} // namespace io +} // namespace xgboost +#endif // XGBOOST_IO_PAGE_ROW_ITER_INL_HPP_ diff --git a/src/io/page_fmatrix-inl.hpp b/src/io/page_fmatrix-inl.hpp new file mode 100644 index 000000000000..0527da827b3c --- /dev/null +++ b/src/io/page_fmatrix-inl.hpp @@ -0,0 +1,382 @@ +#ifndef XGBOOST_IO_PAGE_FMATRIX_INL_HPP_ +#define XGBOOST_IO_PAGE_FMATRIX_INL_HPP_ +/*! + * \file page_fmatrix-inl.hpp + * sparse page manager for fmatrix + * \author Tianqi Chen + */ +#include +#include +#include +#include "../data.h" +#include "../utils/iterator.h" +#include "../utils/io.h" +#include "../utils/matrix_csr.h" +#include "../utils/thread_buffer.h" +namespace xgboost { +namespace io { +class CSCMatrixManager { + public: + /*! \brief in memory page */ + struct Page { + public: + /*! \brief initialize the page */ + explicit Page(size_t size) { + buffer.resize(size); + col_index.reserve(10); + col_data.reserve(10); + } + /*! \brief clear the page */ + inline void Clear(void) { + num_entry = 0; + col_index.clear(); + col_data.clear(); + } + /*! \brief number of used entries */ + size_t num_entry; + /*! \brief column index */ + std::vector col_index; + /*! \brief column data */ + std::vector col_data; + /*! \brief number of free entries */ + inline size_t NumFreeEntry(void) const { + return buffer.size() - num_entry; + } + inline ColBatch::Entry* AllocEntry(size_t len) { + ColBatch::Entry *p_data = &buffer[0] + num_entry; + num_entry += len; + return p_data; + } + /*! \brief get underlying batch */ + inline ColBatch GetBatch(void) const { + ColBatch batch; + batch.size = col_index.size(); + batch.col_index = BeginPtr(col_index); + batch.col_data = BeginPtr(col_data); + return batch; + } + + private: + /*! \brief buffer space, not to be changed since ready */ + std::vector buffer; + }; + /*! \brief define type of page pointer */ + typedef Page *PagePtr; + // constructor + CSCMatrixManager(void) { + fi_ = NULL; + } + /*! \brief get column pointer */ + inline const std::vector &col_ptr(void) const { + return col_ptr_; + } + inline void SetParam(const char *name, const char *val) { + } + inline PagePtr Create(void) { + return new Page(page_size_); + } + inline void FreeSpace(PagePtr &a) { + delete a; + } + inline void Destroy(void) { + } + inline void BeforeFirst(void) { + col_index_ = col_todo_; + read_top_ = 0; + } + inline bool LoadNext(PagePtr &val) { + val->Clear(); + if (read_top_ >= col_index_.size()) return false; + while (read_top_ < col_index_.size()) { + if (!this->TryFill(col_index_[read_top_], val)) { + return true; + } + ++read_top_; + } + return true; + } + inline bool Init(void) { + this->BeforeFirst(); + return true; + } + inline void Setup(utils::ISeekStream *fi, double page_ratio) { + fi_ = fi; + fi_->Read(&begin_meta_ , sizeof(begin_meta_)); + begin_data_ = static_cast(fi->Tell()); + fi_->Seek(begin_meta_); + fi_->Read(&col_ptr_); + size_t psmax = 0; + for (size_t i = 0; i < col_ptr_.size() - 1; ++i) { + psmax = std::max(psmax, col_ptr_[i+1] - col_ptr_[i]); + } + utils::Check(page_ratio >= 1.0f, "col_page_ratio must be at least 1"); + page_size_ = std::max(static_cast(psmax * page_ratio), psmax); + } + inline void SetColSet(const std::vector &cset, bool setall) { + if (!setall) { + col_todo_.resize(0); + for (size_t i = 0; i < cset.size(); ++i) { + if (col_todo_[i] < static_cast(col_ptr_.size() - 1)) { + col_todo_.push_back(cset[i]); + } + } + std::sort(col_todo_.begin(), col_todo_.end()); + } else { + col_todo_.resize(col_ptr_.size()-1); + for (size_t i = 0; i < col_todo_.size(); ++i) { + col_todo_[i] = static_cast(i); + } + } + } + + private: + /*! \brief fill a page with */ + inline bool TryFill(size_t cidx, Page *p_page) { + size_t len = col_ptr_[cidx+1] - col_ptr_[cidx]; + if (p_page->NumFreeEntry() < len) return false; + ColBatch::Entry *p_data = p_page->AllocEntry(len); + fi_->Seek(col_ptr_[cidx] * sizeof(ColBatch::Entry) + begin_data_); + utils::Check(fi_->Read(p_data, sizeof(ColBatch::Entry) * len) != 0, + "invalid column buffer format"); + p_page->col_data.push_back(ColBatch::Inst(p_data, static_cast(len))); + p_page->col_index.push_back(static_cast(cidx)); + return true; + } + // the following are in memory auxiliary data structure + /*! \brief top of reader position */ + size_t read_top_; + /*! \brief size of page */ + size_t page_size_; + /*! \brief column index to be loaded */ + std::vector col_index_; + /*! \brief column index to be after calling before first */ + std::vector col_todo_; + // the following are input content + /*! \brief beginning position of data content */ + size_t begin_data_; + /*! \brief size of data content */ + size_t begin_meta_; + /*! \brief input stream */ + utils::ISeekStream *fi_; + /*! \brief column pointer of CSC format */ + std::vector col_ptr_; +}; + +class ThreadColPageIterator : public utils::IIterator { + public: + explicit ThreadColPageIterator(utils::ISeekStream *fi, + float page_ratio, bool silent) { + itr_.SetParam("buffer_size", "2"); + itr_.get_factory().Setup(fi, page_ratio); + itr_.Init(); + if (!silent) { + utils::Printf("ThreadColPageIterator: finish initialzing, %u columns\n", + static_cast(col_ptr().size() - 1)); + } + } + virtual ~ThreadColPageIterator(void) { + } + virtual void BeforeFirst(void) { + itr_.BeforeFirst(); + } + virtual bool Next(void) { + // page to be loaded + CSCMatrixManager::PagePtr page; + if (!itr_.Next(page)) return false; + out_ = page->GetBatch(); + return true; + } + virtual const ColBatch &Value(void) const { + return out_; + } + inline const std::vector &col_ptr(void) const { + return itr_.get_factory().col_ptr(); + } + inline void SetColSet(const std::vector &cset, + bool setall = false) { + itr_.get_factory().SetColSet(cset, setall); + } + + private: + // output data + ColBatch out_; + // internal iterator + utils::ThreadBuffer itr_; +}; +/*! + * \brief sparse matrix that support column access + */ +class FMatrixPage : public IFMatrix { + public: + /*! \brief constructor */ + FMatrixPage(utils::IIterator *iter, std::string fname_buffer) + : fname_cbuffer_(fname_buffer) { + this->row_iter_ = iter; + this->col_iter_ = NULL; + this->fi_ = NULL; + } + // destructor + virtual ~FMatrixPage(void) { + if (row_iter_ != NULL) delete row_iter_; + if (col_iter_ != NULL) delete col_iter_; + if (fi_ != NULL) { + fi_->Close(); delete fi_; + } + } + /*! \return whether column access is enabled */ + virtual bool HaveColAccess(void) const { + return col_iter_ != NULL; + } + /*! \brief get number of colmuns */ + virtual size_t NumCol(void) const { + utils::Check(this->HaveColAccess(), "NumCol:need column access"); + return col_iter_->col_ptr().size() - 1; + } + /*! \brief get number of buffered rows */ + virtual const std::vector &buffered_rowset(void) const { + return buffered_rowset_; + } + /*! \brief get column size */ + virtual size_t GetColSize(size_t cidx) const { + const std::vector &col_ptr = col_iter_->col_ptr(); + return col_ptr[cidx+1] - col_ptr[cidx]; + } + /*! \brief get column density */ + virtual float GetColDensity(size_t cidx) const { + const std::vector &col_ptr = col_iter_->col_ptr(); + size_t nmiss = buffered_rowset_.size() - (col_ptr[cidx+1] - col_ptr[cidx]); + return 1.0f - (static_cast(nmiss)) / buffered_rowset_.size(); + } + virtual void InitColAccess(const std::vector &enabled, float pkeep = 1.0f) { + if (this->HaveColAccess()) return; + utils::Printf("start to initialize page col access\n"); + if (this->LoadColData()) { + utils::Printf("loading previously saved col data\n"); + return; + } + this->InitColData(pkeep, fname_cbuffer_.c_str(), + 1 << 30, 5); + utils::Check(this->LoadColData(), "fail to read in column data"); + utils::Printf("finish initialize page col access\n"); + } + /*! + * \brief get the row iterator associated with FMatrix + */ + virtual utils::IIterator* RowIterator(void) { + row_iter_->BeforeFirst(); + return row_iter_; + } + /*! + * \brief get the column based iterator + */ + virtual utils::IIterator* ColIterator(void) { + std::vector cset; + col_iter_->SetColSet(cset, true); + col_iter_->BeforeFirst(); + return col_iter_; + } + /*! + * \brief colmun based iterator + */ + virtual utils::IIterator *ColIterator(const std::vector &fset) { + col_iter_->SetColSet(fset, false); + col_iter_->BeforeFirst(); + return col_iter_; + } + + protected: + /*! + * \brief try load column data from file + */ + inline bool LoadColData(void) { + FILE *fp = fopen64(fname_cbuffer_.c_str(), "rb"); + if (fp == NULL) return false; + fi_ = new utils::FileStream(fp); + static_cast(fi_)->Read(&buffered_rowset_); + col_iter_ = new ThreadColPageIterator(fi_, 2.0f, false); + return true; + } + /*! + * \brief intialize column data + * \param pkeep probability to keep a row + */ + inline void InitColData(float pkeep, const char *fname, + size_t buffer_size, size_t col_step) { + buffered_rowset_.clear(); + utils::FileStream fo(utils::FopenCheck(fname, "wb+")); + // use 64M buffer + utils::SparseCSRFileBuilder builder(&fo, buffer_size); + // start working + row_iter_->BeforeFirst(); + while (row_iter_->Next()) { + const RowBatch &batch = row_iter_->Value(); + for (size_t i = 0; i < batch.size; ++i) { + if (pkeep == 1.0f || random::SampleBinary(pkeep)) { + buffered_rowset_.push_back(static_cast(batch.base_rowid+i)); + RowBatch::Inst inst = batch[i]; + for (bst_uint j = 0; j < inst.length; ++j) { + builder.AddBudget(inst[j].index); + } + } + } + } + // write buffered rowset + static_cast(&fo)->Write(buffered_rowset_); + builder.InitStorage(); + row_iter_->BeforeFirst(); + size_t ktop = 0; + while (row_iter_->Next()) { + const RowBatch &batch = row_iter_->Value(); + for (size_t i = 0; i < batch.size; ++i) { + if (ktop < buffered_rowset_.size() && + buffered_rowset_[ktop] == batch.base_rowid + i) { + ++ktop; + RowBatch::Inst inst = batch[i]; + for (bst_uint j = 0; j < inst.length; ++j) { + builder.PushElem(inst[j].index, + ColBatch::Entry((bst_uint)(batch.base_rowid+i), + inst[j].fvalue)); + } + if (ktop % 100000 == 0) { + utils::Printf("\r \r"); + utils::Printf("InitCol: %lu rows ", static_cast(ktop)); + } + } + } + } + builder.Finalize(); + builder.SortRows(ColBatch::Entry::CmpValue, col_step); + fo.Close(); + } + + private: + // row iterator + utils::IIterator *row_iter_; + // column iterator + ThreadColPageIterator *col_iter_; + // file pointer to data + utils::FileStream *fi_; + // file name of column buffer + std::string fname_cbuffer_; + /*! \brief list of row index that are buffered */ + std::vector buffered_rowset_; +}; + +class DMatrixColPage : public DMatrixPageBase<0xffffab03> { + public: + explicit DMatrixColPage(const char *fname) { + fmat_ = new FMatrixPage(iter_, fname); + } + virtual ~DMatrixColPage(void) { + delete fmat_; + } + virtual IFMatrix *fmat(void) const { + return fmat_; + } + /*! \brief the real fmatrix */ + IFMatrix *fmat_; +}; + +} // namespace io +} // namespace xgboost +#endif // XGBOOST_IO_PAGE_FMATRIX_INL_HPP_ diff --git a/src/io/simple_dmatrix-inl.hpp b/src/io/simple_dmatrix-inl.hpp index 374d621e9ab6..a793c779fb28 100644 --- a/src/io/simple_dmatrix-inl.hpp +++ b/src/io/simple_dmatrix-inl.hpp @@ -44,8 +44,8 @@ class DMatrixSimple : public DataMatrix { } /*! \brief copy content data from source matrix */ inline void CopyFrom(const DataMatrix &src) { - this->info = src.info; this->Clear(); + this->info = src.info; // clone data content in thos matrix utils::IIterator *iter = src.fmat()->RowIterator(); iter->BeforeFirst(); @@ -84,7 +84,12 @@ class DMatrixSimple : public DataMatrix { inline void LoadText(const char* fname, bool silent = false) { using namespace std; this->Clear(); - FILE* file = utils::FopenCheck(fname, "r"); + FILE* file; + if (!strcmp(fname, "stdin")) { + file = stdin; + } else { + file = utils::FopenCheck(fname, "r"); + } float label; bool init = true; char tmp[1024]; std::vector feats; @@ -112,7 +117,9 @@ class DMatrixSimple : public DataMatrix { static_cast(info.num_col()), static_cast(row_data_.size()), fname); } - fclose(file); + if (file != stdin) { + fclose(file); + } // try to load in additional file std::string name = fname; std::string gname = name + ".group"; @@ -152,7 +159,7 @@ class DMatrixSimple : public DataMatrix { inline void LoadBinary(utils::IStream &fs, bool silent = false, const char *fname = NULL) { int tmagic; utils::Check(fs.Read(&tmagic, sizeof(tmagic)) != 0, "invalid input file format"); - utils::Check(tmagic == kMagic, "invalid format,magic number mismatch"); + utils::Check(tmagic == kMagic, "\"%s\" invalid format, magic number mismatch", fname == NULL ? "" : fname); info.LoadBinary(fs); FMatrixS::LoadBinary(fs, &row_ptr_, &row_data_); diff --git a/src/io/simple_fmatrix-inl.hpp b/src/io/simple_fmatrix-inl.hpp index 7c8631a29854..08e25e28b781 100644 --- a/src/io/simple_fmatrix-inl.hpp +++ b/src/io/simple_fmatrix-inl.hpp @@ -48,9 +48,10 @@ class FMatrixS : public IFMatrix{ size_t nmiss = buffered_rowset_.size() - (col_ptr_[cidx+1] - col_ptr_[cidx]); return 1.0f - (static_cast(nmiss)) / buffered_rowset_.size(); } - virtual void InitColAccess(float pkeep = 1.0f) { + virtual void InitColAccess(const std::vector &enabled, + float pkeep = 1.0f) { if (this->HaveColAccess()) return; - this->InitColData(pkeep); + this->InitColData(pkeep, enabled); } /*! * \brief get the row iterator associated with FMatrix @@ -75,7 +76,11 @@ class FMatrixS : public IFMatrix{ * \brief colmun based iterator */ virtual utils::IIterator *ColIterator(const std::vector &fset) { - col_iter_.col_index_ = fset; + size_t ncol = this->NumCol(); + col_iter_.col_index_.resize(0); + for (size_t i = 0; i < fset.size(); ++i) { + if (fset[i] < ncol) col_iter_.col_index_.push_back(fset[i]); + } col_iter_.SetBatch(col_ptr_, col_data_); return &col_iter_; } @@ -141,7 +146,7 @@ class FMatrixS : public IFMatrix{ * \brief intialize column data * \param pkeep probability to keep a row */ - inline void InitColData(float pkeep) { + inline void InitColData(float pkeep, const std::vector &enabled) { buffered_rowset_.clear(); // note: this part of code is serial, todo, parallelize this transformer utils::SparseCSRMBuilder builder(col_ptr_, col_data_); @@ -150,12 +155,14 @@ class FMatrixS : public IFMatrix{ iter_->BeforeFirst(); while (iter_->Next()) { const RowBatch &batch = iter_->Value(); - for (size_t i = 0; i < batch.size; ++i) { + for (size_t i = 0; i < batch.size; ++i) { if (pkeep == 1.0f || random::SampleBinary(pkeep)) { buffered_rowset_.push_back(static_cast(batch.base_rowid+i)); RowBatch::Inst inst = batch[i]; for (bst_uint j = 0; j < inst.length; ++j) { - builder.AddBudget(inst[j].index); + if (enabled[inst[j].index]){ + builder.AddBudget(inst[j].index); + } } } } @@ -172,9 +179,11 @@ class FMatrixS : public IFMatrix{ ++ktop; RowBatch::Inst inst = batch[i]; for (bst_uint j = 0; j < inst.length; ++j) { - builder.PushElem(inst[j].index, - Entry((bst_uint)(batch.base_rowid+i), - inst[j].fvalue)); + if (enabled[inst[j].index]) { + builder.PushElem(inst[j].index, + Entry((bst_uint)(batch.base_rowid+i), + inst[j].fvalue)); + } } } } diff --git a/src/learner/evaluation-inl.hpp b/src/learner/evaluation-inl.hpp index fb0b8953daab..7334ad6891c7 100644 --- a/src/learner/evaluation-inl.hpp +++ b/src/learner/evaluation-inl.hpp @@ -11,6 +11,8 @@ #include #include #include +// rabit library for synchronization +#include #include "./evaluation.h" #include "./helper_utils.h" @@ -23,7 +25,8 @@ namespace learner { template struct EvalEWiseBase : public IEvaluator { virtual float Eval(const std::vector &preds, - const MetaInfo &info) const { + const MetaInfo &info, + bool distributed) const { utils::Check(info.labels.size() != 0, "label set cannot be empty"); utils::Check(preds.size() % info.labels.size() == 0, "label and prediction size not match"); @@ -37,7 +40,11 @@ struct EvalEWiseBase : public IEvaluator { sum += Derived::EvalRow(info.labels[i], preds[i]) * wt; wsum += wt; } - return Derived::GetFinal(sum, wsum); + float dat[2]; dat[0] = sum, dat[1] = wsum; + if (distributed) { + rabit::Allreduce(dat, 2); + } + return Derived::GetFinal(dat[0], dat[1]); } /*! * \brief to be implemented by subclass, @@ -113,7 +120,9 @@ struct EvalCTest: public IEvaluator { return name_.c_str(); } virtual float Eval(const std::vector &preds, - const MetaInfo &info) const { + const MetaInfo &info, + bool distributed) const { + utils::Check(!distributed, "metric %s do not support distributed evaluation", name_.c_str()); utils::Check(preds.size() % info.labels.size() == 0, "label and prediction size not match"); size_t ngroup = preds.size() / info.labels.size() - 1; @@ -150,7 +159,9 @@ struct EvalAMS : public IEvaluator { utils::Check(std::sscanf(name, "ams@%f", &ratio_) == 1, "invalid ams format"); } virtual float Eval(const std::vector &preds, - const MetaInfo &info) const { + const MetaInfo &info, + bool distributed) const { + utils::Check(!distributed, "metric AMS do not support distributed evaluation"); using namespace std; const bst_omp_uint ndata = static_cast(info.labels.size()); @@ -212,7 +223,9 @@ struct EvalPrecisionRatio : public IEvaluator{ } } virtual float Eval(const std::vector &preds, - const MetaInfo &info) const { + const MetaInfo &info, + bool distributed) const { + utils::Check(!distributed, "metric %s do not support distributed evaluation", Name()); utils::Check(info.labels.size() != 0, "label set cannot be empty"); utils::Assert(preds.size() % info.labels.size() == 0, "label size predict size not match"); @@ -252,7 +265,8 @@ struct EvalPrecisionRatio : public IEvaluator{ /*! \brief Area under curve, for both classification and rank */ struct EvalAuc : public IEvaluator { virtual float Eval(const std::vector &preds, - const MetaInfo &info) const { + const MetaInfo &info, + bool distributed) const { utils::Check(info.labels.size() != 0, "label set cannot be empty"); utils::Check(preds.size() % info.labels.size() == 0, "label size predict size not match"); @@ -299,8 +313,14 @@ struct EvalAuc : public IEvaluator { sum_auc += sum_pospair / (sum_npos*sum_nneg); } } - // return average AUC over list - return static_cast(sum_auc) / ngroup; + if (distributed) { + float dat[2]; dat[0] = sum_auc; dat[1] = ngroup; + // approximately estimate auc using mean + rabit::Allreduce(dat, 2); + return dat[0] / dat[1]; + } else { + return static_cast(sum_auc) / ngroup; + } } virtual const char *Name(void) const { return "auc"; @@ -311,7 +331,8 @@ struct EvalAuc : public IEvaluator { struct EvalRankList : public IEvaluator { public: virtual float Eval(const std::vector &preds, - const MetaInfo &info) const { + const MetaInfo &info, + bool distributed) const { utils::Check(preds.size() == info.labels.size(), "label size predict size not match"); // quick consistency when group is not available @@ -336,7 +357,14 @@ struct EvalRankList : public IEvaluator { sum_metric += this->EvalMetric(rec); } } - return static_cast(sum_metric) / ngroup; + if (distributed) { + float dat[2]; dat[0] = sum_metric; dat[1] = ngroup; + // approximately estimate auc using mean + rabit::Allreduce(dat, 2); + return dat[0] / dat[1]; + } else { + return static_cast(sum_metric) / ngroup; + } } virtual const char *Name(void) const { return name_.c_str(); diff --git a/src/learner/evaluation.h b/src/learner/evaluation.h index 33370e706f3a..4d59e270aba5 100644 --- a/src/learner/evaluation.h +++ b/src/learner/evaluation.h @@ -19,9 +19,13 @@ struct IEvaluator{ * \brief evaluate a specific metric * \param preds prediction * \param info information, including label etc. + * \param distributed whether a call to Allreduce is needed to gather + * the average statistics across all the node, + * this is only supported by some metrics */ virtual float Eval(const std::vector &preds, - const MetaInfo &info) const = 0; + const MetaInfo &info, + bool distributed = false) const = 0; /*! \return name of metric */ virtual const char *Name(void) const = 0; /*! \brief virtual destructor */ @@ -70,10 +74,11 @@ class EvalSet{ } inline std::string Eval(const char *evname, const std::vector &preds, - const MetaInfo &info) const { + const MetaInfo &info, + bool distributed = false) { std::string result = ""; for (size_t i = 0; i < evals_.size(); ++i) { - float res = evals_[i]->Eval(preds, info); + float res = evals_[i]->Eval(preds, info, distributed); char tmp[1024]; utils::SPrintf(tmp, sizeof(tmp), "\t%s-%s:%f", evname, evals_[i]->Name(), res); result += tmp; diff --git a/src/learner/learner-inl.hpp b/src/learner/learner-inl.hpp index 88026975db3b..5e3622e4dc30 100644 --- a/src/learner/learner-inl.hpp +++ b/src/learner/learner-inl.hpp @@ -10,6 +10,10 @@ #include #include #include +// rabit library for synchronization +#include +#include "../utils/io.h" +#include "../utils/base64.h" #include "./objective.h" #include "./evaluation.h" #include "../gbm/gbm.h" @@ -21,7 +25,7 @@ namespace learner { * \brief learner that takes do gradient boosting on specific objective functions * and do training and prediction */ -class BoostLearner { +class BoostLearner : public rabit::ISerializable { public: BoostLearner(void) { obj_ = NULL; @@ -30,8 +34,13 @@ class BoostLearner { name_gbm_ = "gbtree"; silent= 0; prob_buffer_row = 1.0f; + distributed_mode = 0; + pred_buffer_size = 0; + seed_per_iteration = 0; + seed = 0; + save_base64 = 0; } - ~BoostLearner(void) { + virtual ~BoostLearner(void) { if (obj_ != NULL) delete obj_; if (gbm_ != NULL) delete gbm_; } @@ -44,11 +53,9 @@ class BoostLearner { * \param mats array of pointers to matrix whose prediction result need to be cached */ inline void SetCacheData(const std::vector& mats) { - // estimate feature bound - unsigned num_feature = 0; + utils::Assert(cache_.size() == 0, "can only call cache data once"); // assign buffer index size_t buffer_size = 0; - utils::Assert(cache_.size() == 0, "can only call cache data once"); for (size_t i = 0; i < mats.size(); ++i) { bool dupilicate = false; for (size_t j = 0; j < i; ++j) { @@ -59,19 +66,12 @@ class BoostLearner { mats[i]->cache_learner_ptr_ = this; cache_.push_back(CacheEntry(mats[i], buffer_size, mats[i]->info.num_row())); buffer_size += mats[i]->info.num_row(); - num_feature = std::max(num_feature, static_cast(mats[i]->info.num_col())); } char str_temp[25]; - if (num_feature > mparam.num_feature) { - utils::SPrintf(str_temp, sizeof(str_temp), "%u", num_feature); - this->SetParam("bst:num_feature", str_temp); - } - utils::SPrintf(str_temp, sizeof(str_temp), "%lu", - static_cast(buffer_size)); + utils::SPrintf(str_temp, sizeof(str_temp), "%lu", + static_cast(buffer_size)); this->SetParam("num_pbuffer", str_temp); - if (!silent) { - utils::Printf("buffer_size=%ld\n", static_cast(buffer_size)); - } + this->pred_buffer_size = buffer_size; } /*! * \brief set parameters from outside @@ -86,9 +86,29 @@ class BoostLearner { this->SetParam(n.c_str(), val); } if (!strcmp(name, "silent")) silent = atoi(val); - if (!strcmp(name, "prob_buffer_row")) prob_buffer_row = static_cast(atof(val)); + if (!strcmp(name, "dsplit")) { + if (!strcmp(val, "col")) { + this->SetParam("updater", "distcol"); + distributed_mode = 1; + } else if (!strcmp(val, "row")) { + this->SetParam("updater", "grow_histmaker,prune"); + distributed_mode = 2; + } else { + utils::Error("%s is invalid value for dsplit, should be row or col", val); + } + } + if (!strcmp(name, "prob_buffer_row")) { + prob_buffer_row = static_cast(atof(val)); + utils::Check(distributed_mode == 0, + "prob_buffer_row can only be used in single node mode so far"); + this->SetParam("updater", "grow_colmaker,refresh,prune"); + } if (!strcmp(name, "eval_metric")) evaluator_.AddEval(val); - if (!strcmp("seed", name)) random::Seed(atoi(val)); + if (!strcmp("seed", name)) { + this->seed = seed; random::Seed(atoi(val)); + } + if (!strcmp("seed_per_iter", name)) seed_per_iteration = atoi(val); + if (!strcmp("save_base64", name)) save_base64 = atoi(val); if (!strcmp(name, "num_class")) this->SetParam("num_output_group", val); if (!strcmp(name, "nthread")) { omp_set_num_threads(atoi(val)); @@ -104,10 +124,29 @@ class BoostLearner { cfg_.push_back(std::make_pair(std::string(name), std::string(val))); } } + // this is an internal function + // initialize the trainer, called at InitModel and LoadModel + inline void InitTrainer(bool calc_num_feature = true) { + if (calc_num_feature) { + // estimate feature bound + unsigned num_feature = 0; + for (size_t i = 0; i < cache_.size(); ++i) { + num_feature = std::max(num_feature, + static_cast(cache_[i].mat_->info.num_col())); + } + // run allreduce on num_feature to find the maximum value + rabit::Allreduce(&num_feature, 1); + if (num_feature > mparam.num_feature) mparam.num_feature = num_feature; + } + char str_temp[25]; + utils::SPrintf(str_temp, sizeof(str_temp), "%d", mparam.num_feature); + this->SetParam("bst:num_feature", str_temp); + } /*! * \brief initialize the model */ inline void InitModel(void) { + this->InitTrainer(); // initialize model this->InitObjGBM(); // reset the base score @@ -118,8 +157,10 @@ class BoostLearner { /*! * \brief load model from stream * \param fi input stream + * \param with_pbuffer whether to load with predict buffer + * \param calc_num_feature whether call InitTrainer with calc_num_feature */ - inline void LoadModel(utils::IStream &fi) { + inline void LoadModel(utils::IStream &fi, bool with_pbuffer = true, bool calc_num_feature = true) { utils::Check(fi.Read(&mparam, sizeof(ModelParam)) != 0, "BoostLearner: wrong model format"); utils::Check(fi.Read(&name_obj_), "BoostLearner: wrong model format"); @@ -127,32 +168,83 @@ class BoostLearner { // delete existing gbm if any if (obj_ != NULL) delete obj_; if (gbm_ != NULL) delete gbm_; + this->InitTrainer(calc_num_feature); this->InitObjGBM(); - gbm_->LoadModel(fi); + gbm_->LoadModel(fi, with_pbuffer); + if (!with_pbuffer || distributed_mode == 2) { + gbm_->ResetPredBuffer(pred_buffer_size); + } + } + // rabit load model from rabit checkpoint + virtual void Load(rabit::IStream &fi) { + RabitStreamAdapter fs(fi); + // for row split, we should not keep pbuffer + this->LoadModel(fs, distributed_mode != 2, false); + } + // rabit save model to rabit checkpoint + virtual void Save(rabit::IStream &fo) const { + RabitStreamAdapter fs(fo); + // for row split, we should not keep pbuffer + this->SaveModel(fs, distributed_mode != 2); } /*! * \brief load model from file * \param fname file name */ inline void LoadModel(const char *fname) { - utils::FileStream fi(utils::FopenCheck(fname, "rb")); + FILE *fp = utils::FopenCheck(fname, "rb"); + std::string header; header.resize(4); + utils::FileStream fi(fp); + // check header for different binary encode + // can be base64 or binary + if (fi.Read(&header[0], 4) != 0) { + // base64 format + if (header == "bs64") { + utils::Base64InStream bsin(fp); + bsin.InitPosition(); + this->LoadModel(bsin); + fclose(fp); + return; + } + if (header == "binf") { + this->LoadModel(fi); + fclose(fp); + return; + } + } + fi.Seek(0); this->LoadModel(fi); - fi.Close(); + fclose(fp); } - inline void SaveModel(utils::IStream &fo) const { + inline void SaveModel(utils::IStream &fo, bool with_pbuffer = true) const { fo.Write(&mparam, sizeof(ModelParam)); fo.Write(name_obj_); fo.Write(name_gbm_); - gbm_->SaveModel(fo); + gbm_->SaveModel(fo, with_pbuffer); } /*! * \brief save model into file * \param fname file name */ inline void SaveModel(const char *fname) const { - utils::FileStream fo(utils::FopenCheck(fname, "wb")); - this->SaveModel(fo); - fo.Close(); + FILE *fp; + if (!strcmp(fname, "stdout")) { + fp = stdout; + } else { + fp = utils::FopenCheck(fname, "wb"); + } + utils::FileStream fo(fp); + std::string header; + if (save_base64 != 0|| fp == stdout) { + fo.Write("bs64\t", 5); + utils::Base64OutStream bout(fp); + this->SaveModel(bout); + bout.Finish('\n'); + } else { + fo.Write("binf", 4); + this->SaveModel(fo); + } + if (fp != stdout) fclose(fp); } /*! * \brief check if data matrix is ready to be used by training, @@ -160,7 +252,10 @@ class BoostLearner { * \param p_train pointer to the matrix used by training */ inline void CheckInit(DMatrix *p_train) { - p_train->fmat()->InitColAccess(prob_buffer_row); + int ncol = static_cast(p_train->info.info.num_col); + std::vector enabled(ncol, true); + // initialize column access + p_train->fmat()->InitColAccess(enabled, prob_buffer_row); } /*! * \brief update the model for one iteration @@ -168,9 +263,18 @@ class BoostLearner { * \param p_train pointer to the data matrix */ inline void UpdateOneIter(int iter, const DMatrix &train) { + if (seed_per_iteration || rabit::IsDistributed()) { + random::Seed(this->seed * kRandSeedMagic); + } this->PredictRaw(train, &preds_); obj_->GetGradient(preds_, train.info, iter, &gpair_); - gbm_->DoBoost(train.fmat(), train.info.info, &gpair_); + gbm_->DoBoost(train.fmat(), this->FindBufferOffset(train), train.info.info, &gpair_); + } + /*! + * \brief whether model allow lazy checkpoint + */ + inline bool AllowLazyCheckPoint(void) const { + return gbm_->AllowLazyCheckPoint(); } /*! * \brief evaluate the model for specific iteration @@ -189,7 +293,7 @@ class BoostLearner { for (size_t i = 0; i < evals.size(); ++i) { this->PredictRaw(*evals[i], &preds_); obj_->EvalTransform(&preds_); - res += evaluator_.Eval(evname[i].c_str(), preds_, evals[i]->info); + res += evaluator_.Eval(evname[i].c_str(), preds_, evals[i]->info, distributed_mode == 2); } return res; } @@ -219,10 +323,16 @@ class BoostLearner { inline void Predict(const DMatrix &data, bool output_margin, std::vector *out_preds, - unsigned ntree_limit = 0) const { - this->PredictRaw(data, out_preds, ntree_limit); - if (!output_margin) { - obj_->PredTransform(out_preds); + unsigned ntree_limit = 0, + bool pred_leaf = false + ) const { + if (pred_leaf) { + gbm_->PredictLeaf(data.fmat(), data.info.info, out_preds, ntree_limit); + } else { + this->PredictRaw(data, out_preds, ntree_limit); + if (!output_margin) { + obj_->PredTransform(out_preds); + } } } /*! \brief dump model out */ @@ -240,6 +350,7 @@ class BoostLearner { utils::Assert(gbm_ == NULL, "GBM and obj should be NULL"); obj_ = CreateObjFunction(name_obj_.c_str()); gbm_ = gbm::CreateGradBooster(name_gbm_.c_str()); + for (size_t i = 0; i < cfg_.size(); ++i) { obj_->SetParam(cfg_[i].first.c_str(), cfg_[i].second.c_str()); gbm_->SetParam(cfg_[i].first.c_str(), cfg_[i].second.c_str()); @@ -287,7 +398,7 @@ class BoostLearner { /* \brief number of class, if it is multi-class classification */ int num_class; /*! \brief reserved field */ - int reserved[32]; + int reserved[31]; /*! \brief constructor */ ModelParam(void) { base_score = 0.5f; @@ -308,14 +419,26 @@ class BoostLearner { } }; // data fields + // stored random seed + int seed; + // whether seed the PRNG each iteration + // this is important for restart from existing iterations + // default set to no, but will auto switch on in distributed mode + int seed_per_iteration; + // save model in base64 encoding + int save_base64; // silent during training int silent; + // distributed learning mode, if any, 0:none, 1:col, 2:row + int distributed_mode; + // cached size of predict buffer + size_t pred_buffer_size; // maximum buffred row value float prob_buffer_row; // evaluation set EvalSet evaluator_; // model parameter - ModelParam mparam; + ModelParam mparam; // gbm model that back everything gbm::IGradBooster *gbm_; // name of gbm model used for training @@ -331,7 +454,9 @@ class BoostLearner { // gradient pairs std::vector gpair_; - private: + protected: + // magic number to transform random seed + const static int kRandSeedMagic = 127; // cache entry object that helps handle feature caching struct CacheEntry { const DMatrix *mat_; @@ -354,6 +479,23 @@ class BoostLearner { // data structure field /*! \brief the entries indicates that we have internal prediction cache */ std::vector cache_; + + private: + // adapt rabit stream to utils stream + struct RabitStreamAdapter : public utils::IStream { + // rabit stream + rabit::IStream &fs; + // constructr + RabitStreamAdapter(rabit::IStream &fs) : fs(fs) {} + // destructor + virtual ~RabitStreamAdapter(void){} + virtual size_t Read(void *ptr, size_t size) { + return fs.Read(ptr, size); + } + virtual void Write(const void *ptr, size_t size) { + fs.Write(ptr, size); + } + }; }; } // namespace learner } // namespace xgboost diff --git a/src/learner/objective-inl.hpp b/src/learner/objective-inl.hpp index 96aacf12d05d..69b7ae4fd446 100644 --- a/src/learner/objective-inl.hpp +++ b/src/learner/objective-inl.hpp @@ -41,6 +41,25 @@ struct LossType { default: utils::Error("unknown loss_type"); return 0.0f; } } + /*! + * \brief check if label range is valid + */ + inline bool CheckLabel(float x) const { + if (loss_type != kLinearSquare) { + return x >= 0.0f && x <= 1.0f; + } + return true; + } + /*! + * \brief error message displayed when check label fail + */ + inline const char * CheckLabelErrorMsg(void) const { + if (loss_type != kLinearSquare) { + return "label must be in [0,1] for logistic regression"; + } else { + return ""; + } + } /*! * \brief calculate first order gradient of loss, given transformed prediction * \param predt transformed prediction @@ -115,6 +134,8 @@ class RegLossObj : public IObjFunction{ "labels are not correctly provided"); std::vector &gpair = *out_gpair; gpair.resize(preds.size()); + // check if label in range + bool label_correct = true; // start calculating gradient const unsigned nstep = static_cast(info.labels.size()); const bst_omp_uint ndata = static_cast(preds.size()); @@ -124,9 +145,11 @@ class RegLossObj : public IObjFunction{ float p = loss.PredTransform(preds[i]); float w = info.GetWeight(j); if (info.labels[j] == 1.0f) w *= scale_pos_weight; + if (!loss.CheckLabel(info.labels[j])) label_correct = false; gpair[i] = bst_gpair(loss.FirstOrderGradient(p, info.labels[j]) * w, loss.SecondOrderGradient(p, info.labels[j]) * w); } + utils::Check(label_correct, loss.CheckLabelErrorMsg()); } virtual const char* DefaultEvalMetric(void) const { return loss.DefaultEvalMetric(); diff --git a/src/tree/model.h b/src/tree/model.h index aa9ad2794b85..f3575488acbf 100644 --- a/src/tree/model.h +++ b/src/tree/model.h @@ -68,8 +68,9 @@ class TreeModel { } }; /*! \brief tree node */ - class Node{ + class Node { public: + Node(void) : sindex_(0) {} /*! \brief index of left child */ inline int cleft(void) const { return this->cleft_; @@ -110,6 +111,10 @@ class TreeModel { inline bool is_left_child(void) const { return (parent_ & (1U << 31)) != 0; } + /*! \brief whether this node is deleted */ + inline bool is_deleted(void) const { + return sindex_ == std::numeric_limits::max(); + } /*! \brief whether current node is root */ inline bool is_root(void) const { return parent_ == -1; @@ -144,7 +149,11 @@ class TreeModel { this->cleft_ = -1; this->cright_ = right; } - + /*! \brief mark that this node is deleted */ + inline void mark_delete(void) { + this->sindex_ = std::numeric_limits::max(); + } + private: friend class TreeModel; /*! @@ -197,11 +206,11 @@ class TreeModel { leaf_vector.resize(param.num_nodes * param.size_leaf_vector); return nd; } - // delete a tree node + // delete a tree node, keep the parent field to allow trace back inline void DeleteNode(int nid) { utils::Assert(nid >= param.num_roots, "can not delete root"); deleted_nodes.push_back(nid); - nodes[nid].set_parent(-1); + nodes[nid].mark_delete(); ++param.num_deleted; } @@ -296,11 +305,12 @@ class TreeModel { } // chg deleted nodes deleted_nodes.resize(0); - for (int i = param.num_roots; i < param.num_nodes; i ++) { - if (nodes[i].is_root()) deleted_nodes.push_back(i); + for (int i = param.num_roots; i < param.num_nodes; ++i) { + if (nodes[i].is_deleted()) deleted_nodes.push_back(i); } utils::Assert(static_cast(deleted_nodes.size()) == param.num_deleted, - "number of deleted nodes do not match"); + "number of deleted nodes do not match, num_deleted=%d, dnsize=%lu, num_nodes=%d", + param.num_deleted, deleted_nodes.size(), param.num_nodes); } /*! * \brief save model to stream diff --git a/src/tree/param.h b/src/tree/param.h index 04ea5277f9b7..701721c177f9 100644 --- a/src/tree/param.h +++ b/src/tree/param.h @@ -36,8 +36,14 @@ struct TrainParam{ float colsample_bytree; // speed optimization for dense column float opt_dense_col; + // accuracy of sketch + float sketch_eps; + // accuracy of sketch + float sketch_ratio; // leaf vector size - int size_leaf_vector; + int size_leaf_vector; + // option for parallelization + int parallel_option; // number of threads to be used for tree construction, // if OpenMP is enabled, if equals 0, use system default int nthread; @@ -55,6 +61,9 @@ struct TrainParam{ opt_dense_col = 1.0f; nthread = 0; size_leaf_vector = 0; + parallel_option = 2; + sketch_eps = 0.1f; + sketch_ratio = 2.0f; } /*! * \brief set parameters from outside @@ -76,10 +85,13 @@ struct TrainParam{ if (!strcmp(name, "subsample")) subsample = static_cast(atof(val)); if (!strcmp(name, "colsample_bylevel")) colsample_bylevel = static_cast(atof(val)); if (!strcmp(name, "colsample_bytree")) colsample_bytree = static_cast(atof(val)); + if (!strcmp(name, "sketch_eps")) sketch_eps = static_cast(atof(val)); + if (!strcmp(name, "sketch_ratio")) sketch_ratio = static_cast(atof(val)); if (!strcmp(name, "opt_dense_col")) opt_dense_col = static_cast(atof(val)); if (!strcmp(name, "size_leaf_vector")) size_leaf_vector = atoi(val); if (!strcmp(name, "max_depth")) max_depth = atoi(val); if (!strcmp(name, "nthread")) nthread = atoi(val); + if (!strcmp(name, "parallel_option")) parallel_option = atoi(val); if (!strcmp(name, "default_direction")) { if (!strcmp(val, "learn")) default_direction = 0; if (!strcmp(val, "left")) default_direction = 1; @@ -132,6 +144,12 @@ struct TrainParam{ inline bool cannot_split(double sum_hess, int depth) const { return sum_hess < this->min_child_weight * 2.0; } + /*! \brief maximum sketch size */ + inline unsigned max_sketch_size(void) const { + unsigned ret = static_cast(sketch_ratio / sketch_eps); + utils::Check(ret > 0, "sketch_ratio/sketch_eps must be bigger than 1"); + return ret; + } protected: // functions for L1 cost @@ -186,6 +204,10 @@ struct GradStats { inline void Add(const GradStats &b) { this->Add(b.sum_grad, b.sum_hess); } + /*! \brief same as add, reduce is used in All Reduce */ + inline void Reduce(const GradStats &b) { + this->Add(b); + } /*! \brief set current value to a - b */ inline void SetSubstract(const GradStats &a, const GradStats &b) { sum_grad = a.sum_grad - b.sum_grad; @@ -262,6 +284,10 @@ struct CVGradStats : public GradStats { valid[i].Add(b.valid[i]); } } + /*! \brief same as add, reduce is used in All Reduce */ + inline void Reduce(const CVGradStats &b) { + this->Add(b); + } /*! \brief set current value to a - b */ inline void SetSubstract(const CVGradStats &a, const CVGradStats &b) { GradStats::SetSubstract(a, b); @@ -341,6 +367,10 @@ struct SplitEntry{ return false; } } + /*! \brief same as update, used by AllReduce*/ + inline void Reduce(const SplitEntry &e) { + this->Update(e); + } /*!\return feature index to split on */ inline unsigned split_index(void) const { return sindex & ((1U << 31) - 1U); diff --git a/src/tree/updater.cpp b/src/tree/updater.cpp index 2cb6552fe8ed..ca19b93b35a5 100644 --- a/src/tree/updater.cpp +++ b/src/tree/updater.cpp @@ -1,18 +1,28 @@ #define _CRT_SECURE_NO_WARNINGS #define _CRT_SECURE_NO_DEPRECATE +#define NOMINMAX #include #include "./updater.h" +#include "./updater_sync-inl.hpp" #include "./updater_prune-inl.hpp" #include "./updater_refresh-inl.hpp" #include "./updater_colmaker-inl.hpp" +#include "./updater_distcol-inl.hpp" +#include "./updater_histmaker-inl.hpp" +//#include "./updater_skmaker-inl.hpp" namespace xgboost { namespace tree { IUpdater* CreateUpdater(const char *name) { using namespace std; if (!strcmp(name, "prune")) return new TreePruner(); + if (!strcmp(name, "sync")) return new TreeSyncher(); if (!strcmp(name, "refresh")) return new TreeRefresher(); if (!strcmp(name, "grow_colmaker")) return new ColMaker(); + if (!strcmp(name, "grow_histmaker")) return new CQHistMaker(); + //if (!strcmp(name, "grow_skmaker")) return new SketchMaker(); + if (!strcmp(name, "distcol")) return new DistColMaker(); + utils::Error("unknown updater:%s", name); return NULL; } diff --git a/src/tree/updater.h b/src/tree/updater.h index e3a05c84f46a..49adc8dca27c 100644 --- a/src/tree/updater.h +++ b/src/tree/updater.h @@ -37,6 +37,16 @@ class IUpdater { IFMatrix *p_fmat, const BoosterInfo &info, const std::vector &trees) = 0; + + /*! + * \brief this is simply a function for optimizing performance + * this function asks the updater to return the leaf position of each instance in the p_fmat, + * if it is cached in the updater, if it is not available, return NULL + * \return array of leaf position of each instance in the last updated tree + */ + virtual const int* GetLeafPosition(void) const { + return NULL; + } // destructor virtual ~IUpdater(void) {} }; diff --git a/src/tree/updater_basemaker-inl.hpp b/src/tree/updater_basemaker-inl.hpp new file mode 100644 index 000000000000..851811eaee80 --- /dev/null +++ b/src/tree/updater_basemaker-inl.hpp @@ -0,0 +1,409 @@ +#ifndef XGBOOST_TREE_UPDATER_BASEMAKER_INL_HPP_ +#define XGBOOST_TREE_UPDATER_BASEMAKER_INL_HPP_ +/*! + * \file updater_basemaker-inl.hpp + * \brief implement a common tree constructor + * \author Tianqi Chen + */ +#include +#include +#include +#include +#include "../utils/random.h" +#include "../utils/quantile.h" + +namespace xgboost { +namespace tree { +/*! + * \brief base tree maker class that defines common operation + * needed in tree making + */ +class BaseMaker: public IUpdater { + public: + // destructor + virtual ~BaseMaker(void) {} + // set training parameter + virtual void SetParam(const char *name, const char *val) { + param.SetParam(name, val); + } + + protected: + // helper to collect and query feature meta information + struct FMetaHelper { + public: + /*! \brief find type of each feature, use column format */ + inline void InitByCol(IFMatrix *p_fmat, + const RegTree &tree) { + fminmax.resize(tree.param.num_feature * 2); + std::fill(fminmax.begin(), fminmax.end(), + -std::numeric_limits::max()); + // start accumulating statistics + utils::IIterator *iter = p_fmat->ColIterator(); + iter->BeforeFirst(); + while (iter->Next()) { + const ColBatch &batch = iter->Value(); + for (bst_uint i = 0; i < batch.size; ++i) { + const bst_uint fid = batch.col_index[i]; + const ColBatch::Inst &c = batch[i]; + if (c.length != 0) { + fminmax[fid * 2 + 0] = std::max(-c[0].fvalue, fminmax[fid * 2 + 0]); + fminmax[fid * 2 + 1] = std::max(c[c.length - 1].fvalue, fminmax[fid * 2 + 1]); + } + } + } + rabit::Allreduce(BeginPtr(fminmax), fminmax.size()); + } + // get feature type, 0:empty 1:binary 2:real + inline int Type(bst_uint fid) const { + utils::Assert(fid * 2 + 1 < fminmax.size(), + "FeatHelper fid exceed query bound "); + bst_float a = fminmax[fid * 2]; + bst_float b = fminmax[fid * 2 + 1]; + if (a == -std::numeric_limits::max()) return 0; + if (-a == b) return 1; + else return 2; + } + inline bst_float MaxValue(bst_uint fid) const { + return fminmax[fid *2 + 1]; + } + inline void SampleCol(float p, std::vector *p_findex) const { + std::vector &findex = *p_findex; + findex.clear(); + for (size_t i = 0; i < fminmax.size(); i += 2) { + const bst_uint fid = static_cast(i / 2); + if (this->Type(fid) != 0) findex.push_back(fid); + } + unsigned n = static_cast(p * findex.size()); + random::Shuffle(findex); + findex.resize(n); + // sync the findex if it is subsample + std::string s_cache; + utils::MemoryBufferStream fc(&s_cache); + utils::IStream &fs = fc; + if (rabit::GetRank() == 0) { + fs.Write(findex); + } + rabit::Broadcast(&s_cache, 0); + fs.Read(&findex); + } + + private: + std::vector fminmax; + }; + // ------static helper functions ------ + // helper function to get to next level of the tree + /*! \brief this is helper function for row based data*/ + inline static int NextLevel(const RowBatch::Inst &inst, const RegTree &tree, int nid) { + const RegTree::Node &n = tree[nid]; + bst_uint findex = n.split_index(); + for (unsigned i = 0; i < inst.length; ++i) { + if (findex == inst[i].index) { + if (inst[i].fvalue < n.split_cond()) { + return n.cleft(); + } else { + return n.cright(); + } + } + } + return n.cdefault(); + } + /*! \brief get number of omp thread in current context */ + inline static int get_nthread(void) { + int nthread; + #pragma omp parallel + { + nthread = omp_get_num_threads(); + } + return nthread; + } + // ------class member helpers--------- + /*! \brief initialize temp data structure */ + inline void InitData(const std::vector &gpair, + const IFMatrix &fmat, + const std::vector &root_index, + const RegTree &tree) { + utils::Assert(tree.param.num_nodes == tree.param.num_roots, + "TreeMaker: can only grow new tree"); + {// setup position + position.resize(gpair.size()); + if (root_index.size() == 0) { + std::fill(position.begin(), position.end(), 0); + } else { + for (size_t i = 0; i < position.size(); ++i) { + position[i] = root_index[i]; + utils::Assert(root_index[i] < (unsigned)tree.param.num_roots, + "root index exceed setting"); + } + } + // mark delete for the deleted datas + for (size_t i = 0; i < position.size(); ++i) { + if (gpair[i].hess < 0.0f) position[i] = ~position[i]; + } + // mark subsample + if (param.subsample < 1.0f) { + for (size_t i = 0; i < position.size(); ++i) { + if (gpair[i].hess < 0.0f) continue; + if (random::SampleBinary(param.subsample) == 0) position[i] = ~position[i]; + } + } + } + {// expand query + qexpand.reserve(256); qexpand.clear(); + for (int i = 0; i < tree.param.num_roots; ++i) { + qexpand.push_back(i); + } + this->UpdateNode2WorkIndex(tree); + } + } + /*! \brief update queue expand add in new leaves */ + inline void UpdateQueueExpand(const RegTree &tree) { + std::vector newnodes; + for (size_t i = 0; i < qexpand.size(); ++i) { + const int nid = qexpand[i]; + if (!tree[nid].is_leaf()) { + newnodes.push_back(tree[nid].cleft()); + newnodes.push_back(tree[nid].cright()); + } + } + // use new nodes for qexpand + qexpand = newnodes; + this->UpdateNode2WorkIndex(tree); + } + // return decoded position + inline int DecodePosition(bst_uint ridx) const{ + const int pid = position[ridx]; + return pid < 0 ? ~pid : pid; + } + // encode the encoded position value for ridx + inline void SetEncodePosition(bst_uint ridx, int nid) { + if (position[ridx] < 0) { + position[ridx] = ~nid; + } else { + position[ridx] = nid; + } + } + /*! + * \brief this is helper function uses column based data structure, + * reset the positions to the lastest one + * \param nodes the set of nodes that contains the split to be used + * \param p_fmat feature matrix needed for tree construction + * \param tree the regression tree structure + */ + inline void ResetPositionCol(const std::vector &nodes, IFMatrix *p_fmat, const RegTree &tree) { + // set the positions in the nondefault + this->SetNonDefaultPositionCol(nodes, p_fmat, tree); + // set rest of instances to default position + const std::vector &rowset = p_fmat->buffered_rowset(); + // set default direct nodes to default + // for leaf nodes that are not fresh, mark then to ~nid, + // so that they are ignored in future statistics collection + const bst_omp_uint ndata = static_cast(rowset.size()); + + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < ndata; ++i) { + const bst_uint ridx = rowset[i]; + const int nid = this->DecodePosition(ridx); + if (tree[nid].is_leaf()) { + // mark finish when it is not a fresh leaf + if (tree[nid].cright() == -1) { + position[ridx] = ~nid; + } + } else { + // push to default branch + if (tree[nid].default_left()) { + this->SetEncodePosition(ridx, tree[nid].cleft()); + } else { + this->SetEncodePosition(ridx, tree[nid].cright()); + } + } + } + } + /*! + * \brief this is helper function uses column based data structure, + * update all positions into nondefault branch, if any, ignore the default branch + * \param nodes the set of nodes that contains the split to be used + * \param p_fmat feature matrix needed for tree construction + * \param tree the regression tree structure + */ + virtual void SetNonDefaultPositionCol(const std::vector &nodes, + IFMatrix *p_fmat, const RegTree &tree) { + // step 1, classify the non-default data into right places + std::vector fsplits; + for (size_t i = 0; i < nodes.size(); ++i) { + const int nid = nodes[i]; + if (!tree[nid].is_leaf()) { + fsplits.push_back(tree[nid].split_index()); + } + } + std::sort(fsplits.begin(), fsplits.end()); + fsplits.resize(std::unique(fsplits.begin(), fsplits.end()) - fsplits.begin()); + + utils::IIterator *iter = p_fmat->ColIterator(fsplits); + while (iter->Next()) { + const ColBatch &batch = iter->Value(); + for (size_t i = 0; i < batch.size; ++i) { + ColBatch::Inst col = batch[i]; + const bst_uint fid = batch.col_index[i]; + const bst_omp_uint ndata = static_cast(col.length); + #pragma omp parallel for schedule(static) + for (bst_omp_uint j = 0; j < ndata; ++j) { + const bst_uint ridx = col[j].index; + const float fvalue = col[j].fvalue; + const int nid = this->DecodePosition(ridx); + // go back to parent, correct those who are not default + if (!tree[nid].is_leaf() && tree[nid].split_index() == fid) { + if(fvalue < tree[nid].split_cond()) { + this->SetEncodePosition(ridx, tree[nid].cleft()); + } else { + this->SetEncodePosition(ridx, tree[nid].cright()); + } + } + } + } + } + } + /*! \brief helper function to get statistics from a tree */ + template + inline void GetNodeStats(const std::vector &gpair, + const IFMatrix &fmat, + const RegTree &tree, + const BoosterInfo &info, + std::vector< std::vector > *p_thread_temp, + std::vector *p_node_stats) { + std::vector< std::vector > &thread_temp = *p_thread_temp; + thread_temp.resize(this->get_nthread()); + p_node_stats->resize(tree.param.num_nodes); + #pragma omp parallel + { + const int tid = omp_get_thread_num(); + thread_temp[tid].resize(tree.param.num_nodes, TStats(param)); + for (size_t i = 0; i < qexpand.size(); ++i) { + const unsigned nid = qexpand[i]; + thread_temp[tid][nid].Clear(); + } + } + const std::vector &rowset = fmat.buffered_rowset(); + // setup position + const bst_omp_uint ndata = static_cast(rowset.size()); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < ndata; ++i) { + const bst_uint ridx = rowset[i]; + const int nid = position[ridx]; + const int tid = omp_get_thread_num(); + if (nid >= 0) { + thread_temp[tid][nid].Add(gpair, info, ridx); + } + } + // sum the per thread statistics together + for (size_t j = 0; j < qexpand.size(); ++j) { + const int nid = qexpand[j]; + TStats &s = (*p_node_stats)[nid]; + s.Clear(); + for (size_t tid = 0; tid < thread_temp.size(); ++tid) { + s.Add(thread_temp[tid][nid]); + } + } + } + /*! \brief common helper data structure to build sketch*/ + struct SketchEntry { + /*! \brief total sum of amount to be met */ + bst_float sum_total; + /*! \brief statistics used in the sketch */ + bst_float rmin, wmin; + /*! \brief