DF-LDA-0.1.2/ 0000755 0003103 0000163 00000000000 11374252176 012405 5 ustar andrzeje medinfo DF-LDA-0.1.2/multinode.cpp 0000644 0003103 0000163 00000020450 11363057152 015105 0 ustar andrzeje medinfo /**
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
*/
#include "multinode.h"
#include "dtnode.h"
/**
* Return the log of the weight for a given variant
* (the weight is equal to the sum of all leaf weights
* under the likely internal node for this variant)
*/
double multinode :: var_logweight(int given_y)
{
return variant_logweights[given_y];
}
/**
* Calculate logpwz term for a given y value
*/
double multinode :: calc_logpwz(int given_y)
{
dtnode* dtv = dynamic_cast(variants[given_y]);
return (*dtv).calc_logpwz();
}
/**
* Set a new y value
*/
void multinode :: set_y(int given_y)
{
y = given_y;
}
/**
* Return y value (which variant is 'active') for this multinode
*/
int multinode :: get_y()
{
return y;
}
/**
* Return number of variants represented by this multinode
*/
int multinode :: num_variants()
{
return variants.size();
}
/**
* Calculate topic-word term of Gibbs sampling eqn
*/
double multinode :: calc_wordterm(double val, int leaf)
{
// Is this leaf under one of our interior children?
double newval;
for(unsigned int i = 0; i < ichildren.size(); i++)
{
if(leaf <= maxind[i])
{
// Update value with call to appropriate variant
newval = (*(variants[y])).calc_wordterm(val,fake_leafmaps[y][i]);
// Pass value down to actual child
return (*(this->ichildren[i])).calc_wordterm(newval,leaf);
}
}
// No, this leaf must be one of our leaves
// Get leaf edge index
int ei = ichildren.size() + (leaf - this->leafstart);
// Update value with call to appropriate variant
return (*(variants[y])).calc_wordterm(val,fake_leafmaps[y][ei]);
}
/**
* Update all branch counts associated with a given leaf
*/
void multinode :: modify_count(double val, int leaf)
{
// Is this leaf under one of our interior children?
for(unsigned int i = 0; i < ichildren.size(); i++)
{
if(leaf <= maxind[i])
{
// Update all variants
for(unsigned int v = 0; v < variants.size(); v++)
{
(*(variants[v])).modify_count(val,fake_leafmaps[v][i]);
}
// Recursively update child cts
(*(ichildren[i])).modify_count(val,leaf);
return;
}
}
// No, this leaf must be one of our leaves
// Get leaf edge index
int ei = ichildren.size() + (leaf - leafstart);
// Update all variants
for(unsigned int v = 0; v < variants.size(); v++)
{
(*(variants[v])).modify_count(val,fake_leafmaps[v][ei]);
}
return;
}
/**
* Constructor
*/
multinode :: multinode(PyObject* arg_children, PyObject* arg_maxind,
int arg_leafstart, PyObject* arg_variants)
{
// Do some argument checking
if(!PyList_Check(arg_children) || !PyList_Check(arg_maxind))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Non-List children/maxind element in dirtree node");
throw 1;
}
if(arg_leafstart < 0)
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Negative leafstart value in dirtree node");
throw 1;
}
if(PyList_Size(arg_children) != PyList_Size(arg_maxind))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Different sizes for children/maxind in dirtree node");
throw 1;
}
// Populate data members
int nci = PyList_Size(arg_children);
this->leafstart = arg_leafstart;
// Build variants
for(unsigned int v = 0; v < PyList_Size(arg_variants); v++)
{
// Unpack variant dtnode and fake leafmap
PyObject* newvar = PyList_GetItem(arg_variants,v);
PyObject* vdtn = PyTuple_GetItem(newvar,0);
PyObject* flm = PyTuple_GetItem(newvar,1);
try {
// Extract fake leaf map for this variant
vector cur_flm;
for(unsigned int fli = 0; fli < PyList_Size(flm); fli++)
cur_flm.push_back(PyInt_AsLong(PyList_GetItem(flm,fli)));
fake_leafmaps.push_back(cur_flm);
// Unpack tuple contents
PyArrayObject* vedges = (PyArrayObject*) PyTuple_GetItem(vdtn,0);
PyObject* vchildren = PyTuple_GetItem(vdtn,1);
PyObject* vmaxind = PyTuple_GetItem(vdtn,2);
int vleafstart = PyInt_AsLong(PyTuple_GetItem(vdtn,3));
// Build child dtnode
dtnode* newvardt = new dtnode(vedges,vchildren,
vmaxind,vleafstart);
variants.push_back(newvardt);
// Calculate log of weighing term (for y-sampling)
PyArrayObject* likely_edges = (PyArrayObject*)
PyTuple_GetItem(PyList_GetItem(vchildren,0),0);
double le_sum = PyFloat_AsDouble(PyArray_Sum(likely_edges,NPY_MAXDIMS,
PyArray_DOUBLE,NULL));
variant_logweights.push_back(log(le_sum));
}
catch (int err)
{
throw 1;
}
}
// Init y to -1
// (will raise errors if not properly initialized elsewhere)
y = -1;
// Recursively build children
//
int c;
for(c = 0; c < nci; c++)
{
// Get max leaf index under this child, check it
maxind.push_back(PyInt_AsLong(PyList_GetItem(arg_maxind,c)));
if(maxind[c] < 0)
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Negative maxind value in dirtree node");
throw 1;
}
// exploiting boolean short-circuit here...
if(c > 0 && maxind[c] <= maxind[c-1])
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Non-increasing maxind value in dirtree node");
throw 1;
}
// Recursively build child
//
try{
// Check that node tuple size is correct
PyObject* pynode = PyList_GetItem(arg_children,c);
// Is child a dtnode, or a multinode?
//
if(4 == PyTuple_Size(pynode))
{
// dtnode
// Unpack tuple contents
PyArrayObject* cedges = (PyArrayObject*) PyTuple_GetItem(pynode,0);
PyObject* cchildren = PyTuple_GetItem(pynode,1);
PyObject* cmaxind = PyTuple_GetItem(pynode,2);
int cleafstart = PyInt_AsLong(PyTuple_GetItem(pynode,3));
// Build child dtnode
dtnode* newchild = new dtnode(cedges,cchildren,
cmaxind,cleafstart);
ichildren.push_back(newchild);
}
else if(5 == PyTuple_Size(pynode))
{
// multinode
// Unpack tuple contents
PyObject* cchildren = PyTuple_GetItem(pynode,1);
PyObject* cmaxind = PyTuple_GetItem(pynode,2);
int cleafstart = PyInt_AsLong(PyTuple_GetItem(pynode,3));
PyObject* cvariants = PyTuple_GetItem(pynode,4);
// Build child multinode
node* newchild = new multinode(cchildren,cmaxind,
cleafstart,cvariants);
ichildren.push_back(newchild);
}
else{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Node has wrong number of elements");
throw 1;
}
}
catch (int err)
{
throw 1;
}
}
}
/**
* Destuctor
*/
multinode :: ~multinode()
{
//
// Delete variants, ichildren will be
// deleted by base class destructor (node)
//
for(unsigned int vi = 0; vi < variants.size(); vi++)
delete variants[vi];
}
DF-LDA-0.1.2/Reporter.py 0000644 0003103 0000163 00000021701 11374251606 014557 0 ustar andrzeje medinfo """
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
"""
""" Contains Reporter class """
class Reporter:
""" Implements various reporting functions """
#
# Basic LDA reporting
#
def printTopics(self,N=0):
"""
Print Top N words by P(w|z) for each topic
Also print out q-samples for each topic
"""
(T,W) = self.phi.shape
(T,P) = self.qsamp.shape
N = N if N>0 else W # If no N arg, print all words foreach topic
for ti in range(T):
qstr = str([self.qsamp[ti,pi] for pi in range(P)])
print '\nTopic %d (q=%s)' % (ti,qstr)
top = range(W)
top.sort(key=lambda wi:-self.phi[ti,wi])
for wi in range(N):
print '%s = %f' % (self.vocab[top[wi]],self.phi[ti,top[wi]])
print ''
def printParams(self):
""" Print experimental parameters """
print ''
print 'Parameters'
print '----------------------'
print 'D = %d' % (self.D)
vals = (self.alpha,self.beta,self.eta)
print 'alpha = %f\nbeta = %f\neta = %f' % vals
print 'randseed = %d' % (self.randseed)
print 'burnin = %d\ninterval = %d' % (self.burnin,self.interval)
print 'numsamp = %d' % (self.numsamp)
print ''
#
# Report details of Dirichlet Forest data structure, CL/ML graphs, etc
#
def reportTree(self):
"""
Display the resulting Dirichlet Forest data structure which
will be passed to the C++ inference code
Schema:
root = (edges,ichildren,maxind,leafstart)
multinode = (None,ichildren,maxind,leafstart,variants)
variant = (fakeroot,fakeleafmap)
mustlink = (edges,[],[],leafstart)
NOTE: fakeroot is a typical dtnode (like root)
Permutation Mappings:
leafmap[word_index] = leaf_index
fakeleafmap[index_within_multinode] = fake_leaf_index
(wrt *this* variant)
That is, the procedure for querying a multinode is:
1) Get leaf index from leafmap
2) Find appropriate multinode
3) Find appropriate leaf/internal child of multinode
4) Call variant with the index of the leaf/internal (wrt *multinode*)
5) Pass along to true internal if necessary
"""
(root,leafmap,mlcc,W) = (self.root,self.leafmap,self.mlcc,self.W)
# Get a leaf-mapper function for reporting
# (will map leaf indices back to original word indices, or actual words)
lm = getLeafMapper(leafmap)
print '\nFINAL TREE STRUCTURE'
print '----------------------\n'
for (ichild,ind) in zip(root[1],range(len(root[1]))):
if(ichild[0] == None):
# This root child is a multinode
print 'Root Edge %f' % (root[0][ind])
print 'Root child %d - multinode' % (ind)
# Print the must-link children
for mlc in ichild[1]:
print '\tmust-link child'
(lstart,nch) = (mlc[3],len(mlc[0]))
#print ('\t\tleafstart index = %d'
# % (lstart))
#print ('\t\tnum children = %d'
# % (nch))
print '\t\tML-edges = %s' % (str(mlc[0]))
print ('\t\twords = %s'
% (str(lm(range(lstart,lstart+nch)))))
# Print the leaf children
nlch = len(ichild[4][0][1]) - len(ichild[1])
print '\tnum leaf children = %d' % (nlch)
print ('\tleaf words = %s'
% str(lm(range(ichild[3],ichild[3]+nlch))))
print ''
# Enumerate agreeable sets for this multinode
print '\tagreeable sets'
for (fakeroot,fakeleafmap) in ichild[4]:
print '\tAgree Edge = %f' % (fakeroot[0][0])
# For this agreeable set, see which children
# lie under the 'likely' internal node
likely_maxind = fakeroot[2][0]
agreeable = [fakeleafmap.index(i)
for i in range(likely_maxind+1)]
aset = []
for ai in agreeable:
if(ai >= len(ichild[1])):
# leaf child of multinode
aset.append(ichild[3]+ai-len(ichild[1]))
else:
# mustlink child of multinode
maxind = ichild[2][ai]
nch = len(ichild[1][ai][0])
aset.append(range((maxind+1-nch),maxind+1))
print '\tLeaf Edges = %s' % (str(fakeroot[0][1:]))
print '\t\tleaf indices = %s' % str(aset)
print '\t\tword indices = %s' % str(lm(aset))
print ''
else:
# This root child is a must-link node
print 'Root Edge = %f' % (root[0][ind])
print 'Root child %d - mustlink node' % (ind)
(lstart,nch) = (ichild[3],len(ichild[0]))
#print '\tleafstart = %d' % (lstart)
print '\tML-edges = %s' % (str(ichild[0]))
print '\tnum children = %d' % (nch)
print ('\twords = %s'
% (str(lm(range(lstart,lstart+nch)))))
def reportGraph(self):
"""
Display some information about the graph implied
by a given set of constraints
"""
(mlcc,clcc,allowable,W) = (self.mlcc,self.clcc,self.allowable,self.W)
print '\nCONSTRAINT GRAPH ANALYSIS'
print '----------------------------\n'
# must-link connected components
print 'Must-Link Connected Components'
print '----------' * 4
for i in range(len(mlcc)):
