pax_global_header 0000666 0000000 0000000 00000000064 11645337641 0014524 g ustar 00root root 0000000 0000000 52 comment=fa6bed74ec94d2f776947723e6587595ee843a50
sgd-2.0/ 0000775 0000000 0000000 00000000000 11645337641 0012142 5 ustar 00root root 0000000 0000000 sgd-2.0/COPYING 0000664 0000000 0000000 00000016727 11645337641 0013212 0 ustar 00root root 0000000 0000000 GNU LESSER GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc.
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
This version of the GNU Lesser General Public License incorporates
the terms and conditions of version 3 of the GNU General Public
License, supplemented by the additional permissions listed below.
0. Additional Definitions.
As used herein, "this License" refers to version 3 of the GNU Lesser
General Public License, and the "GNU GPL" refers to version 3 of the GNU
General Public License.
"The Library" refers to a covered work governed by this License,
other than an Application or a Combined Work as defined below.
An "Application" is any work that makes use of an interface provided
by the Library, but which is not otherwise based on the Library.
Defining a subclass of a class defined by the Library is deemed a mode
of using an interface provided by the Library.
A "Combined Work" is a work produced by combining or linking an
Application with the Library. The particular version of the Library
with which the Combined Work was made is also called the "Linked
Version".
The "Minimal Corresponding Source" for a Combined Work means the
Corresponding Source for the Combined Work, excluding any source code
for portions of the Combined Work that, considered in isolation, are
based on the Application, and not on the Linked Version.
The "Corresponding Application Code" for a Combined Work means the
object code and/or source code for the Application, including any data
and utility programs needed for reproducing the Combined Work from the
Application, but excluding the System Libraries of the Combined Work.
1. Exception to Section 3 of the GNU GPL.
You may convey a covered work under sections 3 and 4 of this License
without being bound by section 3 of the GNU GPL.
2. Conveying Modified Versions.
If you modify a copy of the Library, and, in your modifications, a
facility refers to a function or data to be supplied by an Application
that uses the facility (other than as an argument passed when the
facility is invoked), then you may convey a copy of the modified
version:
a) under this License, provided that you make a good faith effort to
ensure that, in the event an Application does not supply the
function or data, the facility still operates, and performs
whatever part of its purpose remains meaningful, or
b) under the GNU GPL, with none of the additional permissions of
this License applicable to that copy.
3. Object Code Incorporating Material from Library Header Files.
The object code form of an Application may incorporate material from
a header file that is part of the Library. You may convey such object
code under terms of your choice, provided that, if the incorporated
material is not limited to numerical parameters, data structure
layouts and accessors, or small macros, inline functions and templates
(ten or fewer lines in length), you do both of the following:
a) Give prominent notice with each copy of the object code that the
Library is used in it and that the Library and its use are
covered by this License.
b) Accompany the object code with a copy of the GNU GPL and this license
document.
4. Combined Works.
You may convey a Combined Work under terms of your choice that,
taken together, effectively do not restrict modification of the
portions of the Library contained in the Combined Work and reverse
engineering for debugging such modifications, if you also do each of
the following:
a) Give prominent notice with each copy of the Combined Work that
the Library is used in it and that the Library and its use are
covered by this License.
b) Accompany the Combined Work with a copy of the GNU GPL and this license
document.
c) For a Combined Work that displays copyright notices during
execution, include the copyright notice for the Library among
these notices, as well as a reference directing the user to the
copies of the GNU GPL and this license document.
d) Do one of the following:
0) Convey the Minimal Corresponding Source under the terms of this
License, and the Corresponding Application Code in a form
suitable for, and under terms that permit, the user to
recombine or relink the Application with a modified version of
the Linked Version to produce a modified Combined Work, in the
manner specified by section 6 of the GNU GPL for conveying
Corresponding Source.
1) Use a suitable shared library mechanism for linking with the
Library. A suitable mechanism is one that (a) uses at run time
a copy of the Library already present on the user's computer
system, and (b) will operate properly with a modified version
of the Library that is interface-compatible with the Linked
Version.
e) Provide Installation Information, but only if you would otherwise
be required to provide such information under section 6 of the
GNU GPL, and only to the extent that such information is
necessary to install and execute a modified version of the
Combined Work produced by recombining or relinking the
Application with a modified version of the Linked Version. (If
you use option 4d0, the Installation Information must accompany
the Minimal Corresponding Source and Corresponding Application
Code. If you use option 4d1, you must provide the Installation
Information in the manner specified by section 6 of the GNU GPL
for conveying Corresponding Source.)
5. Combined Libraries.
You may place library facilities that are a work based on the
Library side by side in a single library together with other library
facilities that are not Applications and are not covered by this
License, and convey such a combined library under terms of your
choice, if you do both of the following:
a) Accompany the combined library with a copy of the same work based
on the Library, uncombined with any other library facilities,
conveyed under the terms of this License.
b) Give prominent notice with the combined library that part of it
is a work based on the Library, and explaining where to find the
accompanying uncombined form of the same work.