last seen feature value */ + bst_float last_fvalue; + /*! \brief current size of sketch */ + bst_float next_goal; + // pointer to the sketch to put things in + utils::WXQuantileSketch *sketch; + // initialize the space + inline void Init(unsigned max_size) { + next_goal = -1.0f; + rmin = wmin = 0.0f; + sketch->temp.Reserve(max_size + 1); + sketch->temp.size = 0; + } + /*! + * \brief push a new element to sketch + * \param fvalue feature value, comes in sorted ascending order + * \param w weight + * \param max_size + */ + inline void Push(bst_float fvalue, bst_float w, unsigned max_size) { + if (next_goal == -1.0f) { + next_goal = 0.0f; + last_fvalue = fvalue; + wmin = w; + return; + } + if (last_fvalue != fvalue) { + bst_float rmax = rmin + wmin; + if (rmax >= next_goal) { + if (sketch->temp.size == 0 || last_fvalue > sketch->temp.data[sketch->temp.size-1].value) { + // push to sketch + sketch->temp.data[sketch->temp.size] = + utils::WXQuantileSketch:: + Entry(rmin, rmax, wmin, last_fvalue); + utils::Assert(sketch->temp.size < max_size, + "invalid maximum size max_size=%u, stemp.size=%lu\n", + max_size, sketch->temp.size); + ++sketch->temp.size; + } + if (sketch->temp.size == max_size) { + next_goal = sum_total * 2.0f + 1e-5f; + } else{ + next_goal = static_cast(sketch->temp.size * sum_total / max_size); + } + } + rmin = rmax; + wmin = w; + last_fvalue = fvalue; + } else { + wmin += w; + } + } + /*! \brief push final unfinished value to the sketch */ + inline void Finalize(unsigned max_size) { + bst_float rmax = rmin + wmin; + if (sketch->temp.size == 0 || last_fvalue > sketch->temp.data[sketch->temp.size-1].value) { + utils::Assert(sketch->temp.size <= max_size, + "Finalize: invalid maximum size, max_size=%u, stemp.size=%lu", + sketch->temp.size, max_size ); + // push to sketch + sketch->temp.data[sketch->temp.size] = + utils::WXQuantileSketch:: + Entry(rmin, rmax, wmin, last_fvalue); + ++sketch->temp.size; + } + sketch->PushTemp(); + } + }; + /*! \brief training parameter of tree grower */ + TrainParam param; + /*! \brief queue of nodes to be expanded */ + std::vector qexpand; + /*! + * \brief map active node to is working index offset in qexpand, + * can be -1, which means the node is node actively expanding + */ + std::vector node2workindex; + /*! + * \brief position of each instance in the tree + * can be negative, which means this position is no longer expanding + * see also Decode/EncodePosition + */ + std::vector position; + + private: + inline void UpdateNode2WorkIndex(const RegTree &tree) { + // update the node2workindex + std::fill(node2workindex.begin(), node2workindex.end(), -1); + node2workindex.resize(tree.param.num_nodes); + for (size_t i = 0; i < qexpand.size(); ++i) { + node2workindex[qexpand[i]] = static_cast(i); + } + } +}; +} // namespace tree +} // namespace xgboost +#endif // XGBOOST_TREE_UPDATER_BASEMAKER_INL_HPP_ diff --git a/src/tree/updater_colmaker-inl.hpp b/src/tree/updater_colmaker-inl.hpp index 2d7c5311e2f1..61b7fbf7a3ea 100644 --- a/src/tree/updater_colmaker-inl.hpp +++ b/src/tree/updater_colmaker-inl.hpp @@ -14,7 +14,7 @@ namespace xgboost { namespace tree { -/*! \brief pruner that prunes a tree after growing finishs */ +/*! \brief colunwise update to construct a tree */ template class ColMaker: public IUpdater { public: @@ -36,24 +36,29 @@ class ColMaker: public IUpdater { Builder builder(param); builder.Update(gpair, p_fmat, info, trees[i]); } + param.learning_rate = lr; } - private: + protected: // training parameter TrainParam param; // data structure /*! \brief per thread x per node entry to store tmp data */ struct ThreadEntry { - /*! \brief statistics of data*/ + /*! \brief statistics of data */ TStats stats; + /*! \brief extra statistics of data */ + TStats stats_extra; /*! \brief last feature value scanned */ float last_fvalue; + /*! \brief first feature value scanned */ + float first_fvalue; /*! \brief current best solution */ SplitEntry best; // constructor explicit ThreadEntry(const TrainParam ¶m) - : stats(param) { + : stats(param), stats_extra(param) { } }; struct NodeEntry { @@ -104,7 +109,7 @@ class ColMaker: public IUpdater { } } - private: + protected: // initialize temp data structure inline void InitData(const std::vector &gpair, const IFMatrix &fmat, @@ -127,17 +132,17 @@ class ColMaker: public IUpdater { // mark delete for the deleted datas for (size_t i = 0; i < rowset.size(); ++i) { const bst_uint ridx = rowset[i]; - if (gpair[ridx].hess < 0.0f) position[ridx] = -1; + if (gpair[ridx].hess < 0.0f) position[ridx] = ~position[ridx]; } // mark subsample if (param.subsample < 1.0f) { for (size_t i = 0; i < rowset.size(); ++i) { const bst_uint ridx = rowset[i]; if (gpair[ridx].hess < 0.0f) continue; - if (random::SampleBinary(param.subsample) == 0) position[ridx] = -1; + if (random::SampleBinary(param.subsample) == 0) position[ridx] = ~position[ridx]; } } - } + } { // initialize feature index unsigned ncol = static_cast(fmat.NumCol()); @@ -219,7 +224,138 @@ class ColMaker: public IUpdater { } // use new nodes for qexpand qexpand = newnodes; - } + } + // parallel find the best split of current fid + // this function does not support nested functions + inline void ParallelFindSplit(const ColBatch::Inst &col, + bst_uint fid, + const IFMatrix &fmat, + const std::vector &gpair, + const BoosterInfo &info) { + bool need_forward = param.need_forward_search(fmat.GetColDensity(fid)); + bool need_backward = param.need_backward_search(fmat.GetColDensity(fid)); + const std::vector &qexpand = qexpand_; + int nthread; + #pragma omp parallel + { + const int tid = omp_get_thread_num(); + std::vector &temp = stemp[tid]; + // cleanup temp statistics + for (size_t j = 0; j < qexpand.size(); ++j) { + temp[qexpand[j]].stats.Clear(); + } + nthread = omp_get_num_threads(); + bst_uint step = (col.length + nthread - 1) / nthread; + bst_uint end = std::min(col.length, step * (tid + 1)); + for (bst_uint i = tid * step; i < end; ++i) { + const bst_uint ridx = col[i].index; + const int nid = position[ridx]; + if (nid < 0) continue; + const float fvalue = col[i].fvalue; + if (temp[nid].stats.Empty()) { + temp[nid].first_fvalue = fvalue; + } + temp[nid].stats.Add(gpair, info, ridx); + temp[nid].last_fvalue = fvalue; + } + } + // start collecting the partial sum statistics + bst_omp_uint nnode = static_cast(qexpand.size()); + #pragma omp parallel for schedule(static) + for (bst_omp_uint j = 0; j < nnode; ++j) { + const int nid = qexpand[j]; + TStats sum(param), tmp(param), c(param); + for (int tid = 0; tid < nthread; ++tid) { + tmp = stemp[tid][nid].stats; + stemp[tid][nid].stats = sum; + sum.Add(tmp); + if (tid != 0) { + std::swap(stemp[tid - 1][nid].last_fvalue, stemp[tid][nid].first_fvalue); + } + } + for (int tid = 0; tid < nthread; ++tid) { + stemp[tid][nid].stats_extra = sum; + ThreadEntry &e = stemp[tid][nid]; + float fsplit; + if (tid != 0) { + if(fabsf(stemp[tid - 1][nid].last_fvalue - e.first_fvalue) > rt_2eps) { + fsplit = (stemp[tid - 1][nid].last_fvalue - e.first_fvalue) * 0.5f; + } else { + continue; + } + } else { + fsplit = e.first_fvalue - rt_eps; + } + if (need_forward && tid != 0) { + c.SetSubstract(snode[nid].stats, e.stats); + if (c.sum_hess >= param.min_child_weight && e.stats.sum_hess >= param.min_child_weight) { + bst_float loss_chg = static_cast(e.stats.CalcGain(param) + c.CalcGain(param) - snode[nid].root_gain); + e.best.Update(loss_chg, fid, fsplit, false); + } + } + if (need_backward) { + tmp.SetSubstract(sum, e.stats); + c.SetSubstract(snode[nid].stats, tmp); + if (c.sum_hess >= param.min_child_weight && tmp.sum_hess >= param.min_child_weight) { + bst_float loss_chg = static_cast(tmp.CalcGain(param) + c.CalcGain(param) - snode[nid].root_gain); + e.best.Update(loss_chg, fid, fsplit, true); + } + } + } + if (need_backward) { + tmp = sum; + ThreadEntry &e = stemp[nthread-1][nid]; + c.SetSubstract(snode[nid].stats, tmp); + if (c.sum_hess >= param.min_child_weight && tmp.sum_hess >= param.min_child_weight) { + bst_float loss_chg = static_cast(tmp.CalcGain(param) + c.CalcGain(param) - snode[nid].root_gain); + e.best.Update(loss_chg, fid, e.last_fvalue + rt_eps, true); + } + } + } + // rescan, generate candidate split + #pragma omp parallel + { + TStats c(param), cright(param); + const int tid = omp_get_thread_num(); + std::vector &temp = stemp[tid]; + nthread = static_cast(omp_get_num_threads()); + bst_uint step = (col.length + nthread - 1) / nthread; + bst_uint end = std::min(col.length, step * (tid + 1)); + for (bst_uint i = tid * step; i < end; ++i) { + const bst_uint ridx = col[i].index; + const int nid = position[ridx]; + if (nid < 0) continue; + const float fvalue = col[i].fvalue; + // get the statistics of nid + ThreadEntry &e = temp[nid]; + if (e.stats.Empty()) { + e.stats.Add(gpair, info, ridx); + e.first_fvalue = fvalue; + } else { + // forward default right + if (fabsf(fvalue - e.first_fvalue) > rt_2eps){ + if (need_forward) { + c.SetSubstract(snode[nid].stats, e.stats); + if (c.sum_hess >= param.min_child_weight && e.stats.sum_hess >= param.min_child_weight) { + bst_float loss_chg = static_cast(e.stats.CalcGain(param) + c.CalcGain(param) - snode[nid].root_gain); + e.best.Update(loss_chg, fid, (fvalue + e.first_fvalue) * 0.5f, false); + } + } + if (need_backward) { + cright.SetSubstract(e.stats_extra, e.stats); + c.SetSubstract(snode[nid].stats, cright); + if (c.sum_hess >= param.min_child_weight && cright.sum_hess >= param.min_child_weight) { + bst_float loss_chg = static_cast(cright.CalcGain(param) + c.CalcGain(param) - snode[nid].root_gain); + e.best.Update(loss_chg, fid, (fvalue + e.first_fvalue) * 0.5f, true); + } + } + } + e.stats.Add(gpair, info, ridx); + e.first_fvalue = fvalue; + } + } + } + } // enumerate the split values of specific feature inline void EnumerateSplit(const ColBatch::Entry *begin, const ColBatch::Entry *end, @@ -273,6 +409,42 @@ class ColMaker: public IUpdater { } } } + // update the solution candidate + virtual void UpdateSolution(const ColBatch &batch, + const std::vector &gpair, + const IFMatrix &fmat, + const BoosterInfo &info) { + // start enumeration + const bst_omp_uint nsize = static_cast(batch.size); + #if defined(_OPENMP) + const int batch_size = std::max(static_cast(nsize / this->nthread / 32), 1); + #endif + int poption = param.parallel_option; + if (poption == 2) { + poption = nsize * 2 < nthread ? 1 : 0; + } + if (poption == 0) { + #pragma omp parallel for schedule(dynamic, batch_size) + for (bst_omp_uint i = 0; i < nsize; ++i) { + const bst_uint fid = batch.col_index[i]; + const int tid = omp_get_thread_num(); + const ColBatch::Inst c = batch[i]; + if (param.need_forward_search(fmat.GetColDensity(fid))) { + this->EnumerateSplit(c.data, c.data + c.length, +1, + fid, gpair, info, stemp[tid]); + } + if (param.need_backward_search(fmat.GetColDensity(fid))) { + this->EnumerateSplit(c.data + c.length - 1, c.data - 1, -1, + fid, gpair, info, stemp[tid]); + } + } + } else { + for (bst_omp_uint i = 0; i < nsize; ++i) { + this->ParallelFindSplit(batch[i], batch.col_index[i], + fmat, gpair, info); + } + } + } // find splits at current level, do split per level inline void FindSplit(int depth, const std::vector &qexpand, @@ -289,66 +461,76 @@ class ColMaker: public IUpdater { } utils::IIterator *iter = p_fmat->ColIterator(feat_set); while (iter->Next()) { - const ColBatch &batch = iter->Value(); - // start enumeration - const bst_omp_uint nsize = static_cast(batch.size); - #if defined(_OPENMP) - const int batch_size = std::max(static_cast(nsize / this->nthread / 32), 1); - #endif - #pragma omp parallel for schedule(dynamic, batch_size) - for (bst_omp_uint i = 0; i < nsize; ++i) { - const bst_uint fid = batch.col_index[i]; - const int tid = omp_get_thread_num(); - const ColBatch::Inst c = batch[i]; - if (param.need_forward_search(p_fmat->GetColDensity(fid))) { - this->EnumerateSplit(c.data, c.data + c.length, +1, - fid, gpair, info, stemp[tid]); - } - if (param.need_backward_search(p_fmat->GetColDensity(fid))) { - this->EnumerateSplit(c.data + c.length - 1, c.data - 1, -1, - fid, gpair, info, stemp[tid]); - } - } + this->UpdateSolution(iter->Value(), gpair, *p_fmat, info); } // after this each thread's stemp will get the best candidates, aggregate results + this->SyncBestSolution(qexpand); + // get the best result, we can synchronize the solution for (size_t i = 0; i < qexpand.size(); ++i) { const int nid = qexpand[i]; - NodeEntry &e = snode[nid]; - for (int tid = 0; tid < this->nthread; ++tid) { - e.best.Update(stemp[tid][nid].best); - } + NodeEntry &e = snode[nid]; // now we know the solution in snode[nid], set split if (e.best.loss_chg > rt_eps) { p_tree->AddChilds(nid); (*p_tree)[nid].set_split(e.best.split_index(), e.best.split_value, e.best.default_left()); + // mark right child as 0, to indicate fresh leaf + (*p_tree)[(*p_tree)[nid].cleft()].set_leaf(0.0f, 0); + (*p_tree)[(*p_tree)[nid].cright()].set_leaf(0.0f, 0); } else { (*p_tree)[nid].set_leaf(e.weight * param.learning_rate); } - } + } } // reset position of each data points after split is created in the tree inline void ResetPosition(const std::vector &qexpand, IFMatrix *p_fmat, const RegTree &tree) { + // set the positions in the nondefault + this->SetNonDefaultPosition(qexpand, p_fmat, tree); + // set rest of instances to default position const std::vector &rowset = p_fmat->buffered_rowset(); - // step 1, set default direct nodes to default, and leaf nodes to -1 + // set default direct nodes to default + // for leaf nodes that are not fresh, mark then to ~nid, + // so that they are ignored in future statistics collection const bst_omp_uint ndata = static_cast(rowset.size()); + #pragma omp parallel for schedule(static) for (bst_omp_uint i = 0; i < ndata; ++i) { const bst_uint ridx = rowset[i]; - const int nid = position[ridx]; - if (nid >= 0) { - if (tree[nid].is_leaf()) { - position[ridx] = -1; + const int nid = this->DecodePosition(ridx); + if (tree[nid].is_leaf()) { + // mark finish when it is not a fresh leaf + if (tree[nid].cright() == -1) { + position[ridx] = ~nid; + } + } else { + // push to default branch + if (tree[nid].default_left()) { + this->SetEncodePosition(ridx, tree[nid].cleft()); } else { - // push to default branch, correct latter - position[ridx] = tree[nid].default_left() ? tree[nid].cleft(): tree[nid].cright(); + this->SetEncodePosition(ridx, tree[nid].cright()); } } } - // step 2, classify the non-default data into right places + } + // customization part + // synchronize the best solution of each node + virtual void SyncBestSolution(const std::vector &qexpand) { + for (size_t i = 0; i < qexpand.size(); ++i) { + const int nid = qexpand[i]; + NodeEntry &e = snode[nid]; + for (int tid = 0; tid < this->nthread; ++tid) { + e.best.Update(stemp[tid][nid].best); + } + } + } + virtual void SetNonDefaultPosition(const std::vector &qexpand, + IFMatrix *p_fmat, const RegTree &tree) { + // step 1, classify the non-default data into right places std::vector fsplits; for (size_t i = 0; i < qexpand.size(); ++i) { const int nid = qexpand[i]; - if (!tree[nid].is_leaf()) fsplits.push_back(tree[nid].split_index()); + if (!tree[nid].is_leaf()) { + fsplits.push_back(tree[nid].split_index()); + } } std::sort(fsplits.begin(), fsplits.end()); fsplits.resize(std::unique(fsplits.begin(), fsplits.end()) - fsplits.begin()); @@ -364,21 +546,33 @@ class ColMaker: public IUpdater { for (bst_omp_uint j = 0; j < ndata; ++j) { const bst_uint ridx = col[j].index; const float fvalue = col[j].fvalue; - int nid = position[ridx]; - if (nid == -1) continue; + const int nid = this->DecodePosition(ridx); // go back to parent, correct those who are not default - nid = tree[nid].parent(); - if (tree[nid].split_index() == fid) { - if (fvalue < tree[nid].split_cond()) { - position[ridx] = tree[nid].cleft(); + if (!tree[nid].is_leaf() && tree[nid].split_index() == fid) { + if(fvalue < tree[nid].split_cond()) { + this->SetEncodePosition(ridx, tree[nid].cleft()); } else { - position[ridx] = tree[nid].cright(); + this->SetEncodePosition(ridx, tree[nid].cright()); } } } } } } + // utils to get/set position, with encoded format + // return decoded position + inline int DecodePosition(bst_uint ridx) const{ + const int pid = position[ridx]; + return pid < 0 ? ~pid : pid; + } + // encode the encoded position value for ridx + inline void SetEncodePosition(bst_uint ridx, int nid) { + if (position[ridx] < 0) { + position[ridx] = ~nid; + } else { + position[ridx] = nid; + } + } //--data fields-- const TrainParam ¶m; // number of omp thread used during training diff --git a/src/tree/updater_distcol-inl.hpp b/src/tree/updater_distcol-inl.hpp new file mode 100644 index 000000000000..658fbe2b169e --- /dev/null +++ b/src/tree/updater_distcol-inl.hpp @@ -0,0 +1,169 @@ +#ifndef XGBOOST_TREE_UPDATER_DISTCOL_INL_HPP_ +#define XGBOOST_TREE_UPDATER_DISTCOL_INL_HPP_ +/*! + * \file updater_distcol-inl.hpp + * \brief beta distributed version that takes a sub-column + * and construct a tree + * \author Tianqi Chen + */ +#include +#include "../utils/bitmap.h" +#include "../utils/io.h" +#include "./updater_colmaker-inl.hpp" +#include "./updater_prune-inl.hpp" + +namespace xgboost { +namespace tree { +template +class DistColMaker : public ColMaker { + public: + DistColMaker(void) : builder(param) {} + virtual ~DistColMaker(void) {} + // set training parameter + virtual void SetParam(const char *name, const char *val) { + param.SetParam(name, val); + pruner.SetParam(name, val); + } + virtual void Update(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &trees) { + TStats::CheckInfo(info); + utils::Check(trees.size() == 1, "DistColMaker: only support one tree at a time"); + // build the tree + builder.Update(gpair, p_fmat, info, trees[0]); + //// prune the tree, note that pruner will sync the tree + pruner.Update(gpair, p_fmat, info, trees); + // update position after the tree is pruned + builder.UpdatePosition(p_fmat, *trees[0]); + } + virtual const int* GetLeafPosition(void) const { + return builder.GetLeafPosition(); + } + private: + struct Builder : public ColMaker::Builder { + public: + Builder(const TrainParam ¶m) + : ColMaker::Builder(param) { + } + inline void UpdatePosition(IFMatrix *p_fmat, const RegTree &tree) { + const std::vector &rowset = p_fmat->buffered_rowset(); + const bst_omp_uint ndata = static_cast(rowset.size()); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < ndata; ++i) { + const bst_uint ridx = rowset[i]; + int nid = this->DecodePosition(ridx); + while (tree[nid].is_deleted()) { + nid = tree[nid].parent(); + utils::Assert(nid >=0, "distributed learning error"); + } + this->position[ridx] = nid; + } + } + virtual const int* GetLeafPosition(void) const { + return BeginPtr(this->position); + } + protected: + virtual void SetNonDefaultPosition(const std::vector &qexpand, + IFMatrix *p_fmat, const RegTree &tree) { + // step 2, classify the non-default data into right places + std::vector fsplits; + for (size_t i = 0; i < qexpand.size(); ++i) { + const int nid = qexpand[i]; + if (!tree[nid].is_leaf()) { + fsplits.push_back(tree[nid].split_index()); + } + } + // get the candidate split index + std::sort(fsplits.begin(), fsplits.end()); + fsplits.resize(std::unique(fsplits.begin(), fsplits.end()) - fsplits.begin()); + while (fsplits.size() != 0 && fsplits.back() >= p_fmat->NumCol()) { + fsplits.pop_back(); + } + // bitmap is only word concurrent, set to bool first + { + bst_omp_uint ndata = static_cast(this->position.size()); + boolmap.resize(ndata); + #pragma omp parallel for schedule(static) + for (bst_omp_uint j = 0; j < ndata; ++j) { + boolmap[j] = 0; + } + } + utils::IIterator *iter = p_fmat->ColIterator(fsplits); + while (iter->Next()) { + const ColBatch &batch = iter->Value(); + for (size_t i = 0; i < batch.size; ++i) { + ColBatch::Inst col = batch[i]; + const bst_uint fid = batch.col_index[i]; + const bst_omp_uint ndata = static_cast(col.length); + #pragma omp parallel for schedule(static) + for (bst_omp_uint j = 0; j < ndata; ++j) { + const bst_uint ridx = col[j].index; + const float fvalue = col[j].fvalue; + const int nid = this->DecodePosition(ridx); + if (!tree[nid].is_leaf() && tree[nid].split_index() == fid) { + if (fvalue < tree[nid].split_cond()) { + if (!tree[nid].default_left()) boolmap[ridx] = 1; + } else { + if (tree[nid].default_left()) boolmap[ridx] = 1; + } + } + } + } + } + + bitmap.InitFromBool(boolmap); + // communicate bitmap + rabit::Allreduce(BeginPtr(bitmap.data), bitmap.data.size()); + const std::vector &rowset = p_fmat->buffered_rowset(); + // get the new position + const bst_omp_uint ndata = static_cast(rowset.size()); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < ndata; ++i) { + const bst_uint ridx = rowset[i]; + const int nid = this->DecodePosition(ridx); + if (bitmap.Get(ridx)) { + utils::Assert(!tree[nid].is_leaf(), "inconsistent reduce information"); + if (tree[nid].default_left()) { + this->SetEncodePosition(ridx, tree[nid].cright()); + } else { + this->SetEncodePosition(ridx, tree[nid].cleft()); + } + } + } + } + // synchronize the best solution of each node + virtual void SyncBestSolution(const std::vector &qexpand) { + std::vector vec; + for (size_t i = 0; i < qexpand.size(); ++i) { + const int nid = qexpand[i]; + for (int tid = 0; tid < this->nthread; ++tid) { + this->snode[nid].best.Update(this->stemp[tid][nid].best); + } + vec.push_back(this->snode[nid].best); + } + // TODO, lazy version + // communicate best solution + reducer.Allreduce(BeginPtr(vec), vec.size()); + // assign solution back + for (size_t i = 0; i < qexpand.size(); ++i) { + const int nid = qexpand[i]; + this->snode[nid].best = vec[i]; + } + } + + private: + utils::BitMap bitmap; + std::vector boolmap; + rabit::Reducer reducer; + }; + // we directly introduce pruner here + TreePruner pruner; + // training parameter + TrainParam param; + // pointer to the builder + Builder builder; +}; +} // namespace tree +} // namespace xgboost +#endif diff --git a/src/tree/updater_histmaker-inl.hpp b/src/tree/updater_histmaker-inl.hpp new file mode 100644 index 000000000000..9f35d57318c4 --- /dev/null +++ b/src/tree/updater_histmaker-inl.hpp @@ -0,0 +1,701 @@ +#ifndef XGBOOST_TREE_UPDATER_HISTMAKER_INL_HPP_ +#define XGBOOST_TREE_UPDATER_HISTMAKER_INL_HPP_ +/*! + * \file updater_histmaker-inl.hpp + * \brief use histogram counting to construct a tree + * \author Tianqi Chen + */ +#include +#include +#include +#include "../utils/quantile.h" +#include "../utils/group_data.h" +#include "./updater_basemaker-inl.hpp" + +namespace xgboost { +namespace tree { +template +class HistMaker: public BaseMaker { + public: + virtual ~HistMaker(void) {} + virtual void Update(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &trees) { + TStats::CheckInfo(info); + // rescale learning rate according to size of trees + float lr = param.learning_rate; + param.learning_rate = lr / trees.size(); + // build tree + for (size_t i = 0; i < trees.size(); ++i) { + this->Update(gpair, p_fmat, info, trees[i]); + } + param.learning_rate = lr; + } + + protected: + /*! \brief a single histogram */ + struct HistUnit { + /*! \brief cutting point of histogram, contains maximum point */ + const bst_float *cut; + /*! \brief content of statistics data */ + TStats *data; + /*! \brief size of histogram */ + unsigned size; + // default constructor + HistUnit(void) {} + // constructor + HistUnit(const bst_float *cut, TStats *data, unsigned size) + : cut(cut), data(data), size(size) {} + /*! \brief add a histogram to data */ + inline void Add(bst_float fv, + const std::vector &gpair, + const BoosterInfo &info, + const bst_uint ridx) { + unsigned i = std::upper_bound(cut, cut + size, fv) - cut; + utils::Assert(size != 0, "try insert into size=0"); + utils::Assert(i < size, + "maximum value must be in cut, fv = %g, cutmax=%g", fv, cut[size-1]); + data[i].Add(gpair, info, ridx); + } + }; + /*! \brief a set of histograms from different index */ + struct HistSet { + /*! \brief the index pointer of each histunit */ + const unsigned *rptr; + /*! \brief cutting points in each histunit */ + const bst_float *cut; + /*! \brief data in different hist unit */ + std::vector data; + /*! \brief */ + inline HistUnit operator[](size_t fid) { + return HistUnit(cut + rptr[fid], + &data[0] + rptr[fid], + rptr[fid+1] - rptr[fid]); + } + }; + // thread workspace + struct ThreadWSpace { + /*! \brief actual unit pointer */ + std::vector rptr; + /*! \brief cut field */ + std::vector cut; + // per thread histset + std::vector hset; + // initialize the hist set + inline void Init(const TrainParam ¶m, int nthread) { + hset.resize(nthread); + // cleanup statistics + for (int tid = 0; tid < nthread; ++tid) { + for (size_t i = 0; i < hset[tid].data.size(); ++i) { + hset[tid].data[i].Clear(); + } + hset[tid].rptr = BeginPtr(rptr); + hset[tid].cut = BeginPtr(cut); + hset[tid].data.resize(cut.size(), TStats(param)); + } + } + // aggregate all statistics to hset[0] + inline void Aggregate(void) { + bst_omp_uint nsize = static_cast(cut.size()); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < nsize; ++i) { + for (size_t tid = 1; tid < hset.size(); ++tid) { + hset[0].data[i].Add(hset[tid].data[i]); + } + } + } + /*! \brief clear the workspace */ + inline void Clear(void) { + cut.clear(); rptr.resize(1); rptr[0] = 0; + } + /*! \brief total size */ + inline size_t Size(void) const { + return rptr.size() - 1; + } + }; + // workspace of thread + ThreadWSpace