print 'MLCC %d = %s' % (W+i,str(self.vocMap(mlcc[i])))
# get an expander function for expanding must-link components
expander = self.getExpander(mlcc,W)
# cannot-link connected components
print '\nCannot-Link Connected Components'
print '----------' * 4
for (cc,allow) in zip(clcc,allowable):
print 'CLCC = %s' % str(self.vocMap(expander(cc)))
print 'Number of allowable sets = %d' % len(allow)
for a in allow:
print '\t%s' % str(self.vocMap(expander(a)))
print ''
def vocMap(self,val):
"""
If we have vocab List, map a word index --> actual word
"""
vocab = self.vocab
if(vocab != None):
if(type(val) == int):
return vocab[val]
elif(type(val) == list):
return [self.vocMap(v) for v in val]
else:
raise RuntimeError('Bad vocMap() call, type = %s' %
(str(type(val))))
else:
return val
def getExpander(self):
"""
Creates a function to expand indices into must-link components
"""
(mlcc,W) = (self.mlcc,self.W)
def expander(val):
if(type(val) == list):
return [expander(v) for v in val]
elif(type(val) == int):
if(val < W):
return val
else:
return mlcc[val-W]
else:
raise RuntimeError('Bad expander() call, type = %s' %
(str(type(val))))
return expander
def getLeafMapper(self):
"""
Creates a function to do reverse leaf-mapping
That is, map leaf indices -> word indices
(or actual words if vocab available)
"""
leafmap = self.leafmap
def leafmapper(val):
if(type(val) == list):
return [leafmapper(v) for v in val]
elif(type(val) == int):
return self.vocMap(leafmap.index(val))
else:
raise RuntimeError('Bad leafmapper() call, type = %s' %
(str(type(val))))
return leafmapper
DF-LDA-0.1.2/intLDA.h 0000644 0003103 0000163 00000004111 11363057154 013663 0 ustar andrzeje medinfo /**
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
*/
#include
#include
#include
#include
#define PY_ARRAY_UNIQUE_SYMBOL PyArray_API_intLDA
#include
#include
#include "dtnode.h"
#include "multinode.h"
#define OK 0
#define BAD 1
typedef struct {
vector topics;
PyArrayObject* nd;
} counts;
typedef struct {
PyArrayObject* alpha;
PyArrayObject* alphasum;
vector leafmap;
int T;
} model_params;
typedef struct {
int D;
int W;
vector doclens;
vector > docs;
vector > sample;
vector f;
} dataset;
static PyObject* intLDA(PyObject *self, PyObject *args, PyObject* keywds);
static int convert_args(PyObject* docs_arg, PyArrayObject* alpha,
PyObject* dirtree, PyObject* leafmap,
PyObject* f_arg,
model_params* mp, dataset* ds, counts* c);
static int init_model(model_params* mp, dataset* ds,counts* c,
int randseed, PyObject* zinit, PyArrayObject* qinit);
static void gibbs_chain(model_params* mp, dataset* ds, counts* c);
static PyArrayObject* est_phi(model_params* mp, dataset* ds, counts* c);
static PyArrayObject* est_theta(model_params* mp, dataset* ds, counts* c);
DF-LDA-0.1.2/DirichletForest.py 0000644 0003103 0000163 00000020444 11365346116 016053 0 ustar andrzeje medinfo """
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
"""
""" Contains the DirichletForest class """
from numpy import *
from Reporter import Reporter
import ConstraintCompiler as CC
from intLDA import intLDA
def readDocs(doc_fn):
""" Read docs in sparse SVMLight-style format """
docs = []
for line in open(doc_fn):
doc = []
pairs = line.split()
for pair in pairs:
(wi,ct) = [int(v) for v in pair.split(':')]
doc.extend([wi] * ct)
docs.append(doc)
return docs
def combine(*constraints):
"""
Given Lists of constraints, combine into a single set
constraints = varargs Tuple of Lists of constraints
each constraint = 2-element Tuples
constraint[0] = List of 2-element must-link constraint Tuples
constraint[1] = List of 2-element cannot-link constraint Tuples
"""
combo = (set([]),set([]))
for constr in constraints:
combo[0].update(constr[0])
combo[1].update(constr[1])
return combo
class DirichletForest(Reporter):
""" Dirichlet Forest Prior for Topic Modeling """
def __init__(self,alpha,beta,eta,T,W,vocab=None):
""" Constructor """
# Basic parameters
self.vocab = vocab
(self.T,self.W) = (T,W)
self.alpha = alpha * ones((1,self.T))
self.beta = beta
self.eta = eta
# Constraints and DF structures
self.constr = (set(),set())
(self.root,self.leafmap,
self.mlcc,self.clcc,self.allowable) = 5 * (None,)
# Outputs of inference
(self.zsamp,self.qsamp,
self.phi,self.theta) = 4 * (None,)
# Docs
self.docs = None
def inference(self,docs,numsamp,randseed,**kwargs):
"""
Do inference via collapsed Gibbs sampling,
estimate phi/theta from final sample
"""
# Compile preferences, if we haven't already
if(self.root == None):
self.compile()
# Save numsamp and randseed arguments
(self.numsamp,self.randseed) = (numsamp,randseed)
# Do inference, initializing sampler if applicable
if(self.zsamp != None and self.qsamp != None):
(self.phi,self.theta,self.zsamp,self.qsamp) = intLDA(
docs,self.alpha,self.root,
self.leafmap,numsamp,randseed,
init=self.zsamp,qinit=self.qsamp,
**kwargs)
elif(self.zsamp != None):
(self.phi,self.theta,self.zsamp,self.qsamp) = intLDA(
docs,self.alpha,self.root,
self.leafmap,numsamp,randseed,
init=self.zsamp,
**kwargs)
elif(self.qsamp != None):
(self.phi,self.theta,self.zsamp,self.qsamp) = intLDA(
docs,self.alpha,self.root,
self.leafmap,numsamp,randseed,
qinit=self.qsamp,
**kwargs)
else:
(self.phi,self.theta,self.zsamp,self.qsamp) = intLDA(
docs,self.alpha,self.root,
self.leafmap,numsamp,randseed,
**kwargs)
def compile(self):
""" Compile constraints into Dirichlet Forest data structure """
(self.mlcc,self.clcc,self.allowable) = CC.processPairwise(
self.constr,self.W,self.vocab)
(self.root,self.leafmap) = CC.buildTree(self.mlcc,self.clcc,
self.allowable,self.W,
self.beta,self.eta)
#
# For tricky descriptor-related reasons, these decorators
# MUST be applied to functions in the current order (!)
#
class Listify(object):
"""
Decorator to wrap constraint operation functions, wraps non-List
arguments as 1-element Lists if necessary
"""
def __init__(self,fcn):
self.fcn = fcn
self.instance = None
def __get__(self,instance,cls):
""" Need this to 'save' instance reference (self) """
self.instance = instance
return self
def __call__(self,*args):
""" Prepend instance ref (self) to args, pass to inner calls """
newargs = (self.instance,) + tuple([arg if type(arg)==list
else [arg] for arg in args])
return self.fcn(*newargs)
class VocabConvert(object):
"""
Decorator to wrap constraint operation functions, converts List of
words to List of integers (using vocab) if neccesary
"""
def __init__(self,fcn):
self.fcn = fcn
def __call__(self,*args):
""" Assumes 0th arg is instance reference (self) """
newargs = (args[0],) + tuple([[args[0].vocab.index(w)
if type(w) != int else w
for w in a] for a in args[1:]])
return self.fcn(*newargs)
#
# These functions build up the preferences for this Dirichlet Forest
#
def ML(self,w0,w1):
""" Add a simple must-link between single words w0 and w1 """
self.constr = combine(self.constr,self.merge(w0,w1))
def CL(self,w0,w1):
""" Add a simple cannot-link between single words w0 and w1 """
self.constr = combine(self.constr,self.split(w0,w1))
@Listify
@VocabConvert
def merge(self,w0,w1):
"""
Merge wordsets w0 and w1
(place Must-Links among and between)
"""
newML0 = [(w0[wi],w0[wi+1]) for wi in range(len(w0)-1)]
newML1 = [(w1[wi],w1[wi+1]) for wi in range(len(w1)-1)]
newML2 = [(w0[0],w1[0])]
self.constr = combine(self.constr,(newML0 + newML1 + newML2,[]))
@Listify
@VocabConvert
def split(self,*wsets):
"""
Split all wordsets in wsets
(place Must-Links among, Cannot-Links between)
"""
# Must-Link within wordsets
newML = reduce(lambda x,y: x+y,
[[(ws[wi],ws[wi+1])
for wi in range(len(ws)-1)]
for ws in wsets])
# Cannot-Link across all pairs of wordsets
newCL = []
for i in range(len(wsets)):
for j in range(i+1,len(wsets)):
newCL.append((wsets[i][0],wsets[j][0]))
self.constr = combine(self.constr,(newML, newCL))
@Listify
@VocabConvert
def stopword1(self,*args):
"""
Args is a Tuple of Lists, final List are stopwords by convention
-mustlink existing topics
-mustlink stopwords
-cannotlink stopwords vs all existing topics
"""
topics = args[:-1]
stop = args[-1]
newML = []
newCL = []
for t in topics:
(tml,tcl) = split(t,stop)
newML = newML + tml
newCL = newCL + tcl
self.constr = combine(self.constr,(newML,newCL))
@Listify
@VocabConvert
def stopword2(self,*args):
"""
Args is a Tuple of Lists, final List are stopwords by convention
-mustlink stopwords
-cannotlink stopwords vs all existing topics
"""
topics = args[:-1]
stop = args[-1]
newML = [(stop[wi],stop[wi+1]) for wi in range(len(stop)-1)]
newCL = []
alltopicwords = list(set(reduce(lambda x,y: x+y,topics)))
for ws in stop:
for wt in alltopicwords:
newCL.append((ws,wt))
self.constr = combine(self.constr,(newML,newCL))
DF-LDA-0.1.2/example/ 0000755 0003103 0000163 00000000000 11374252176 014040 5 ustar andrzeje medinfo DF-LDA-0.1.2/example/example.docs 0000644 0003103 0000163 00000000040 11303055575 016333 0 ustar andrzeje medinfo 0:1 2:1
1:1 2:1
0:1 2:1
1:1 2:1
DF-LDA-0.1.2/example/example.py 0000644 0003103 0000163 00000002615 11303055575 016045 0 ustar andrzeje medinfo """ Simple example demonstrating usage of Dirichlet Forest """
from numpy import *
import DirichletForest as DF
# Model parameters (see paper for meanings)
alpha = 1
beta = .01
eta = 100
# Number of topics, size of vocab
(T,W) = (2,3)
# Vocabulary
vocab = ['apple','banana','motorcycle']
# Documents, can be:
# -List of Lists of integers
# docs = [[0,2],
# [1,2],
# [0,2],
# [1,2]]
# -read in from SVMLight-style sparse text file
docs = DF.readDocs('example.docs')
# Build DF
df = DF.DirichletForest(alpha,beta,eta,T,W,vocab)
# Apply constraints
# Must-Link between apple and banana
df.merge('apple','banana')
# Note that any of these would be OK as well:
# df.merge('apple',[1])
# df.merge('apple',['banana'])
# df.merge(0,1)
# df.merge([0],[1])
# etc...
# Cannot-Link between apple and motorcycle
df.split('apple','motorcycle')
# Do inference on docs (take numsamp samples before estimating phi/theta)
(numsamp, randseed) = (50, 821945)
df.inference(docs,numsamp,randseed)
# Output results
print '\nTop 3 words from learned topics'
df.printTopics(N=3)
print '\nRaw inference output\n'
print 'phi = T-by-W NumPy matrix of topic-word probabilities'
print str(df.phi) + '\n'
print 'theta = D-by-T NumPy matrix of doc-topic probabilities'
print str(df.theta) + '\n'
print 'zsamp = final z-sample'
print str(df.zsamp) + '\n'
print 'qsamp = final q-sample'
print str(df.qsamp) + '\n'
DF-LDA-0.1.2/README 0000644 0003103 0000163 00000005553 11374245161 013271 0 ustar andrzeje medinfo DIRICHLET FOREST TOPIC MODELING (DF-LDA)
Version 0.1.2
David Andrzejewski (andrzeje@cs.wisc.edu)
Department of Computer Sciences
University of Wisconsin-Madison, USA
This software implements the Dirichlet Forest (DF) Prior [1] within
the Latent Dirichlet Allocation (LDA) model for discrete count data.