6. Revised Versions of the GNU Lesser General Public License.
The Free Software Foundation may publish revised and/or new versions
of the GNU Lesser General Public License from time to time. Such new
versions will be similar in spirit to the present version, but may
differ in detail to address new problems or concerns.
Each version is given a distinguishing version number. If the
Library as you received it specifies that a certain numbered version
of the GNU Lesser General Public License "or any later version"
applies to it, you have the option of following the terms and
conditions either of that published version or of any later version
published by the Free Software Foundation. If the Library as you
received it does not specify a version number of the GNU Lesser
General Public License, you may choose any version of the GNU Lesser
General Public License ever published by the Free Software Foundation.
If the Library as you received it specifies that a proxy can decide
whether future versions of the GNU Lesser General Public License shall
apply, that proxy's public statement of acceptance of any version is
permanent authorization for you to choose that version for the
Library.
sgd-2.0/Makefile 0000664 0000000 0000000 00000001321 11645337641 0013577 0 ustar 00root root 0000000 0000000 # General makefile
MAKE=make
SHELL=/bin/sh
SUBDIRS=svm crf
world: check all
@echo "================================================"
@echo "CONGRATULATIONS: The compilation was successful."
@echo "To know what to do next, check the README file."
@echo "================================================"
all clean:
@for n in ${SUBDIRS} ; \
do ( cd $$n && ${MAKE} ${@}) || exit ; done
all: check
check:
@if [ -r data/rcv1/rcv1-v2.topics.qrels.gz ] ; then : ; else \
echo "=======================================" ; \
echo "ATTENTION: Missing data files" ; \
echo "You should have read the README file!" ; \
echo "=======================================" ; \
fi
.PHONY: world all depend check sgd-2.0/README.txt 0000664 0000000 0000000 00000010652 11645337641 0013644 0 ustar 00root root 0000000 0000000
SGD-2.0
-------
L. Bottou, October 2011
1. INTRODUCTION
The goal of this package is to illustrate the efficiency of stochastic
gradient descent for large-scale learning tasks.
Two algorithms,
* Stochastic gradient descent (SGD), and
* Averaged stochastic gradient descent (ASGD),
are applied to two well known problems
* Linear Support Vector Machines, and
* Conditional Random Fields.
The corresponding programs are designed for simplicity and readability. In
particular they avoid optimizations that would made the programs less
readable. The sole exception is the handling of sparse training data.
2. DATASETS
The programs are demonstrated using a number of standard datasets.
* The RCV1-V2 dataset.
* The ALPHA and WEBSPAM datasets from the first Pascal Large Scale Learning Challenge.
* The dataset associated with the CONLL2000 chunking task.
These datasets are available from the web. File "data/README" contains
instructions for downloading. The Pascal datasets must be
preprocessed using a relatively slow python script.
3. ALGORITHMS
Unlike most optimization algorithm, each iteration of these stochastic
algorithms process a single example and update the parameters. Although the
theory calls for picking a random example at each iteration, this
implementation performs sequential passes over randomly shuffled training
examples. This process is in fact more effective in practice. Each pass is
called an epoch.
Assume we have an objective function of the form
Obj(w) = 1/2 lambda w^2 + 1/n sum_i=1^n L(z_i,w)
where w is the parameter, {z_1,...,z_n} are the training examples, 1/2 \lambda
w^2 a regularization term, and L(z,w) is the loss function. Each iteration of
the SGD algorithm picks a single example z and updates the parameter vector
using the formula:
SGD: w := (1 - lambda eta_t) w - eta_t dL/dw(z,w)
The trick of course is to choose the gain sequence eta_t wisely. We use the
formula eta_t = eta_0 / (1 + lambda eta_0 t), and we pick eta_0 by trying
several gain values on a subset of the training data. In order to leverage
sparse dataset, we represent vector w as the ratio of a vector W and a scalar
wDivisor, that is, w = W / wDivisor. Each iteration effectively becomes:
SGD: wDivisor = wDivisor / (1 - lambda eta_t)
W = W - eta_t wDivisor dL/dw(z,w)
The ASGD algorithm maintains two parameter vectors. The first parameter
vector, w, is updated like the SGD parameter. However, the output of the
algorithm is the second parameter vector, a, which computes an average of
all the previous values of w.
ASGD: w := (1 - lambda eta_t) w - eta_t dL/dw(z,w)
a := a + mu_t [ w - a ]
This algorithm has been shown to work extremely well (Polyak and Juditsky,
1992) provided that the sequence eta_t decreases with exactly the right speed.
We follow (Xu, 2010) and choose eta_t = eta_0 / (1 + lambda eta0 t) ^ 0.75.
We select eta_0 by trying several gain values on a subset of the training
data, and we start the averaging process after a certain time, that is,
mu_t = 1/max(1,t-t0). Following (Xu, 2010), sparse training data is treated
using the substitutions w = W / wDivisor and a = (A + wFraction W) / aDivisor.