wspace; + // reducer for histogram + rabit::Reducer histred; + // set of working features + std::vector fwork_set; + // update function implementation + virtual void Update(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + RegTree *p_tree) { + this->InitData(gpair, *p_fmat, info.root_index, *p_tree); + this->InitWorkSet(p_fmat, *p_tree, &fwork_set); + for (int depth = 0; depth < param.max_depth; ++depth) { + // reset and propose candidate split + this->ResetPosAndPropose(gpair, p_fmat, info, fwork_set, *p_tree); + // create histogram + this->CreateHist(gpair, p_fmat, info, fwork_set, *p_tree); + // find split based on histogram statistics + this->FindSplit(depth, gpair, p_fmat, info, fwork_set, p_tree); + // reset position after split + this->ResetPositionAfterSplit(p_fmat, *p_tree); + this->UpdateQueueExpand(*p_tree); + // if nothing left to be expand, break + if (qexpand.size() == 0) break; + } + for (size_t i = 0; i < qexpand.size(); ++i) { + const int nid = qexpand[i]; + (*p_tree)[nid].set_leaf(p_tree->stat(nid).base_weight * param.learning_rate); + } + } + // this function does two jobs + // (1) reset the position in array position, to be the latest leaf id + // (2) propose a set of candidate cuts and set wspace.rptr wspace.cut correctly + virtual void ResetPosAndPropose(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &fset, + const RegTree &tree) = 0; + // initialize the current working set of features in this round + virtual void InitWorkSet(IFMatrix *p_fmat, + const RegTree &tree, + std::vector *p_fset) { + p_fset->resize(tree.param.num_feature); + for (size_t i = 0; i < p_fset->size(); ++i) { + (*p_fset)[i] = static_cast(i); + } + } + // reset position after split, this is not a must, depending on implementation + virtual void ResetPositionAfterSplit(IFMatrix *p_fmat, + const RegTree &tree) { + } + virtual void CreateHist(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &fset, + const RegTree &tree) = 0; + private: + inline void EnumerateSplit(const HistUnit &hist, + const TStats &node_sum, + bst_uint fid, + SplitEntry *best, + TStats *left_sum) { + if (hist.size == 0) return; + + double root_gain = node_sum.CalcGain(param); + TStats s(param), c(param); + for (bst_uint i = 0; i < hist.size; ++i) { + s.Add(hist.data[i]); + if (s.sum_hess >= param.min_child_weight) { + c.SetSubstract(node_sum, s); + if (c.sum_hess >= param.min_child_weight) { + double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain; + if (best->Update((float)loss_chg, fid, hist.cut[i], false)) { + *left_sum = s; + } + } + } + } + s.Clear(); + for (bst_uint i = hist.size - 1; i != 0; --i) { + s.Add(hist.data[i]); + if (s.sum_hess >= param.min_child_weight) { + c.SetSubstract(node_sum, s); + if (c.sum_hess >= param.min_child_weight) { + double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain; + if (best->Update((float)loss_chg, fid, hist.cut[i-1], true)) { + *left_sum = c; + } + } + } + } + } + inline void FindSplit(int depth, + const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &fset, + RegTree *p_tree) { + const size_t num_feature = fset.size(); + // get the best split condition for each node + std::vector sol(qexpand.size()); + std::vector left_sum(qexpand.size()); + bst_omp_uint nexpand = static_cast(qexpand.size()); + #pragma omp parallel for schedule(dynamic, 1) + for (bst_omp_uint wid = 0; wid < nexpand; ++ wid) { + const int nid = qexpand[wid]; + utils::Assert(node2workindex[nid] == static_cast(wid), + "node2workindex inconsistent"); + SplitEntry &best = sol[wid]; + TStats &node_sum = wspace.hset[0][num_feature + wid * (num_feature + 1)].data[0]; + for (size_t i = 0; i < fset.size(); ++ i) { + EnumerateSplit(this->wspace.hset[0][i + wid * (num_feature+1)], + node_sum, fset[i], &best, &left_sum[wid]); + } + } + // get the best result, we can synchronize the solution + for (bst_omp_uint wid = 0; wid < nexpand; ++ wid) { + const int nid = qexpand[wid]; + const SplitEntry &best = sol[wid]; + const TStats &node_sum = wspace.hset[0][num_feature + wid * (num_feature + 1)].data[0]; + this->SetStats(p_tree, nid, node_sum); + // set up the values + p_tree->stat(nid).loss_chg = best.loss_chg; + // now we know the solution in snode[nid], set split + if (best.loss_chg > rt_eps) { + p_tree->AddChilds(nid); + (*p_tree)[nid].set_split(best.split_index(), + best.split_value, best.default_left()); + // mark right child as 0, to indicate fresh leaf + (*p_tree)[(*p_tree)[nid].cleft()].set_leaf(0.0f, 0); + (*p_tree)[(*p_tree)[nid].cright()].set_leaf(0.0f, 0); + // right side sum + TStats right_sum; + right_sum.SetSubstract(node_sum, left_sum[wid]); + this->SetStats(p_tree, (*p_tree)[nid].cleft(), left_sum[wid]); + this->SetStats(p_tree, (*p_tree)[nid].cright(), right_sum); + } else { + (*p_tree)[nid].set_leaf(p_tree->stat(nid).base_weight * param.learning_rate); + } + } + } + + inline void SetStats(RegTree *p_tree, int nid, const TStats &node_sum) { + p_tree->stat(nid).base_weight = static_cast(node_sum.CalcWeight(param)); + p_tree->stat(nid).sum_hess = static_cast(node_sum.sum_hess); + node_sum.SetLeafVec(param, p_tree->leafvec(nid)); + } +}; + +template +class CQHistMaker: public HistMaker { + protected: + struct HistEntry { + typename HistMaker::HistUnit hist; + unsigned istart; + /*! + * \brief add a histogram to data, + * do linear scan, start from istart + */ + inline void Add(bst_float fv, + const std::vector &gpair, + const BoosterInfo &info, + const bst_uint ridx) { + while (istart < hist.size && !(fv < hist.cut[istart])) ++istart; + utils::Assert(istart != hist.size, "the bound variable must be max"); + hist.data[istart].Add(gpair, info, ridx); + } + }; + // sketch type used for this + typedef utils::WXQuantileSketch WXQSketch; + // initialize the work set of tree + virtual void InitWorkSet(IFMatrix *p_fmat, + const RegTree &tree, + std::vector *p_fset) { + feat_helper.InitByCol(p_fmat, tree); + feat_helper.SampleCol(this->param.colsample_bytree, p_fset); + } + // code to create histogram + virtual void CreateHist(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &fset, + const RegTree &tree) { + // fill in reverse map + feat2workindex.resize(tree.param.num_feature); + std::fill(feat2workindex.begin(), feat2workindex.end(), -1); + for (size_t i = 0; i < fset.size(); ++i) { + feat2workindex[fset[i]] = static_cast(i); + } + // start to work + this->wspace.Init(this->param, 1); + // if it is C++11, use lazy evaluation for Allreduce, + // to gain speedup in recovery +#if __cplusplus >= 201103L + auto lazy_get_hist = [&]() +#endif + { + thread_hist.resize(this->get_nthread()); + // start accumulating statistics + utils::IIterator *iter = p_fmat->ColIterator(fset); + iter->BeforeFirst(); + while (iter->Next()) { + const ColBatch &batch = iter->Value(); + // start enumeration + const bst_omp_uint nsize = static_cast(batch.size); + #pragma omp parallel for schedule(dynamic, 1) + for (bst_omp_uint i = 0; i < nsize; ++i) { + int offset = feat2workindex[batch.col_index[i]]; + if (offset >= 0) { + this->UpdateHistCol(gpair, batch[i], info, tree, + fset, offset, + &thread_hist[omp_get_thread_num()]); + } + } + } + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const int nid = this->qexpand[i]; + const int wid = this->node2workindex[nid]; + this->wspace.hset[0][fset.size() + wid * (fset.size()+1)] + .data[0] = node_stats[nid]; + } + }; + // sync the histogram + // if it is C++11, use lazy evaluation for Allreduce +#if __cplusplus >= 201103L + this->histred.Allreduce(BeginPtr(this->wspace.hset[0].data), + this->wspace.hset[0].data.size(), lazy_get_hist); +#else + this->histred.Allreduce(BeginPtr(this->wspace.hset[0].data), this->wspace.hset[0].data.size()); +#endif + } + virtual void ResetPositionAfterSplit(IFMatrix *p_fmat, + const RegTree &tree) { + this->ResetPositionCol(this->qexpand, p_fmat, tree); + } + virtual void ResetPosAndPropose(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &fset, + const RegTree &tree) { + // fill in reverse map + feat2workindex.resize(tree.param.num_feature); + std::fill(feat2workindex.begin(), feat2workindex.end(), -1); + freal_set.clear(); + for (size_t i = 0; i < fset.size(); ++i) { + if (feat_helper.Type(fset[i]) == 2) { + feat2workindex[fset[i]] = static_cast(freal_set.size()); + freal_set.push_back(fset[i]); + } else { + feat2workindex[fset[i]] = -2; + } + } + this->GetNodeStats(gpair, *p_fmat, tree, info, + &thread_stats, &node_stats); + sketchs.resize(this->qexpand.size() * freal_set.size()); + for (size_t i = 0; i < sketchs.size(); ++i) { + sketchs[i].Init(info.num_row, this->param.sketch_eps); + } + // intitialize the summary array + summary_array.resize(sketchs.size()); + // setup maximum size + unsigned max_size = this->param.max_sketch_size(); + for (size_t i = 0; i < sketchs.size(); ++i) { + summary_array[i].Reserve(max_size); + } + // if it is C++11, use lazy evaluation for Allreduce +#if __cplusplus >= 201103L + auto lazy_get_summary = [&]() +#endif + {// get smmary + thread_sketch.resize(this->get_nthread()); + // number of rows in + const size_t nrows = p_fmat->buffered_rowset().size(); + // start accumulating statistics + utils::IIterator *iter = p_fmat->ColIterator(freal_set); + iter->BeforeFirst(); + while (iter->Next()) { + const ColBatch &batch = iter->Value(); + // start enumeration + const bst_omp_uint nsize = static_cast(batch.size); + #pragma omp parallel for schedule(dynamic, 1) + for (bst_omp_uint i = 0; i < nsize; ++i) { + int offset = feat2workindex[batch.col_index[i]]; + if (offset >= 0) { + this->UpdateSketchCol(gpair, batch[i], tree, + node_stats, + freal_set, offset, + batch[i].length == nrows, + &thread_sketch[omp_get_thread_num()]); + } + } + } + for (size_t i = 0; i < sketchs.size(); ++i) { + utils::WXQuantileSketch::SummaryContainer out; + sketchs[i].GetSummary(&out); + summary_array[i].SetPrune(out, max_size); + } + utils::Assert(summary_array.size() == sketchs.size(), "shape mismatch"); + }; + if (summary_array.size() != 0) { + size_t nbytes = WXQSketch::SummaryContainer::CalcMemCost(max_size); +#if __cplusplus >= 201103L + sreducer.Allreduce(BeginPtr(summary_array), nbytes, summary_array.size(), lazy_get_summary); +#else + sreducer.Allreduce(BeginPtr(summary_array), nbytes, summary_array.size()); +#endif + } + // now we get the final result of sketch, setup the cut + this->wspace.cut.clear(); + this->wspace.rptr.clear(); + this->wspace.rptr.push_back(0); + for (size_t wid = 0; wid < this->qexpand.size(); ++wid) { + for (size_t i = 0; i < fset.size(); ++i) { + int offset = feat2workindex[fset[i]]; + if (offset >= 0) { + const WXQSketch::Summary &a = summary_array[wid * freal_set.size() + offset]; + for (size_t i = 1; i < a.size; ++i) { + bst_float cpt = a.data[i].value - rt_eps; + if (i == 1 || cpt > this->wspace.cut.back()) { + this->wspace.cut.push_back(cpt); + } + } + // push a value that is greater than anything + if (a.size != 0) { + bst_float cpt = a.data[a.size - 1].value; + // this must be bigger than last value in a scale + bst_float last = cpt + fabs(cpt) + rt_eps; + this->wspace.cut.push_back(last); + } + this->wspace.rptr.push_back(static_cast(this->wspace.cut.size())); + } else { + utils::Assert(offset == -2, "BUG in mark"); + bst_float cpt = feat_helper.MaxValue(fset[i]); + this->wspace.cut.push_back(cpt + fabs(cpt) + rt_eps); + this->wspace.rptr.push_back(static_cast(this->wspace.cut.size())); + } + } + // reserve last value for global statistics + this->wspace.cut.push_back(0.0f); + this->wspace.rptr.push_back(static_cast(this->wspace.cut.size())); + } + utils::Assert(this->wspace.rptr.size() == + (fset.size() + 1) * this->qexpand.size() + 1, + "cut space inconsistent"); + } + + private: + inline void UpdateHistCol(const std::vector &gpair, + const ColBatch::Inst &c, + const BoosterInfo &info, + const RegTree &tree, + const std::vector &fset, + bst_uint fid_offset, + std::vector *p_temp) { + if (c.length == 0) return; + // initialize sbuilder for use + std::vector &hbuilder = *p_temp; + hbuilder.resize(tree.param.num_nodes); + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const unsigned nid = this->qexpand[i]; + const unsigned wid = this->node2workindex[nid]; + hbuilder[nid].istart = 0; + hbuilder[nid].hist = this->wspace.hset[0][fid_offset + wid * (fset.size()+1)]; + } + for (bst_uint j = 0; j < c.length; ++j) { + const bst_uint ridx = c[j].index; + const int nid = this->position[ridx]; + if (nid >= 0) { + hbuilder[nid].Add(c[j].fvalue, gpair, info, ridx); + } + } + } + inline void UpdateSketchCol(const std::vector &gpair, + const ColBatch::Inst &c, + const RegTree &tree, + const std::vector &nstats, + const std::vector &frealset, + bst_uint offset, + bool col_full, + std::vector *p_temp) { + if (c.length == 0) return; + // initialize sbuilder for use + std::vector &sbuilder = *p_temp; + sbuilder.resize(tree.param.num_nodes); + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const unsigned nid = this->qexpand[i]; + const unsigned wid = this->node2workindex[nid]; + sbuilder[nid].sum_total = 0.0f; + sbuilder[nid].sketch = &sketchs[wid * frealset.size() + offset]; + } + + if (!col_full) { + // first pass, get sum of weight, TODO, optimization to skip first pass + for (bst_uint j = 0; j < c.length; ++j) { + const bst_uint ridx = c[j].index; + const int nid = this->position[ridx]; + if (nid >= 0) { + sbuilder[nid].sum_total += gpair[ridx].hess; + } + } + } else { + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const unsigned nid = this->qexpand[i]; + sbuilder[nid].sum_total = static_cast(nstats[nid].sum_hess); + } + } + // if only one value, no need to do second pass + if (c[0].fvalue == c[c.length-1].fvalue) { + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const int nid = this->qexpand[i]; + sbuilder[nid].sketch->Push(c[0].fvalue, sbuilder[nid].sum_total); + } + return; + } + // two pass scan + unsigned max_size = this->param.max_sketch_size(); + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const int nid = this->qexpand[i]; + sbuilder[nid].Init(max_size); + } + // second pass, build the sketch + for (bst_uint j = 0; j < c.length; ++j) { + const bst_uint ridx = c[j].index; + const int nid = this->position[ridx]; + if (nid >= 0) { + sbuilder[nid].Push(c[j].fvalue, gpair[ridx].hess, max_size); + } + } + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const int nid = this->qexpand[i]; + sbuilder[nid].Finalize(max_size); + } + } + // feature helper + BaseMaker::FMetaHelper feat_helper; + // temp space to map feature id to working index + std::vector feat2workindex; + // set of index from fset that are real + std::vector freal_set; + // thread temp data + std::vector< std::vector > thread_sketch; + // used to hold statistics + std::vector< std::vector > thread_stats; + // used to hold start pointer + std::vector< std::vector > thread_hist; + // node statistics + std::vector node_stats; + // summary array + std::vector summary_array; + // reducer for summary + rabit::SerializeReducer sreducer; + // per node, per feature sketch + std::vector< utils::WXQuantileSketch > sketchs; +}; + +template +class QuantileHistMaker: public HistMaker { + protected: + typedef utils::WXQuantileSketch WXQSketch; + virtual void ResetPosAndPropose(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &fset, + const RegTree &tree) { + // initialize the data structure + int nthread = BaseMaker::get_nthread(); + sketchs.resize(this->qexpand.size() * tree.param.num_feature); + for (size_t i = 0; i < sketchs.size(); ++i) { + sketchs[i].Init(info.num_row, this->param.sketch_eps); + } + // start accumulating statistics + utils::IIterator *iter = p_fmat->RowIterator(); + iter->BeforeFirst(); + while (iter->Next()) { + const RowBatch &batch = iter->Value(); + // parallel convert to column major format + utils::ParallelGroupBuilder builder(&col_ptr, &col_data, &thread_col_ptr); + builder.InitBudget(tree.param.num_feature, nthread); + + const bst_omp_uint nbatch = static_cast(batch.size); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < nbatch; ++i) { + RowBatch::Inst inst = batch[i]; + const bst_uint ridx = static_cast(batch.base_rowid + i); + int nid = this->position[ridx]; + if (nid >= 0) { + if (!tree[nid].is_leaf()) { + this->position[ridx] = nid = HistMaker::NextLevel(inst, tree, nid); + } + if (this->node2workindex[nid] < 0) { + this->position[ridx] = ~nid; + } else{ + for (bst_uint j = 0; j < inst.length; ++j) { + builder.AddBudget(inst[j].index, omp_get_thread_num()); + } + } + } + } + builder.InitStorage(); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < nbatch; ++i) { + RowBatch::Inst inst = batch[i]; + const bst_uint ridx = static_cast(batch.base_rowid + i); + const int nid = this->position[ridx]; + if (nid >= 0) { + for (bst_uint j = 0; j < inst.length; ++j) { + builder.Push(inst[j].index, + SparseBatch::Entry(nid, inst[j].fvalue), + omp_get_thread_num()); + } + } + } + // start putting things into sketch + const bst_omp_uint nfeat = col_ptr.size() - 1; + #pragma omp parallel for schedule(dynamic, 1) + for (bst_omp_uint k = 0; k < nfeat; ++k) { + for (size_t i = col_ptr[k]; i < col_ptr[k+1]; ++i) { + const SparseBatch::Entry &e = col_data[i]; + const int wid = this->node2workindex[e.index]; + sketchs[wid * tree.param.num_feature + k].Push(e.fvalue, gpair[e.index].hess); + } + } + } + // setup maximum size + unsigned max_size = this->param.max_sketch_size(); + // synchronize sketch + summary_array.resize(sketchs.size()); + for (size_t i = 0; i < sketchs.size(); ++i) { + utils::WQuantileSketch::SummaryContainer out; + sketchs[i].GetSummary(&out); + summary_array[i].Reserve(max_size); + summary_array[i].SetPrune(out, max_size); + } + + size_t nbytes = WXQSketch::SummaryContainer::CalcMemCost(max_size); + sreducer.Allreduce(BeginPtr(summary_array), nbytes, summary_array.size()); + // now we get the final result of sketch, setup the cut + this->wspace.cut.clear(); + this->wspace.rptr.clear(); + this->wspace.rptr.push_back(0); + for (size_t wid = 0; wid < this->qexpand.size(); ++wid) { + for (int fid = 0; fid < tree.param.num_feature; ++fid) { + const WXQSketch::Summary &a = summary_array[wid * tree.param.num_feature + fid]; + for (size_t i = 1; i < a.size; ++i) { + bst_float cpt = a.data[i].value - rt_eps; + if (i == 1 || cpt > this->wspace.cut.back()) { + this->wspace.cut.push_back(cpt); + } + } + // push a value that is greater than anything + if (a.size != 0) { + bst_float cpt = a.data[a.size - 1].value; + // this must be bigger than last value in a scale + bst_float last = cpt + fabs(cpt) + rt_eps; + this->wspace.cut.push_back(last); + } + this->wspace.rptr.push_back(this->wspace.cut.size()); + } + // reserve last value for global statistics + this->wspace.cut.push_back(0.0f); + this->wspace.rptr.push_back(this->wspace.cut.size()); + } + utils::Assert(this->wspace.rptr.size() == + (tree.param.num_feature + 1) * this->qexpand.size() + 1, + "cut space inconsistent"); + } + + private: + // summary array + std::vector summary_array; + // reducer for summary + rabit::SerializeReducer sreducer; + // local temp column data structure + std::vector col_ptr; + // local storage of column data + std::vector col_data; + std::vector< std::vector > thread_col_ptr; + // per node, per feature sketch + std::vector< utils::WQuantileSketch > sketchs; +}; + +} // namespace tree +} // namespace xgboost +#endif // XGBOOST_TREE_UPDATER_HISTMAKER_INL_HPP_ diff --git a/src/tree/updater_prune-inl.hpp b/src/tree/updater_prune-inl.hpp index 726999f55059..e7e5f9f0b752 100644 --- a/src/tree/updater_prune-inl.hpp +++ b/src/tree/updater_prune-inl.hpp @@ -8,6 +8,7 @@ #include #include "./param.h" #include "./updater.h" +#include "./updater_sync-inl.hpp" namespace xgboost { namespace tree { @@ -19,6 +20,7 @@ class TreePruner: public IUpdater { virtual void SetParam(const char *name, const char *val) { using namespace std; param.SetParam(name, val); + syncher.SetParam(name, val); if (!strcmp(name, "silent")) silent = atoi(val); } // update the tree, do pruning @@ -33,8 +35,8 @@ class TreePruner: public IUpdater { this->DoPrune(*trees[i]); } param.learning_rate = lr; + syncher.Update(gpair, p_fmat, info, trees); } - private: // try to prune off current leaf inline int TryPruneLeaf(RegTree &tree, int nid, int depth, int npruned) { @@ -70,6 +72,8 @@ class TreePruner: public IUpdater { } private: + // synchronizer + TreeSyncher syncher; // shutup int silent; // training parameter diff --git a/src/tree/updater_refresh-inl.hpp b/src/tree/updater_refresh-inl.hpp index a37630333611..0285f07713c0 100644 --- a/src/tree/updater_refresh-inl.hpp +++ b/src/tree/updater_refresh-inl.hpp @@ -7,6 +7,7 @@ */ #include #include +#include #include "./param.h" #include "./updater.h" #include "../utils/omp.h" @@ -26,7 +27,7 @@ class TreeRefresher: public IUpdater { virtual void Update(const std::vector &gpair, IFMatrix *p_fmat, const BoosterInfo &info, - const std::vector &trees) { + const std::vector &trees) { if (trees.size() == 0) return; // number of threads // thread temporal space @@ -39,54 +40,71 @@ class TreeRefresher: public IUpdater { nthread = omp_get_num_threads(); } fvec_temp.resize(nthread, RegTree::FVec()); - stemp.resize(trees.size() * nthread, std::vector()); + stemp.resize(nthread, std::vector()); #pragma omp parallel { int tid = omp_get_thread_num(); + int num_nodes = 0; for (size_t i = 0; i < trees.size(); ++i) { - std::vector &vec = stemp[tid * trees.size() + i]; - vec.resize(trees[i]->param.num_nodes, TStats(param)); - std::fill(vec.begin(), vec.end(), TStats(param)); + num_nodes += trees[i]->param.num_nodes; } + stemp[tid].resize(num_nodes, TStats(param)); + std::fill(stemp[tid].begin(), stemp[tid].end(), TStats(param)); fvec_temp[tid].Init(trees[0]->param.num_feature); } - // start accumulating statistics - utils::IIterator *iter = p_fmat->RowIterator(); - iter->BeforeFirst(); - while (iter->Next()) { - const RowBatch &batch = iter->Value(); - utils::Check(batch.size < std::numeric_limits::max(), - "too large batch size "); - const bst_omp_uint nbatch = static_cast(batch.size); - #pragma omp parallel for schedule(static) - for (bst_omp_uint i = 0; i < nbatch; ++i) { - RowBatch::Inst inst = batch[i]; - const int tid = omp_get_thread_num(); - const bst_uint ridx = static_cast(batch.base_rowid + i); - RegTree::FVec &feats = fvec_temp[tid]; - feats.Fill(inst); - for (size_t j = 0; j < trees.size(); ++j) { - AddStats(*trees[j], feats, gpair, info, ridx, - &stemp[tid * trees.size() + j]); + // if it is C++11, use lazy evaluation for Allreduce, + // to gain speedup in recovery +#if __cplusplus >= 201103L + auto lazy_get_stats = [&]() +#endif + { + // start accumulating statistics + utils::IIterator *iter = p_fmat->RowIterator(); + iter->BeforeFirst(); + while (iter->Next()) { + const RowBatch &batch = iter->Value(); + utils::Check(batch.size < std::numeric_limits::max(), + "too large batch size "); + const bst_omp_uint nbatch = static_cast(batch.size); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < nbatch; ++i) { + RowBatch::Inst inst = batch[i]; + const int tid = omp_get_thread_num(); + const bst_uint ridx = static_cast(batch.base_rowid + i); + RegTree::FVec &feats = fvec_temp[tid]; + feats.Fill(inst); + int offset = 0; + for (size_t j = 0; j < trees.size(); ++j) { + AddStats(*trees[j], feats, gpair, info, ridx, + BeginPtr(stemp[tid]) + offset); + offset += trees[j]->param.num_nodes; + } + feats.Drop(inst); } - feats.Drop(inst); } - } - // start update the trees using the statistics - // rescale learning rate according to size of trees - float lr = param.learning_rate; - param.learning_rate = lr / trees.size(); - for (size_t i = 0; i < trees.size(); ++i) { - // aggregate + // aggregate the statistics + int num_nodes = static_cast(stemp[0].size()); #pragma omp parallel for schedule(static) - for (int nid = 0; nid < trees[i]->param.num_nodes; ++nid) { + for (int nid = 0; nid < num_nodes; ++nid) { for (int tid = 1; tid < nthread; ++tid) { - stemp[i][nid].Add(stemp[tid * trees.size() + i][nid]); + stemp[0][nid].Add(stemp[tid][nid]); } } + }; +#if __cplusplus >= 201103L + reducer.Allreduce(BeginPtr(stemp[0]), stemp[0].size(), lazy_get_stats); +#else + reducer.Allreduce(BeginPtr(stemp[0]), stemp[0].size()); +#endif + // rescale learning rate according to size of trees + float lr = param.learning_rate; + param.learning_rate = lr / trees.size(); + int offset = 0; + for (size_t i = 0; i < trees.size(); ++i) { for (int rid = 0; rid < trees[i]->param.num_roots; ++rid) { - this->Refresh(stemp[i], rid, trees[i]); + this->Refresh(BeginPtr(stemp[0]) + offset, rid, trees[i]); } + offset += trees[i]->param.num_nodes; } // set learning rate back param.learning_rate = lr; @@ -98,8 +116,7 @@ class TreeRefresher: public IUpdater { const std::vector &gpair, const BoosterInfo &info, const bst_uint ridx, - std::vector *p_gstats) { - std::vector &gstats = *p_gstats; + TStats *gstats) { // start from groups that belongs to current data int pid = static_cast(info.GetRoot(ridx)); gstats[pid].Add(gpair, info, ridx); @@ -110,7 +127,7 @@ class TreeRefresher: public IUpdater { gstats[pid].Add(gpair, info, ridx); } } - inline void Refresh(const std::vector &gstats, + inline void Refresh(const TStats *gstats, int nid, RegTree *p_tree) { RegTree &tree = *p_tree; tree.stat(nid).base_weight = static_cast(gstats[nid].CalcWeight(param)); @@ -129,6 +146,8 @@ class TreeRefresher: public IUpdater { } // training parameter TrainParam param; + // reducer + rabit::Reducer reducer; }; } // namespace tree diff --git a/src/tree/updater_skmaker-inl.hpp b/src/tree/updater_skmaker-inl.hpp new file mode 100644 index 000000000000..45202273a796 --- /dev/null +++ b/src/tree/updater_skmaker-inl.hpp @@ -0,0 +1,393 @@ +#ifndef XGBOOST_TREE_UPDATER_SKMAKER_INL_HPP_ +#define XGBOOST_TREE_UPDATER_SKMAKER_INL_HPP_ +/*! + * \file updater_skmaker-inl.hpp + * \brief use approximation sketch to construct a tree, + a refresh is needed to make the statistics exactly correct + * \author Tianqi Chen + */ +#include +#include +#include +#include "../utils/quantile.h" +#include "./updater_basemaker-inl.hpp" + +namespace xgboost { +namespace tree { +class SketchMaker: public BaseMaker { + public: + virtual ~SketchMaker(void) {} + virtual void Update(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &trees) { + // rescale learning rate according to size of trees + float lr = param.learning_rate; + param.learning_rate = lr / trees.size(); + // build tree + for (size_t i = 0; i < trees.size(); ++i) { + this->Update(gpair, p_fmat, info, trees[i]); + } + param.learning_rate = lr; + } + + protected: + inline void Update(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + RegTree *p_tree) { + this->InitData(gpair, *p_fmat, info.root_index, *p_tree); + for (int depth = 0; depth < param.max_depth; ++depth) { + this->GetNodeStats(gpair, *p_fmat, *p_tree, info, + &thread_stats, &node_stats); + this->BuildSketch(gpair, p_fmat, info, *p_tree); + this->SyncNodeStats(); + this->FindSplit(depth, gpair, p_fmat, info, p_tree); + this->ResetPositionCol(qexpand, p_fmat, *p_tree); + this->UpdateQueueExpand(*p_tree); + // if nothing left to be expand, break + if (qexpand.size() == 0) break; + } + if (qexpand.size() != 0) { + this->GetNodeStats(gpair, *p_fmat, *p_tree, info, + &thread_stats, &node_stats); + this->SyncNodeStats(); + } + // set all statistics correctly + for (int nid = 0; nid < p_tree->param.num_nodes; ++nid) { + this->SetStats(nid, node_stats[nid], p_tree); + if (!