When combined with LDA [2], the Dirichlet Forest Prior allows the user
to encode domain knowledge into the prior on topic-word
multinomials. The inference method is Collapsed Gibbs sampling [3].
This code can also be used to do "standard" LDA by applying no domain
knowledge, or setting the "strength" parameter eta to 1, equivalent to
[3].
MCMC inference is implemented within a Python C++ extension module.
BUILD/INSTALL
Building this module requires installation of Python and NumPy, as
well as a C++ compiler. From the command-line, do:
% python setup.py install
(Note that if NumPy or Python are installed to non-standard locations,
you may need to make the appropriate changes in setup.py)
You can then test the installation with:
% python test/testPrefGraph.py -v
% python test/testDF.py -v
There is also a simple example scipt showing how to use the software:
% cd example
% python example.py
LOCAL INSTALL
If you do not have write access to your Python installation directories,
you will need to tell setup.py to install this module somewhere else.
For example:
% python setup.py install --prefix=~/local
will install the module under a subdirectory of your home directory called
"local".
It may then be necessary to let Python know where that is by setting
the PYTHONPATH environment variable (e.g., in .bashrc or .cshrc). For
our example this might involve adding something like the line:
setenv PYTHONPATH ~/local/lib/python2.5/site-packages
HOW TO USE
The commenting in the example/example.py script explains the meanings and
types of all input and return arguments. For more detailed examples
of good/bad inputs, you can examine the tests in test/testDF.py.
LICENSE
This software is open-source, released under the terms of the GNU
General Public License version 3, or any later version of the GPL (see
COPYING).
REFERENCES
[1] Andrzejewski, D., Zhu, X., and Craven, M.
Incorporating Domain Knowedge into Topic Modeling via Dirichlet Forest Priors
Proceedings of the 26th International Conference on Machine Learning
(ICML 2009)
[2] Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent dirichlet
allocation. Journal of Machine Learning Research (JMLR) 3
(Mar. 2003), 993-1022.
[3] Griffiths, T., and Steyvers, M. (2004). Finding Scientific
Topics. Proceedings of the National Academy of Sciences (PNAS), 101
(suppl. 1), 5228-5235.
VERSION HISTORY
0.1 Initial release
0.1.1 Fix in PrefGraph constructor for compatibility with older Pythons
(Python < 2.6 doesn't have tuple.index()...)
0.1.2 uint --> unsigned int for better compiler compatibility
DF-LDA-0.1.2/setup.py 0000644 0003103 0000163 00000001764 11374245100 014114 0 ustar andrzeje medinfo from distutils.core import Extension, setup
import os
from numpy.distutils.misc_util import *
numpyincl = get_numpy_include_dirs()
cpp_mod = Extension('intLDA',
sources = ['intLDA.cpp','dtnode.cpp',
'multinode.cpp','node.cpp'],
include_dirs = [os.getcwd()] + numpyincl,
library_dirs = [],
libraries = [],
extra_compile_args = ['-O3','-Wall'],
extra_link_args = [])
py_mods = ['ConstraintCompiler','DirichletForest',
'DirichletTreeNodes','PrefGraph','Reporter']
setup(name = 'DF-LDA',
description = 'LDA with Dirichlet Forest Prior',
version = '0.1.2',
author = 'David Andrzejewski',
author_email = 'andrzeje@cs.wisc.edu',
license = 'GNU General Public License (Version 3 or later)',
url = 'http://pages.cs.wisc.edu/~andrzeje/research/df_lda.html',
ext_modules = [cpp_mod],
py_modules = py_mods)
DF-LDA-0.1.2/PrefGraph.py 0000644 0003103 0000163 00000014076 11303055576 014642 0 ustar andrzeje medinfo """
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
"""
""" Contains PrefGraph class """
class PrefGraph:
"""
Graph represented as adjacency list
(self.adj: node index --> List of neighbors' indices)
"""
def __init__(self,nodes,edges):
"""
nodes = List of (hashable) nodes
edges = List of tuples (order doesn't matter, this is undirected graph)
"""
self.adj = dict([(n,set([(e[0] if n==e[1] else e[1])
for e in edges if n in e]))
for n in nodes])
def connected_components(self):
""" Connected components as List of Lists """
(remaining,visited) = (set(self.adj.keys()),set([]))
cc = []
while(remaining):
nbd = set([remaining.pop()])
while(nbd):
cur = nbd.pop()
visited.add(cur)
nbd = (set(self.adj[cur]) | nbd) - visited
cc.append(list(visited))
remaining = remaining - visited
visited = set([])
return cc
def subgraph(self,sg_nodes):
""" Subgraph """
return PrefGraph(sg_nodes,
set(reduce(lambda x,y: x+y,
[[(n,a) for a in self.adj[n]
if a in sg_nodes]
for n in sg_nodes])))
def complement(self):
""" Graph complement """
return PrefGraph(self.adj.keys(),
set(reduce(lambda x,y: x+y,
[[(n,a) for a in self.adj.keys()
if a != n and a not in self.adj[n]]
for n in self.adj.keys()])))
def cliques(self):
"""
Bron-Kerbosch maximal clique detection algorithm
(implementation based on ACM Algorithm 457 note)
"""
return self._extend([],self.adj.keys(),[])
def _extend(self,compsub,candlist,notlist):
""" Recursive clique extenstion fcn for branch-and-bound """
# Check recursive base cases
if((not candlist) and (not notlist)):
# Max clique condition
return [compsub]
elif([n for n in notlist
if self.connToAll(n,candlist)]):
# Bound (stop) condition
return []
# Otherwise recursively extend
cliques = []
# Fix point for B&K w/ "clever" cand select
#
fixpt = self.get_fixpt(candlist,notlist)
while(candlist):
# Clever candidate selection
#
if(fixpt in candlist):
cand = fixpt
candlist.remove(fixpt)
else:
disconnected_candidates = [c for c in candlist
if not self.connected(c,fixpt)]
if(disconnected_candidates):
cand = disconnected_candidates[0]
candlist.remove(cand)
else:
# If all remaining candidates are connected
# to the fixedpt in notlist, all recursive extensions
# will hit bound/terminate condition and return
return cliques
# Naive candidate selection
#
#cand = candlist.pop()
# Construct updated lists
r_compsub = compsub + [cand]
r_candlist = [n for n in candlist
if self.connected(n,cand)]
r_notlist = [n for n in notlist
if self.connected(n,cand)]
# Recursive extension call
cliques.extend(self._extend(r_compsub,r_candlist,r_notlist))
# Add extended candidate to notlist
notlist.append(cand)
return cliques
def get_fixpt(self,candlist,notlist):
"""
For Bron & Kerbosch Version 2 extension
Find candidate/notlist point with *fewest*
disconnections in candidate set
Picking pts disconnected to this fixed point
will get us to bound/termination condition fastest
"""
cscores = [len([c for c in candlist
if not self.connected(fp,c)])
for fp in candlist]
# Note: have to give notlist nodes an extra -1
# (nodes in candlist will have to be chosen before
# starting to decrement their counters)
nscores = [len([c for c in candlist
if not self.connected(fp,c)]) - 1
for fp in notlist]
# Find all points w/ minimal score, arbitrarily
# choose final point in this list
fixes = [fp for (fp,s) in zip(candlist+notlist,cscores+nscores)
if s==min(cscores+nscores)]
return fixes[-1]
def connToAll(self,node,rest):
""" Is node connected to all nodes in rest? """
if(rest):
return all([self.connected(node,other)
for other in rest])
else:
return True
def connected(self,a,b):
""" Are nodes A and B connected (or the same)? """
return b in self.adj[a] or b is a
DF-LDA-0.1.2/PKG-INFO 0000644 0003103 0000163 00000000474 11374252176 013507 0 ustar andrzeje medinfo Metadata-Version: 1.0
Name: DF-LDA
Version: 0.1.2
Summary: LDA with Dirichlet Forest Prior
Home-page: http://pages.cs.wisc.edu/~andrzeje/research/df_lda.html
Author: David Andrzejewski
Author-email: andrzeje@cs.wisc.edu
License: GNU General Public License (Version 3 or later)
Description: UNKNOWN
Platform: UNKNOWN
DF-LDA-0.1.2/multinode.h 0000644 0003103 0000163 00000003410 11363057154 014551 0 ustar andrzeje medinfo /**
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
*/
#include
#include
#include
#include
#include
#include
#define PY_ARRAY_UNIQUE_SYMBOL PyArray_API_intLDA
#define NO_IMPORT_ARRAY
#include
#include "node.h"
#ifndef unif
#define unif() ((double) rand()) / ((double) RAND_MAX)
#endif
using namespace std;
#ifndef H_MULTINODE
#define H_MULTINODE
/**
* Multinode
*/
class multinode: public node
{
protected:
vector variants;
vector > fake_leafmaps;
vector variant_logweights;
int y;
public:
multinode(PyObject* children, PyObject* maxind,
int leafstart, PyObject* variants);
virtual ~multinode();
virtual void modify_count(double val, int leaf);
virtual double calc_wordterm(double val, int leaf);
int num_variants();
int get_y();
double var_logweight(int given_y);
void set_y(int given_y);
double calc_logpwz(int given_y);
};
#endif
DF-LDA-0.1.2/dtnode.h 0000644 0003103 0000163 00000003345 11363057154 014035 0 ustar andrzeje medinfo /**
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
*/
#include
#include
#include
#include
#include
#include
#define PY_ARRAY_UNIQUE_SYMBOL PyArray_API_intLDA
#define NO_IMPORT_ARRAY
#include
using namespace std;
#include "node.h"
#ifndef H_DTNODE
#define H_DTNODE
#define OK 0
#define BAD 1
/**
* Dirichlet Tree node
*/
class dtnode: public node
{
protected:
PyArrayObject* edges;
double edgesum;
PyArrayObject* orig_edges;
double orig_edgesum;
public:
dtnode(PyArrayObject* edges, PyObject* children,
PyObject* maxind, int leafstart);
dtnode() {;}
void report();
void qreport();
vector get_yvec();
void y_sampling();
int y_init(PyArrayObject* yinit, int ti);
virtual void modify_count(double val, int leaf);
virtual double calc_wordterm(double val, int leaf);
virtual double calc_logpwz();
virtual ~dtnode();
};
#endif
DF-LDA-0.1.2/DirichletTreeNodes.py 0000644 0003103 0000163 00000005704 11303055576 016502 0 ustar andrzeje medinfo """
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
"""
"""
Contains DTNode, MLNode, MultiNode classes
(used by ConstraintCompiler to build DirichletForest data structure)
"""
class DTNode(object):
""" Represent a Dirichlet Tree node """
def __init__(self,edges,ichildren,maxind,leafstart):
"""
How to do this exactly...