The algorithm effectively becomes:
ASGD: wDivisor = wDivisor / (1 - eta_t * lambda)
W = W - eta_t wDivisor dL/dw(z,w)
A = A + eta_t wFraction wDivisor dL/dw(z,w)
aDivisor = aDivisor / (1 - mu_t)
wFraction = wFraction + mu_t aDivisor / wDivisor
4. SUPPORT VECTOR MACHINES
The directory "svm" contains programs to train a L2-regularized linear model
for binary classification tasks. Compilation time switches determine whether
the models include a bias term, whether the bias term is regularized, and
which loss function should be used. The default is to use an unregularized
bias term using the log-loss function L(x,y,w) = log(1+exp(-ywx)). See file
"svm/README" for details about these programs and their usage for each of the
datasets.
5. CONDITIONAL RANDOM FIELDS
The directory "crf" contains programs "crfsgd" and "crfasgd" for training
conditional random fields for sequences. Both programs take data files and
template files and produces tagging files similar to those of Taku Kudo's
CRF++ program described at . However they also
accepts gzipped data files instead of plain files. See the file "crf/README"
for detailed information about these programs and their usage.
sgd-2.0/crf/ 0000775 0000000 0000000 00000000000 11645337641 0012714 5 ustar 00root root 0000000 0000000 sgd-2.0/crf/Makefile 0000664 0000000 0000000 00000003545 11645337641 0014363 0 ustar 00root root 0000000 0000000 # CRF with stochastic gradient
# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 2 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program; if not, write to the Free Software
# Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111, USA
L=../lib
CXX=g++
OPTS=-g -O2
CXXFLAGS= ${OPTS} -Wall -I$L
LIBS = -lz -lm
PROGRAMS = crfsgd crfasgd
OBJS = vectors.o matrices.o gzstream.o pstream.o timer.o
all: ${PROGRAMS}
clean:
-rm ${PROGRAMS} 2>/dev/null
-rm *.o 2>/dev/null
crfsgd: crfsgd.o ${OBJS}
-rm $@ 2>/dev/null
${CXX} ${CXXFLAGS} -o $@ crfsgd.o ${OBJS} ${LIBS}
crfasgd: crfasgd.o ${OBJS}
-rm $@ 2>/dev/null
${CXX} ${CXXFLAGS} -o $@ crfasgd.o ${OBJS} ${LIBS}
crfsgd.o: crfsgd.cpp $L/vectors.h $L/gzstream.h $L/timer.h
${CXX} ${CXXFLAGS} -c -o $@ crfsgd.cpp
crfasgd.o: crfasgd.cpp $L/vectors.h $L/gzstream.h $L/timer.h
${CXX} ${CXXFLAGS} -c -o $@ crfasgd.cpp
vectors.o: $L/vectors.cpp $L/vectors.h $L/wrapper.h $L/assert.h
${CXX} ${CXXFLAGS} -c -o $@ $L/vectors.cpp
matrices.o: $L/matrices.cpp $L/matrices.h $L/vectors.h $L/wrapper.h $L/assert.h
${CXX} ${CXXFLAGS} -c -o $@ $L/matrices.cpp
gzstream.o: $L/gzstream.cpp $L/gzstream.h $L/assert.h
${CXX} ${CXXFLAGS} -c -o $@ $L/gzstream.cpp
pstream.o: $L/pstream.cpp $L/pstream.h $L/assert.h
${CXX} ${CXXFLAGS} -c -o $@ $L/pstream.cpp
timer.o: $L/timer.cpp $L/timer.h $L/assert.h
${CXX} ${CXXFLAGS} -c -o $@ $L/timer.cpp
sgd-2.0/crf/README.txt 0000664 0000000 0000000 00000022540 11645337641 0014415 0 ustar 00root root 0000000 0000000
1. COMPILING CRFSGD AND CRFASGD
Compiling under Unix is achieved using the traditional command "make".
The compilation requires the libz library. This library usually comes
preinstalled on most Linux distributions, and is otherwise available
from from http://www.zlib.org.
Compiling under Windows is possible using Cygwin, using MSYS, or using the
MSVC project files provided in the subdirectory "win" of the sgd distribution.
Make sure to read the instructions as you need to compile zlib adequately.
You then need to copy the executable files in this directory.
2. USAGE
Synopsis (training):
crfsgd [options] model template traindata [validdata]
crfasgd [options] model template traindata [validdata]
Synopsis (tagging):
crfsgd -t model testdata
crfasgd -t model testdata
Program "crfsgd" and "crfasgd" implement stochastic gradient algorithms for
training conditional random field models. The program inputs are modeled after
the well known CRF++ program (http://crfpp.sourceforge.net). In particular
these programs use the same format for the template files and the data
files. These formats are well documented on the CRF++ web page. Programs
crfsgd and crfasgd can also directly read gzipped data files, provided that
the file name ends with suffix ".gz".