(*p_tree)[nid].is_leaf()) { + p_tree->stat(nid).loss_chg = + node_stats[(*p_tree)[nid].cleft()].CalcGain(param) + + node_stats[(*p_tree)[nid].cright()].CalcGain(param) - + node_stats[nid].CalcGain(param); + } + } + // set left leaves + for (size_t i = 0; i < qexpand.size(); ++i) { + const int nid = qexpand[i]; + (*p_tree)[nid].set_leaf(p_tree->stat(nid).base_weight * param.learning_rate); + } + } + // define the sketch we want to use + typedef utils::WXQuantileSketch WXQSketch; + + private: + // statistics needed in the gradient calculation + struct SKStats { + /*! \brief sum of all positive gradient */ + double pos_grad; + /*! \brief sum of all negative gradient */ + double neg_grad; + /*! \brief sum of hessian statistics */ + double sum_hess; + explicit SKStats(void) {} + // constructor + explicit SKStats(const TrainParam ¶m) { + this->Clear(); + } + /*! \brief clear the statistics */ + inline void Clear(void) { + neg_grad = pos_grad = sum_hess = 0.0f; + } + // accumulate statistics + inline void Add(const std::vector &gpair, + const BoosterInfo &info, + bst_uint ridx) { + const bst_gpair &b = gpair[ridx]; + if (b.grad >= 0.0f) { + pos_grad += b.grad; + } else { + neg_grad -= b.grad; + } + sum_hess += b.hess; + } + /*! \brief calculate gain of the solution */ + inline double CalcGain(const TrainParam ¶m) const { + return param.CalcGain(pos_grad - neg_grad, sum_hess); + } + /*! \brief set current value to a - b */ + inline void SetSubstract(const SKStats &a, const SKStats &b) { + pos_grad = a.pos_grad - b.pos_grad; + neg_grad = a.neg_grad - b.neg_grad; + sum_hess = a.sum_hess - b.sum_hess; + } + // calculate leaf weight + inline double CalcWeight(const TrainParam ¶m) const { + return param.CalcWeight(pos_grad - neg_grad, sum_hess); + } + /*! \brief add statistics to the data */ + inline void Add(const SKStats &b) { + pos_grad += b.pos_grad; + neg_grad += b.neg_grad; + sum_hess += b.sum_hess; + } + /*! \brief same as add, reduce is used in All Reduce */ + inline void Reduce(const SKStats &b) { + this->Add(b); + } + /*! \brief set leaf vector value based on statistics */ + inline void SetLeafVec(const TrainParam ¶m, bst_float *vec) const { + } + }; + inline void BuildSketch(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const RegTree &tree) { + sketchs.resize(this->qexpand.size() * tree.param.num_feature * 3); + for (size_t i = 0; i < sketchs.size(); ++i) { + sketchs[i].Init(info.num_row, this->param.sketch_eps); + } + thread_sketch.resize(this->get_nthread()); + // number of rows in + const size_t nrows = p_fmat->buffered_rowset().size(); + // start accumulating statistics + utils::IIterator *iter = p_fmat->ColIterator(); + iter->BeforeFirst(); + while (iter->Next()) { + const ColBatch &batch = iter->Value(); + // start enumeration + const bst_omp_uint nsize = static_cast(batch.size); + #pragma omp parallel for schedule(dynamic, 1) + for (bst_omp_uint i = 0; i < nsize; ++i) { + this->UpdateSketchCol(gpair, batch[i], tree, + node_stats, + batch.col_index[i], + batch[i].length == nrows, + &thread_sketch[omp_get_thread_num()]); + } + } + // setup maximum size + unsigned max_size = param.max_sketch_size(); + // synchronize sketch + summary_array.Init(sketchs.size(), max_size); + for (size_t i = 0; i < sketchs.size(); ++i) { + utils::WXQuantileSketch::SummaryContainer out; + sketchs[i].GetSummary(&out); + summary_array.Set(i, out); + } + size_t nbytes = summary_array.MemSize();; + sketch_reducer.Allreduce(&summary_array, nbytes); + } + // update sketch information in column fid + inline void UpdateSketchCol(const std::vector &gpair, + const ColBatch::Inst &c, + const RegTree &tree, + const std::vector &nstats, + bst_uint fid, + bool col_full, + std::vector *p_temp) { + if (c.length == 0) return; + // initialize sbuilder for use + std::vector &sbuilder = *p_temp; + sbuilder.resize(tree.param.num_nodes * 3); + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const unsigned nid = this->qexpand[i]; + const unsigned wid = this->node2workindex[nid]; + for (int k = 0; k < 3; ++k) { + sbuilder[3 * nid + k].sum_total = 0.0f; + sbuilder[3 * nid + k].sketch = &sketchs[(wid * tree.param.num_feature + fid) * 3 + k]; + } + } + if (!col_full) { + for (bst_uint j = 0; j < c.length; ++j) { + const bst_uint ridx = c[j].index; + const int nid = this->position[ridx]; + if (nid >= 0) { + const bst_gpair &e = gpair[ridx]; + if (e.grad >= 0.0f) { + sbuilder[3 * nid + 0].sum_total += e.grad; + } else { + sbuilder[3 * nid + 1].sum_total -= e.grad; + } + sbuilder[3 * nid + 2].sum_total += e.hess; + } + } + } else { + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const unsigned nid = this->qexpand[i]; + sbuilder[3 * nid + 0].sum_total = nstats[nid].pos_grad; + sbuilder[3 * nid + 1].sum_total = nstats[nid].neg_grad; + sbuilder[3 * nid + 2].sum_total = nstats[nid].sum_hess; + } + } + // if only one value, no need to do second pass + if (c[0].fvalue == c[c.length-1].fvalue) { + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const int nid = this->qexpand[i]; + for (int k = 0; k < 3; ++k) { + sbuilder[3 * nid + k].sketch->Push(c[0].fvalue, sbuilder[3 * nid + k].sum_total); + } + } + return; + } + // two pass scan + unsigned max_size = param.max_sketch_size(); + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const int nid = this->qexpand[i]; + for (int k = 0; k < 3; ++k) { + sbuilder[3 * nid + k].Init(max_size); + } + } + // second pass, build the sketch + for (bst_uint j = 0; j < c.length; ++j) { + const bst_uint ridx = c[j].index; + const int nid = this->position[ridx]; + if (nid >= 0) { + const bst_gpair &e = gpair[ridx]; + if (e.grad >= 0.0f) { + sbuilder[3 * nid + 0].Push(c[j].fvalue, e.grad, max_size); + } else { + sbuilder[3 * nid + 1].Push(c[j].fvalue, -e.grad, max_size); + } + sbuilder[3 * nid + 2].Push(c[j].fvalue, e.hess, max_size); + } + } + for (size_t i = 0; i < this->qexpand.size(); ++i) { + const int nid = this->qexpand[i]; + for (int k = 0; k < 3; ++k) { + sbuilder[3 * nid + k].Finalize(max_size); + } + } + } + inline void SyncNodeStats(void) { + utils::Assert(qexpand.size() != 0, "qexpand must not be empty"); + std::vector tmp(qexpand.size()); + for (size_t i = 0; i < qexpand.size(); ++i) { + tmp[i] = node_stats[qexpand[i]]; + } + stats_reducer.Allreduce(BeginPtr(tmp), tmp.size()); + for (size_t i = 0; i < qexpand.size(); ++i) { + node_stats[qexpand[i]] = tmp[i]; + } + } + inline void FindSplit(int depth, + const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + RegTree *p_tree) { + const bst_uint num_feature = p_tree->param.num_feature; + // get the best split condition for each node + std::vector sol(qexpand.size()); + bst_omp_uint nexpand = static_cast(qexpand.size()); + #pragma omp parallel for schedule(dynamic, 1) + for (bst_omp_uint wid = 0; wid < nexpand; ++ wid) { + const int nid = qexpand[wid]; + utils::Assert(node2workindex[nid] == static_cast(wid), + "node2workindex inconsistent"); + SplitEntry &best = sol[wid]; + for (bst_uint fid = 0; fid < num_feature; ++ fid) { + unsigned base = (wid * p_tree->param.num_feature + fid) * 3; + EnumerateSplit(summary_array[base + 0], + summary_array[base + 1], + summary_array[base + 2], + node_stats[nid], fid, &best); + } + } + // get the best result, we can synchronize the solution + for (bst_omp_uint wid = 0; wid < nexpand; ++ wid) { + const int nid = qexpand[wid]; + const SplitEntry &best = sol[wid]; + // set up the values + p_tree->stat(nid).loss_chg = best.loss_chg; + this->SetStats(nid, node_stats[nid], p_tree); + // now we know the solution in snode[nid], set split + if (best.loss_chg > rt_eps) { + p_tree->AddChilds(nid); + (*p_tree)[nid].set_split(best.split_index(), + best.split_value, best.default_left()); + // mark right child as 0, to indicate fresh leaf + (*p_tree)[(*p_tree)[nid].cleft()].set_leaf(0.0f, 0); + (*p_tree)[(*p_tree)[nid].cright()].set_leaf(0.0f, 0); + } else { + (*p_tree)[nid].set_leaf(p_tree->stat(nid).base_weight * param.learning_rate); + } + } + } + // set statistics on ptree + inline void SetStats(int nid, const SKStats &node_sum, RegTree *p_tree) { + p_tree->stat(nid).base_weight = node_sum.CalcWeight(param); + p_tree->stat(nid).sum_hess = static_cast(node_sum.sum_hess); + node_sum.SetLeafVec(param, p_tree->leafvec(nid)); + } + inline void EnumerateSplit(const WXQSketch::Summary &pos_grad, + const WXQSketch::Summary &neg_grad, + const WXQSketch::Summary &sum_hess, + const SKStats &node_sum, + bst_uint fid, + SplitEntry *best) { + if (sum_hess.size == 0) return; + double root_gain = node_sum.CalcGain(param); + std::vector fsplits; + for (size_t i = 0; i < pos_grad.size; ++i) { + fsplits.push_back(pos_grad.data[i].value); + } + for (size_t i = 0; i < neg_grad.size; ++i) { + fsplits.push_back(neg_grad.data[i].value); + } + for (size_t i = 0; i < sum_hess.size; ++i) { + fsplits.push_back(sum_hess.data[i].value); + } + std::sort(fsplits.begin(), fsplits.end()); + fsplits.resize(std::unique(fsplits.begin(), fsplits.end()) - fsplits.begin()); + // sum feature + SKStats feat_sum; + feat_sum.pos_grad = pos_grad.data[pos_grad.size - 1].rmax; + feat_sum.neg_grad = neg_grad.data[neg_grad.size - 1].rmax; + feat_sum.sum_hess = sum_hess.data[sum_hess.size - 1].rmax; + size_t ipos = 0, ineg = 0, ihess = 0; + for (size_t i = 1; i < fsplits.size(); ++i) { + WXQSketch::Entry pos = pos_grad.Query(fsplits[i], ipos); + WXQSketch::Entry neg = neg_grad.Query(fsplits[i], ineg); + WXQSketch::Entry hess = sum_hess.Query(fsplits[i], ihess); + SKStats s, c; + s.pos_grad = 0.5f * (pos.rmin + pos.rmax - pos.wmin); + s.neg_grad = 0.5f * (neg.rmin + neg.rmax - neg.wmin); + s.sum_hess = 0.5f * (hess.rmin + hess.rmax - hess.wmin); + c.SetSubstract(node_sum, s); + // forward + if (s.sum_hess >= param.min_child_weight && + c.sum_hess >= param.min_child_weight) { + double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain; + best->Update(loss_chg, fid, fsplits[i], false); + } + // backward + c.SetSubstract(feat_sum, s); + s.SetSubstract(node_sum, c); + if (s.sum_hess >= param.min_child_weight && + c.sum_hess >= param.min_child_weight) { + double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain; + best->Update(loss_chg, fid, fsplits[i], true); + } + } + {// all including + SKStats s = feat_sum, c; + c.SetSubstract(node_sum, s); + if (s.sum_hess >= param.min_child_weight && + c.sum_hess >= param.min_child_weight) { + bst_float cpt = fsplits.back(); + double loss_chg = s.CalcGain(param) + c.CalcGain(param) - root_gain; + best->Update(loss_chg, fid, cpt + fabsf(cpt) + 1.0f, true); + } + } + } + + // thread temp data + // used to hold temporal sketch + std::vector< std::vector > thread_sketch; + // used to hold statistics + std::vector< std::vector > thread_stats; + // node statistics + std::vector node_stats; + // summary array + WXQSketch::SummaryArray summary_array; + // reducer for summary + rabit::Reducer stats_reducer; + // reducer for summary + rabit::SerializeReducer sketch_reducer; + // per node, per feature sketch + std::vector< utils::WXQuantileSketch > sketchs; +}; +} // tree +} // xgboost +#endif diff --git a/src/tree/updater_sync-inl.hpp b/src/tree/updater_sync-inl.hpp new file mode 100644 index 000000000000..0cbbb4eede68 --- /dev/null +++ b/src/tree/updater_sync-inl.hpp @@ -0,0 +1,53 @@ +#ifndef XGBOOST_TREE_UPDATER_SYNC_INL_HPP_ +#define XGBOOST_TREE_UPDATER_SYNC_INL_HPP_ +/*! + * \file updater_sync-inl.hpp + * \brief synchronize the tree in all distributed nodes + * \author Tianqi Chen + */ +#include +#include +#include +#include "./updater.h" + +namespace xgboost { +namespace tree { +/*! + * \brief syncher that synchronize the tree in all distributed nodes + * can implement various strategies, so far it is always set to node 0's tree + */ +class TreeSyncher: public IUpdater { + public: + virtual ~TreeSyncher(void) {} + virtual void SetParam(const char *name, const char *val) { + } + // update the tree, do pruning + virtual void Update(const std::vector &gpair, + IFMatrix *p_fmat, + const BoosterInfo &info, + const std::vector &trees) { + this->SyncTrees(trees); + } + + private: + // synchronize the trees in different nodes, take tree from rank 0 + inline void SyncTrees(const std::vector &trees) { + if (rabit::GetWorldSize() == 1) return; + std::string s_model; + utils::MemoryBufferStream fs(&s_model); + int rank = rabit::GetRank(); + if (rank == 0) { + for (size_t i = 0; i < trees.size(); ++i) { + trees[i]->SaveModel(fs); + } + } + fs.Seek(0); + rabit::Broadcast(&s_model, 0); + for (size_t i = 0; i < trees.size(); ++i) { + trees[i]->LoadModel(fs); + } + } +}; +} // namespace tree +} // namespace xgboost +#endif // XGBOOST_TREE_UPDATER_SYNC_INL_HPP_ diff --git a/src/utils/base64.h b/src/utils/base64.h new file mode 100644 index 000000000000..36699199f7d6 --- /dev/null +++ b/src/utils/base64.h @@ -0,0 +1,205 @@ +#ifndef XGBOOST_UTILS_BASE64_H_ +#define XGBOOST_UTILS_BASE64_H_ +/*! + * \file base64.h + * \brief data stream support to input and output from/to base64 stream + * base64 is easier to store and pass as text format in mapreduce + * \author Tianqi Chen + */ +#include +#include +#include "./utils.h" +#include "./io.h" + +namespace xgboost { +namespace utils { +/*! \brief namespace of base64 decoding and encoding table */ +namespace base64 { +const char DecodeTable[] = { + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 62, // '+' + 0, 0, 0, + 63, // '/' + 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, // '0'-'9' + 0, 0, 0, 0, 0, 0, 0, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, + 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, // 'A'-'Z' + 0, 0, 0, 0, 0, 0, + 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, + 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, // 'a'-'z' +}; +static const char EncodeTable[] = + "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/"; +} // namespace base64 +/*! \brief the stream that reads from base64, note we take from file pointers */ +class Base64InStream: public IStream { + public: + explicit Base64InStream(FILE *fp) : fp(fp) { + num_prev = 0; tmp_ch = 0; + } + /*! + * \brief initialize the stream position to beginning of next base64 stream + * call this function before actually start read + */ + inline void InitPosition(void) { + // get a charater + do { + tmp_ch = fgetc(fp); + } while (isspace(tmp_ch)); + } + /*! \brief whether current position is end of a base64 stream */ + inline bool IsEOF(void) const { + return num_prev == 0 && (tmp_ch == EOF || isspace(tmp_ch)); + } + virtual size_t Read(void *ptr, size_t size) { + using base64::DecodeTable; + if (size == 0) return 0; + // use tlen to record left size + size_t tlen = size; + unsigned char *cptr = static_cast(ptr); + // if anything left, load from previous buffered result + if (num_prev != 0) { + if (num_prev == 2) { + if (tlen >= 2) { + *cptr++ = buf_prev[0]; + *cptr++ = buf_prev[1]; + tlen -= 2; + num_prev = 0; + } else { + // assert tlen == 1 + *cptr++ = buf_prev[0]; --tlen; + buf_prev[0] = buf_prev[1]; + num_prev = 1; + } + } else { + // assert num_prev == 1 + *cptr++ = buf_prev[0]; --tlen; num_prev = 0; + } + } + if (tlen == 0) return size; + int nvalue; + // note: everything goes with 4 bytes in Base64 + // so we process 4 bytes a unit + while (tlen && tmp_ch != EOF && !isspace(tmp_ch)) { + // first byte + nvalue = DecodeTable[tmp_ch] << 18; + { + // second byte + Check((tmp_ch = fgetc(fp), tmp_ch != EOF && !isspace(tmp_ch)), + "invalid base64 format"); + nvalue |= DecodeTable[tmp_ch] << 12; + *cptr++ = (nvalue >> 16) & 0xFF; --tlen; + } + { + // third byte + Check((tmp_ch = fgetc(fp), tmp_ch != EOF && !isspace(tmp_ch)), + "invalid base64 format"); + // handle termination + if (tmp_ch == '=') { + Check((tmp_ch = fgetc(fp), tmp_ch == '='), "invalid base64 format"); + Check((tmp_ch = fgetc(fp), tmp_ch == EOF || isspace(tmp_ch)), + "invalid base64 format"); + break; + } + nvalue |= DecodeTable[tmp_ch] << 6; + if (tlen) { + *cptr++ = (nvalue >> 8) & 0xFF; --tlen; + } else { + buf_prev[num_prev++] = (nvalue >> 8) & 0xFF; + } + } + { + // fourth byte + Check((tmp_ch = fgetc(fp), tmp_ch != EOF && !isspace(tmp_ch)), + "invalid base64 format"); + if (tmp_ch == '=') { + Check((tmp_ch = fgetc(fp), tmp_ch == EOF || isspace(tmp_ch)), + "invalid base64 format"); + break; + } + nvalue |= DecodeTable[tmp_ch]; + if (tlen) { + *cptr++ = nvalue & 0xFF; --tlen; + } else { + buf_prev[num_prev ++] = nvalue & 0xFF; + } + } + // get next char + tmp_ch = fgetc(fp); + } + if (kStrictCheck) { + Check(tlen == 0, "Base64InStream: read incomplete"); + } + return size - tlen; + } + virtual void Write(const void *ptr, size_t size) { + utils::Error("Base64InStream do not support write"); + } + + private: + FILE *fp; + unsigned char tmp_ch; + int num_prev; + unsigned char buf_prev[2]; + // whether we need to do strict check + static const bool kStrictCheck = false; +}; +/*! \brief the stream that write to base64, note we take from file pointers */ +class Base64OutStream: public IStream { + public: + explicit Base64OutStream(FILE *fp) : fp(fp) { + buf_top = 0; + } + virtual void Write(const void *ptr, size_t size) { + using base64::EncodeTable; + size_t tlen = size; + const unsigned char *cptr = static_cast(ptr); + while (tlen) { + while (buf_top < 3 && tlen != 0) { + buf[++buf_top] = *cptr++; --tlen; + } + if (buf_top == 3) { + // flush 4 bytes out + fputc(EncodeTable[buf[1] >> 2], fp); + fputc(EncodeTable[((buf[1] << 4) | (buf[2] >> 4)) & 0x3F], fp); + fputc(EncodeTable[((buf[2] << 2) | (buf[3] >> 6)) & 0x3F], fp); + fputc(EncodeTable[buf[3] & 0x3F], fp); + buf_top = 0; + } + } + } + virtual size_t Read(void *ptr, size_t size) { + Error("Base64OutStream do not support read"); + return 0; + } + /*! + * \brief finish writing of all current base64 stream, do some post processing + * \param endch charater to put to end of stream, if it is EOF, then nothing will be done + */ + inline void Finish(char endch = EOF) { + using base64::EncodeTable; + if (buf_top == 1) { + fputc(EncodeTable[buf[1] >> 2], fp); + fputc(EncodeTable[(buf[1] << 4) & 0x3F], fp); + fputc('=', fp); + fputc('=', fp); + } + if (buf_top == 2) { + fputc(EncodeTable[buf[1] >> 2], fp); + fputc(EncodeTable[((buf[1] << 4) | (buf[2] >> 4)) & 0x3F], fp); + fputc(EncodeTable[(buf[2] << 2) & 0x3F], fp); + fputc('=', fp); + } + buf_top = 0; + if (endch != EOF) fputc(endch, fp); + } + + private: + FILE *fp; + int buf_top; + unsigned char buf[4]; +}; +} // namespace utils +} // namespace xgboost +#endif // XGBOOST_UTILS_BASE64_H_ diff --git a/src/utils/bitmap.h b/src/utils/bitmap.h new file mode 100644 index 000000000000..ba12caf41d4c --- /dev/null +++ b/src/utils/bitmap.h @@ -0,0 +1,66 @@ +#ifndef XGBOOST_UTILS_BITMAP_H_ +#define XGBOOST_UTILS_BITMAP_H_ +/*! + * \file bitmap.h + * \brief a simple implement of bitmap + * NOTE: bitmap is only threadsafe per word access, remember this when using bitmap + * \author Tianqi Chen + */ +#include +#include "./utils.h" +#include "./omp.h" + +namespace xgboost { +namespace utils { +/*! \brief bit map that contains set of bit indicators */ +struct BitMap { + /*! \brief internal data structure */ + std::vector data; + /*! + * \brief resize the bitmap to be certain size + * \param size the size of bitmap + */ + inline void Resize(size_t size) { + data.resize((size + 31U) >> 5, 0); + } + /*! + * \brief query the i-th position of bitmap + * \param i the position in + */ + inline bool Get(size_t i) const { + return (data[i >> 5] >> (i & 31U)) & 1U; + } + /*! + * \brief set i-th position to true + * \param i position index + */ + inline void SetTrue(size_t i) { + data[i >> 5] |= (1 << (i & 31U)); + } + /*! \brief initialize the value of bit map from vector of bool*/ + inline void InitFromBool(const std::vector &vec) { + this->Resize(vec.size()); + // parallel over the full cases + bst_omp_uint nsize = static_cast(vec.size() / 32); + #pragma omp parallel for schedule(static) + for (bst_omp_uint i = 0; i < nsize; ++i) { + uint32_t res = 0; + for (int k = 0; k < 32; ++k) { + int bit = vec[(i << 5) | k]; + res |= (bit << k); + } + data[i] = res; + } + if (nsize != vec.size()) data.back() = 0; + for (size_t i = nsize; i < vec.size(); ++i) { + if (vec[i]) this->SetTrue(i); + } + } + /*! \brief clear the bitmap, set all places to false */ + inline void Clear(void) { + std::fill(data.begin(), data.end(), 0U); + } +}; +} // namespace utils +} // namespace xgboost +#endif diff --git a/src/utils/group_data.h b/src/utils/group_data.h new file mode 100644 index 000000000000..a25eb1eddc9b --- /dev/null +++ b/src/utils/group_data.h @@ -0,0 +1,111 @@ +#ifndef XGBOOST_UTILS_GROUP_DATA_H_ +#define XGBOOST_UTILS_GROUP_DATA_H_ +/*! + * \file group_data.h + * \brief this file defines utils to group data by integer keys + * Input: given input sequence (key,value), (k1,v1), (k2,v2) + * Ouptupt: an array of values data = [v1,v2,v3 .. vn] + * and a group pointer ptr, + * data[ptr[k]:ptr[k+1]] contains values that corresponds to key k + * + * This can be used to construct CSR/CSC matrix from un-ordered input + * The major algorithm is a two pass linear scan algorithm that requires two pass scan over the data + * \author Tianqi Chen + */ +namespace xgboost { +namespace utils { +/*! + * \brief multi-thread version of group builder + * \tparam ValueType type of entries in the sparse matrix + * \tparam SizeType type of the index range holder + */ +template +struct ParallelGroupBuilder { + public: + // parallel group builder of data + ParallelGroupBuilder(std::vector *p_rptr, + std::vector *p_data) + : rptr(*p_rptr), data(*p_data), thread_rptr(tmp_thread_rptr) { + } + ParallelGroupBuilder(std::vector *p_rptr, + std::vector *p_data, + std::vector< std::vector > *p_thread_rptr) + : rptr(*p_rptr), data(*p_data), thread_rptr(*p_thread_rptr) { + } + + public: + /*! + * \brief step 1: initialize the helper, with hint of number keys + * and thread used in the construction + * \param nkeys number of keys in the matrix, can be smaller than expected + * \param nthread number of thread that will be used in construction + */ + inline void InitBudget(size_t nkeys = 0, int nthread = 1) { + thread_rptr.resize(nthread); + for (size_t i = 0; i < thread_rptr.size(); ++i) { + thread_rptr[i].resize(nkeys); + std::fill(thread_rptr[i].begin(), thread_rptr[i].end(), 0); + } + } + /*! + * \brief step 2: add budget to each key + * \param key the key + * \param threadid the id of thread that calls this function + * \param nelem number of element budget add to this row + */ + inline void AddBudget(size_t key, int threadid = 0, SizeType nelem = 1) { + std::vector &trptr = thread_rptr[threadid]; + if (trptr.size() < key + 1) { + trptr.resize(key + 1, 0); + } + trptr[key] += nelem; + } + /*! \brief step 3: initialize the necessary storage */ + inline void InitStorage(void) { + // set rptr to correct size + for (size_t tid = 0; tid < thread_rptr.size(); ++tid) { + if (rptr.size() <= thread_rptr[tid].size()) { + rptr.resize(thread_rptr[tid].size()+1); + } + } + // initialize rptr to be beginning of each segment + size_t start = 0; + for (size_t i = 0; i + 1 < rptr.size(); ++i) { + for (size_t tid = 0; tid < thread_rptr.size(); ++tid) { + std::vector &trptr = thread_rptr[tid]; + if (i < trptr.size()) { + size_t ncnt = trptr[i]; + trptr[i] = start; + start += ncnt; + } + } + rptr[i + 1] = start; + } + data.resize(start); + } + /*! + * \brief step 4: add data to the allocated space, + * the calls to this function should be