"""
self.edges = edges
self.ichildren = ichildren
self.maxind = maxind
self.leafstart = leafstart
def numLeaves(self):
""" Get the number of leaves below this node """
n = sum([ic.numLeaves() for ic in self.ichildren])
return n + len(self.edges) - len(self.ichildren)
def tupleConvert(self):
"""
Given a Dirichlet Tree in node=List form,
recursively convert to node=Tuple form
"""
tchildren = [ic.tupleConvert() for ic in self.ichildren]
return (self.edges,tchildren,self.maxind,self.leafstart)
class MLNode(DTNode):
""" Represent a MustLink-node """
def __init__(self,edges,ichildren,maxind,leafstart,words):
self.edges = edges
self.ichildren = ichildren
self.maxind = maxind
self.leafstart = leafstart
self.words = words
class MultiNode(DTNode):
"""
Represent a MultiNode
(can take on different subtree structures)
"""
def __init__(self,edges,ichildren,maxind,leafstart,words,variants):
self.edges = edges
self.ichildren = ichildren
self.maxind = maxind
self.leafstart = leafstart
self.words = words
self.variants = variants
def tupleConvert(self):
""" For MultiNode, will need to also convert variants """
tchildren = [ic.tupleConvert() for ic in self.ichildren]
tvars = [(fakeroot.tupleConvert(),fake_leafmap)
for (fakeroot,fake_leafmap) in self.variants]
return (None,tchildren,self.maxind,self.leafstart,tvars)
def numLeaves(self):
""" Can't take len(edges) b/c edges==None by convention """
n = sum([ic.numLeaves() for ic in self.ichildren])
return n + len(self.words)
DF-LDA-0.1.2/ConstraintCompiler.py 0000644 0003103 0000163 00000022176 11374252151 016577 0 ustar andrzeje medinfo """
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
"""
"""
ConstraintCompiler module
Builds pairwise ML/CL constraints into DirichletForest
data structure (for passing along to C++ inference code)
"""
import operator as OP
from numpy import *
import PrefGraph as PG
from DirichletTreeNodes import *
def buildTree(mlcc,clcc,allowable,W,beta,eta):
"""
Given:
mlcc = must-link clusters of words (List of Lists)
clcc = cannot-link clusters of must-link clusters and words (List of Tuples)
allowable = within each clcc, the allowable combinations of mlcc/words
W = total number of words (connect extra words directly to root)
Return Dirichlet Forest data structure
(Dirichlet Tree plus special multinodes for structure variants)
PROCEDURE
1) construct tree structure
2) populate indices
3) tuple-ify tree for pass to intLDA
"""
# Build M-nodes for each mlcc (will have only leaf children)
ml_nodes = [MLNode(eta*beta*ones((len(Mi))),[],[],None,Mi) for Mi in mlcc]
# Build multinodes for each clcc
multinodes = []
for (cc,allow) in zip(clcc,allowable):
# these are the true output connections of the multinode
# ensure that ML-nodes always appear 1st
icids = [ci for ci in cc if ci >= W]
lcids = [ci for ci in cc if ci < W]
fake_words = icids + lcids
# construct a variation for each allowable set
variations = []
for aset in allow:
# split btwn good/bad (allowable/not)
good = aset
bad = [fl for fl in fake_words if fl not in good]
# scale the edges coming out of the likely internal node
aedges = []
for a in good:
if(a >= W):
# M-node? --> beta * num leaves
aedges.append(beta * ml_nodes[a-W].numLeaves())
else:
# Leaf? --> beta
aedges.append(beta)
# scale the edges coming out of the fakeroot
fedges = [eta * sum(aedges)]
for b in bad:
if(b >= W):
# M-node? --> beta * num leaves
fedges.append(beta * ml_nodes[b-W].numLeaves())
else:
# Leaf? --> beta
fedges.append(beta)
# variant subtree
likely_internal = DTNode(array(aedges),[],[],0)
fakeroot = DTNode(array(fedges),[likely_internal],
[len(good)-1],
len(good))
# fake leafmap permutation for this variant
fake_wordmap = good + bad
fake_leafmap = [fake_wordmap.index(wi)
for wi in fake_words]
# save it
variations.append((fakeroot,fake_leafmap))
# build the multinode
ichildren = [ml_nodes[ci-W] for ci in cc if ci >= W]
lchildren = [ci for ci in cc if ci < W]
multinodes.append(MultiNode(None,ichildren,[],0,lchildren,variations))
# Create empty root node
root = DTNode([],[],[],0)
# Start connecting things and populating indices
#
cur_ind = 0 # current leaf index
wordmap = [] # maps leaf index --> word index
# MultiNodes connected to root
#
for multinode in multinodes:
# Map ML-node children
for ml_child in multinode.ichildren:
# Get word indices under this M-node
wordmap = wordmap + ml_child.words
# Save leafstart index for this M-node
ml_child.leafstart = cur_ind
# Record maxind in multinode
cur_ind += len(ml_child.words)
multinode.maxind.append(cur_ind-1)
# Does this multinode connect directly to any leaves?
if(len(multinode.words) > 0):
# Record leaf start index
multinode.leafstart = cur_ind
cur_ind += len(multinode.words)
# Record word indices
wordmap = wordmap + multinode.words
# Put this C-node and maxindex in root
root.edges.append(beta*multinode.numLeaves())
root.ichildren.append(multinode)
root.maxind.append(cur_ind-1)
# ML-nodes connected directly to root
#
for mln in ml_nodes:
# Leaf start index == None --> has not been assigned yet
if(mln.leafstart == None):
# Set leaf start index for ML-node
mln.leafstart = cur_ind
cur_ind += mln.numLeaves()
# Put ML-node and its maxind in root
root.edges.append(beta*mln.numLeaves())
root.ichildren.append(mln)
root.maxind.append(cur_ind-1)
# Record word indices under this M-node
wordmap = wordmap + mln.words
# Finally, rest of words must be connected directly to root
root.edges = array(root.edges + [beta for i in range(W - cur_ind)])
root.leafstart = cur_ind
wordmap = wordmap + [wi for wi in range(W) if wi not in wordmap]
# Convert everything to tuples
root = root.tupleConvert()
# Recover leafmap by inverting wordmap permutation
leafmap = [wordmap.index(wi) for wi in range(W)]
return (root,leafmap)
def processPairwise(constr,W,vocab):
# Unpack pairwise to construct graphs
(mustlinks,cannotlinks) = constr
# Get ML connected components
mlcc = getML(mustlinks)
# Invert this mapping
if(len(mlcc) > 0):
invmlcc = dict(reduce(OP.concat,[[(wi,mli+W) for wi in mlcc[mli]]
for mli in range(len(mlcc))],[]))
else:
invmlcc = {}
# Translate cannot-links between words to cannot-links
# between ML components (when applicable)
newcannotlinks = map(lambda x: tuple(map(lambda y: invmlcc.get(y,y),x)),
cannotlinks)
# Check for conflicting constraints
#
# if 2 words are connected (transitively or directly) by must-links,
# they will be in the same ML connected component
#
# if 2 words also have a pairwise cannot-link btwn them, then after
# the mapping we will have a a pairwise cannot-link between an ML
# connected component and *itself*
#
conflicts = [cl[0] for cl in newcannotlinks if cl[0]==cl[1]]
if(len(conflicts) > 0):
print ''
# Identify the conflicting constraints
print '!!! CONFLICTING CONSTRAINTS !!!'
for conflict in conflicts:
print 'Following pairs must *and* cannot-link:'
mli = conflict - W
mlw = mlcc[mli]
for cl in cannotlinks:
if(cl[0] in mlw and cl[1] in mlw):
# Print words if available, else word indices
if(vocab != None):
print '\t%s - %s' % (vocab[cl[0]],vocab[cl[1]])
else:
print '\t%d - %d' % (cl[0],cl[1])
print ''
raise RuntimeError('Conflicting constraints, see err msg')
# Get CL graph connected components and allowable sets
(clcc,allowable) = getCL(newcannotlinks)
return (mlcc,clcc,allowable)
def getML(mustlinks):
"""
Get must-link connected components from pairwise mustlink constraints
"""
# Check if mustlinks is empty
if(len(mustlinks) == 0):
return []
# Get indices of all involved nodes
nodes = reduce(lambda x,y: x.union(y),
[set(ml) for ml in mustlinks],
set([]))
# Get must-link connected components
return (PG.PrefGraph(nodes,mustlinks)).connected_components()
def getCL(cannotlinks):
"""
Get cannot-link graph connected components, along with maximal independent
sets for each (the 'allowable' word sets)
"""
# Check if cannotlinks is empty
if(len(cannotlinks) == 0):
return ([],[])
# First get indices of all involved nodes
nodes = reduce(lambda x,y: x.union(y),
[set(cl) for cl in cannotlinks],set())
# Get cannot-link connected components
cl = PG.PrefGraph(nodes,cannotlinks)
clcc = cl.connected_components()
# For each connected component, enumerate all maximal independent sets
# (equivalent to all maximal cliques in the complement graph)
allowable = [((cl.subgraph(cc)).complement()).cliques()
for cc in clcc]
return (clcc,allowable)
DF-LDA-0.1.2/COPYING 0000644 0003103 0000163 00000104513 11374245113 013435 0 ustar andrzeje medinfo GNU 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.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU 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
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
Copyright (C)
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 3 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, see .
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
Copyright (C)
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
.
DF-LDA-0.1.2/node.h 0000644 0003103 0000163 00000003117 11363057154 013502 0 ustar andrzeje medinfo /**
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
*/
#include
#include
#include
#include
#include
using namespace std;
#ifndef H_NODE
#define H_NODE
// Uniform rand between [0,1] (inclusive)
#define unif() ((double) rand()) / ((double) RAND_MAX)
class node
{
protected:
vector ichildren;
vector maxind;
int leafstart;
public:
static double unif01();
static int mult_sample(double* vals, double sum);
static int mult_sample(vector vals, double sum);
static int log_mult_sample(vector vals);
static void set_randseed(unsigned int randseed);
vector get_multinodes();
virtual void modify_count(double val, int leaf);
virtual double calc_wordterm(double val, int leaf);
virtual ~node();
};
#endif
DF-LDA-0.1.2/node.cpp 0000644 0003103 0000163 00000006653 11363057152 014043 0 ustar andrzeje medinfo /**
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
*/
#include "node.h"
/**
* Uniform rand between [0,1] (inclusive)
*/
double node :: unif01()
{
return unif();
}
/**
* Use the randseed to init random number generator for node
*/
void node :: set_randseed(unsigned int randseed)
{
srand(randseed);
}
/**
* Draw a multinomial sample propto exp(vals)
* (i.e., given log-values, sample wrt actual values)
*
* Use logsumexp trick variant to avoid numerical problems
*
*/
int node :: log_mult_sample(vector vals)
{
// First, get max value
double maxval = vals[0];
for(unsigned int vi = 0; vi < vals.size(); vi++)
{
if(vals[vi] > maxval)
{
maxval = vals[vi];
}
}
// Now use max val to normalize in log-domain
vector newvals;
double normsum = 0;
for(unsigned int vi = 0; vi < vals.size(); vi++)
{
newvals.push_back(exp(vals[vi] - maxval));
normsum += exp(vals[vi] - maxval);
}
// Now just call standard multinomial sampler
return node::mult_sample(newvals,normsum);
}
/**
* Draw a multinomial sample propto vals
*
* (!!! we're assuming sum is the correct sum for vals !!!)