When operating in training mode, these program construct a CRF according to
the template file and perform a predefined number of training epochs on the
training data. Every so many epochs are followed by a performance evaluation
pass over the training set and optionally over a validation set. The
performance evaluation procedure can pipe the tags into an external evaluation
program such as the standard CONLL evaluation script conlleval. The training
set performance is useful to monitor the progress of the optimization. The
validation set performance is useful to estimate the generalization
performance. The recommended procedure is to monitor the validation
performance and stop the algorithm when the validation metrics no longer
improve. In the limit of large number of examples, program "crfasgd" should
reach this point after one or two epochs only. The model is saved in the
specified model file (which in fact is a compressed text file.)
Both programs accept the same options when used in training mode.
-c : capacity control parameter (1.0)
-f : threshold on the occurences of each feature (3)
-r : total number of epochs (50)
-h : epochs between each testing phase (5)
-e : performance evaluation command (conlleval -q)
-s : initial learning rate
-q : silent mode
Program crfasgd accepts one additional option:
-a d : starts averaging after d iterations (default: 1.0.)
Using option -t switches to the tagging mode. When operating in tagging mode, the
program reads the model, tags every sentence from the provided test data file,
and outputs the tags on the standard output using a format suitable for the
standard evaluation script conlleval.
2. RUNNING THE STOCHASTIC GRADIENT CRF ON THE CONLL CHUNKING TASK
Please follow the instructions in file "data/README.txt" to populate the
directories "data/conll". The gzipped files are directly usable.
No further preprocessing is necessary.
Training a model using stochastic gradient descent.
$ ./crfsgd -c 1.0 -f 3 model.gz template \
../data/conll2000/train.txt.gz ../data/conll2000/test.txt.gz
Reading template file template.
...
Reading and preprocessing ../data/conll2000/train.txt.gz.
...
Reading and preprocessing ../data/conll2000/test.txt.gz.
...
[Calibrating] -- 1000 samples
...
[Epoch 1] -- wnorm=3428.22 time=15.66s.
[Epoch 2] -- wnorm=4981.97 time=21.59s.
[Epoch 3] -- wnorm=6099.82 time=27.5s.
[Epoch 4] -- wnorm=6888.25 time=33.41s.
[Epoch 5] -- wnorm=7465.87 time=39.29s.
Training perf: sentences=8936 loss=0.8069 obj=1.22464 err=2454 (1.15904%)
accuracy: 98.84%; precision: 97.95%; recall: 98.04%; FB1: 98.00
Testing perf: sentences=2012 loss=2.35348 obj=2.77122 err=1997 (4.21513%)
accuracy: 95.78%; precision: 93.31%; recall: 93.47%; FB1: 93.39
[Epoch 6] -- wnorm=7904.99 time=45.19s.
[Epoch 7] -- wnorm=8238.51 time=51.07s.
[Epoch 8] -- wnorm=8494.34 time=56.96s.
[Epoch 9] -- wnorm=8695.67 time=62.84s.
[Epoch 10] -- wnorm=8859.06 time=68.73s.
Training perf: sentences=8936 loss=0.592674 obj=1.08837 err=1492 (0.704681%)
accuracy: 99.30%; precision: 98.81%; recall: 98.64%; FB1: 98.72
Testing perf: sentences=2012 loss=2.27945 obj=2.77514 err=1950 (4.11592%)
accuracy: 95.88%; precision: 93.60%; recall: 93.43%; FB1: 93.51
...
[Epoch 46] -- wnorm=9670.23 time=281.04s.
[Epoch 47] -- wnorm=9670.91 time=286.93s.
[Epoch 48] -- wnorm=9670.6 time=292.83s.
[Epoch 49] -- wnorm=9670.94 time=298.74s.
[Epoch 50] -- wnorm=9669.67 time=304.64s.
Training perf: sentences=8936 loss=0.476964 obj=1.01802 err=692 (0.326836%)
accuracy: 99.67%; precision: 99.42%; recall: 99.26%; FB1: 99.34
Testing perf: sentences=2012 loss=2.20519 obj=2.74624 err=1889 (3.98717%)
accuracy: 96.01%; precision: 93.94%; recall: 93.55%; FB1: 93.74
Saving model file model.gz.
Done! 304.64 seconds.
Training a model using averaged stochastic gradient descent.
$ ./crfasgd -c 1.0 -f 3 -r 10 model.gz template \
../data/conll2000/train.txt.gz ../data/conll2000/test.txt.gz
Reading template file template.
...
Reading and preprocessing ../data/conll2000/train.txt.gz.
...
Reading and preprocessing ../data/conll2000/test.txt.gz.
...
[Calibrating] -- 1000 samples
...
[Epoch 1] -- wnorm=3471.6 anorm=3471.6 time=16.88s.
[Epoch 2] -- wnorm=5093.77 anorm=4238.68 time=23.5s.