exactly match previous call to AddBudget + * + * \param key the key of + * \param threadid the id of thread that calls this function + */ + inline void Push(size_t key, ValueType value, int threadid = 0) { + SizeType &rp = thread_rptr[threadid][key]; + data[rp++] = value; + } + + private: + /*! \brief pointer to the beginning and end of each continuous key */ + std::vector &rptr; + /*! \brief index of nonzero entries in each row */ + std::vector &data; + /*! \brief thread local data structure */ + std::vector< std::vector > &thread_rptr; + /*! \brief local temp thread ptr, use this if not specified by the constructor */ + std::vector< std::vector > tmp_thread_rptr; +}; +} // namespace utils +} // namespace xgboost +#endif + diff --git a/src/utils/io.h b/src/utils/io.h index 026e3fec7f3a..dff691ee0cb0 100644 --- a/src/utils/io.h +++ b/src/utils/io.h @@ -88,12 +88,98 @@ class IStream { } }; -/*! \brief implementation of file i/o stream */ -class FileStream : public IStream { +/*! \brief interface of i/o stream that support seek */ +class ISeekStream: public IStream { + public: + /*! \brief seek to certain position of the file */ + virtual void Seek(size_t pos) = 0; + /*! \brief tell the position of the stream */ + virtual size_t Tell(void) = 0; +}; + +/*! \brief fixed size memory buffer */ +struct MemoryFixSizeBuffer : public ISeekStream { + public: + MemoryFixSizeBuffer(void *p_buffer, size_t buffer_size) + : p_buffer_(reinterpret_cast(p_buffer)), buffer_size_(buffer_size) { + curr_ptr_ = 0; + } + virtual ~MemoryFixSizeBuffer(void) {} + virtual size_t Read(void *ptr, size_t size) { + utils::Assert(curr_ptr_ + size <= buffer_size_, + "read can not have position excceed buffer length"); + size_t nread = std::min(buffer_size_ - curr_ptr_, size); + if (nread != 0) memcpy(ptr, p_buffer_ + curr_ptr_, nread); + curr_ptr_ += nread; + return nread; + } + virtual void Write(const void *ptr, size_t size) { + if (size == 0) return; + utils::Assert(curr_ptr_ + size <= buffer_size_, + "write position exceed fixed buffer size"); + memcpy(p_buffer_ + curr_ptr_, ptr, size); + curr_ptr_ += size; + } + virtual void Seek(size_t pos) { + curr_ptr_ = static_cast(pos); + } + virtual size_t Tell(void) { + return curr_ptr_; + } + + private: + /*! \brief in memory buffer */ + char *p_buffer_; + /*! \brief current pointer */ + size_t buffer_size_; + /*! \brief current pointer */ + size_t curr_ptr_; +}; // class MemoryFixSizeBuffer + +/*! \brief a in memory buffer that can be read and write as stream interface */ +struct MemoryBufferStream : public ISeekStream { + public: + MemoryBufferStream(std::string *p_buffer) + : p_buffer_(p_buffer) { + curr_ptr_ = 0; + } + virtual ~MemoryBufferStream(void) {} + virtual size_t Read(void *ptr, size_t size) { + utils::Assert(curr_ptr_ <= p_buffer_->length(), + "read can not have position excceed buffer length"); + size_t nread = std::min(p_buffer_->length() - curr_ptr_, size); + if (nread != 0) memcpy(ptr, &(*p_buffer_)[0] + curr_ptr_, nread); + curr_ptr_ += nread; + return nread; + } + virtual void Write(const void *ptr, size_t size) { + if (size == 0) return; + if (curr_ptr_ + size > p_buffer_->length()) { + p_buffer_->resize(curr_ptr_+size); + } + memcpy(&(*p_buffer_)[0] + curr_ptr_, ptr, size); + curr_ptr_ += size; + } + virtual void Seek(size_t pos) { + curr_ptr_ = static_cast(pos); + } + virtual size_t Tell(void) { + return curr_ptr_; + } + private: - std::FILE *fp; + /*! \brief in memory buffer */ + std::string *p_buffer_; + /*! \brief current pointer */ + size_t curr_ptr_; +}; // class MemoryBufferStream + +/*! \brief implementation of file i/o stream */ +class FileStream : public ISeekStream { public: - explicit FileStream(std::FILE *fp) : fp(fp) { + explicit FileStream(FILE *fp) : fp(fp) {} + explicit FileStream(void) { + this->fp = NULL; } virtual size_t Read(void *ptr, size_t size) { return std::fread(ptr, size, 1, fp); @@ -101,14 +187,21 @@ class FileStream : public IStream { virtual void Write(const void *ptr, size_t size) { std::fwrite(ptr, size, 1, fp); } - inline void Seek(size_t pos) { - std::fseek(fp, 0, SEEK_SET); + virtual void Seek(size_t pos) { + std::fseek(fp, static_cast(pos), SEEK_SET); + } + virtual size_t Tell(void) { + return std::ftell(fp); } inline void Close(void) { - std::fclose(fp); + if (fp != NULL){ + std::fclose(fp); fp = NULL; + } } -}; + private: + FILE *fp; +}; } // namespace utils } // namespace xgboost #endif diff --git a/src/utils/matrix_csr.h b/src/utils/matrix_csr.h index 0f3b20a14a49..bc9479cc3945 100644 --- a/src/utils/matrix_csr.h +++ b/src/utils/matrix_csr.h @@ -6,8 +6,11 @@ * \author Tianqi Chen */ #include +#include #include +#include "./io.h" #include "./utils.h" +#include "./omp.h" namespace xgboost { namespace utils { @@ -118,6 +121,141 @@ struct SparseCSRMBuilder { } }; +/*! + * \brief a class used to help construct CSR format matrix file + * \tparam IndexType type of index used to store the index position + * \tparam SizeType type of size used in row pointer + */ +template +struct SparseCSRFileBuilder { + public: + explicit SparseCSRFileBuilder(utils::ISeekStream *fo, size_t buffer_size) + : fo(fo), buffer_size(buffer_size) { + } + /*! + * \brief step 1: initialize the number of rows in the data, not necessary exact + * \nrows number of rows in the matrix, can be smaller than expected + */ + inline void InitBudget(size_t nrows = 0) { + rptr.clear(); + rptr.resize(nrows + 1, 0); + } + /*! + * \brief step 2: add budget to each rows + * \param row_id the id of the row + * \param nelem number of element budget add to this row + */ + inline void AddBudget(size_t row_id, SizeType nelem = 1) { + if (rptr.size() < row_id + 2) { + rptr.resize(row_id + 2, 0); + } + rptr[row_id + 1] += nelem; + } + /*! \brief step 3: initialize the necessary storage */ + inline void InitStorage(void) { + SizeType nelem = 0; + for (size_t i = 1; i < rptr.size(); i++) { + nelem += rptr[i]; + rptr[i] = nelem; + } + begin_data = static_cast(fo->Tell()) + sizeof(SizeType); + SizeType begin_meta = begin_data + nelem * sizeof(IndexType); + fo->Write(&begin_meta, sizeof(begin_meta)); + fo->Seek(begin_meta); + fo->Write(rptr); + // setup buffer space + buffer_rptr.resize(rptr.size()); + buffer_temp.reserve(buffer_size); + buffer_data.resize(buffer_size); + saved_offset = rptr; + saved_offset.resize(rptr.size() - 1); + this->ClearBuffer(); + } + /*! \brief step 4: push element into buffer */ + inline void PushElem(SizeType row_id, IndexType col_id) { + if (buffer_temp.size() == buffer_size) { + this->WriteBuffer(); + this->ClearBuffer(); + } + buffer_rptr[row_id + 1] += 1; + buffer_temp.push_back(std::make_pair(row_id, col_id)); + } + /*! \brief finalize the construction */ + inline void Finalize(void) { + this->WriteBuffer(); + for (size_t i = 0; i < saved_offset.size(); ++i) { + utils::Assert(saved_offset[i] == rptr[i+1], "some block not write out"); + } + } + /*! \brief content must be in wb+ */ + template + inline void SortRows(Comparator comp, size_t step) { + for (size_t i = 0; i < rptr.size() - 1; i += step) { + bst_omp_uint begin = static_cast(i); + bst_omp_uint end = static_cast(std::min(rptr.size() - 1, i + step)); + if (rptr[end] != rptr[begin]) { + fo->Seek(begin_data + rptr[begin] * sizeof(IndexType)); + buffer_data.resize(rptr[end] - rptr[begin]); + fo->Read(BeginPtr(buffer_data), (rptr[end] - rptr[begin]) * sizeof(IndexType)); + // do parallel sorting + #pragma omp parallel for schedule(static) + for (bst_omp_uint j = begin; j < end; ++j) { + std::sort(&buffer_data[0] + rptr[j] - rptr[begin], + &buffer_data[0] + rptr[j+1] - rptr[begin], + comp); + } + fo->Seek(begin_data + rptr[begin] * sizeof(IndexType)); + fo->Write(BeginPtr(buffer_data), (rptr[end] - rptr[begin]) * sizeof(IndexType)); + } + } + printf("CSV::begin_dat=%lu\n", begin_data); + } + protected: + inline void WriteBuffer(void) { + SizeType start = 0; + for (size_t i = 1; i < buffer_rptr.size(); ++i) { + size_t rlen = buffer_rptr[i]; + buffer_rptr[i] = start; + start += rlen; + } + for (size_t i = 0; i < buffer_temp.size(); ++i) { + SizeType &rp = buffer_rptr[buffer_temp[i].first + 1]; + buffer_data[rp++] = buffer_temp[i].second; + } + // write out + for (size_t i = 0; i < buffer_rptr.size() - 1; ++i) { + size_t nelem = buffer_rptr[i+1] - buffer_rptr[i]; + if (nelem != 0) { + utils::Assert(saved_offset[i] + nelem <= rptr[i+1], "data exceed bound"); + fo->Seek(saved_offset[i] * sizeof(IndexType) + begin_data); + fo->Write(&buffer_data[0] + buffer_rptr[i], nelem * sizeof(IndexType)); + saved_offset[i] += nelem; + } + } + } + inline void ClearBuffer(void) { + buffer_temp.clear(); + std::fill(buffer_rptr.begin(), buffer_rptr.end(), 0); + } + private: + /*! \brief output file pointer the data */ + utils::ISeekStream *fo; + /*! \brief pointer to each of the row */ + std::vector rptr; + /*! \brief saved top space of each item */ + std::vector saved_offset; + /*! \brief beginning position of data */ + size_t begin_data; + // ----- the following are buffer space + /*! \brief maximum size of content buffer*/ + size_t buffer_size; + /*! \brief store the data content */ + std::vector< std::pair > buffer_temp; + /*! \brief saved top space of each item */ + std::vector buffer_rptr; + /*! \brief saved top space of each item */ + std::vector buffer_data; +}; } // namespace utils } // namespace xgboost #endif diff --git a/src/utils/quantile.h b/src/utils/quantile.h new file mode 100644 index 000000000000..bc76f40171ab --- /dev/null +++ b/src/utils/quantile.h @@ -0,0 +1,747 @@ +#ifndef XGBOOST_UTILS_QUANTILE_H_ +#define XGBOOST_UTILS_QUANTILE_H_ +/*! + * \file quantile.h + * \brief util to compute quantiles + * \author Tianqi Chen + */ +#include +#include +#include +#include +#include +#include "./io.h" +#include "./utils.h" + +namespace xgboost { +namespace utils { +/*! + * \brief experimental wsummary + * \tparam DType type of data content + * \tparam RType type of rank + */ +template +struct WQSummary { + /*! \brief an entry in the sketch summary */ + struct Entry { + /*! \brief minimum rank */ + RType rmin; + /*! \brief maximum rank */ + RType rmax; + /*! \brief maximum weight */ + RType wmin; + /*! \brief the value of data */ + DType value; + // constructor + Entry(void) {} + // constructor + Entry(RType rmin, RType rmax, RType wmin, DType value) + : rmin(rmin), rmax(rmax), wmin(wmin), value(value) {} + /*! + * \brief debug function, check Valid + * \param eps the tolerate level for violating the relation + */ + inline void CheckValid(RType eps = 0) const { + utils::Assert(rmin >= 0 && rmax >= 0 && wmin >= 0, "nonneg constraint"); + utils::Assert(rmax- rmin - wmin > -eps, "relation constraint: min/max"); + } + /*! \return rmin estimation for v strictly bigger than value */ + inline RType rmin_next(void) const { + return rmin + wmin; + } + /*! \return rmax estimation for v strictly smaller than value */ + inline RType rmax_prev(void) const { + return rmax - wmin; + } + }; + /*! \brief input data queue before entering the summary */ + struct Queue { + // entry in the queue + struct QEntry { + // value of the instance + DType value; + // weight of instance + RType weight; + // default constructor + QEntry(void) {} + // constructor + QEntry(DType value, RType weight) + : value(value), weight(weight) {} + // comparator on value + inline bool operator<(const QEntry &b) const { + return value < b.value; + } + }; + // the input queue + std::vector queue; + // end of the queue + size_t qtail; + // push data to the queue + inline void Push(DType x, RType w) { + if (qtail == 0 || queue[qtail - 1].value != x) { + queue[qtail++] = QEntry(x, w); + } else { + queue[qtail - 1].weight += w; + } + } + inline void MakeSummary(WQSummary *out) { + std::sort(queue.begin(), queue.begin() + qtail); + out->size = 0; + // start update sketch + RType wsum = 0; + // construct data with unique weights + for (size_t i = 0; i < qtail;) { + size_t j = i + 1; + RType w = queue[i].weight; + while (j < qtail && queue[j].value == queue[i].value) { + w += queue[j].weight; ++j; + } + out->data[out->size++] = Entry(wsum, wsum + w, w, queue[i].value); + wsum += w; i = j; + } + } + }; + /*! \brief data field */ + Entry *data; + /*! \brief number of elements in the summary */ + size_t size; + // constructor + WQSummary(Entry *data, size_t size) + : data(data), size(size) {} + /*! + * \return the maximum error of the Summary + */ + inline RType MaxError(void) const { + RType res = data[0].rmax - data[0].rmin - data[0].wmin; + for (size_t i = 1; i < size; ++i) { + res = std::max(data[i].rmax_prev() - data[i - 1].rmin_next(), res); + res = std::max(data[i].rmax - data[i].rmin - data[i].wmin, res); + } + return res; + } + /*! + * \brief query qvalue, start from istart + * \param qvalue the value we query for + * \param istart starting position + */ + inline Entry Query(DType qvalue, size_t &istart) const { + while (istart < size && qvalue > data[istart].value) { + ++istart; + } + if (istart == size) { + RType rmax = data[size - 1].rmax; + return Entry(rmax, rmax, 0.0f, qvalue); + } + if (qvalue == data[istart].value) { + return data[istart]; + } else { + if (istart == 0) { + return Entry(0.0f, 0.0f, 0.0f, qvalue); + } else { + return Entry(data[istart - 1].rmin_next(), + data[istart].rmax_prev(), + 0.0f, qvalue); + } + } + } + /*! \return maximum rank in the summary */ + inline RType MaxRank(void) const { + return data[size - 1].rmax; + } + /*! + * \brief copy content from src + * \param src source sketch + */ + inline void CopyFrom(const WQSummary &src) { + size = src.size; + std::memcpy(data, src.data, sizeof(Entry) * size); + } + /*! + * \brief debug function, validate whether the summary + * run consistency check to check if it is a valid summary + * \param eps the tolerate error level, used when RType is floating point and + * some inconsistency could occur due to rounding error + */ + inline void CheckValid(RType eps) const { + for (size_t i = 0; i < size; ++i) { + data[i].CheckValid(eps); + if (i != 0) { + utils::Assert(data[i].rmin >= data[i - 1].rmin + data[i - 1].wmin, "rmin range constraint"); + utils::Assert(data[i].rmax >= data[i - 1].rmax + data[i].wmin, "rmax range constraint"); + } + } + } + /*! \brief used for debug purpose, print the summary */ + inline void Print(void) const { + for (size_t i = 0; i < size; ++i) { + std::cout << "x=" << data[i].value << "\t" + << "[" << data[i].rmin << "," << data[i].rmax << "]" + << " wmin=" << data[i].wmin << std::endl; + } + } + /*! + * \brief set current summary to be pruned summary of src + * assume data field is already allocated to be at least maxsize + * \param src source summary + * \param maxsize size we can afford in the pruned sketch + */ + + inline void SetPrune(const WQSummary &src, size_t maxsize) { + if (src.size <= maxsize) { + this->CopyFrom(src); return; + } + const RType begin = src.data[0].rmax; + const RType range = src.data[src.size - 1].rmin - src.data[0].rmax; + const size_t n = maxsize - 1; + data[0] = src.data[0]; + this->size = 1; + // lastidx is used to avoid duplicated records + size_t i = 1, lastidx = 0; + for (size_t k = 1; k < n; ++k) { + RType dx2 = 2 * ((k * range) / n + begin); + // find first i such that d < (rmax[i+1] + rmin[i+1]) / 2 + while (i < src.size - 1 + && dx2 >= src.data[i + 1].rmax + src.data[i + 1].rmin) ++i; + utils::Assert(i != src.size - 1, "this cannot happen"); + if (dx2 < src.data[i].rmin_next() + src.data[i + 1].rmax_prev()) { + if (i != lastidx) { + data[size++] = src.data[i]; lastidx = i; + } + } else { + if (i + 1 != lastidx) { + data[size++] = src.data[i + 1]; lastidx = i + 1; + } + } + } + if (lastidx != src.size - 1) { + data[size++] = src.data[src.size - 1]; + } + } + /*! + * \brief set current summary to be merged summary of sa and sb + * \param sa first input summary to be merged + * \param sb second input summar to be merged + */ + inline void SetCombine(const WQSummary &sa, + const WQSummary &sb) { + if (sa.size == 0) { + this->CopyFrom(sb); return; + } + if (sb.size == 0) { + this->CopyFrom(sa); return; + } + utils::Assert(sa.size > 0 && sb.size > 0, "invalid input for merge"); + const Entry *a = sa.data, *a_end = sa.data + sa.size; + const Entry *b = sb.data, *b_end = sb.data + sb.size; + // extended rmin value + RType aprev_rmin = 0, bprev_rmin = 0; + Entry *dst = this->data; + while (a != a_end && b != b_end) { + // duplicated value entry + if (a->value == b->value) { + *dst = Entry(a->rmin + b->rmin, + a->rmax + b->rmax, + a->wmin + b->wmin, a->value); + aprev_rmin = a->rmin_next(); + bprev_rmin = b->rmin_next(); + ++dst; ++a; ++b; + } else if (a->value < b->value) { + *dst = Entry(a->rmin + bprev_rmin, + a->rmax + b->rmax_prev(), + a->wmin, a->value); + aprev_rmin = a->rmin_next(); + ++dst; ++a; + } else { + *dst = Entry(b->rmin + aprev_rmin, + b->rmax + a->rmax_prev(), + b->wmin, b->value); + bprev_rmin = b->rmin_next(); + ++dst; ++b; + } + } + if (a != a_end) { + RType brmax = (b_end - 1)->rmax; + do { + *dst = Entry(a->rmin + bprev_rmin, a->rmax + brmax, a->wmin, a->value); + ++dst; ++a; + } while (a != a_end); + } + if (b != b_end) { + RType armax = (a_end - 1)->rmax; + do { + *dst = Entry(b->rmin + aprev_rmin, b->rmax + armax, b->wmin, b->value); + ++dst; ++b; + } while (b != b_end); + } + this->size = dst - data; + utils::Assert(size <= sa.size + sb.size, "bug in combine"); + } +}; +/*! \brief try to do efficient prunning */ +template +struct WXQSummary : public WQSummary { + // redefine entry type + typedef typename WQSummary::Entry Entry; + // constructor + WXQSummary(Entry *data, size_t size) + : WQSummary(data, size) {} + // check if the block is large chunk + inline static bool CheckLarge(const Entry &e, RType chunk) { + return e.rmin_next() > e.rmax_prev() + chunk; + } + // set prune + inline void SetPrune(const WQSummary &src, size_t maxsize) { + if (src.size <= maxsize) { + this->CopyFrom(src); return; + } + RType begin = src.data[0].rmax; + size_t n = maxsize - 1, nbig = 0; + const RType range = src.data[src.size - 1].rmin - begin; + const RType chunk = 2 * range / n; + // minimized range + RType mrange = 0; + { + // first scan, grab all the big chunk + // moviing block index + size_t bid = 0; + for (size_t i = 1; i < src.size; ++i) { + if (CheckLarge(src.data[i], chunk)) { + if (bid != i - 1) { + mrange += src.data[i].rmax_prev() - src.data[bid].rmin_next(); + } + bid = i; ++nbig; + } + } + if (bid != src.size - 2) { + mrange += src.data[src.size-1].rmax_prev() - src.data[bid].rmin_next(); + } + } + utils::Assert(nbig < n - 1, "too many large chunk"); + this->data[0] = src.data[0]; + this->size = 1; + // use smaller size + n = n - nbig; + // find the rest of point + size_t bid = 0, k = 1, lastidx = 0; + for (size_t end = 1; end < src.size; ++end) { + if (end == src.size - 1 || CheckLarge(src.data[end], chunk)) { + if (bid != end - 1) { + size_t i = bid; + RType maxdx2 = src.data[end].rmax_prev() * 2; + for (; k < n; ++k) { + RType dx2 = 2 * ((k * mrange) / n + begin); + if (dx2 >= maxdx2) break; + while (i < end && + dx2 >= src.data[i + 1].rmax + src.data[i + 1].rmin) ++i; + if (dx2 < src.data[i].rmin_next() + src.data[i + 1].rmax_prev()) { + if (i != lastidx) { + this->data[this->size++] = src.data[i]; lastidx = i; + } + } else { + if (i + 1 != lastidx) { + this->data[this->size++] = src.data[i + 1]; lastidx = i + 1; + } + } + } + } + if (lastidx != end) { + this->data[this->size++] = src.data[end]; + lastidx = end; + } + bid = end; + // shift base by the gap + begin += src.data[bid].rmin_next() - src.data[bid].rmax_prev(); + } + } + } +}; +/*! + * \brief traditional GK summary + */ +template +struct GKSummary { + /*! \brief an entry in the sketch summary */ + struct Entry { + /*! \brief minimum rank */ + RType rmin; + /*! \brief maximum rank */ + RType rmax; + /*! \brief the value of data */ + DType value; + // constructor + Entry(void) {} + // constructor + Entry(RType rmin, RType rmax, DType value) + : rmin(rmin), rmax(rmax), value(value) {} + }; + /*! \brief input data queue before entering the summary */ + struct Queue { + // the input queue + std::vector queue; + // end of the queue + size_t qtail; + // push data to the queue + inline void Push(DType x, RType w) { + queue[qtail++] = x; + } + inline void MakeSummary(GKSummary *out) { + std::sort(queue.begin(), queue.begin() + qtail); + out->size = qtail; + for (size_t i = 0; i < qtail; ++i) { + out->data[i] = Entry(i + 1, i + 1, queue[i]); + } + } + }; + /*! \brief data field */ + Entry *data; + /*! \brief number of elements in the summary */ + size_t size; + GKSummary(Entry *data, size_t size) + : data(data), size(size) {} + /*! \brief the maximum error of the summary */ + inline RType MaxError(void) const { + RType res = 0; + for (size_t i = 1; i < size; ++i) { + res = std::max(data[i].rmax - data[i-1].rmin, res); + } + return res; + } + /*! \return maximum rank in the summary */ + inline RType MaxRank(void) const { + return data[size - 1].rmax; + } + /*! + * \brief copy content from src + * \param src source sketch + */ + inline void CopyFrom(const GKSummary &src) { + size = src.size; + std::memcpy(data, src.data, sizeof(Entry) * size); + } + inline void CheckValid(RType eps) const { + // assume always valid + } + /*! \brief used for debug purpose, print the summary */ + inline void Print(void) const { + for (size_t i = 0; i < size; ++i) { + std::cout << "x=" << data[i].value << "\t" + << "[" << data[i].rmin << "," << data[i].rmax << "]" + << std::endl; + } + } + /*! + * \brief set current summary to be pruned summary of src + * assume data field is already allocated to be at least maxsize + * \param src source summary + * \param maxsize size we can afford in the pruned sketch + */ + inline void SetPrune(const GKSummary &src, size_t maxsize) { + if (src.size <= maxsize) { + this->CopyFrom(src); return; + } + const RType max_rank = src.MaxRank(); + this->size = maxsize; + data[0] = src.data[0]; + size_t n = maxsize - 1; + RType top = 1; + for (size_t i = 1; i < n; ++i) { + RType k = (i * max_rank) / n; + while (k > src.data[top + 1].rmax) ++top; + // assert src.data[top].rmin <= k + // because k > src.data[top].rmax >= src.data[top].rmin + if ((k - src.data[top].rmin) < (src.data[top+1].rmax - k)) { + data[i] = src.data[top]; + } else { + data[i] = src.data[top + 1]; + } + } + data[n] = src.data[src.size - 1]; + } + inline void SetCombine(const GKSummary &sa, + const GKSummary &sb) { + if (sa.size == 0) { + this->CopyFrom(sb); return; + } + if (sb.size == 0) { + this->CopyFrom(sa); return; + } + utils::Assert(sa.size > 0 && sb.size > 0, "invalid input for merge"); + const Entry *a = sa.data, *a_end = sa.data + sa.size; + const Entry *b = sb.data, *b_end = sb.data + sb.size; + this->size = sa.size + sb.size; + RType aprev_rmin = 0, bprev_rmin = 0; + Entry *dst = this->data; + while (a != a_end && b != b_end) { + if (a->value < b->value) { + *dst = Entry(bprev_rmin + a->rmin, + a->rmax + b->rmax - 1, a->value); + aprev_rmin = a->rmin; + ++dst; ++a; + } else { + *dst = Entry(aprev_rmin + b->rmin, + b->rmax + a->rmax - 1, b->value); + bprev_rmin = b->rmin; + ++dst; ++b; + } + } + if (a != a_end) { + RType bprev_rmax = (b_end - 1)->rmax; + do { + *dst = Entry(bprev_rmin + a->rmin, bprev_rmax + a->rmax, a->value); + ++dst; ++a; + } while (a != a_end); + } + if (b != b_end) { + RType aprev_rmax = (a_end - 1)->rmax; + do { + *dst = Entry(aprev_rmin + b->rmin, aprev_rmax + b->rmax, b->value); + ++dst; ++b; + } while (b != b_end); + } + utils::Assert(dst == data + size, "bug in combine"); + } +}; + +/*! + * \brief template for all quantle sketch algorithm + * that uses merge/prune scheme + * \tparam DType type of data content + * \tparam RType type of rank + * \tparam TSummary actual summary data structure it uses + */ +template +class QuantileSketchTemplate { + public: + /*! \brief type of summary type */ + typedef TSummary Summary; + /*! \brief the entry type */ + typedef typename Summary::Entry Entry; + /*! \brief same as summary, but use STL to backup the space */ + struct SummaryContainer : public Summary { + std::vector space; + SummaryContainer(const SummaryContainer &src) : Summary(NULL, src.size) { + this->space = src.space; + this->data = BeginPtr(this->space); + } + SummaryContainer(void) : Summary(NULL, 0) { + } + /*! \brief reserve space for summary */ + inline void Reserve(size_t size) { + if (size > space.size()) { + space.resize(size); + this->data = BeginPtr(space); + } + } + /*! + * \brief set the space to be merge of all Summary arrays + * \param begin begining position in th summary array + * \param end ending position in the Summary array + */ + inline void SetMerge(const Summary *begin, + const Summary *end) { + utils::Assert(begin < end, "can not set combine to empty instance"); + size_t len = end - begin; + if (len == 1) { + this->Reserve(begin[0].size); + this->CopyFrom(begin[0]); + } else if (len == 2) { + this->Reserve(begin[0].size + begin[1].size); + this->SetMerge(begin[0], begin[1]); + } else { + // recursive merge + SummaryContainer lhs, rhs; + lhs.SetCombine(begin, begin + len / 2); + rhs.SetCombine(begin + len / 2, end); + this->Reserve(lhs.size + rhs.size); + this->SetCombine(lhs, rhs); + } + } + /*! + * \brief do elementwise combination of summary array + * this[i] = combine(this[i], src[i]) for each i + * \param src the source summary + * \param max_nbyte, maximum number of byte allowed in here + */ + inline void Reduce(const