*
*/
int node :: mult_sample(double* vals, double norm_sum)
{
double rand_sample = unif() * norm_sum;
double tmp_sum = 0;
int j = 0;
while(tmp_sum < rand_sample || j == 0) {
tmp_sum += vals[j];
j++;
}
return j - 1;
}
// Simple overload for vector instead of array...
int node :: mult_sample(vector vals, double norm_sum)
{
double rand_sample = unif() * norm_sum;
double tmp_sum = 0;
int j = 0;
while(tmp_sum < rand_sample || j == 0) {
tmp_sum += vals[j];
j++;
}
return j - 1;
}
vector node :: get_multinodes()
{
vector multinodes;
for(unsigned int i = 0; i < ichildren.size(); i++)
{
// 1st, is this child a multinode?
string childtype(typeid((*(ichildren[i]))).name());
if(childtype.find("multinode") != string::npos)
{
multinodes.push_back(ichildren[i]);
}
// Next, recursively have children check for multinodes
vector childmulti =
(*(ichildren[i])).get_multinodes();
for(unsigned int ci = 0; ci < childmulti.size(); ci++)
{
multinodes.push_back(childmulti[ci]);
}
}
return multinodes;
}
void node :: modify_count(double val, int leaf)
{
return;
}
double node :: calc_wordterm(double val, int leaf)
{
return 0;
}
node :: ~node() {
//
// Since node is base class, this will be called by descendants
//
for(unsigned int ci = 0; ci < ichildren.size(); ci++)
delete ichildren[ci];
}
DF-LDA-0.1.2/intLDA.cpp 0000644 0003103 0000163 00000053620 11374252034 014223 0 ustar andrzeje medinfo /**
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
*/
#include "intLDA.h"
/**
* This is the exposed method which is called from Python
*/
static PyObject * intLDA(PyObject *self, PyObject *args, PyObject* keywds)
{
// Null-terminated list of arg keywords
//
static char* kwlist[] = {"docs","alpha","dirtree","leafmap","numsamp",
"randseed","f","init","verbose","sampout",
"burnin","interval","qinit",NULL};
// Required args
//
PyObject* docs_arg; // List of Lists
PyArrayObject* alpha; // NumPy Array
PyObject* dirtree; // Tuple
PyObject* leafmap; // List
int numsamp;
int randseed;
// Optional args
//
PyObject* f_arg = NULL; // List of f-labels foreach doc
PyObject* init = NULL; // List of Lists to initialize Gibbs
int verbose = 0; // 1 = verbose output (sampler progress)
int sampout = 0; // 1 = output actual samples
int burnin = 0; // Number of initial samples to throw away
int interval = 1; // Number of samples between printed samples
PyArrayObject* qinit = NULL; // NumPy Array
// Parse function args
//
if(!PyArg_ParseTupleAndKeywords(args,keywds,"O!O!O!O!ii|O!O!iiiiO!",kwlist,
&PyList_Type,&docs_arg,
&PyArray_Type,&alpha,
&PyTuple_Type,&dirtree,
&PyList_Type,&leafmap,
&numsamp,&randseed,
&PyList_Type,&f_arg,
&PyList_Type,&init,
&verbose,&sampout,
&burnin,&interval,
&PyArray_Type,&qinit))
// ERROR - bad args
return NULL;
// Init random number generator
//
node::set_randseed(static_cast(randseed));
// Use args to populate structs
// (also check for *validity*)
//
model_params* mp = new model_params;
dataset* ds = new dataset;
counts* c = new counts;
if(BAD == convert_args(docs_arg,alpha,dirtree,
leafmap,f_arg,mp,ds,c))
{
// Args bad! Return to Python...error condition should be set
delete mp;
delete ds;
delete c;
return NULL;
}
// Initialize model, and make sure init went OK
//
int initstatus = init_model(mp,ds,c,randseed,init,qinit);
if(initstatus == BAD)
{
// ERROR - something went wrong with our init
for(int ti=0; ti < mp->T; ti++)
{
delete c->topics[ti];
}
delete c;
Py_DECREF(mp->alphasum);
delete mp;
delete ds;
return NULL;
}
// Do burn-in
//
for(int si=0; si < burnin; si++)
{
gibbs_chain(mp,ds,c);
}
// Do the requested number of Gibbs samples
//
for(int si=0; si < numsamp; si++)
{
// In order to make it easier to test against deltaLDA
node::set_randseed(static_cast(randseed+si));
if(verbose == 1)
{
printf("Gibbs sample %d of %d\n",si,numsamp);
for(int ti=0; ti < mp->T; ti++)
{
printf("Topic %d q=",ti);
(*(c->topics[ti])).qreport();
printf("\n");
}
}
// Sample z values for each word,
// then sample y values for each topic
//
// (repeat times for good sampling practice)
//
for(int ii=0; ii < interval; ii++)
{
gibbs_chain(mp,ds,c);
for(int ti=0; ti < mp->T; ti++)
{
(*(c->topics[ti])).y_sampling();
}
}
// If sampout flag set, output
// -z-sample
// -q-sample
// -phi estimate
//
if(sampout == 1)
{
// z-sample
printf("z=");
int d,i;
for(d = 0; d < ds->D; d++)
{
for(i = 0; i < ds->doclens[d]; i++)
{
printf("%d",ds->sample[d][i]);
}
}
printf(" ");
// q-sample
printf("q=");
for(int ti=0; ti < mp->T; ti++)
{
(*(c->topics[ti])).qreport();
if(ti < (mp->T-1))
printf(",");
}
printf("\n");
// phi estimate
PyArrayObject* phi = est_phi(mp, ds, c);
printf("[");
for(int pt = 0; pt < mp->T; pt++)
{
printf("[");
for(int pw = 0; pw < ds->W; pw++)
{
double val = *((double*)PyArray_GETPTR2(phi,pt,pw));
if(pw < (ds->W - 1))
printf("%f,",val);
else
printf("%f",val);
}
if(pt < (mp->T -1))
printf("],");
else
printf("]");
}
printf("]\n");
Py_DECREF(phi);
}
}
// Estimate phi and theta
//
PyArrayObject* phi = est_phi(mp, ds, c);
PyArrayObject* theta = est_theta(mp, ds, c);
// Convert final z-sample back to List of Lists for return
//
PyObject* finalsamp = PyList_New(ds->D);
for(int d = 0; d < ds->D; d++)
{
PyObject* docz = PyList_New(ds->doclens[d]);
for(int i = 0; i < ds->doclens[d]; i++)
PyList_SetItem(docz,i,PyInt_FromLong(ds->sample[d][i]));
PyList_SetItem(finalsamp,d,docz);
}
// Get final q-sample
//
int P = ((*(c->topics[0])).get_multinodes()).size();
npy_intp* qdims = new npy_intp[2];
qdims[0] = mp->T;
qdims[1] = P;
PyArrayObject* qsamp = (PyArrayObject*)
PyArray_ZEROS(2,qdims,PyArray_INT,0);
delete[] qdims;
for(int ti = 0; ti < mp->T; ti++)
{
vector q = (*(c->topics[ti])).get_yvec();
for(int yi = 0; yi < P; yi++)
{
*((int*)PyArray_GETPTR2(qsamp,ti,yi)) = q[yi];
}
}
// Package phi, theta, and final sample in tuple for return
//
PyObject* retval = PyTuple_New(4);
PyTuple_SetItem(retval,0,(PyObject*) phi);
PyTuple_SetItem(retval,1,(PyObject*) theta);
PyTuple_SetItem(retval,2,finalsamp);
PyTuple_SetItem(retval,3,(PyObject*) qsamp);
// Do some final memory cleanup...
//
// Counts
Py_DECREF(c->nd);
for(int ti=0; ti < mp->T; ti++)
{
delete c->topics[ti];
}
delete c;
// Model params
Py_DECREF(mp->alphasum);
delete mp;
// Dataset
delete ds;
return (PyObject*) retval;
}
/**
* Do an "online" init of Gibbs chain, adding one word
* position at a time and then sampling for each new position
*/
static int init_model(model_params* mp, dataset* ds, counts* c,
int randseed,PyObject* zinit, PyArrayObject* qinit)
{
// Init y-values for each topic
//
int status = OK;
if(qinit != NULL)
{
// Initialize from user-provided values
// (checking these as we go)
if(mp->T != PyArray_DIM(qinit,0))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"q-init dimensionality mismatch - wrong # topics");
return BAD;
}
for(int ti=0; ti < mp->T; ti++)
{
status = (*(c->topics[ti])).y_init(qinit,ti);
if(status == BAD)
{
// ERROR - flag should already be set
return BAD;
}
}
}
else
{
// Initialize by sampling proportionally
//
for(int ti=0; ti < mp->T; ti++)
{
(*(c->topics[ti])).y_sampling();
}
}
// Re-seed random number generator
// (done to ensure standard LDA recovers identical results
// to DeltaLDA implementation for easier testing)
//
node::set_randseed(static_cast(randseed));
// If z-sample provided, has correct number of docs
if(zinit != NULL && ds->D != PyList_Size(zinit))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Number of docs / number of init samples mismatch");
return BAD;
}
// Init ND count matrix
npy_intp* nddims = new npy_intp[2];
nddims[0] = ds->D;
nddims[1] = mp->T;
c->nd = (PyArrayObject*) PyArray_ZEROS(2,nddims,PyArray_INT,0);
delete[] nddims;
// Build init z sample, one word at a time
//
ds->sample.resize(ds->D);
// Temporary array used for sampling
double* num = new double[mp->T];
// For each doc in corpus
for(int d = 0; d < ds->D; d++)
{
// Create this sample
vector newsample;
newsample.resize(ds->doclens[d]);
ds->sample[d] = newsample;
// Do we have a z-init or not?
//
if(zinit != NULL)
{
//
// INIT FROM Z-INIT
//
PyObject* docinit = PyList_GetItem(zinit,d);
// verify that this is a List
if(!PyList_Check(docinit))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Non-List element in initial sample");
Py_DECREF(c->nd);
delete[] num;
return BAD;
}
// verify that it has the right length
else if(ds->doclens[d] != PyList_Size(docinit))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Init sample/doc length mismatch");
Py_DECREF(c->nd);
delete[] num;
return BAD;
}
// Copy over values from List to our sample[]
for(int i = 0; i < ds->doclens[d]; i++)
{
// Get topic from init sample
int zi = PyInt_AsLong(PyList_GetItem(docinit,i));
if(zi < 0 || zi > (mp->T-1))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Bad init sample value");
Py_DECREF(c->nd);
delete[] num;
return BAD;
}
// Get the corresponding word
int wi = ds->docs[d][i];
// Store topic and increment counts
ds->sample[d][i] = zi;
(*((int*)PyArray_GETPTR2(c->nd,d,zi)))++;
(*(c->topics[zi])).modify_count(1,mp->leafmap[wi]);
}
}
else
{
//
// INIT "ONLINE"
//
for(int i = 0; i < ds->doclens[d]; i++)
{
int w_i = ds->docs[d][i];
// For each topic, calculate numerators
double norm_sum = 0;
for(int j = 0; j < mp->T; j++)
{
// Calculate numerator for this topic
// (NOTE: alpha denom omitted, since same for all topics)
double alpha_j =
*((double*)PyArray_GETPTR2(mp->alpha,ds->f[d],j));
double wordterm =
(*(c->topics[j])).calc_wordterm(1,mp->leafmap[w_i]);
num[j] =
wordterm * (*((int*)PyArray_GETPTR2(c->nd,d,j))+alpha_j);
norm_sum += num[j];
}
// Draw a sample
//
int newz = node::mult_sample(num,norm_sum);
// Update counts and initial sample vec
//
ds->sample[d][i] = newz;
(*((int*)PyArray_GETPTR2(c->nd,d,newz)))++;
(*(c->topics[newz])).modify_count(1,mp->leafmap[w_i]);
}
}
}
// Cleanup and return
//
delete[] num;
return OK;
}
/**
* Run Gibbs chain to get a new full sample
*/
void gibbs_chain(model_params* mp, dataset* ds, counts* c)
{
// Use Gibbs sampling to get a new z
// sample, one position at a time
//
// Temporary array used for sampling
double* num = new double[mp->T];
// For each doc in corpus
int d,j,i;
for(d = 0; d < ds->D; d++)
{
// For each word in doc
for(i = 0; i < ds->doclens[d]; i++)
{
// remove this w/z pair from all count/cache matrices
int z_i = ds->sample[d][i];
int w_i = ds->docs[d][i];
(*(c->topics[z_i])).modify_count(-1,mp->leafmap[w_i]);
(*((int*)PyArray_GETPTR2(c->nd,d,z_i)))--;
// For each topic, calculate numerators
double norm_sum = 0;