[Epoch 3] -- wnorm=6281.55 anorm=4871.78 time=30.11s.
[Epoch 4] -- wnorm=7128.27 anorm=5400.67 time=36.75s.
[Epoch 5] -- wnorm=7748.73 anorm=5837.29 time=43.34s.
Training perf: sentences=8936 loss=0.879526 obj=1.20614 err=2945 (1.39%)
accuracy: 98.61%; precision: 97.66%; recall: 97.51%; FB1: 97.58
Testing perf: sentences=2012 loss=2.23476 obj=2.56138 err=1895 (3.99%)
accuracy: 96.00%; precision: 93.75%; recall: 93.59%; FB1: 93.67
[Epoch 6] -- wnorm=8219.54 anorm=6203.39 time=50.05s.
[Epoch 7] -- wnorm=8569.77 anorm=6514.33 time=56.84s.
[Epoch 8] -- wnorm=8858.12 anorm=6790.36 time=63.46s.
[Epoch 9] -- wnorm=9059.93 anorm=7026.73 time=70.42s.
[Epoch 10] -- wnorm=9230.78 anorm=7235.45 time=77.04s.
Training perf: sentences=8936 loss=0.68643 obj=1.09128 err=1977 (0.93%)
accuracy: 99.07%; precision: 98.40%; recall: 98.24%; FB1: 98.32
Testing perf: sentences=2012 loss=2.21381 obj=2.61866 err=1872 (3.95%)
accuracy: 96.05%; precision: 93.84%; recall: 93.67%; FB1: 93.75
Saving model file model.gz.
Done! 77.04 seconds.
Testing the final model (using crfsgd or crfasgd is equivalent.)
$ ./crfsgd -t model.gz ../data/conll2000/test.txt.gz | ./conlleval
processed 47377 tokens with 23852 phrases; found: 23809 phrases; correct: 22342.
accuracy: 96.05%; precision: 93.84%; recall: 93.67%; FB1: 93.75
ADJP: precision: 79.66%; recall: 74.20%; FB1: 76.83 408
ADVP: precision: 82.96%; recall: 80.95%; FB1: 81.94 845
CONJP: precision: 55.56%; recall: 55.56%; FB1: 55.56 9
INTJ: precision: 100.00%; recall: 50.00%; FB1: 66.67 1
LST: precision: 0.00%; recall: 0.00%; FB1: 0.00 0
NP: precision: 94.36%; recall: 94.08%; FB1: 94.22 12385
PP: precision: 96.61%; recall: 97.80%; FB1: 97.20 4870
PRT: precision: 77.45%; recall: 74.53%; FB1: 75.96 102
SBAR: precision: 88.33%; recall: 84.86%; FB1: 86.56 514
VP: precision: 93.80%; recall: 94.14%; FB1: 93.97 4675
Comparing with CRF++ (on a different machine, about twice slower.)
$ crf_learn -c 1.0 -f 3 template train.txt model
...
Number of sentences: 8936
Number of features: 1679700
... iter=18 terr=0.04522 serr=0.45636 act=1679700 obj=24917.57905 diff=0.02882
... iter=36 terr=0.02188 serr=0.27775 act=1679700 obj=13697.78077 diff=0.01717
... iter=71 terr=0.00518 serr=0.09109 act=1679700 obj=9654.43394 diff=0.00167
... iter=142 terr=0.00340 serr=0.06256 act=1679700 obj=9042.07254 diff=0.00007
Done!4335.34 s
$ crf_test -m model test.txt | tr '\t' ' ' | ./conlleval
processed 47377 tokens with 23852 phrases; found: 23799 phrases; correct: 22334.
accuracy: 96.02%; precision: 93.84%; recall: 93.64%; FB1: 93.74
ADJP: precision: 79.71%; recall: 74.43%; FB1: 76.98 409
ADVP: precision: 83.18%; recall: 81.06%; FB1: 82.11 844
CONJP: precision: 55.56%; recall: 55.56%; FB1: 55.56 9
INTJ: precision: 100.00%; recall: 50.00%; FB1: 66.67 1
LST: precision: 0.00%; recall: 0.00%; FB1: 0.00 0
NP: precision: 94.36%; recall: 94.03%; FB1: 94.19 12378
PP: precision: 96.71%; recall: 97.82%; FB1: 97.26 4866
PRT: precision: 79.05%; recall: 78.30%; FB1: 78.67 105
SBAR: precision: 88.65%; recall: 84.67%; FB1: 86.62 511
VP: precision: 93.63%; recall: 93.99%; FB1: 93.81 4676
sgd-2.0/crf/conlleval 0000775 0000000 0000000 00000031175 11645337641 0014630 0 ustar 00root root 0000000 0000000 #!/usr/bin/perl -w
