def __init__(self, sModelName, sModelDir,
iBlockVisibility = None,
iLineVisibility = None,
sComment = None,
C=None, tol=None, njobs=None, max_iter=None,
inference_cache=None):
DU_ABPTableRCut.iBlockVisibility = iBlockVisibility
DU_ABPTableRCut.iLineVisibility = iLineVisibility
DU_CRF_Task.__init__(self
, sModelName, sModelDir
, dFeatureConfig = {'row_row':{} , 'row_hdr':{}, 'row_sepH':{},
'hdr_row':{} , 'hdr_hdr':{}, 'hdr_sepH':{},
'sepH_row':{},'sepH_hdr':{}, 'sepH_sepH':{},
'row':{} , 'hdr':{} , 'sepH':{}}
, dLearnerConfig = {
'C' : .1 if C is None else C
, 'njobs' : 4 if njobs is None else njobs
, 'inference_cache' : 50 if inference_cache is None else inference_cache
#, 'tol' : .1
, 'tol' : .05 if tol is None else tol
, 'save_every' : 50 #save every 50 iterations,for warm start
, 'max_iter' : 10 if max_iter is None else max_iter
}
, sComment=sComment
#,cFeatureDefinition=FeatureDefinition_PageXml_StandardOnes_noText
,cFeatureDefinition=My_FeatureDefinition_v2
)
def __init__(self, sModelName, sModelDir, sComment=None):
DU_CRF_Task.__init__(self
, sModelName, sModelDir
, dFeatureConfig = {
'n_tfidf_node' : 10
, 't_ngrams_node' : (2,2)
, 'b_tfidf_node_lc' : False
, 'n_tfidf_edge' : 10
, 't_ngrams_edge' : (2,2)
, 'b_tfidf_edge_lc' : False
}
, dLearnerConfig = {
'C' : .1
, 'njobs' : 2
, 'inference_cache' : 10
, 'tol' : .1
, 'save_every' : 5 #save every 50 iterations,for warm start
#, 'max_iter' : 1000
, 'max_iter' : 2
}
, sComment=sComment
)
self.addBaseline_LogisticRegression() #use a LR model as baseline
def __init__(self, sModelName, sModelDir, sComment=None, C=None, tol=None, njobs=None, max_iter=None, inference_cache=None):
#another way to specify the graph class
# defining a getConfiguredGraphClass is preferred
self.configureGraphClass(self.DU_GRAPH)
DU_CRF_Task.__init__(self
, sModelName, sModelDir
, dLearnerConfig = {
'C' : .1 if C is None else C
, 'njobs' : 8 if njobs is None else njobs
, 'inference_cache' : 50 if inference_cache is None else inference_cache
#, 'tol' : .1
, 'tol' : .05 if tol is None else tol
, 'save_every' : 50 #save every 50 iterations,for warm start
, 'max_iter' : 1000 if max_iter is None else max_iter
}
, sComment=sComment
, cFeatureDefinition=FeatureDefinition_T_PageXml_StandardOnes_noText_v3
, dFeatureConfig = {
#config for the extractor of nodes of each type
"text": None,
"sprtr": None,
#config for the extractor of edges of each type
"text_text": None,
"text_sprtr": None,
"sprtr_text": None,
"sprtr_sprtr": None
}
)
traceln("- classes: ", self.DU_GRAPH.getLabelNameList())
self.bsln_mdl = self.addBaseline_LogisticRegression() #use a LR model trained by GridSearch as baseline
def __init__(self,
sModelName,
sModelDir,
iBlockVisibility=None,
iLineVisibility=None,
fCutHeight=None,
bCutAbove=None,
lRadAngle=None,
bTxt=None,
sComment=None,
cFeatureDefinition=None,
dFeatureConfig={},
C=None,
tol=None,
njobs=None,
max_iter=None,
inference_cache=None):
DU_ABPTableSkewedRowCutLine.iBlockVisibility = iBlockVisibility
DU_ABPTableSkewedRowCutLine.iLineVisibility = iLineVisibility
DU_ABPTableSkewedRowCutLine.fCutHeight = fCutHeight
DU_ABPTableSkewedRowCutLine.bCutAbove = bCutAbove
DU_ABPTableSkewedRowCutLine.lRadAngle = lRadAngle
DU_ABPTableSkewedRowCutLine.bTxt = bTxt