for(j = 0; j < mp->T; j++)
{
// Calculate numerator for each topic
// (NOTE: alpha denom omitted, since same for all topics)
double alpha_j =
*((double*)PyArray_GETPTR2(mp->alpha,ds->f[d],j));
double wordterm =
(*(c->topics[j])).calc_wordterm(1,mp->leafmap[w_i]);
num[j] =
wordterm * (*((int*)PyArray_GETPTR2(c->nd,d,j))+alpha_j);
norm_sum += num[j];
}
// Draw a sample
//
j = node::mult_sample(num,norm_sum);
// update count/cache matrices and sample vec
//
ds->sample[d][i] = j;
(*((int*)PyArray_GETPTR2(c->nd,d,j)))++;
(*(c->topics[j])).modify_count(1,mp->leafmap[w_i]);
}
}
// Just cleanup and return
// (new sample will be returned in ds->sample
//
delete[] num;
return;
}
/**
* Use final sample to estimate theta = P(z|d)
*/
PyArrayObject* est_theta(model_params* mp, dataset* ds, counts* c)
{
npy_intp* tdims = new npy_intp[2];
tdims[0] = ds->D;
tdims[1] = mp->T;
PyArrayObject* theta =
(PyArrayObject*) PyArray_ZEROS(2,tdims,PyArray_DOUBLE,0);
delete[] tdims;
PyArrayObject* rowsums =
(PyArrayObject*) PyArray_Sum(c->nd,1,PyArray_DOUBLE,NULL);
int d,t;
for(d = 0; d < ds->D; d++)
{
double rowsum = *((double*)PyArray_GETPTR1(rowsums,d));
int f = ds->f[d];
double alphasum = *((double*)PyArray_GETPTR1(mp->alphasum,f));
for(t = 0; t < mp->T; t++)
{
double alpha_t = *((double*)PyArray_GETPTR2(mp->alpha,f,t));
int ndct = *((int*)PyArray_GETPTR2(c->nd,d,t));
// Calc and assign theta entry
double newval = (ndct + alpha_t) / (rowsum+alphasum);
*((double*)PyArray_GETPTR2(theta,d,t)) = newval;
}
}
return theta;
}
/**
* Use final sample to estimate phi = P(w|z)
*/
PyArrayObject* est_phi(model_params* mp, dataset* ds, counts* c)
{
npy_intp* pdims = new npy_intp[2];
pdims[0] = mp->T;
pdims[1] = ds->W;
PyArrayObject* phi =
(PyArrayObject*) PyArray_ZEROS(2,pdims,PyArray_DOUBLE,0);
delete[] pdims;
int t,w;
for(t = 0; t < mp->T; t++)
{
for(w = 0; w < ds->W; w++)
{
*((double*)PyArray_GETPTR2(phi,t,w)) =
(*(c->topics[t])).calc_wordterm(1,mp->leafmap[w]);
}
}
return phi;
}
/**
* Simultaneously check args and populate structs
*/
static int convert_args(PyObject* docs_arg, PyArrayObject* alpha,
PyObject* dirtree, PyObject* leafmap,
PyObject* f_arg,
model_params* mp, dataset* ds, counts* c)
{
int i;
int fmax = 0; // Will need to ensure alpha dim is this big
int D = PyList_Size(docs_arg);
ds->D = D;
// If f-labels were not specified, just set all to zero
if(f_arg == NULL)
{
ds->f.assign(D,0);
}
// Otherwise convert and check
else
{
if(D != PyList_Size(f_arg))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"f and docs have different lengths");
return BAD;
}
for(i = 0; i < D; i++)
{
ds->f.push_back(PyInt_AsLong(PyList_GetItem(f_arg,i)));
// If PyInt_AsLong fails, it returns -1
// (also should not be getting neg f-values anyway)
if(ds->f[i] < 0)
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Non-numeric or negative f-value");
return BAD;
}
else if(ds->f[i] > fmax)
{
fmax = ds->f[i];
}
}
}
// Get info from parameters
int W = PyList_Size(leafmap);
ds->W = W;
int F = PyArray_DIM(alpha,0);
int T = PyArray_DIM(alpha,1);
mp->alpha = alpha;
mp->T = T;
mp->alphasum = (PyArrayObject*) PyArray_Sum(alpha,1,PyArray_DOUBLE,NULL);
// Check max F
if(fmax != (F - 1))
{
// ERROR
Py_DECREF(mp->alphasum);
PyErr_SetString(PyExc_RuntimeError,
"Alpha/f dimensionality mismatch");
return BAD;
}
// Check that all alpha values are non-negative
//
double alphamin = PyFloat_AsDouble(PyArray_Min(alpha,NPY_MAXDIMS,NULL));
if(alphamin < 0)
{
// ERROR
Py_DECREF(mp->alphasum);
PyErr_SetString(PyExc_RuntimeError,
"Negative Alpha value");
return BAD;
}
// Read in and check leafmap (must be a valid permutation)
//
int* valid = new int[W];
for(i = 0; i < W; i++)
valid[i] = 0;
mp->leafmap.reserve(W);
for(i = 0; i < W; i++)
{
mp->leafmap[i] = PyInt_AsLong(PyList_GetItem(leafmap,i));
if(mp->leafmap[i] < 0 || mp->leafmap[i] > (W-1))
{
// ERROR
Py_DECREF(mp->alphasum);
PyErr_SetString(PyExc_RuntimeError,
"Out-of-range value in leafmap");
delete[] valid;
return BAD;
}
valid[mp->leafmap[i]] = 1;
}
int sum = 0;
for(i = 0; i < W; i++)
sum += valid[i];
if(sum != W)
{
// ERROR
Py_DECREF(mp->alphasum);
PyErr_SetString(PyExc_RuntimeError,
"Out-of-range value in leafmap");
delete[] valid;
return BAD;
}
delete[] valid;
// Use dirtree arg to build Dirichlet Tree
// data structures for each topic
//
// Populate counts struct
// Build Dirichlet Tree for each topic
for(int ti = 0; ti < T; ti++)
{
try{
if(4 == PyTuple_Size(dirtree))
{
// Unpack tuple contents
PyArrayObject* cedges = (PyArrayObject*)
PyTuple_GetItem(dirtree,0);
PyObject* cchildren = PyTuple_GetItem(dirtree,1);
PyObject* cmaxind = PyTuple_GetItem(dirtree,2);
int cleafstart = PyInt_AsLong(PyTuple_GetItem(dirtree,3));
// Build root node of Dirichlet Tree
dtnode* newtopic = new dtnode(cedges,cchildren,
cmaxind,cleafstart);
c->topics.push_back(newtopic);
}
else{
// ERROR
Py_DECREF(mp->alphasum);
PyErr_SetString(PyExc_RuntimeError,
"Node has wrong number of elements");
return BAD;
}
}
catch (int err)
{
// ERROR - called fcn has set err condition
Py_DECREF(mp->alphasum);
return BAD;
}
}
// Convert documents from PyObject* to vector
//
int d;
ds->doclens.resize(D);
ds->docs.resize(D);
for(d = 0; d < D; d++)
{
PyObject* doc = PyList_GetItem(docs_arg,d);
if(!PyList_Check(doc))
{
// ERROR
for(int ti=0; ti < mp->T; ti++)
{
delete c->topics[ti];
}
Py_DECREF(mp->alphasum);
PyErr_SetString(PyExc_RuntimeError,
"Non-List element in docs List");
return BAD;
}
ds->doclens[d] = PyList_Size(doc);
vector newdoc;
newdoc.resize(ds->doclens[d]);
ds->docs[d] = newdoc;
for(i = 0; i < ds->doclens[d]; i++)
{
ds->docs[d][i] = PyInt_AsLong(PyList_GetItem(doc,i));
if(ds->docs[d][i] < 0 || ds->docs[d][i] > (W - 1))
{
// ERROR
for(int ti=0; ti < mp->T; ti++)
{
delete c->topics[ti];
}
Py_DECREF(mp->alphasum);
PyErr_SetString(PyExc_RuntimeError,
"Non-numeric or out of range word");
return BAD;
}
}
}
// Everything went fine, return OK code
//
return OK;
}
//
// PYTHON EXTENSION BOILERPLATE BELOW
//
// Defines the module method table
PyMethodDef methods[] =
{
{"intLDA", (PyCFunction) intLDA,
METH_VARARGS | METH_KEYWORDS, "Run Interactive LDA"},
{NULL, NULL, 0, NULL} // Is a 'sentinel' (?)
};
// This is a macro that does stuff for us (linkage, declaration, etc)
PyMODINIT_FUNC
initintLDA() // Passes method table to init our module
{
(void) Py_InitModule("intLDA", methods);
import_array(); // Must do this to satisfy NumPy (!)
}
DF-LDA-0.1.2/test/ 0000755 0003103 0000163 00000000000 11374252176 013364 5 ustar andrzeje medinfo DF-LDA-0.1.2/test/testPrefGraph.py 0000644 0003103 0000163 00000012605 11374251551 016514 0 ustar andrzeje medinfo import unittest
import pdb
import operator
import random
import time
import PrefGraph as PG
def exists(cond, seq):
""" For existential queries """
for val in seq:
if(cond(val)):
return True
return False
def forall(cond, seq):
""" For universal queries """
for val in seq:
if(not cond(val)):
return False
return True
def listeq(seq1,seq2):
""" Rough list equality operator """
return (len(seq1) == len(seq2)
and forall(lambda x: x in seq1,seq2))
def listcheck(results,targets):
# Forall lists in targets,
# exists a 'list-equal' list in results
return forall(
lambda y: exists(
lambda x: listeq(x,y),
results),
targets)
# Size of star graph for complementary cliques test
STAR_SIZE = 100
class TestPrefGraph(unittest.TestCase):
def setUp(self):
"""
Set up base data/parameter values
"""
nodes = range(10)
edges = [(0,1),
(1,2),
(3,4),
(4,5),
(5,6),
(8,9)]
self.pg = PG.PrefGraph(nodes,edges)
def testConnectedComponents(self):
""" Test connected_components """
cc = self.pg.connected_components()
target = [[0,1,2],[3,4,5,6],[7],[8,9]]
# Forall connected components in target,
# there exists a 'list-equal' component in cc
self.assert_(listcheck(cc,target))
def testSubgraphComplement(self):
""" Test complement and subgraph """
cc = self.pg.connected_components()
for c in cc:
csg = (self.pg.subgraph(c)).complement()
if(0 in c):
self.assert_(len(csg.adj.keys()) == 3 and
csg.connected(0,2) and
not csg.connected(0,1) and
not csg.connected(1,2))
elif(4 in c):
self.assert_(len(csg.adj.keys()) == 4 and
csg.connected(3,5) and
csg.connected(4,6) and
csg.connected(3,6) and
not csg.connected(3,4) and
not csg.connected(4,5) and
not csg.connected(5,6))
elif(7 in c):
self.assert_(len(csg.adj.keys()) == 1 and
len(csg.adj[7]) == 0)
elif(8 in c):
self.assert_(len(csg.adj.keys()) == 2 and
not csg.connected(8,9))
else:
# Should not reach here
self.assert_(1 == 0)
def testCliques(self):
""" Test Bron-Kerbosh max clique algorithm """
cc = self.pg.connected_components()
for c in cc:
if(0 in c):
targets = [[0,2],[1]]
elif(4 in c):
targets = [[3,5],[3,6],[4,6]]
elif(7 in c):
targets = [[7]]
elif(8 in c):
targets = [[8],[9]]
else:
# Should not reach here
self.assert_(1 == 0)
# Find max-cliques of complement subgraph
sgc = self.pg.subgraph(c).complement()
cliques = sgc.cliques()
# Forall maxcliques in target,
# exists a 'list-equal' clique in cliques
self.assert_(listcheck(cliques,targets))
def testLollipopClique(self):
""" Test cliques on lollipop graph """
nodes = range(5) + range(5,9)
edges = [(0,1),(1,2),(3,4),
(4,5),(4,6),(4,7),(4,8),
(5,6),(5,7),(5,8),
(6,7),(6,8),
(7,8)]
pg = PG.PrefGraph(nodes,edges)
cliques = pg.cliques()
targets = [[0,1],[1,2],[3,4],
[4,5,6,7,8]]
self.assert_(listcheck(cliques,targets))
def testStarClique(self):
""" Test cliques on complement of star graph """
start = time.clock()
nodes = range(STAR_SIZE)
edges = [(0,n) for n in range(1,STAR_SIZE)]
cpg = (PG.PrefGraph(nodes,edges)).complement()
cliques = cpg.cliques()
targets = [[0],range(1,STAR_SIZE)]
# Forall maxcliques in target,
# exists a 'list-equal' clique in cliques
self.assert_(listcheck(cliques,targets))
print '%f sec' % (time.clock() - start)
def testConnectedClique(self):
""" Test cliques on complete graph """
csize = 10
nodes = range(csize)
edges = reduce(lambda x,y: x+y,
[[(i,j) for j in range(i+1,csize)]
for i in range(csize-1)])
pg = PG.PrefGraph(nodes,edges)
cliques = pg.cliques()
targets = [range(csize)]
# Forall maxcliques in target,
# exists a 'list-equal' clique in cliques
self.assert_(listcheck(cliques,targets))
def testNoEdgeClique(self):
""" Test cliques on edge-free graph """
csize = 10
nodes = range(csize)
edges = []
pg = PG.PrefGraph(nodes,edges)
cliques = pg.cliques()
targets = [[i] for i in range(csize)]
# Forall maxcliques in target,
# exists a 'list-equal' clique in cliques
self.assert_(listcheck(cliques,targets))