# conlleval: evaluate result of processing CoNLL-2000 shared task
# usage: conlleval [-l] [-r] [-d delimiterTag] [-o oTag] < file
# README: http://cnts.uia.ac.be/conll2000/chunking/output.html
# options: l: generate LaTeX output for tables like in
# http://cnts.uia.ac.be/conll2003/ner/example.tex
# r: accept raw result tags (without B- and I- prefix;
# assumes one word per chunk)
# d: alternative delimiter tag (default is single space)
# o: alternative outside tag (default is O)
# note: the file should contain lines with items separated
# by $delimiter characters (default space). The final
# two items should contain the correct tag and the
# guessed tag in that order. Sentences should be
# separated from each other by empty lines or lines
# with $boundary fields (default -X-).
# url: http://lcg-www.uia.ac.be/conll2000/chunking/
# started: 1998-09-25
# version: 2004-01-26
# author: Erik Tjong Kim Sang
use strict;
my $false = 0;
my $true = 42;
my $boundary = "-X-"; # sentence boundary
my $correct; # current corpus chunk tag (I,O,B)
my $correctChunk = 0; # number of correctly identified chunks
my $correctTags = 0; # number of correct chunk tags
my $correctType; # type of current corpus chunk tag (NP,VP,etc.)
my $delimiter = " "; # field delimiter
my $FB1 = 0.0; # FB1 score (Van Rijsbergen 1979)
my $firstItem; # first feature (for sentence boundary checks)
my $foundCorrect = 0; # number of chunks in corpus
my $foundGuessed = 0; # number of identified chunks
my $guessed; # current guessed chunk tag
my $guessedType; # type of current guessed chunk tag
my $i; # miscellaneous counter
my $inCorrect = $false; # currently processed chunk is correct until now
my $lastCorrect = "O"; # previous chunk tag in corpus
my $latex = 0; # generate LaTeX formatted output
my $lastCorrectType = ""; # type of previously identified chunk tag
my $lastGuessed = "O"; # previously identified chunk tag
my $lastGuessedType = ""; # type of previous chunk tag in corpus
my $lastType; # temporary storage for detecting duplicates
my $line; # line
my $nbrOfFeatures = -1; # number of features per line
my $precision = 0.0; # precision score
my $oTag = "O"; # outside tag, default O
my $raw = 0; # raw input: add B to every token
my $quiet = 0; # only display summary line
my $recall = 0.0; # recall score
my $tokenCounter = 0; # token counter (ignores sentence breaks)
my %correctChunk = (); # number of correctly identified chunks per type
my %foundCorrect = (); # number of chunks in corpus per type
my %foundGuessed = (); # number of identified chunks per type
my @features; # features on line
my @sortedTypes; # sorted list of chunk type names
# sanity check
while (@ARGV and $ARGV[0] =~ /^-/) {
if ($ARGV[0] eq "-l") { $latex = 1; shift(@ARGV); }
elsif ($ARGV[0] eq "-r") { $raw = 1; shift(@ARGV); }
elsif ($ARGV[0] eq "-q") { $quiet = 1; shift(@ARGV); }
elsif ($ARGV[0] eq "-d") {
shift(@ARGV);
if (not defined $ARGV[0]) {
die "conlleval: -d requires delimiter character";
}
$delimiter = shift(@ARGV);
} elsif ($ARGV[0] eq "-o") {
shift(@ARGV);
if (not defined $ARGV[0]) {
die "conlleval: -o requires delimiter character";
}
$oTag = shift(@ARGV);
} else { die "conlleval: unknown argument $ARGV[0]\n"; }
}
if (@ARGV) { die "conlleval: unexpected command line argument\n"; }
# process input
while () {
chomp($line = $_);
@features = split(/$delimiter/,$line);
if ($nbrOfFeatures < 0) { $nbrOfFeatures = $#features; }
elsif ($nbrOfFeatures != $#features and @features != 0) {
printf STDERR "unexpected number of features: %d (%d)\n",
$#features+1,$nbrOfFeatures+1;
exit(1);
}
if (@features == 0 or
$features[0] eq $boundary) { @features = ($boundary,"O","O"); }
if (@features < 2) {
die "conlleval: unexpected number of features in line $line\n";
}
if ($raw) {
if ($features[$#features] eq $oTag) { $features[$#features] = "O"; }
if ($features[$#features-1] eq $oTag) { $features[$#features-1] = "O"; }
if ($features[$#features] ne "O") {
$features[$#features] = "B-$features[$#features]";
}
if ($features[$#features-1] ne "O") {
$features[$#features-1] = "B-$features[$#features-1]";
}
}
# 20040126 ET code which allows hyphens in the types
if ($features[$#features] =~ /^([^-]*)-(.*)$/) {
$guessed = $1;
$guessedType = $2;
} else {
$guessed = $features[$#features];
$guessedType = "";
}
pop(@features);
if ($features[$#features] =~ /^([^-]*)-(.*)$/) {
$correct = $1;
$correctType = $2;
} else {
$correct = $features[$#features];
$correctType = "";
}
pop(@features);
# ($guessed,$guessedType) = split(/-/,pop(@features));
# ($correct,$correctType) = split(/-/,pop(@features));
$guessedType = $guessedType ? $guessedType : "";
$correctType = $correctType ? $correctType : "";
$firstItem = shift(@features);
# 1999-06-26 sentence breaks should always be counted as out of chunk
if ( $firstItem eq $boundary ) { $guessed = "O"; }
if ($inCorrect) {
if ( &endOfChunk($lastCorrect,$correct,$lastCorrectType,$correctType) and
&endOfChunk($lastGuessed,$guessed,$lastGuessedType,$guessedType) and
$lastGuessedType eq $lastCorrectType) {
$inCorrect=$false;
$correctChunk++;
$correctChunk{$lastCorrectType} = $correctChunk{$lastCorrectType} ?