DU_CRF_Task.__init__(
self,
sModelName,
sModelDir,
dFeatureConfig={
'row_row': {},
'row_sepH': {},
'sepH_row': {},
'sepH_sepH': {},
'sepH': {},
'row': {}
},
dLearnerConfig={
'C':
.1 if C is None else C,
'njobs':
4 if njobs is None else njobs,
'inference_cache':
50 if inference_cache is None else inference_cache
#, 'tol' : .1
,
'tol':
.05 if tol is None else tol,
'save_every':
50 #save every 50 iterations,for warm start
,
'max_iter':
10 if max_iter is None else max_iter
},
sComment=sComment
#,cFeatureDefinition=FeatureDefinition_PageXml_StandardOnes_noText
,
cFeatureDefinition=My_FeatureDefinition_v3_txt
if self.bTxt else My_FeatureDefinition_v3)
def __init__(self,
sModelName,
sModelDir,
sComment=None,
C=None,
tol=None,
njobs=None,
max_iter=None,
inference_cache=None):
#NOTE: we might get a list in C tol max_iter inference_cache (in case of gridsearch)
DU_CRF_Task.__init__(
self,
sModelName,
sModelDir,
DU_GRAPH,
dFeatureConfig={
'nbClass': nbClass,
't_ngrams_node': (2, 4),
'b_node_lc': False,
't_ngrams_edge': (2, 4),
'b_edge_lc': False,
'n_jobs':
5 #n_jobs when fitting the internal Logit feat extractor model by grid search
},
dLearnerConfig={
'C':
.1 if C is None else C,
'njobs':
5 if njobs is None else njobs,
'inference_cache':
50 if inference_cache is None else inference_cache
#, 'tol' : .1
,
'tol':
.05 if tol is None else tol,
'save_every':
50 #save every 50 iterations,for warm start
,
'max_iter':
1000 if max_iter is None else max_iter
},
sComment=sComment,
cFeatureDefinition=FeatureDefinition_PageXml_LogitExtractorV2)
self.setNbClass(
nbClass
) #so that we check if all classes are represented in the training set
self.bsln_mdl = self.addBaseline_LogisticRegression(
) #use a LR model trained by GridSearch as baseline
def __init__(self,
sModelName,
sModelDir,
iGridVisibility=None,
iBlockVisibility=None,
sComment=None,
C=None,
tol=None,
njobs=None,
max_iter=None,
inference_cache=None):
DU_ABPTableRG.iGridVisibility = iGridVisibility
DU_ABPTableRG.iBlockVisibility = iBlockVisibility
DU_CRF_Task.__init__(
self,
sModelName,
sModelDir,
dFeatureConfig={
'row_row': {},
'row_gh': {},
'gh_row': {},
'gh_gh': {},
'gh': {},
'row': {}
},
dLearnerConfig={
'C':
.1 if C is None else C,
'njobs':
4 if njobs is None else njobs,
'inference_cache':
50 if inference_cache is None else inference_cache
#, 'tol' : .1
,
'tol':
.05 if tol is None else tol,
'save_every':
50 #save every 50 iterations,for warm start
,
'max_iter':
10 if max_iter is None else max_iter
},
sComment=sComment
#,cFeatureDefinition=FeatureDefinition_PageXml_StandardOnes_noText
,
cFeatureDefinition=My_FeatureDefinition)
self.cModelClass.setBalancedWeights(True)
def __init__(self, sModelName, sModelDir, sComment=None, C=None, tol=None, njobs=None, max_iter=None, inference_cache=None):
DU_CRF_Task.__init__(self
, sModelName, sModelDir
, dFeatureConfig = { }
, dLearnerConfig = {
'C' : .1 if C is None else C
, 'njobs' : 4 if njobs is None else njobs
, 'inference_cache' : 50 if inference_cache is None else inference_cache
#, 'tol' : .1
, 'tol' : .05 if tol is None else tol
, 'save_every' : 50 #save every 50 iterations,for warm start
, 'max_iter' : 10 if max_iter is None else max_iter
}
, sComment=sComment
#,cFeatureDefinition=FeatureDefinition_PageXml_StandardOnes_noText