# Run the unit tests!
if __name__ == '__main__':
unittest.main()
DF-LDA-0.1.2/test/testDF.py 0000644 0003103 0000163 00000041622 11303055575 015130 0 ustar andrzeje medinfo import unittest
import pdb
import operator
import random
import time
from numpy import *
import DirichletForest as DF
# If deltaLDA extension module present,
# we will sanity check against it
hasDelta = True
try:
from deltaLDA import deltaLDA
except:
hasDelta = False
class TestDF(unittest.TestCase):
def setUp(self):
""" Set up base data/parameter values """
self.T = 3
(self.alpha,self.beta,self.eta) = (.1,.1,10000)
self.ldaalpha = self.alpha * ones((1,self.T))
self.docs = [[1,1,2],
[1,1,1,1,2],
[3,3,3,4],
[3,3,4,4,3,3],
[0,0,0,0,0],
[0,0,0,0]]
self.init = [[0 for w in doc] for doc in self.docs]
self.f = [0 for doc in self.docs]
self.W = 7
self.numsamp = 50
self.randseed = 821945
# Build DF object
self.df = DF.DirichletForest(self.alpha,self.beta,self.eta,
self.T,self.W)
# Default constraint set will create a 'full' tree
# (ie, exhibits all variety of substructures)
self.df.split([1,2],[3])
self.df.split([0],[3])
self.df.merge([4],[5])
# equality tolerance for testing
self.tol = 1e-6
def matAgree(self,mat1,mat2):
"""
Given 2 NumPy matrices,
check that all values agree within self.tol
"""
return (abs(mat1-mat2)).sum(axis=None) < self.tol
def matProb(self,mat):
"""
Given a NumPy matrix,
check that all values >= 0, sum to 1 (valid prob dist)
"""
sumto1 = all([abs(val - float(1)) < self.tol
for val in mat.sum(axis=1)])
geq0 = all([val >= 0 for val in
mat.reshape(mat.size)])
return sumto1 and geq0
#
# These are to test *correct* operation
#
def testCC(self):
""" Test basic ConstraintCompiler operation """
W = 8
df = DF.DirichletForest(self.alpha,self.beta,self.eta,
self.T,W)
df.split([0,1,2,3],[4,5])
df.split([0,1,2,3],[6])
df.split([4,5],[7])
df.compile()
# ML clusters should be: 0123, 45
self.assert_(len(df.mlcc) == 2)
self.assert_(set(df.mlcc[0]) == set([0,1,2,3]))
self.assert_(set(df.mlcc[1]) == set([4,5]))
# CL contention set should be: 8,9,6,7
self.assert_(len(df.clcc) == 1)
self.assert_(set(df.clcc[0]) == set([8,9,6,7]))
# Agreeable sets should be: [0123]7,[45]6,67
self.assert_(len(df.allowable) == 1)
allowsets = [set(a) for a in df.allowable[0]]
correctsets = [set([8,7]),set([9,6]),set([6,7])]
matching = [a in correctsets for a in allowsets]
self.assert_(reduce(operator.__or__,matching))
def testSanity1(self):
""" Test no constraints vs deltaLDA implementation """
# Don't even try unless deltaLDA module present
if(not hasDelta):
return
# Randomly generated docs
docs = [[random.randint(self.W) for i in range(1000)]
for j in range(100)]
# Set beta for standard LDA
ldabeta = self.df.beta * ones((self.T,self.W))
# Run standard LDA
(sphi,stheta,ssample) = deltaLDA(docs,self.ldaalpha,
ldabeta,self.numsamp,self.randseed)
# Run Interactive LDA with empty constraint set
df = DF.DirichletForest(self.alpha,self.beta,self.eta,
self.T,self.W)
df.inference(docs,self.numsamp,self.randseed)
# Assert matrix agreement, valid prob dists
self.assert_(self.matAgree(df.phi,sphi))
self.assert_(self.matProb(df.phi))
self.assert_(self.matProb(sphi))
self.assert_(self.matAgree(df.theta,stheta))
self.assert_(self.matProb(df.theta))
self.assert_(self.matProb(stheta))
def testSanity2(self):
""" Test eta=1 against deltaLDA implementation """
# Don't even try unless deltaLDA module present
if(not hasDelta):
return
# Randomly generated docs
docs = [[random.randint(self.W) for i in range(1000)]
for j in range(100)]
# Set beta for standard LDA
ldabeta = self.beta * ones((self.T,self.W))
# Run standard LDA
start = time.time()
(sphi,stheta,ssample) = deltaLDA(docs,self.ldaalpha,
ldabeta,self.numsamp,self.randseed)
ldatime = time.time() - start
# Run Interactive LDA with the baseline (full-tree) constraints,
# but temporarily set constraint strength to 1 to build tree
eta = 1
df = DF.DirichletForest(self.alpha,self.beta,eta,
self.T,self.W)
df.split([1,2],[3])
df.split([0],[3])
df.merge([4],[5])
df.inference(docs,self.numsamp,self.randseed)
# Assert matrix agreement, valid prob dists
self.assert_(self.matAgree(df.phi,sphi))
self.assert_(self.matProb(df.phi))
self.assert_(self.matProb(sphi))
self.assert_(self.matAgree(df.theta,stheta))
self.assert_(self.matProb(df.theta))
self.assert_(self.matProb(stheta))
def testSanity3(self):
""" Test beta*eta=X vs deltaLDA with beta=X """
# Don't even try unless deltaLDA module present
if(not hasDelta):
return
# 'magic' X parameter
X = 50
# Randomly generated docs
W = 2
docs = [[random.randint(W) for i in range(1000)]
for j in range(100)]
# Set beta for standard LDA
ldabeta = X * ones((self.T,W))
# Run standard LDA
(sphi,stheta,ssample) = deltaLDA(docs,self.ldaalpha,
ldabeta,self.numsamp,self.randseed)
# Run Interactive LDA with the baseline (full-tree) constraints,
# but temporarily set constraint strength to 1 to build tree
eta = X
beta = 1
df = DF.DirichletForest(self.alpha,beta,eta,
self.T,W)
df.merge([0],[1])
df.inference(docs,self.numsamp,self.randseed)
# Assert matrix agreement, valid prob dists
self.assert_(self.matAgree(df.phi,sphi))
self.assert_(self.matProb(df.phi))
self.assert_(self.matProb(sphi))
self.assert_(self.matAgree(df.theta,stheta))
self.assert_(self.matProb(df.theta))
self.assert_(self.matProb(stheta))
def testBalance(self):
""" Test the 'balance' properties of constructed DF """
# Use baseline (full-tree) constraints, but temporarily set
# constraint strength eta=1 to build tree
eta = 1
df = DF.DirichletForest(self.alpha,self.beta,eta,
self.T,self.W)
df.split([1,2],[3])
df.split([0],[3])
df.merge([4],[5])
df.compile()
# Test balance condition for all internal edges
for (edge,ich) in zip(df.root[0],df.root[1]):
if(ich[0] == None):
# Is child a multinode?
for (fakeroot,fake_leafmap) in ich[4]:
self.assert_(edge == sum(fakeroot[0]))
# Check likely internal node
likeint = fakeroot[1][0]
self.assert_(fakeroot[0][0] == sum(likeint[0]))
# Check all ML-node children
for (iich,i) in zip(ich[1],range(len(ich[1]))):
fi = fake_leafmap[i]
# Under likely internal node, or not?
if(fi >= len(likeint[0])):
tmpi = fi - len(likeint[0]) + 1
self.assert_(fakeroot[0][tmpi] == sum(iich[0]))
else:
self.assert_(likeint[0][fi] == sum(iich[0]))
else:
# Otherwise it's an ML-node
self.assert_(edge == sum(ich[0]))
def testConstrained(self):
""" Test *split* constraint """
# True topics = [0,1,2,3],[4,5]
docs = [[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[4,5,4,5,4,5,4,5,4,5,4,5,0,1,2,3],
[4,5,4,5,4,5,4,5,4,5,4,5,0,1,2,3],
[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[4,5,4,5,4,5,4,5,4,5,4,5,0,1,2,3],
[4,5,4,5,4,5,4,5,4,5,4,5,0,1,2,3],
[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[0,1,2,3,0,1,2,3,0,1,2,3,4,5],
[4,5,4,5,4,5,4,5,4,5,4,5,0,1,2,3],
[4,5,4,5,4,5,4,5,4,5,4,5,0,1,2,3]]
# Now we want to 'force' recovery of
# 2 distinct topics [0,1] and [2,3]
# Build constraints
(T,W) = (2,6)
df = DF.DirichletForest(self.alpha,self.beta,self.eta,T,W)
df.split([0,1],[2,3])
randseed = 1
numsamp = 200
df.inference(docs,numsamp,randseed)
# Verify that we find:
# 1) a topics with high [0,1] and low [2,3]
# 2) a topics with low [0,1] and high [2,3]
(hit1,hit2) = (False,False)
phimax = df.phi.max(axis=1)
for t in range(df.phi.shape[0]):
tmax = phimax[t]
if(df.phi[t][0] >= .5*tmax and
df.phi[t][1] >= .5*tmax and
df.phi[t][2] < .5*tmax and
df.phi[t][3] < .5*tmax):
hit1 = True
elif(df.phi[t][2] >= .5*tmax and
df.phi[t][3] >= .5*tmax and
df.phi[t][0] < .5*tmax and
df.phi[t][1] < .5*tmax):
hit2 = True
self.assert_(hit1 and hit2)
# Check that phi/theta form valid probability distributions
self.assert_(self.matProb(df.phi))
self.assert_(self.matProb(df.theta))
def testStandard(self):
""" Test no constraints mode """
# Don't even try unless deltaLDA module present
if(not hasDelta):
return
# Temporarily shrink vocab
W = 5
# Set beta for standard LDA
ldabeta = self.beta * ones((self.T,W))
# Run standard LDA
(sphi,stheta,ssample) = deltaLDA(self.docs,self.ldaalpha,ldabeta,
self.numsamp,self.randseed)
# Run Interactive LDA with empty constraint set
df = DF.DirichletForest(self.alpha,self.beta,self.eta,
self.T,W)
df.inference(self.docs,self.numsamp,self.randseed)
#
# First, validate correctness of recovered topics
#
# theta should clust docs [0,1], [2,3], [4,5]
maxtheta = argmax(df.theta,axis=1)
self.assert_(maxtheta[0] == maxtheta[1])
self.assert_(maxtheta[2] == maxtheta[3])
self.assert_(maxtheta[4] == maxtheta[5])
# corresponding phi should emph [1,2], [3,4], [0]
maxphi = argmax(df.phi,axis=1)
self.assert_(maxphi[maxtheta[0]] == 1)
self.assert_(maxphi[maxtheta[2]] == 3)
self.assert_(maxphi[maxtheta[4]] == 0)
# Assert matrix agreement, valid prob dists
self.assert_(self.matAgree(df.phi,sphi))
self.assert_(self.matProb(df.phi))
self.assert_(self.matProb(sphi))
self.assert_(self.matAgree(df.theta,stheta))
self.assert_(self.matProb(df.theta))
self.assert_(self.matProb(stheta))
#
# These test that the program fails correctly on *bad input*
#
#
# Bad constraints
#
def testConstraintConflict(self):
""" Test conflicting constraints """
df = DF.DirichletForest(self.alpha,self.beta,self.eta,
self.T,self.W)
df.split([0,1],[2,3])
df.merge(1,2)
self.assertRaises(RuntimeError,df.compile)
#
# Bad data
#
def testNonListDoc(self):
""" Test non-list doc """
self.docs[0] = None
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testNegWord(self):
""" Test bad word (negative) """
self.docs[0][-1] = -1
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testBigWord(self):
""" Test bad word (too big) """
self.docs[0][-1] = self.W # should fail b/c of zero-based indexing
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testNonNumWord(self):
""" Test bad word (non-numeric) """
self.docs[0][-1] = ''
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
#
# Bad parameters
#
def testBadAlpha(self):
""" Test neg alpha """
self.df.alpha[0,1] = -1;