Summary &src, size_t max_nbyte) { + this->Reserve((max_nbyte - sizeof(this->size)) / sizeof(Entry)); + SummaryContainer temp; + temp.Reserve(this->size + src.size); + temp.SetCombine(*this, src); + this->SetPrune(temp, space.size()); + } + /*! \brief return the number of bytes this data structure cost in serialization */ + inline static size_t CalcMemCost(size_t nentry) { + return sizeof(size_t) + sizeof(Entry) * nentry; + } + /*! \brief save the data structure into stream */ + template + inline void Save(TStream &fo) const { + fo.Write(&(this->size), sizeof(this->size)); + if (this->size != 0) { + fo.Write(this->data, this->size * sizeof(Entry)); + } + } + /*! \brief load data structure from input stream */ + template + inline void Load(TStream &fi) { + utils::Check(fi.Read(&this->size, sizeof(this->size)) != 0, "invalid SummaryArray 1"); + this->Reserve(this->size); + if (this->size != 0) { + utils::Check(fi.Read(this->data, this->size * sizeof(Entry)) != 0, "invalid SummaryArray 2"); + } + } + }; + /*! + * \brief intialize the quantile sketch, given the performance specification + * \param maxn maximum number of data points can be feed into sketch + * \param eps accuracy level of summary + */ + inline void Init(size_t maxn, double eps) { + nlevel = 1; + while (true) { + limit_size = static_cast(ceil(nlevel / eps)) + 1; + size_t n = (1UL << nlevel); + if (n * limit_size >= maxn) break; + ++nlevel; + } + // check invariant + size_t n = (1UL << nlevel); + utils::Assert(n * limit_size >= maxn, "invalid init parameter"); + utils::Assert(nlevel <= limit_size * eps, "invalid init parameter"); + // lazy reserve the space, if there is only one value, no need to allocate space + inqueue.queue.resize(1); + inqueue.qtail = 0; + data.clear(); + level.clear(); + } + /*! + * \brief add an element to a sketch + * \param x the elemented added to the sketch + */ + inline void Push(DType x, RType w = 1) { + if (inqueue.qtail == inqueue.queue.size()) { + // jump from lazy one value to limit_size * 2 + if (inqueue.queue.size() == 1) { + inqueue.queue.resize(limit_size * 2); + } else { + temp.Reserve(limit_size * 2); + inqueue.MakeSummary(&temp); + // cleanup queue + inqueue.qtail = 0; + this->PushTemp(); + } + } + inqueue.Push(x, w); + } + /*! \brief push up temp */ + inline void PushTemp(void) { + temp.Reserve(limit_size * 2); + for (size_t l = 1; true; ++l) { + this->InitLevel(l + 1); + // check if level l is empty + if (level[l].size == 0) { + level[l].SetPrune(temp, limit_size); + break; + } else { + // level 0 is actually temp space + level[0].SetPrune(temp, limit_size); + temp.SetCombine(level[0], level[l]); + if (temp.size > limit_size) { + // try next level + level[l].size = 0; + } else { + // if merged record is still smaller, no need to send to next level + level[l].CopyFrom(temp); break; + } + } + } + } + /*! \brief get the summary after finalize */ + inline void GetSummary(SummaryContainer *out) { + if (level.size() != 0) { + out->Reserve(limit_size * 2); + } else { + out->Reserve(inqueue.queue.size()); + } + inqueue.MakeSummary(out); + if (level.size() != 0) { + level[0].SetPrune(*out, limit_size); + for (size_t l = 1; l < level.size(); ++l) { + if (level[l].size == 0) continue; + if (level[0].size == 0) { + level[0].CopyFrom(level[l]); + } else { + out->SetCombine(level[0], level[l]); + level[0].SetPrune(*out, limit_size); + } + } + out->CopyFrom(level[0]); + } else { + if (out->size > limit_size) { + temp.Reserve(limit_size); + temp.SetPrune(*out, limit_size); + out->CopyFrom(temp); + } + } + } + // used for debug, check if the sketch is valid + inline void CheckValid(RType eps) const { + for (size_t l = 1; l < level.size(); ++l) { + level[l].CheckValid(eps); + } + } + // initialize level space to at least nlevel + inline void InitLevel(size_t nlevel) { + if (level.size() >= nlevel) return; + data.resize(limit_size * nlevel); + level.resize(nlevel, Summary(NULL, 0)); + for (size_t l = 0; l < level.size(); ++l) { + level[l].data = BeginPtr(data) + l * limit_size; + } + } + // input data queue + typename Summary::Queue inqueue; + // number of levels + size_t nlevel; + // size of summary in each level + size_t limit_size; + // the level of each summaries + std::vector level; + // content of the summary + std::vector data; + // temporal summary, used for temp-merge + SummaryContainer temp; +}; + +/*! + * \brief Quantile sketch use WQSummary + * \tparam DType type of data content + * \tparam RType type of rank + */ +template +class WQuantileSketch : + public QuantileSketchTemplate >{ +}; + +/*! + * \brief Quantile sketch use WXQSummary + * \tparam DType type of data content + * \tparam RType type of rank + */ +template +class WXQuantileSketch : + public QuantileSketchTemplate >{ +}; +/*! + * \brief Quantile sketch use WQSummary + * \tparam DType type of data content + * \tparam RType type of rank + */ +template +class GKQuantileSketch : + public QuantileSketchTemplate >{ +}; + +} // utils +} // xgboost +#endif diff --git a/src/utils/socket.h b/src/utils/socket.h new file mode 100644 index 000000000000..73caa2feabb1 --- /dev/null +++ b/src/utils/socket.h @@ -0,0 +1,389 @@ +#ifndef XGBOOST_UTILS_SOCKET_H +#define XGBOOST_UTILS_SOCKET_H +/*! + * \file socket.h + * \brief this file aims to provide a wrapper of sockets + * \author Tianqi Chen + */ +#if defined(_WIN32) +#include +#include +#else +#include +#include +#include +#include +#include +#include +#include +#include +#endif +#include +#include +#include "./utils.h" + +namespace xgboost { +namespace utils { +#if defined(_WIN32) +typedef int ssize_t; +typedef int sock_size_t; +#else +typedef int SOCKET; +typedef size_t sock_size_t; +const int INVALID_SOCKET = -1; +#endif + +/*! \brief data structure for network address */ +struct SockAddr { + sockaddr_in addr; + // constructor + SockAddr(void) {} + SockAddr(const char *url, int port) { + this->Set(url, port); + } + inline static std::string GetHostName(void) { + std::string buf; buf.resize(256); + utils::Check(gethostname(&buf[0], 256) != -1, "fail to get host name"); + return std::string(buf.c_str()); + } + /*! + * \brief set the address + * \param url the url of the address + * \param port the port of address + */ + inline void Set(const char *host, int port) { + hostent *hp = gethostbyname(host); + Check(hp != NULL, "cannot obtain address of %s", host); + memset(&addr, 0, sizeof(addr)); + addr.sin_family = AF_INET; + addr.sin_port = htons(port); + memcpy(&addr.sin_addr, hp->h_addr_list[0], hp->h_length); + } + /*! \brief return port of the address*/ + inline int port(void) const { + return ntohs(addr.sin_port); + } + /*! \return a string representation of the address */ + inline std::string AddrStr(void) const { + std::string buf; buf.resize(256); +#ifdef _WIN32 + const char *s = inet_ntop(AF_INET, (PVOID)&addr.sin_addr, &buf[0], buf.length()); +#else + const char *s = inet_ntop(AF_INET, &addr.sin_addr, &buf[0], buf.length()); +#endif + Assert(s != NULL, "cannot decode address"); + return std::string(s); + } +}; +/*! + * \brief a wrapper of TCP socket that hopefully be cross platform + */ +class TCPSocket { + public: + /*! \brief the file descriptor of socket */ + SOCKET sockfd; + // constructor + TCPSocket(void) : sockfd(INVALID_SOCKET) { + } + explicit TCPSocket(SOCKET sockfd) : sockfd(sockfd) { + } + ~TCPSocket(void) { + // do nothing in destructor + // user need to take care of close + } + // default conversion to int + inline operator SOCKET() const { + return sockfd; + } + /*! + * \brief create the socket, call this before using socket + * \param af domain + */ + inline void Create(int af = PF_INET) { + sockfd = socket(PF_INET, SOCK_STREAM, 0); + if (sockfd == INVALID_SOCKET) { + SockError("Create"); + } + } + /*! + * \brief start up the socket module + * call this before using the sockets + */ + inline static void Startup(void) { +#ifdef _WIN32 + WSADATA wsa_data; + if (WSAStartup(MAKEWORD(2, 2), &wsa_data) != -1) { + SockError("Startup"); + } + if (LOBYTE(wsa_data.wVersion) != 2 || HIBYTE(wsa_data.wVersion) != 2) { + WSACleanup(); + utils::Error("Could not find a usable version of Winsock.dll\n"); + } +#endif + } + /*! + * \brief shutdown the socket module after use, all sockets need to be closed + */ + inline static void Finalize(void) { +#ifdef _WIN32 + WSACleanup(); +#endif + } + /*! + * \brief set this socket to use non-blocking mode + * \param non_block whether set it to be non-block, if it is false + * it will set it back to block mode + */ + inline void SetNonBlock(bool non_block) { +#ifdef _WIN32 + u_long mode = non_block ? 1 : 0; + if (ioctlsocket(sockfd, FIONBIO, &mode) != NO_ERROR) { + SockError("SetNonBlock"); + } +#else + int flag = fcntl(sockfd, F_GETFL, 0); + if (flag == -1) { + SockError("SetNonBlock-1"); + } + if (non_block) { + flag |= O_NONBLOCK; + } else { + flag &= ~O_NONBLOCK; + } + if (fcntl(sockfd, F_SETFL, flag) == -1) { + SockError("SetNonBlock-2"); + } +#endif + } + /*! + * \brief perform listen of the socket + * \param backlog backlog parameter + */ + inline void Listen(int backlog = 16) { + listen(sockfd, backlog); + } + /*! \brief get a new connection */ + TCPSocket Accept(void) { + SOCKET newfd = accept(sockfd, NULL, NULL); + if (newfd == INVALID_SOCKET) { + SockError("Accept"); + } + return TCPSocket(newfd); + } + /*! + * \brief bind the socket to an address + * \param addr + */ + inline void Bind(const SockAddr &addr) { + if (bind(sockfd, (sockaddr*)&addr.addr, sizeof(addr.addr)) == -1) { + SockError("Bind"); + } + } + /*! + * \brief try bind the socket to host, from start_port to end_port + * \param start_port starting port number to try + * \param end_port ending port number to try + * \param out_addr the binding address, if successful + * \return whether the binding is successful + */ + inline int TryBindHost(int start_port, int end_port) { + for (int port = start_port; port < end_port; ++port) { + SockAddr addr("0.0.0.0", port); + if (bind(sockfd, (sockaddr*)&addr.addr, sizeof(addr.addr)) == 0) { + return port; + } + if (errno != EADDRINUSE) { + SockError("TryBindHost"); + } + } + return -1; + } + /*! + * \brief connect to an address + * \param addr the address to connect to + */ + inline void Connect(const SockAddr &addr) { + if (connect(sockfd, (sockaddr*)&addr.addr, sizeof(addr.addr)) == -1) { + SockError("Connect"); + } + } + /*! \brief close the connection */ + inline void Close(void) { + if (sockfd != -1) { +#ifdef _WIN32 + closesocket(sockfd); +#else + close(sockfd); +#endif + sockfd = INVALID_SOCKET; + } else { + Error("TCPSocket::Close double close the socket or close without create"); + } + } + /*! + * \brief send data using the socket + * \param buf the pointer to the buffer + * \param len the size of the buffer + * \param flags extra flags + * \return size of data actually sent + */ + inline size_t Send(const void *buf_, size_t len, int flag = 0) { + const char *buf = reinterpret_cast(buf_); + if (len == 0) return 0; + ssize_t ret = send(sockfd, buf, static_cast(len), flag); + if (ret == -1) { + if (errno == EAGAIN || errno == EWOULDBLOCK) return 0; + SockError("Send"); + } + return ret; + } + /*! + * \brief receive data using the socket + * \param buf_ the pointer to the buffer + * \param len the size of the buffer + * \param flags extra flags + * \return size of data actually received + */ + inline size_t Recv(void *buf_, size_t len, int flags = 0) { + char *buf = reinterpret_cast(buf_); + if (len == 0) return 0; + ssize_t ret = recv(sockfd, buf, static_cast(len), flags); + if (ret == -1) { + if (errno == EAGAIN || errno == EWOULDBLOCK) return 0; + SockError("Recv"); + } + return ret; + } + /*! + * \brief peform block write that will attempt to send all data out + * can still return smaller than request when error occurs + * \param buf the pointer to the buffer + * \param len the size of the buffer + * \return size of data actually sent + */ + inline size_t SendAll(const void *buf_, size_t len) { + const char *buf = reinterpret_cast(buf_); + size_t ndone = 0; + while (ndone < len) { + ssize_t ret = send(sockfd, buf, static_cast(len - ndone), 0); + if (ret == -1) { + if (errno == EAGAIN || errno == EWOULDBLOCK) return ndone; + SockError("Recv"); + } + buf += ret; + ndone += ret; + } + return ndone; + } + /*! + * \brief peforma block read that will attempt to read all data + * can still return smaller than request when error occurs + * \param buf_ the buffer pointer + * \param len length of data to recv + * \return size of data actually sent + */ + inline size_t RecvAll(void *buf_, size_t len) { + char *buf = reinterpret_cast(buf_); + size_t ndone = 0; + while (ndone < len) { + ssize_t ret = recv(sockfd, buf, static_cast(len - ndone), MSG_WAITALL); + if (ret == -1) { + if (errno == EAGAIN || errno == EWOULDBLOCK) return ndone; + SockError("Recv"); + } + if (ret == 0) return ndone; + buf += ret; + ndone += ret; + } + return ndone; + } + + private: + // report an socket error + inline static void SockError(const char *msg) { + int errsv = errno; + Error("Socket %s Error:%s", msg, strerror(errsv)); + } +}; +/*! \brief helper data structure to perform select */ +struct SelectHelper { + public: + SelectHelper(void) { + this->Clear(); + } + /*! + * \brief add file descriptor to watch for read + * \param fd file descriptor to be watched + */ + inline void WatchRead(SOCKET fd) { + read_fds.push_back(fd); + if (fd > maxfd) maxfd = fd; + } + /*! + * \brief add file descriptor to watch for write + * \param fd file descriptor to be watched + */ + inline void WatchWrite(SOCKET fd) { + write_fds.push_back(fd); + if (fd > maxfd) maxfd = fd; + } + /*! + * \brief Check if the descriptor is ready for read + * \param fd file descriptor to check status + */ + inline bool CheckRead(SOCKET fd) const { + return FD_ISSET(fd, &read_set) != 0; + } + /*! + * \brief Check if the descriptor is ready for write + * \param fd file descriptor to check status + */ + inline bool CheckWrite(SOCKET fd) const { + return FD_ISSET(fd, &write_set) != 0; + } + /*! + * \brief clear all the monitored descriptors + */ + inline void Clear(void) { + read_fds.clear(); + write_fds.clear(); + maxfd = 0; + } + /*! + * \brief peform select on the set defined + * \param timeout specify timeout in micro-seconds(ms) if equals 0, means select will always block + * \return number of active descriptors selected + */ + inline int Select(long timeout = 0) { + FD_ZERO(&read_set); + FD_ZERO(&write_set); + for (size_t i = 0; i < read_fds.size(); ++i) { + FD_SET(read_fds[i], &read_set); + } + for (size_t i = 0; i < write_fds.size(); ++i) { + FD_SET(write_fds[i], &write_set); + } + int ret; + if (timeout == 0) { + ret = select(static_cast(maxfd + 1), &read_set, &write_set, NULL, NULL); + } else { + timeval tm; + tm.tv_usec = (timeout % 1000) * 1000; + tm.tv_sec = timeout / 1000; + ret = select(static_cast(maxfd + 1), &read_set, &write_set, NULL, &tm); + } + if (ret == -1) { + int errsv = errno; + Error("Select Error: %s", strerror(errsv)); + } + return ret; + } + + private: + SOCKET maxfd; + fd_set read_set, write_set; + std::vector read_fds, write_fds; +}; +} +} +#endif diff --git a/src/utils/thread.h b/src/utils/thread.h new file mode 100644 index 000000000000..830b21cbeb5a --- /dev/null +++ b/src/utils/thread.h @@ -0,0 +1,146 @@ +#ifndef XGBOOST_UTILS_THREAD_H +#define XGBOOST_UTILS_THREAD_H +/*! + * \file thread.h + * \brief this header include the minimum necessary resource for multi-threading + * \author Tianqi Chen + * Acknowledgement: this file is adapted from SVDFeature project, by same author. + * The MAC support part of this code is provided by Artemy Kolchinsky + */ +#ifdef _MSC_VER +#include "utils.h" +#include +#include +namespace xgboost { +namespace utils { +/*! \brief simple semaphore used for synchronization */ +class Semaphore { + public : + inline void Init(int init_val) { + sem = CreateSemaphore(NULL, init_val, 10, NULL); + utils::Assert(sem != NULL, "create Semaphore error"); + } + inline void Destroy(void) { + CloseHandle(sem); + } + inline void Wait(void) { + utils::Assert(WaitForSingleObject(sem, INFINITE) == WAIT_OBJECT_0, "WaitForSingleObject error"); + } + inline void Post(void) { + utils::Assert(ReleaseSemaphore(sem, 1, NULL) != 0, "ReleaseSemaphore error"); + } + private: + HANDLE sem; +}; +/*! \brief simple thread that wraps windows thread */ +class Thread { + private: + HANDLE thread_handle; + unsigned thread_id; + public: + inline void Start(unsigned int __stdcall entry(void*), void *param) { + thread_handle = (HANDLE)_beginthreadex(NULL, 0, entry, param, 0, &thread_id); + } + inline int Join(void) { + WaitForSingleObject(thread_handle, INFINITE); + return 0; + } +}; +/*! \brief exit function called from thread */ +inline void ThreadExit(void *status) { + _endthreadex(0); +} +#define XGBOOST_THREAD_PREFIX unsigned int __stdcall +} // namespace utils +} // namespace xgboost +#else +// thread interface using g++ +#include +#include +namespace xgboost { +namespace utils { +/*!\brief semaphore class */ +class Semaphore { + #ifdef __APPLE__ + private: + sem_t* semPtr; + char sema_name[20]; + private: + inline void GenRandomString(char *s, const int len) { + static const char alphanum[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" ; + for (int i = 0; i < len; ++i) { + s[i] = alphanum[rand() % (sizeof(alphanum) - 1)]; + } + s[len] = 0; + } + public: + inline void Init(int init_val) { + sema_name[0]='/'; + sema_name[1]='s'; + sema_name[2]='e'; + sema_name[3]='/'; + GenRandomString(&sema_name[4], 16); + if((semPtr = sem_open(sema_name, O_CREAT, 0644, init_val)) == SEM_FAILED) { + perror("sem_open"); + exit(1); + } + utils::Assert(semPtr != NULL, "create Semaphore error"); + } + inline void Destroy(void) { + if (sem_close(semPtr) == -1) { + perror("sem_close"); + exit(EXIT_FAILURE); + } + if (sem_unlink(sema_name) == -1) { + perror("sem_unlink"); + exit(EXIT_FAILURE); + } + } + inline void Wait(void) { + sem_wait(semPtr); + } + inline void Post(void) { + sem_post(semPtr); + } + #else + private: + sem_t sem; + public: + inline void Init(int init_val) { + sem_init(&sem, 0, init_val); + } + inline void Destroy(void) { + sem_destroy(&sem); + } + inline void Wait(void) { + sem_wait(&sem); + } + inline void Post(void) { + sem_post(&sem); + } + #endif +}; +/*!\brief simple thread class */ +class Thread { + private: + pthread_t thread; + public : + inline void Start(void * entry(void*), void *param) { + pthread_attr_t attr; + pthread_attr_init(&attr); + pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE); + pthread_create(&thread, &attr, entry, param); + } + inline int Join(void) { + void *status; + return pthread_join(thread, &status); + } +}; +inline void ThreadExit(void *status) { + pthread_exit(status); +} +} // namespace utils +} // namespace xgboost +#define XGBOOST_THREAD_PREFIX void * +#endif +#endif diff --git a/src/utils/thread_buffer.h b/src/utils/thread_buffer.h new file mode 100644 index 000000000000..ed36e1b43551 --- /dev/null +++ b/src/utils/thread_buffer.h @@ -0,0 +1,203 @@ +#ifndef XGBOOST_UTILS_THREAD_BUFFER_H_ +#define XGBOOST_UTILS_THREAD_BUFFER_H_ +/*! + * \file thread_buffer.h + * \brief multi-thread buffer, iterator, can be used to create parallel pipeline + * \author Tianqi Chen + */ +#include +#include +#include +#include "./utils.h" +#include "./thread.h" +namespace xgboost { +namespace utils { +/*! + * \brief buffered loading iterator that uses multithread + * this template method will assume the following paramters + * \tparam Elem elememt type to be buffered + * \tparam ElemFactory factory type to implement in order to use thread buffer + */ +template +class ThreadBuffer { + public: + /*!\brief constructor */ + ThreadBuffer(void) { + this->init_end = false; + this->buf_size = 30; + } + ~ThreadBuffer(void) { + if(init_end) this->Destroy(); + } + /*!\brief set parameter, will also pass the parameter to factory */ + inline void SetParam(const char *name, const char *val) { + if (!strcmp( name, "buffer_size")) buf_size = atoi(val); + factory.SetParam(name, val); + } + /*! + * \brief initalize the buffered iterator + * \param param a initialize parameter that will pass to factory, ignore it if not necessary + * \return false if the initlization can't be done, e.g. buffer file hasn't been created + */ + inline bool Init(void) { + if (!factory.Init()) return false; + for (int i = 0; i < buf_size; ++i) { + bufA.push_back(factory.Create()); + bufB.push_back(factory.Create()); + } + this->init_end = true; + this->StartLoader(); + return true; + } + /*!\brief place the iterator before first value */ + inline void BeforeFirst(void) { + // wait till last loader end + loading_end.Wait(); + // critcal zone + current_buf = 1; + factory.BeforeFirst(); + // reset terminate limit + endA = endB = buf_size; + // wake up loader for first part + loading_need.Post(); + // wait til first part is loaded + loading_end.Wait(); + // set current buf to right value + current_buf = 0; + // wake loader for next part + data_loaded = false; + loading_need.Post(); + // set buffer value + buf_index = 0; + } + /*! \brief destroy the buffer iterator, will deallocate the buffer */ + inline void Destroy(void) { + // wait until the signal is consumed + this->destroy_signal = true; + loading_need.Post(); + loader_thread.Join(); + loading_need.Destroy(); + loading_end.Destroy(); + for (size_t i = 0; i < bufA.size(); ++i) { + factory.FreeSpace(bufA[i]); + } + for (size_t i = 0; i < bufB.size(); ++i) { + factory.FreeSpace(bufB[i]); + } + bufA.clear(); bufB.clear(); + factory.Destroy(); + this->init_end = false; + } + /*! + * \brief get the next element needed in buffer + * \param elem element to store into + * \return whether reaches end of data + */ + inline bool Next(Elem &elem) { + // end of buffer try to switch + if (buf_index == buf_size) { + this->SwitchBuffer(); + buf_index = 0; + } + if (buf_index >= (current_buf ? endA : endB)) { + return false; + } + std::vector &buf = current_buf ? bufA : bufB; + elem = buf[buf_index]; + ++buf_index; + return true; + } + /*! + * \brief get the factory object + */ + inline ElemFactory &get_factory(void) { + return factory; + } + inline const ElemFactory &get_factory(void) const{ + return factory; + } + // size of buffer + int buf_size; + private: + // factory object used to load configures + ElemFactory factory; + // index in current buffer + int buf_index; + // indicate which one is current buffer + int current_buf; + // max limit of visit, also marks termination + int endA, endB; + // double buffer, one is accessed by loader + // the other is accessed by consumer + // buffer of the data + std::vector bufA, bufB; + // initialization end + bool init_end; + // singal whether the data is loaded + bool data_loaded; + // signal to kill the thread + bool destroy_signal; + // thread object + Thread loader_thread; + // signal of the buffer + Semaphore loading_end, loading_need; + /*! + * \brief slave thread + * this implementation is like producer-consumer style + */ + inline void RunLoader(void) { + while(!destroy_signal) { + // sleep until loading is needed + loading_need.Wait(); + std::vector &buf = current_buf ? bufB : bufA; + int i; + for (i = 0; i < buf_size ; ++i) { + if (!factory.LoadNext(buf[i])) { + int &end = current_buf ? endB : endA; + end = i; // marks the termination + break; + } + } + // signal that loading is done + data_loaded = true; + loading_end.Post(); + } + } + /*!\brief entry point of loader thread */ + inline static XGBOOST_THREAD_PREFIX LoaderEntry(void *pthread) { + static_cast< ThreadBuffer* >(pthread)->RunLoader(); + ThreadExit(NULL); + return NULL; + } + /*!\brief start loader thread */ + inline void StartLoader(void) { + destroy_signal = false; + // set param + current_buf = 1; + loading_need.Init(1); + loading_end .Init(0); + // reset terminate limit + endA = endB = buf_size; + loader_thread.Start(LoaderEntry, this); + // wait until first part of data is loaded + loading_end.Wait(); + // set current buf to right value + current_buf = 0; + // wake loader for next part + data_loaded = false; + loading_need.Post(); + buf_index = 0; + } + /*!