,cFeatureDefinition=My_FeatureDefinition_v2
)
def __init__(self, sModelName, sModelDir, feat_select=None, sComment=None):
if feat_select == 'chi2':
DU_CRF_Task.__init__(self,
sModelName,
sModelDir,
DU_GRAPH,
dFeatureConfig=dFeatureConfig_FeatSelect,
dLearnerConfig=dLearnerConfig,
sComment=sComment)
else:
DU_CRF_Task.__init__(self,
sModelName,
sModelDir,
DU_GRAPH,
dFeatureConfig=dFeatureConfig_Baseline,
dLearnerConfig=dLearnerConfig,
sComment=sComment)
self.addBaseline_LogisticRegression() #use a LR model as baseline
def __init__(self,
sModelName,
sModelDir,
sComment=None,
C=None,
tol=None,
njobs=None,
max_iter=None,
inference_cache=None):
if sComment is None: sComment = sModelName
DU_CRF_Task.__init__(
self,
sModelName,
sModelDir,
dFeatureConfig={},
dLearnerConfig={
'C':
.1 if C is None else C,
'njobs':
2 if njobs is None else njobs,
'inference_cache':
50 if inference_cache is None else inference_cache
#, 'tol' : .1
,
'tol':
.05 if tol is None else tol,
'save_every':
50 #save every 50 iterations,for warm start
,
'max_iter':
200 if max_iter is None else max_iter
},
sComment=sComment
#,cFeatureDefinition=FeatureDefinition_PageXml_StandardOnes_noText
,
cFeatureDefinition=FeatureDefinition_PageXml_StandardOnes_noText_v3
)
#self.setNbClass(3) #so that we check if all classes are represented in the training set
if options.bBaseline:
self.bsln_mdl = self.addBaseline_LogisticRegression(
) #use a LR model trained by GridSearch as baseline
def __init__(self, sModelName, sModelDir, sComment=None, C=None, tol=None, njobs=None, max_iter=None, inference_cache=None):
if self.bHTR:
cFeatureDefinition = FeatureDefinition_PageXml_StandardOnes
dFeatureConfig = {
'n_tfidf_node':100, 't_ngrams_node':(1,2), 'b_tfidf_node_lc':False
, 'n_tfidf_edge':100, 't_ngrams_edge':(1,2), 'b_tfidf_edge_lc':False }
else:
cFeatureDefinition = FeatureDefinition_PageXml_StandardOnes_noText
dFeatureConfig = { }
#'n_tfidf_node':None, 't_ngrams_node':None, 'b_tfidf_node_lc':None
#, 'n_tfidf_edge':None, 't_ngrams_edge':None, 'b_tfidf_edge_lc':None }
DU_CRF_Task.__init__(self
, sModelName, sModelDir
, dFeatureConfig = dFeatureConfig
, dLearnerConfig = {
'C' : .1 if C is None else C
, 'njobs' : 16 if njobs is None else njobs
, 'inference_cache' : 50 if inference_cache is None else inference_cache
#, 'tol' : .1
, 'tol' : .05 if tol is None else tol
, 'save_every' : 50 #save every 50 iterations,for warm start
, 'max_iter' : 1000 if max_iter is None else max_iter
}
, sComment=sComment
, cFeatureDefinition=cFeatureDefinition
# , cFeatureDefinition=FeatureDefinition_T_PageXml_StandardOnes_noText
# , dFeatureConfig = {
# #config for the extractor of nodes of each type
# "text": None,
# "sprtr": None,
# #config for the extractor of edges of each type
# "text_text": None,
# "text_sprtr": None,
# "sprtr_text": None,
# "sprtr_sprtr": None
# }
)
traceln("- classes: ", self.getGraphClass().getLabelNameList())
self.bsln_mdl = self.addBaseline_LogisticRegression() #use a LR model trained by GridSearch as baseline
def runForExternalMLMethod(self,
lsColDir,
storeX,
applyY,
bRevertEdges=False):
"""
Return the list of produced files
"""
self.sXmlFilenamePattern = "*.mpxml"
return DU_CRF_Task.runForExternalMLMethod(self, lsColDir, storeX,