$correctChunk{$lastCorrectType}+1 : 1;
} elsif (
&endOfChunk($lastCorrect,$correct,$lastCorrectType,$correctType) !=
&endOfChunk($lastGuessed,$guessed,$lastGuessedType,$guessedType) or
$guessedType ne $correctType ) {
$inCorrect=$false;
}
}
if ( &startOfChunk($lastCorrect,$correct,$lastCorrectType,$correctType) and
&startOfChunk($lastGuessed,$guessed,$lastGuessedType,$guessedType) and
$guessedType eq $correctType) { $inCorrect = $true; }
if ( &startOfChunk($lastCorrect,$correct,$lastCorrectType,$correctType) ) {
$foundCorrect++;
$foundCorrect{$correctType} = $foundCorrect{$correctType} ?
$foundCorrect{$correctType}+1 : 1;
}
if ( &startOfChunk($lastGuessed,$guessed,$lastGuessedType,$guessedType) ) {
$foundGuessed++;
$foundGuessed{$guessedType} = $foundGuessed{$guessedType} ?
$foundGuessed{$guessedType}+1 : 1;
}
if ( $firstItem ne $boundary ) {
if ( $correct eq $guessed and $guessedType eq $correctType ) {
$correctTags++;
}
$tokenCounter++;
}
$lastGuessed = $guessed;
$lastCorrect = $correct;
$lastGuessedType = $guessedType;
$lastCorrectType = $correctType;
}
if ($inCorrect) {
$correctChunk++;
$correctChunk{$lastCorrectType} = $correctChunk{$lastCorrectType} ?
$correctChunk{$lastCorrectType}+1 : 1;
}
if (not $latex) {
# compute overall precision, recall and FB1 (default values are 0.0)
$precision = 100*$correctChunk/$foundGuessed if ($foundGuessed > 0);
$recall = 100*$correctChunk/$foundCorrect if ($foundCorrect > 0);
$FB1 = 2*$precision*$recall/($precision+$recall)
if ($precision+$recall > 0);
# print overall performance
if (not $quiet) {
printf "processed $tokenCounter tokens with $foundCorrect phrases; ";
printf "found: $foundGuessed phrases; correct: $correctChunk.\n";
}
if ($tokenCounter>0) {
printf "accuracy: %6.2f%%; ",100*$correctTags/$tokenCounter;
printf "precision: %6.2f%%; ",$precision;
printf "recall: %6.2f%%; ",$recall;
printf "FB1: %6.2f\n",$FB1;
}
}
# sort chunk type names
undef($lastType);
@sortedTypes = ();
foreach $i (sort (keys %foundCorrect,keys %foundGuessed)) {
if (not($lastType) or $lastType ne $i) {
push(@sortedTypes,($i));
}
$lastType = $i;
}
# print performance per chunk type
if (not $latex) {
if (not $quiet) {
for $i (@sortedTypes) {
$correctChunk{$i} = $correctChunk{$i} ? $correctChunk{$i} : 0;
if (not($foundGuessed{$i})) { $foundGuessed{$i} = 0; $precision = 0.0; }
else { $precision = 100*$correctChunk{$i}/$foundGuessed{$i}; }
if (not($foundCorrect{$i})) { $recall = 0.0; }
else { $recall = 100*$correctChunk{$i}/$foundCorrect{$i}; }
if ($precision+$recall == 0.0) { $FB1 = 0.0; }
else { $FB1 = 2*$precision*$recall/($precision+$recall); }
printf "%17s: ",$i;
printf "precision: %6.2f%%; ",$precision;
printf "recall: %6.2f%%; ",$recall;
printf "FB1: %6.2f %d\n",$FB1,$foundGuessed{$i};
}
}
} else {
print " & Precision & Recall & F\$_{\\beta=1} \\\\\\hline";
if (not $quiet) {
for $i (@sortedTypes) {
$correctChunk{$i} = $correctChunk{$i} ? $correctChunk{$i} : 0;
if (not($foundGuessed{$i})) { $precision = 0.0; }
else { $precision = 100*$correctChunk{$i}/$foundGuessed{$i}; }
if (not($foundCorrect{$i})) { $recall = 0.0; }
else { $recall = 100*$correctChunk{$i}/$foundCorrect{$i}; }
if ($precision+$recall == 0.0) { $FB1 = 0.0; }
else { $FB1 = 2*$precision*$recall/($precision+$recall); }
printf "\n%-7s & %6.2f\\%% & %6.2f\\%% & %6.2f \\\\",
$i,$precision,$recall,$FB1;
}
print "\\hline\n";
}
$precision = 0.0;
$recall = 0;
$FB1 = 0.0;
$precision = 100*$correctChunk/$foundGuessed if ($foundGuessed > 0);