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
#
# Bad q-init samples
#
# !!! q-init MUST be constructed w/ dtype=int32 !!!
#
def testQInitTooManyT(self):
""" Test q-init w/ too many topics """
self.df.qsamp = zeros((self.T+1,1),int32)
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testQInitTooFewT(self):
""" Test q-init w/ too few topics """
self.df.qsamp = zeros((self.T-1,1),int32)
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testQInitTooManyQ(self):
""" Test q-init w/ too many multinode values """
self.df.qsamp = zeros((self.T,2),int32)
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testQInitTooFewQ(self):
""" Test q-init w/ too few multinode values """
self.df.qsamp = zeros((self.T,0),int32)
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testQInitBadVal(self):
""" Test out of range q-init """
# < 0
self.df.qsamp = zeros((self.T,1),int32)
self.df.qsamp[0,0] = -1
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
# > P
self.df.qsamp = zeros((self.T,1),int32)
self.df.qsamp[0,0] = 2
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
#
# Bad z-init samples
#
def testTooFewInit(self):
""" Test z-init w/ too few docs """
self.init = self.init[:-1]
self.df.zsamp = self.init
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testTooManyInit(self):
""" Test z-init w/ too many docs """
self.init = self.init + [[]]
self.df.zsamp = self.init
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testLenMisInit(self):
""" Test z-init w/ doc len mismatch """
self.init[2].append(0)
self.df.zsamp = self.init
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testNonListInit(self):
""" Test z-init w/ non-List doc """
self.init[0] = None
self.df.zsamp = self.init
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testNegInit(self):
""" Test z-init w/ neg value """
self.init[0][0] = -2
self.df.zsamp = self.init
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testBigInit(self):
""" Test z-init w/ too large value """
self.init[0][0] = 3
self.df.zsamp = self.init
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
def testNonNumInit(self):
""" Test z-init w/ non-numeric value """
self.init[0][0] = ''
self.df.zsamp = self.init
self.assertRaises(RuntimeError,self.df.inference,self.docs,
self.numsamp,self.randseed)
# Run the unit tests!
if __name__ == '__main__':
unittest.main()
DF-LDA-0.1.2/dtnode.cpp 0000644 0003103 0000163 00000024362 11374252104 014364 0 ustar andrzeje medinfo /**
This file is a part of
DF-LDA - Implementation of the Dirichlet Forest (DF) prior
for the Latent Dirichlet Allocation (LDA) model
Copyright (C) 2009 David Andrzejewski (andrzeje@cs.wisc.edu)
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 3 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, see .
*/
#include "dtnode.h"
#include "multinode.h"
/**
* Return the vector of y-values for this topic
*/
vector dtnode :: get_yvec()
{
vector multi = get_multinodes();
vector yvals;
for(unsigned int mi = 0; mi < multi.size(); mi++)
{
multinode* mu = dynamic_cast(ichildren[mi]);
yvals.push_back((*(mu)).get_y());
}
return yvals;
}
/**
* Init from a given set of y-values
*
*/
int dtnode :: y_init(PyArrayObject* yinit, int ti)
{
vector multi = get_multinodes();
// Make sure dimensionality is correct
if(multi.size() != PyArray_DIM(yinit,1))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"q-init dimensionality mismatch - wrong # q-values");
return BAD;
}
for(unsigned int mi = 0; mi < multi.size(); mi++)
{
multinode* mu = dynamic_cast(ichildren[mi]);
int yval = *((int*)PyArray_GETPTR2(yinit,ti,mi));
//printf("initing w/ yval =%d\n",yval);
// Make sure y-value in valid range [0,Q-1]
if(yval < 0 || yval >= (*mu).num_variants())
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Out-of-range value in q-init");
return BAD;
}
else
(*(mu)).set_y(yval);
}
return OK;
}
/**
* Do y-sampling for this topic
*/
void dtnode :: y_sampling()
{
// Sample a new y-value for each multinode
//
vector multi = get_multinodes();
for(unsigned int mi = 0; mi < multi.size(); mi++)
{
multinode* mu = dynamic_cast(ichildren[mi]);
// Get y-sampling values foreach variant
//
vector vals;
int numvar = (*(mu)).num_variants();
for(int vi = 0; vi < numvar; vi++)
{
// Log-contribution of the agreeable subtree
// +
// Log-contribution of likely set weight
//
vals.push_back((*(mu)).calc_logpwz(vi) +
(*(mu)).var_logweight(vi));
}
// Sample a new multinomial value from these log-likelihoods
int y = node::log_mult_sample(vals);
(*(mu)).set_y(y);
}
}
/**
* Do a simple q-report on multinodes
*/
void dtnode :: qreport()
{
vector multi = get_multinodes();
for(unsigned int mi = 0; mi < multi.size(); mi++)
{
multinode* mu = dynamic_cast(ichildren[mi]);
printf("%d",(*mu).get_y());
}
}
/**
* Do a report on multinodes
*/
void dtnode :: report()
{
vector multi = get_multinodes();
for(unsigned int mi = 0; mi < multi.size(); mi++)
{
multinode* mu = dynamic_cast(ichildren[mi]);
double e = *((double*)PyArray_GETPTR1(edges,mi));
printf("Edge = %f\n",e);
printf("Multinode %d, y = %d (of %d variants)\n",
mi,(*mu).get_y(),(*mu).num_variants());
}
}
/**
* Calculate the log of the (uncollapsed) P(w|z) expression
* (used for the y-sampling step)
*/
double dtnode :: calc_logpwz()
{
// Calculate for this node
double logpwz = lgamma(orig_edgesum) - lgamma(edgesum);
for(unsigned int ei = 0; ei < PyArray_DIM(edges,0); ei++)
{
logpwz += lgamma(*((double*)PyArray_GETPTR1(edges,ei))) -
lgamma(*((double*)PyArray_GETPTR1(orig_edges,ei)));
}
// Calculate for all interior children
for(unsigned int i = 0; i < ichildren.size(); i++)
{
dtnode* child = dynamic_cast(this->ichildren[i]);
logpwz += (*child).calc_logpwz();
}
return logpwz;
}
/**
* Calculate topic-word term of Gibbs sampling eqn
*/
double dtnode :: calc_wordterm(double val, int leaf)
{
// Is this leaf under one of our interior children?
double newval;
for(unsigned int i = 0; i < ichildren.size(); i++)
{
if(leaf <= maxind[i])
{
// Update value
newval = ((*((double*)PyArray_GETPTR1(this->edges,i)))
/ this->edgesum);
// Recurisively mult value by child
return (*(this->ichildren[i])).calc_wordterm(newval*val,leaf);
}
}
// No, this leaf must be one of our leaves
// Get leaf edge index
int ei = ichildren.size() + (leaf - this->leafstart);
// Update value
newval = (*((double*)PyArray_GETPTR1(this->edges,ei))) / this->edgesum;
return (val * newval);
}
/**
* Update all branch counts associated with a given leaf
*/
void dtnode :: modify_count(double val, int leaf)
{
// Is this leaf under one of our interior children?
for(unsigned int i = 0; i < ichildren.size(); i++)
{
if(leaf <= maxind[i])
{
// Update edge and sum
(*((double*)PyArray_GETPTR1(this->edges,i))) += val;
edgesum += val;
// Recursively update child cts
(*(ichildren[i])).modify_count(val,leaf);
return;
}
}
// No, this leaf must be one of our leaves
// Get leaf edge index
int ei = ichildren.size() + (leaf - leafstart);
// Update edge and sum
(*((double*)PyArray_GETPTR1(edges,ei))) += val;
edgesum += val;
return;
}
/**
* Constructor
*/
dtnode :: dtnode(PyArrayObject* arg_edges, PyObject* arg_children,
PyObject* arg_maxind, int arg_leafstart)
{
// Do some argument checking
if(!PyList_Check(arg_children) || !PyList_Check(arg_maxind))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Non-List children/maxind element in dirtree node");
throw 1;
}
if(arg_leafstart < 0)
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Negative leafstart value in dirtree node");
throw 1;
}
if(PyList_Size(arg_children) != PyList_Size(arg_maxind))
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Different sizes for children/maxind in dirtree node");
throw 1;
}
double edgemin =
PyFloat_AsDouble(PyArray_Min(arg_edges,NPY_MAXDIMS,NULL));
if(edgemin <= 0)
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Negative/zero edge value in dirtree node");
throw 1;
}
// Populate data members
int nci = PyList_Size(arg_children);
this->leafstart = arg_leafstart;
// Get edge values and sum
// (note that this *must* be a copy!)
npy_intp* edims = new npy_intp[1];
edims[0] = PyArray_DIM(arg_edges,0);
this->edges = (PyArrayObject*) PyArray_ZEROS(1,edims,PyArray_DOUBLE,0);
PyArray_CopyInto(this->edges,arg_edges);
this->edgesum = PyFloat_AsDouble(PyArray_Sum(this->edges,NPY_MAXDIMS,
PyArray_DOUBLE,NULL));
// Also make a copy of *original* edge values
// (not augmented by counts)
this->orig_edges = (PyArrayObject*) PyArray_ZEROS(1,edims,PyArray_DOUBLE,0);
PyArray_CopyInto(this->orig_edges,arg_edges);
this->orig_edgesum =
PyFloat_AsDouble(PyArray_Sum(this->orig_edges,NPY_MAXDIMS,
PyArray_DOUBLE,NULL));
delete[] edims;
// Recursively build children
//
int c;
for(c = 0; c < nci; c++)
{
// Get max leaf index under this child, check it
maxind.push_back(PyInt_AsLong(PyList_GetItem(arg_maxind,c)));
if(maxind[c] < 0)
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Negative maxind value in dirtree node");
throw 1;
}
// exploiting boolean short-circuit here...
if(c > 0 && maxind[c] <= maxind[c-1])
{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Non-increasing maxind value in dirtree node");
throw 1;
}
// Recursively build child
//
try{
// Check that node tuple size is correct
PyObject* pynode = PyList_GetItem(arg_children,c);
// Is child a dtnode, or a multinode?
//
if(4 == PyTuple_Size(pynode))
{
// dtnode
// Unpack tuple contents
PyArrayObject* cedges = (PyArrayObject*) PyTuple_GetItem(pynode,0);
PyObject* cchildren = PyTuple_GetItem(pynode,1);
PyObject* cmaxind = PyTuple_GetItem(pynode,2);
int cleafstart = PyInt_AsLong(PyTuple_GetItem(pynode,3));
// Build child dtnode
node* newchild = new dtnode(cedges,cchildren,
cmaxind,cleafstart);
ichildren.push_back(newchild);
}
else if(5 == PyTuple_Size(pynode))
{
// multinode
// Unpack tuple contents
PyObject* cchildren = PyTuple_GetItem(pynode,1);
PyObject* cmaxind = PyTuple_GetItem(pynode,2);
int cleafstart = PyInt_AsLong(PyTuple_GetItem(pynode,3));
PyObject* variants = PyTuple_GetItem(pynode,4);
// Build child multinode
node* newchild = new multinode(cchildren,cmaxind,
cleafstart,variants);
ichildren.push_back(newchild);
}
else{
// ERROR
PyErr_SetString(PyExc_RuntimeError,
"Node has wrong number of elements");
throw 1;
}
}
catch (int err)
{
throw 1;
}
}
}
/**
* Destuctor
*/
dtnode :: ~dtnode()
{
//
// Decrease ref cts on NumPy arrays, ichildren will be
// deleted by base class destructor (node)
//
Py_DECREF(edges);
Py_DECREF(orig_edges);
}