\brief switch double buffer */ + inline void SwitchBuffer(void) { + loading_end.Wait(); + // loader shall be sleep now, critcal zone! + current_buf = !current_buf; + // wake up loader + data_loaded = false; + loading_need.Post(); + } +}; +} // namespace utils +} // namespace xgboost +#endif diff --git a/src/xgboost_main.cpp b/src/xgboost_main.cpp index 75544dd0ec41..94e6d6bc181b 100644 --- a/src/xgboost_main.cpp +++ b/src/xgboost_main.cpp @@ -1,9 +1,10 @@ #define _CRT_SECURE_NO_WARNINGS #define _CRT_SECURE_NO_DEPRECATE - +#define NOMINMAX #include #include #include +#include #include "io/io.h" #include "utils/utils.h" #include "utils/config.h" @@ -13,13 +14,13 @@ namespace xgboost { /*! * \brief wrapping the training process */ -class BoostLearnTask{ +class BoostLearnTask { public: inline int Run(int argc, char *argv[]) { if (argc < 2) { printf("Usage: \n"); return 0; - } + } utils::ConfigIterator itr(argv[1]); while (itr.Next()) { this->SetParam(itr.name(), itr.val()); @@ -30,8 +31,36 @@ class BoostLearnTask{ this->SetParam(name, val); } } + // do not save anything when save to stdout + if (model_out == "stdout" || name_pred == "stdout") { + this->SetParam("silent", "1"); + save_period = 0; + } + // whether need data rank + bool need_data_rank = strchr(train_path.c_str(), '%') != NULL; + // if need data rank in loading, initialize rabit engine before load data + // otherwise, initialize rabit engine after loading data + // lazy initialization of rabit engine can be helpful in speculative execution + if (need_data_rank) rabit::Init(argc, argv); this->InitData(); - this->InitLearner(); + if (!need_data_rank) rabit::Init(argc, argv); + if (rabit::IsDistributed()) { + std::string pname = rabit::GetProcessorName(); + fprintf(stderr, "start %s:%d\n", pname.c_str(), rabit::GetRank()); + } + if (rabit::IsDistributed() && data_split == "NONE") { + this->SetParam("dsplit", "row"); + } + if (rabit::GetRank() != 0) { + this->SetParam("silent", "2"); + } + if (task == "train") { + // if task is training, will try recover from checkpoint + this->TaskTrain(); + return 0; + } else { + this->InitLearner(); + } if (task == "dump") { this->TaskDump(); return 0; } @@ -40,8 +69,6 @@ class BoostLearnTask{ } if (task == "pred") { this->TaskPred(); - } else { - this->TaskTrain(); } return 0; } @@ -62,6 +89,7 @@ class BoostLearnTask{ if (!strcmp("fmap", name)) name_fmap = val; if (!strcmp("name_dump", name)) name_dump = val; if (!strcmp("name_pred", name)) name_pred = val; + if (!strcmp("dsplit", name)) data_split = val; if (!strcmp("dump_stats", name)) dump_model_stats = atoi(val); if (!strncmp("eval[", name, 5)) { char evname[256]; @@ -89,6 +117,8 @@ class BoostLearnTask{ name_pred = "pred.txt"; name_dump = "dump.txt"; model_dir_path = "./"; + data_split = "NONE"; + load_part = 0; data = NULL; } ~BoostLearnTask(void){ @@ -99,13 +129,20 @@ class BoostLearnTask{ } private: inline void InitData(void) { + if (strchr(train_path.c_str(), '%') != NULL) { + char s_tmp[256]; + utils::SPrintf(s_tmp, sizeof(s_tmp), train_path.c_str(), rabit::GetRank()); + train_path = s_tmp; + load_part = 1; + } + if (name_fmap != "NULL") fmap.LoadText(name_fmap.c_str()); if (task == "dump") return; if (task == "pred") { data = io::LoadDataMatrix(test_path.c_str(), silent != 0, use_buffer != 0); } else { // training - data = io::LoadDataMatrix(train_path.c_str(), silent != 0, use_buffer != 0); + data = io::LoadDataMatrix(train_path.c_str(), silent != 0 && load_part == 0, use_buffer != 0); utils::Assert(eval_data_names.size() == eval_data_paths.size(), "BUG"); for (size_t i = 0; i < eval_data_names.size(); ++i) { deval.push_back(io::LoadDataMatrix(eval_data_paths[i].c_str(), silent != 0, use_buffer != 0)); @@ -120,35 +157,61 @@ class BoostLearnTask{ learner.SetCacheData(dcache); // add training set to evaluation set if needed - if( eval_train != 0 ) { + if (eval_train != 0) { devalall.push_back(data); eval_data_names.push_back(std::string("train")); } } } inline void InitLearner(void) { - if (model_in != "NULL"){ - utils::FileStream fi(utils::FopenCheck(model_in.c_str(), "rb")); - learner.LoadModel(fi); - fi.Close(); + if (model_in != "NULL") { + learner.LoadModel(model_in.c_str()); } else { utils::Assert(task == "train", "model_in not specified"); learner.InitModel(); } } inline void TaskTrain(void) { + int version = rabit::LoadCheckPoint(&learner); + if (version == 0) this->InitLearner(); const time_t start = time(NULL); unsigned long elapsed = 0; learner.CheckInit(data); - for (int i = 0; i < num_round; ++i) { + + bool allow_lazy = learner.AllowLazyCheckPoint(); + for (int i = version / 2; i < num_round; ++i) { elapsed = (unsigned long)(time(NULL) - start); - if (!silent) printf("boosting round %d, %lu sec elapsed\n", i, elapsed); - learner.UpdateOneIter(i, *data); + if (version % 2 == 0) { + if (!silent) printf("boosting round %d, %lu sec elapsed\n", i, elapsed); + learner.UpdateOneIter(i, *data); + if (allow_lazy) { + rabit::LazyCheckPoint(&learner); + } else { + rabit::CheckPoint(&learner); + } + version += 1; + } + utils::Assert(version == rabit::VersionNumber(), "consistent check"); std::string res = learner.EvalOneIter(i, devalall, eval_data_names); - fprintf(stderr, "%s\n", res.c_str()); + if (rabit::IsDistributed()){ + if (rabit::GetRank() == 0) { + rabit::TrackerPrintf("%s\n", res.c_str()); + } + } else { + if (silent < 2) { + fprintf(stderr, "%s\n", res.c_str()); + } + } if (save_period != 0 && (i + 1) % save_period == 0) { this->SaveModel(i); } + if (allow_lazy) { + rabit::LazyCheckPoint(&learner); + } else { + rabit::CheckPoint(&learner); + } + version += 1; + utils::Assert(version == rabit::VersionNumber(), "consistent check"); elapsed = (unsigned long)(time(NULL) - start); } // always save final round @@ -176,9 +239,8 @@ class BoostLearnTask{ fclose(fo); } inline void SaveModel(const char *fname) const { - utils::FileStream fo(utils::FopenCheck(fname, "wb")); - learner.SaveModel(fo); - fo.Close(); + if (rabit::GetRank() != 0) return; + learner.SaveModel(fname); } inline void SaveModel(int i) const { char fname[256]; @@ -189,16 +251,23 @@ class BoostLearnTask{ std::vector preds; if (!silent) printf("start prediction...\n"); learner.Predict(*data, pred_margin != 0, &preds, ntree_limit); - if (!silent) printf("writing prediction to %s\n", name_pred.c_str()); - FILE *fo = utils::FopenCheck(name_pred.c_str(), "w"); - for (size_t i = 0; i < preds.size(); i++) { - fprintf(fo, "%f\n", preds[i]); + if (!silent) printf("writing prediction to %s\n", name_pred.c_str()); + FILE *fo; + if (name_pred != "stdout") { + fo = utils::FopenCheck(name_pred.c_str(), "w"); + } else { + fo = stdout; } - fclose(fo); + for (size_t i = 0; i < preds.size(); ++i) { + fprintf(fo, "%g\n", preds[i]); + } + if (fo != stdout) fclose(fo); } private: /*! \brief whether silent */ int silent; + /*! \brief special load */ + int load_part; /*! \brief whether use auto binary buffer */ int use_buffer; /*! \brief whether evaluate training statistics */ @@ -219,6 +288,8 @@ class BoostLearnTask{ std::string task; /*! \brief name of predict file */ std::string name_pred; + /*! \brief data split mode */ + std::string data_split; /*!\brief limit number of trees in prediction */ int ntree_limit; /*!\brief whether to directly output margin value */ @@ -243,7 +314,9 @@ class BoostLearnTask{ } int main(int argc, char *argv[]){ - xgboost::random::Seed(0); xgboost::BoostLearnTask tsk; - return tsk.Run(argc, argv); + tsk.SetParam("seed", "0"); + int ret = tsk.Run(argc, argv); + rabit::Finalize(); + return ret; } diff --git a/test/Makefile b/test/Makefile new file mode 100644 index 000000000000..a702d073fa1c --- /dev/null +++ b/test/Makefile @@ -0,0 +1,35 @@ +export CC = gcc +export CXX = g++ +export MPICXX = mpicxx +export LDFLAGS= -pthread -lm +export CFLAGS = -Wall -O3 -msse2 -Wno-unknown-pragmas -fPIC -I../src + +ifeq ($(no_omp),1) + CFLAGS += -DDISABLE_OPENMP +else + CFLAGS += -fopenmp +endif + +# specify tensor path +BIN = test_group_data test_quantile test_allreduce +OBJ = sync_tcp.o +.PHONY: clean all + +all: $(BIN) $(MPIBIN) + +sync_tcp.o: ../src/sync/sync_tcp.cpp ../src/utils/*.h + +test_group_data: test_group_data.cpp ../src/utils/*.h +test_quantile: test_quantile.cpp ../src/utils/*.h +test_allreduce: test_allreduce.cpp ../src/utils/*.h ../src/sync/sync.h sync_tcp.o +$(BIN) : + $(CXX) $(CFLAGS) $(LDFLAGS) -o $@ $(filter %.cpp %.o %.c, $^) + +$(OBJ) : + $(CXX) -c $(CFLAGS) -o $@ $(firstword $(filter %.cpp %.c, $^) ) + +$(MPIBIN) : + $(MPICXX) $(CFLAGS) $(LDFLAGS) -o $@ $(filter %.cpp %.o %.c, $^) + +clean: + $(RM) $(BIN) $(MPIBIN) *~ diff --git a/test/mkquantest.py b/test/mkquantest.py new file mode 100755 index 000000000000..70c467b462e4 --- /dev/null +++ b/test/mkquantest.py @@ -0,0 +1,42 @@ +#!/usr/bin/python +import math +import sys +import random +import subprocess + +funcs = { + 'seq': 'lambda n: sorted([(x,1) for x in range(1,n+1)], key = lambda x:random.random())', + 'seqlogw': 'lambda n: sorted([(x, math.log(x)) for x in range(1,n+1)], key = lambda x:random.random())', + 'lots0': 'lambda n: sorted([(max(x - n*3/4,0), 1) for x in range(1,n+1)], key = lambda x:random.random())', + 'lots9': 'lambda n: sorted([(9 if x > n / 4 else x, 1) for x in range(1,n+1)], key = lambda x:random.random())', + 'lotsm': 'lambda n: sorted([(n/8 if x > n / 4 else x, 1) for x in range(1,n+1)], key = lambda x:random.random())', + 'lotsmr': 'lambda n: sorted([( x * 4 / n + n / 20 if x > n / 10 else x, 1) for x in range(1,n+1)], key = lambda x:random.random())', + 'lotsmr2': 'lambda n: sorted([( x * 10 / n + n / 20 if x > n / 10 else x, 1) for x in range(1,n+1)], key = lambda x:random.random())' +} + +if len(sys.argv) < 3: + print 'Usage: python mkquantest.py [generate-type] [ndata]|./test_quantile [solver]' + print 'test_quantile need to be compiled, solver can be gk(GK nonweight version), wq(weighted version), wx(weighthed version, with prune optimized for heavy hitter)' + print 'Possible generate-types:' + for k, v in funcs.items(): + print '\t%s: %s' % (k, v) + print 'Example: ./mkquantest.py 50000 0.3 lotsmr |./test_quantile wq' + exit(-1) +random.seed(0) +maxn = int(sys.argv[1]) +eps = float(sys.argv[2]) +if len(sys.argv) > 3: + method = sys.argv[3] + assert method in funcs, ('cannot find method %s' % method) +else: + method = 'seq' +if len(sys.argv) > 4: + ndata = int(sys.argv[4]) + assert ndata <= maxn, 'ndata must be smaller than maxn' +else: + ndata = maxn + +fo = sys.stdout +fo.write('%d\t%g\n' % (maxn, eps)) +for x, w in eval(funcs[method])(ndata): + fo.write(str(x)+'\t'+str(w)+'\n') diff --git a/test/test_allreduce.cpp b/test/test_allreduce.cpp new file mode 100644 index 000000000000..4a47d7f559bd --- /dev/null +++ b/test/test_allreduce.cpp @@ -0,0 +1,124 @@ +#include +#include +#include +#include +#include + +using namespace xgboost; + +inline void TestMax(size_t n) { + int rank = sync::GetRank(); + int nproc = sync::GetWorldSize(); + + std::vector ndata(n); + for (size_t i = 0; i < ndata.size(); ++i) { + ndata[i] = (i * (rank+1)) % 111; + } + sync::AllReduce(&ndata[0], ndata.size(), sync::kMax); + for (size_t i = 0; i < ndata.size(); ++i) { + float rmax = (i * 1) % 111; + for (int r = 0; r < nproc; ++r) { + rmax = std::max(rmax, (float)((i * (r+1)) % 111)); + } + utils::Check(rmax == ndata[i], "[%d] TestMax check failure", rank); + } +} + +inline void TestSum(size_t n) { + int rank = sync::GetRank(); + int nproc = sync::GetWorldSize(); + const int z = 131; + + std::vector ndata(n); + for (size_t i = 0; i < ndata.size(); ++i) { + ndata[i] = (i * (rank+1)) % z; + } + sync::AllReduce(&ndata[0], ndata.size(), sync::kSum); + for (size_t i = 0; i < ndata.size(); ++i) { + float rsum = 0.0f; + for (int r = 0; r < nproc; ++r) { + rsum += (float)((i * (r+1)) % z); + } + utils::Check(fabsf(rsum - ndata[i]) < 1e-5 , + "[%d] TestSum check failure, local=%g, allreduce=%g", rank, rsum, ndata[i]); + } +} + +struct Rec { + double rmax; + double rmin; + double rsum; + Rec() {} + Rec(double r) { + rmax = rmin = rsum = r; + } + inline void Reduce(const Rec &b) { + rmax = std::max(b.rmax, rmax); + rmin = std::max(b.rmin, rmin); + rsum += b.rsum; + } + inline void CheckSameAs(const Rec &b) { + if (rmax != b.rmax || rmin != b.rmin || fabs(rsum - b.rsum) > 1e-6) { + utils::Error("[%d] TestReducer check failure", sync::GetRank()); + } + } +}; + +inline void TestReducer(int n) { + int rank = sync::GetRank(); + int nproc = sync::GetWorldSize(); + const int z = 131; + sync::Reducer red; + std::vector ndata(n); + for (size_t i = 0; i < ndata.size(); ++i) { + ndata[i] = Rec((i * (rank+1)) % z); + } + red.AllReduce(&ndata[0], ndata.size()); + + for (size_t i = 0; i < ndata.size(); ++i) { + Rec rec((i * 1) % z); + for (int r = 1; r < nproc; ++r) { + rec.Reduce(Rec((i * (r+1)) % z)); + } + rec.CheckSameAs(ndata[i]); + } +} + + +inline void TestBcast(size_t n, int root) { + int rank = sync::GetRank(); + std::string s; s.resize(n); + for (size_t i = 0; i < n; ++i) { + s[i] = char(i % 126 + 1); + } + std::string res; + if (root == rank) { + res = s; + sync::Bcast(&res, root); + } else { + sync::Bcast(&res, root); + } + utils::Check(res == s, "[%d] TestBcast fail", rank); +} + +int main(int argc, char *argv[]) { + if (argc < 2) { + printf("Usage: \n"); + return 0; + } + int n = atoi(argv[1]); + sync::Init(argc, argv); + int rank = sync::GetRank(); + //int nproc = sync::GetWorldSize(); + std::string name = sync::GetProcessorName(); + printf("[%d] start at %s\n", rank, name.c_str()); + TestMax(n); + printf("[%d] TestMax pass\n", rank); + TestSum(n); + printf("[%d] TestSum pass\n", rank); + TestReducer(n); + printf("[%d] TestReducer pass\n", rank); + sync::Finalize(); + printf("[%d] all check pass\n", rank); + return 0; +} diff --git a/test/test_group_data.cpp b/test/test_group_data.cpp new file mode 100644 index 000000000000..676d45e27bf0 --- /dev/null +++ b/test/test_group_data.cpp @@ -0,0 +1,84 @@ +#include +#include +#include +#include +#include +#include +#include +#include +#include + +using namespace xgboost::utils; +using namespace xgboost; + +int main(int argc, char *argv[]) { + if (argc < 3) { + printf("Usage: pnthread]\n"); + return 0; + } + if (argc > 3) { + omp_set_num_threads(atoi(argv[3])); + } + random::Seed(0); + unsigned nkey = static_cast(atoi(argv[1])); + size_t ndata = static_cast(atol(argv[2])); + + std::vector keys; + std::vector< std::pair > raw; + raw.reserve(ndata); keys.reserve(ndata); + for (size_t i = 0; i < ndata; ++i) { + unsigned key = random::NextUInt32(nkey); + utils::Check(key < nkey, "key exceed bound\n"); + raw.push_back(std::make_pair(key, i)); + keys.push_back(key); + } + printf("loading finish, start working\n"); + time_t start_t = time(NULL); + int nthread; + #pragma omp parallel + { + nthread = omp_get_num_threads(); + } + std::vector rptr; + std::vector data; + ParallelGroupBuilder builder(&rptr, &data); + builder.InitBudget(0, nthread); + + size_t nstep = (raw.size() +nthread-1)/ nthread; + #pragma omp parallel + { + int tid = omp_get_thread_num(); + size_t begin = tid * nstep; + size_t end = std::min((tid + 1) * nstep, raw.size()); + for (size_t i = begin; i < end; ++i) { + builder.AddBudget(raw[i].first, tid); + } + } + double first_cost = time(NULL) - start_t; + builder.InitStorage(); + + #pragma omp parallel + { + int tid = omp_get_thread_num(); + size_t begin = tid * nstep; + size_t end = std::min((tid + 1)* nstep, raw.size()); + for (size_t i = begin; i < end; ++i) { + builder.Push(raw[i].first, raw[i].second, tid); + } + } + + double second_cost = time(NULL) - start_t; + printf("all finish, phase1=%g sec, phase2=%g sec\n", first_cost, second_cost); + Check(rptr.size() <= nkey+1, "nkey exceed bound"); + Check(rptr.back() == ndata, "data shape inconsistent"); + for (size_t i = 0; i < rptr.size()-1; ++ i) { + Check(rptr[i] <= rptr[i+1], "rptr error"); + for (size_t j = rptr[i]; j < rptr[i+1]; ++j) { + unsigned pos = data[j]; + Check(pos < keys.size(), "invalid pos"); + Check(keys[pos] == i, "invalid key entry"); + } + } + printf("all check pass\n"); + return 0; +} diff --git a/test/test_quantile.cpp b/test/test_quantile.cpp new file mode 100644 index 000000000000..c1b85668dc7f --- /dev/null +++ b/test/test_quantile.cpp @@ -0,0 +1,92 @@ +#include +#include +#include +using namespace xgboost; + + +struct Entry { + double x, w, rmin; + inline bool operator<(const Entry &e) const { + return x < e.x; + } +}; + +inline void MakeQuantile(std::vector &dat) { + std::sort(dat.begin(), dat.end()); + size_t top = 0; + double wsum = 0.0; + for (size_t i = 0; i < dat.size();) { + size_t j = i + 1; + for (;j < dat.size() && dat[i].x == dat[j].x; ++j) { + dat[i].w += dat[j].w; + } + dat[top] = dat[i]; + dat[top].rmin = wsum; + wsum += dat[top].w; + ++top; + i = j; + } + dat.resize(top); +} + +template +inline void verifyWQ(std::vector &dat, Summary out) { + MakeQuantile(dat); + size_t j = 0; + double err = 0.0; + const double eps = 1e-4; + for (size_t i = 0; i < out.size; ++i) { + while (j < dat.size() && dat[j].x < out.data[i].value) ++j; + utils::Assert(j < dat.size() && fabs(dat[j].x - out.data[i].value) < eps, "bug"); + err = std::min(dat[j].rmin - out.data[i].rmin, err); + err = std::min(out.data[i].rmax - dat[j].rmin + dat[j].w, err); + err = std::min(dat[j].w - out.data[i].wmin, err); + } + if (err < 0.0) err = -err; + printf("verify correctness, max-constraint-violation=%g (0 means perfect, coubld be nonzero due to floating point)\n", err); +} + +template +inline typename Sketch::SummaryContainer test(std::vector &dat) { + Sketch sketch; + size_t n; + double wsum = 0.0; + float eps; + utils::Check(scanf("%lu%f", &n, &eps) == 2, "needs to start with n eps"); + sketch.Init(n, eps); + Entry e; + while (scanf("%lf%lf", &e.x, &e.w) == 2) { + dat.push_back(e); + wsum += e.w; + } + clock_t start = clock(); + for (size_t i = 0; i < dat.size(); ++i) { + sketch.Push(dat[i].x, dat[i].w); + } + double tcost = static_cast(clock() - start) / CLOCKS_PER_SEC; + typename Sketch::SummaryContainer out; + sketch.GetSummary(&out); + double maxerr = static_cast(out.MaxError()); + out.Print(); + printf("-------------------------\n"); + printf("timecost=%g sec\n", tcost); + printf("MaxError=%g/%g = %g\n", maxerr, wsum, maxerr / wsum); + printf("maxlevel = %lu, usedlevel=%lu, limit_size=%lu\n", sketch.nlevel, sketch.level.size(), sketch.limit_size); + return out; +} + +int main(int argc, char *argv[]) { + const char *method = "wq"; + if (argc > 1) method = argv[1]; + std::vector dat; + if (!strcmp(method, "wq")) { + verifyWQ(dat, test, float>(dat)); + } + if (!strcmp(method, "wx")) { + verifyWQ(dat, test, float>(dat)); + } + if (!strcmp(method, "gk")) { + test, unsigned>(dat); + } + return 0; +} diff --git a/windows/xgboost/xgboost.vcxproj b/windows/xgboost/xgboost.vcxproj index 3d303efc4da8..ab6216d74fe9 100644 --- a/windows/xgboost/xgboost.vcxproj +++ b/windows/xgboost/xgboost.vcxproj @@ -21,6 +21,7 @@ + @@ -111,6 +112,7 @@ true true true + ws2_32.lib;%(AdditionalDependencies) diff --git a/windows/xgboost_wrapper/xgboost_wrapper.vcxproj b/windows/xgboost_wrapper/xgboost_wrapper.vcxproj index 6c73e3ceed36..e78c6e38c42c 100644 --- a/windows/xgboost_wrapper/xgboost_wrapper.vcxproj +++ b/windows/xgboost_wrapper/xgboost_wrapper.vcxproj @@ -21,6 +21,7 @@ + @@ -112,6 +113,7 @@ true true true + ws2_32.lib;%(AdditionalDependencies) diff --git a/wrapper/xgboost.py b/wrapper/xgboost.py index 8d87dc8db608..1a2a4e1c2bda 100644 --- a/wrapper/xgboost.py +++ b/wrapper/xgboost.py @@ -34,7 +34,6 @@ xglib.XGBoosterEvalOneIter.restype = ctypes.c_char_p xglib.XGBoosterDumpModel.restype = ctypes.POINTER(ctypes.c_char_p) - def ctypes2numpy(cptr, length, dtype): """convert a ctypes pointer array to numpy array """ assert isinstance(cptr, ctypes.POINTER(ctypes.c_float)) @@ -304,6 +303,7 @@ def boost(self, dtrain, grad, hess): (ctypes.c_float*len(grad))(*grad), (ctypes.c_float*len(hess))(*hess), len(grad)) + def eval_set(self, evals, it = 0, feval = None): """evaluates by metric Args: @@ -332,24 +332,38 @@ def eval_set(self, evals, it = 0, feval = None): return res def eval(self, mat, name = 'eval', it = 0): return self.eval_set( [(mat,name)], it) - def predict(self, data, output_margin=False, ntree_limit=0): + def predict(self, data, output_margin=False, ntree_limit=0, pred_leaf=False): """ predict with data Args: data: DMatrix - the dmatrix storing the input + the dmatrix storing the input output_margin: bool - whether output raw margin value that is untransformed - + whether output raw margin value that is untransformed ntree_limit: int - limit number of trees in prediction, default to 0, 0 means using all the trees + limit number of trees in prediction, default to 0, 0 means using all the trees + pred_leaf: bool + when this option is on, the output will be a matrix of (nsample, ntrees) + with each record indicate the predicted leaf index of each sample in each tree + Note that the leaf index of tree is unique per tree, so you may find leaf 1 in both tree 1 and tree 0 Returns: numpy array of prediction """ + option_mask = 0 + if output_margin: + option_mask += 1 + if pred_leaf: + option_mask += 2 length = ctypes.c_ulong() preds = xglib.XGBoosterPredict(self.handle, data.handle, - int(output_margin), ntree_limit, ctypes.byref(length)) - return ctypes2numpy(preds, length.value, 'float32') + option_mask, ntree_limit, ctypes.byref(length)) + preds = ctypes2numpy(preds, length.value, 'float32') + if pred_leaf: + preds = preds.astype('int32') + nrow = data.num_row() + if preds.size != nrow and preds.size % nrow == 0: + preds = preds.reshape(nrow, preds.size / nrow) + return preds def save_model(self, fname): """ save model to file Args: @@ -542,3 +556,4 @@ def cv(params, dtrain, num_boost_round = 10, nfold=3, metrics=[], \ sys.stderr.write(res+'\n') results.append(res) return results + diff --git a/wrapper/xgboost_wrapper.cpp b/wrapper/xgboost_wrapper.cpp index 5ceb5d79de67..d744c3e22c29 100644 --- a/wrapper/xgboost_wrapper.cpp +++ b/wrapper/xgboost_wrapper.cpp @@ -8,7 +8,9 @@ #include // include all std functions using namespace std; - +#ifdef _MSC_VER +#define isnan(x) (_isnan(x) != 0) +#endif #include "./xgboost_wrapper.h" #include "../src/data.h" #include "../src/learner/learner-inl.hpp" @@ -30,9 +32,9 @@ class Booster: public learner::BoostLearner { this->init_model = false; this->SetCacheData(mats); } - inline const float *Pred(const DataMatrix &dmat, int output_margin, unsigned ntree_limit, bst_ulong *len) { + inline const float *Pred(const DataMatrix &dmat, int option_mask, unsigned ntree_limit, bst_ulong *len) { this->CheckInitModel(); - this->Predict(dmat, output_margin != 0, &this->preds_, ntree_limit); + this->Predict(dmat, (option_mask&1) != 0, &this->preds_, ntree_limit, (option_mask&2) != 0); *len = static_cast(this->preds_.size()); return BeginPtr(this->preds_); } @@ -44,7 +46,7 @@ class Booster: public learner::BoostLearner { for (bst_omp_uint j = 0; j < ndata; ++j) { gpair_[j] = bst_gpair(grad[j], hess[j]); } - gbm_->DoBoost(train.fmat(), train.info.info, &gpair_); + gbm_->DoBoost(train.fmat(), this->FindBufferOffset(train), train.info.info, &gpair_); } inline void CheckInitModel(void) { if (!init_model) { @@ -132,7 +134,7 @@ extern "C"{ bst_ulong nrow, bst_ulong ncol, float missing) { - bool nan_missing = std::isnan(missing); + bool nan_missing = isnan(missing); DMatrixSimple *p_mat = new DMatrixSimple(); DMatrixSimple &mat = *p_mat; mat.info.info.num_row = nrow; @@ -140,7 +142,7 @@ extern "C"{ for (bst_ulong i = 0; i < nrow; ++i, data += ncol) { bst_ulong nelem = 0; for (bst_ulong j = 0; j < ncol; ++j) { - if (std::isnan(data[j])) { + if (isnan(data[j])) { utils::Check(nan_missing, "There are NAN in the matrix, however, you did not set missing=NAN"); } else { if (nan_missing || data[j] != missing) { @@ -284,8 +286,8 @@ extern "C"{ bst->eval_str = bst->EvalOneIter(iter, mats, names); return bst->eval_str.c_str(); } - const float *XGBoosterPredict(void *handle, void *dmat, int output_margin, unsigned ntree_limit, bst_ulong *len) { - return static_cast(handle)->Pred(*static_cast(dmat), output_margin, ntree_limit, len); + const float *XGBoosterPredict(void *handle, void *dmat, int option_mask, unsigned ntree_limit, bst_ulong *len) { + return static_cast(handle)->Pred(*static_cast(dmat), option_mask, ntree_limit, len); } void XGBoosterLoadModel(void *handle, const char *fname) { static_cast(handle)->LoadModel(fname); diff --git a/wrapper/xgboost_wrapper.h b/wrapper/xgboost_wrapper.h index 2555289b1dbe..82fedb9d6db2 100644 --- a/wrapper/xgboost_wrapper.h +++ b/wrapper/xgboost_wrapper.h @@ -17,6 +17,28 @@ typedef unsigned long bst_ulong; #ifdef __cplusplus extern "C" { #endif + /*! + * \brief initialize sync module, this is needed if used in distributed model + * normally, argv need to contain master_uri and master_port + * if start using submit_job_tcp script, then pass args to this will do + * \param argc number of arguments + * \param argv the arguments to be passed in sync module + */ + XGB_DLL void XGSyncInit(int argc, char *argv[]); + /*! + * \brief finalize sync module, call this when everything is done + */ + XGB_DLL void XGSyncFinalize(void); + /*! + * \brief get the rank + * \return return the rank of + */ + XGB_DLL int XGSyncGetRank(void); + /*! + * \brief get the world size from sync + * \return return the number of distributed job ran in the group + */ + XGB_DLL int XGSyncGetWorldSize(void); /*! * \brief load a data matrix * \return a loaded data matrix @@ -41,7 +63,7 @@ extern "C" { * \param col_ptr pointer to col headers * \param indices findex * \param data fvalue - * \param nindptr number of rows in the matix + 1 + * \param nindptr number of rows in the matix + 1 * \param nelem number of nonzero elements in the matrix * \return created dmatrix */ @@ -178,12 +200,18 @@ extern "C" { * \brief make prediction based on dmat * \param handle handle * \param dmat data matrix - * \param output_margin whether only output raw margin value + * \param option_mask bit-mask of options taken in prediction, possible values + * 0:normal prediction + * 1:output margin instead of transformed value + * 2:output leaf index of trees instead of leaf value, note leaf index is unique per tree * \param ntree_limit limit number of trees used for prediction, this is only valid for boosted trees * when the parameter is set to 0, we will use all the trees * \param len used to store length of returning result */ - XGB_DLL const float *XGBoosterPredict(void *handle, void *dmat, int output_margin, unsigned ntree_limit, bst_ulong *len); + XGB_DLL const float *XGBoosterPredict(void *handle, void *dmat, + int option_mask, + unsigned ntree_limit, + bst_ulong *len); /*! * \brief load model from existing file * \param handle handle