$recall = 100*$correctChunk/$foundCorrect if ($foundCorrect > 0);
$FB1 = 2*$precision*$recall/($precision+$recall)
if ($precision+$recall > 0);
printf "Overall & %6.2f\\%% & %6.2f\\%% & %6.2f \\\\\\hline\n",
$precision,$recall,$FB1;
}
exit 0;
# endOfChunk: checks if a chunk ended between the previous and current word
# arguments: previous and current chunk tags, previous and current types
# note: this code is capable of handling other chunk representations
# than the default CoNLL-2000 ones, see EACL'99 paper of Tjong
# Kim Sang and Veenstra http://xxx.lanl.gov/abs/cs.CL/9907006
sub endOfChunk {
my $prevTag = shift(@_);
my $tag = shift(@_);
my $prevType = shift(@_);
my $type = shift(@_);
my $chunkEnd = $false;
if ( $prevTag eq "B" and $tag eq "B" ) { $chunkEnd = $true; }
if ( $prevTag eq "B" and $tag eq "O" ) { $chunkEnd = $true; }
if ( $prevTag eq "I" and $tag eq "B" ) { $chunkEnd = $true; }
if ( $prevTag eq "I" and $tag eq "O" ) { $chunkEnd = $true; }
if ( $prevTag eq "E" and $tag eq "E" ) { $chunkEnd = $true; }
if ( $prevTag eq "E" and $tag eq "I" ) { $chunkEnd = $true; }
if ( $prevTag eq "E" and $tag eq "O" ) { $chunkEnd = $true; }
if ( $prevTag eq "I" and $tag eq "O" ) { $chunkEnd = $true; }
if ($prevTag ne "O" and $prevTag ne "." and $prevType ne $type) {
$chunkEnd = $true;
}
# corrected 1998-12-22: these chunks are assumed to have length 1
if ( $prevTag eq "]" ) { $chunkEnd = $true; }
if ( $prevTag eq "[" ) { $chunkEnd = $true; }
return($chunkEnd);
}
# startOfChunk: checks if a chunk started between the previous and current word
# arguments: previous and current chunk tags, previous and current types
# note: this code is capable of handling other chunk representations
# than the default CoNLL-2000 ones, see EACL'99 paper of Tjong
# Kim Sang and Veenstra http://xxx.lanl.gov/abs/cs.CL/9907006
sub startOfChunk {
my $prevTag = shift(@_);
my $tag = shift(@_);
my $prevType = shift(@_);
my $type = shift(@_);
my $chunkStart = $false;
if ( $prevTag eq "B" and $tag eq "B" ) { $chunkStart = $true; }
if ( $prevTag eq "I" and $tag eq "B" ) { $chunkStart = $true; }
if ( $prevTag eq "O" and $tag eq "B" ) { $chunkStart = $true; }
if ( $prevTag eq "O" and $tag eq "I" ) { $chunkStart = $true; }
if ( $prevTag eq "E" and $tag eq "E" ) { $chunkStart = $true; }
if ( $prevTag eq "E" and $tag eq "I" ) { $chunkStart = $true; }
if ( $prevTag eq "O" and $tag eq "E" ) { $chunkStart = $true; }
if ( $prevTag eq "O" and $tag eq "I" ) { $chunkStart = $true; }
if ($tag ne "O" and $tag ne "." and $prevType ne $type) {
$chunkStart = $true;
}
# corrected 1998-12-22: these chunks are assumed to have length 1
if ( $tag eq "[" ) { $chunkStart = $true; }
if ( $tag eq "]" ) { $chunkStart = $true; }
return($chunkStart);
}
sgd-2.0/crf/crfasgd.cpp 0000664 0000000 0000000 00000141020 11645337641 0015027 0 ustar 00root root 0000000 0000000 // -*- C++ -*-
// CRF with stochastic gradient
// Copyright (C) 2007- Leon Bottou
// This program is free software; you can redistribute it and/or
// modify it under the terms of the GNU Lesser General Public
// License as published by the Free Software Foundation; either
// version 2.1 of the License, or (at your option) any later version.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program; if not, write to the Free Software
// Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111, USA
// $Id$
#include "wrapper.h"
#include "vectors.h"
#include "matrices.h"
#include "gzstream.h"
#include "pstream.h"
#include "timer.h"
#include
#include
#include
#include