applyY, bRevertEdges)
except:
pass #we only need the transformer
lsOutputFilename = doer.runForExternalMLMethod(
lRun, options.storeX, options.applyY, options.bRevertEdges)
else:
doer.load()
lsOutputFilename = doer.predict(lRun)
traceln("Done, see in:\n %s" % lsOutputFilename)
# ----------------------------------------------------------------------------
if __name__ == "__main__":
version = "v.01"
usage, description, parser = DU_CRF_Task.getBasicTrnTstRunOptionParser(
sys.argv[0], version)
# parser.add_option("--annotate", dest='bAnnotate', action="store_true",default=False, help="Annotate the textlines with BIES labels")
#FOR GCN
parser.add_option("--revertEdges",
dest='bRevertEdges',
action="store_true",
help="Revert the direction of the edges")
parser.add_option("--detail",
dest='bDetailedReport',
action="store_true",
default=False,
help="Display detailed reporting (score per document)")
parser.add_option("--baseline",
dest='bBaseline',
action="store_true",
def predict(self, lsColDir):
"""
Return the list of produced files
"""
self.sXmlFilenamePattern = "*.mpxml"
return DU_CRF_Task.predict(self, lsColDir)
def main(DU_BAR):
version = "v.01"
usage, description, parser = DU_CRF_Task.getBasicTrnTstRunOptionParser(
sys.argv[0], version)
parser.add_option("--docid",
dest='docid',
action="store",
default=None,
help="only process docid")
# ---
#parse the command line
(options, args) = parser.parse_args()
# ---
try:
sModelDir, sModelName = args
except Exception as e:
traceln("Specify a model folder and a model name!")
_exit(usage, 1, e)
doer = DU_BAR(sModelName,
sModelDir,
C=options.crf_C,
tol=options.crf_tol,
njobs=options.crf_njobs,
max_iter=options.max_iter,
inference_cache=options.crf_inference_cache)
if options.docid:
sDocId = options.docid
else:
sDocId = None
if options.rm:
doer.rm()
sys.exit(0)
lTrn, lTst, lRun, lFold = [
_checkFindColDir(lsDir)
for lsDir in [options.lTrn, options.lTst, options.lRun, options.lFold]
]
# if options.bAnnotate:
# doer.annotateDocument(lTrn)
# traceln('annotation done')
# sys.exit(0)
## use. a_mpxml files
doer.sXmlFilenamePattern = doer.sLabeledXmlFilenamePattern
if options.iFoldInitNum or options.iFoldRunNum or options.bFoldFinish:
if options.iFoldInitNum:
"""
initialization of a cross-validation
"""
splitter, ts_trn, lFilename_trn = doer._nfold_Init(
lFold, options.iFoldInitNum, bStoreOnDisk=True)
elif options.iFoldRunNum:
"""
Run one fold
"""
oReport = doer._nfold_RunFoldFromDisk(options.iFoldRunNum,
options.warm)
traceln(oReport)
elif options.bFoldFinish:
tstReport = doer._nfold_Finish()
traceln(tstReport)
else:
assert False, "Internal error"
#no more processing!!
exit(0)
#-------------------
if lFold:
loTstRpt = doer.nfold_Eval(lFold, 3, .25, None, options.pkl)
import graph.GraphModel
sReportPickleFilename = os.path.join(sModelDir,
sModelName + "__report.txt")
traceln("Results are in %s" % sReportPickleFilename)
graph.GraphModel.GraphModel.gzip_cPickle_dump(sReportPickleFilename,
loTstRpt)
elif lTrn:
doer.train_save_test(lTrn, lTst, options.warm, options.pkl)
try:
traceln("Baseline best estimator: %s" %
doer.bsln_mdl.best_params_) #for GridSearch
except:
pass
traceln(" --- CRF Model ---")
traceln(doer.getModel().getModelInfo())
elif lTst:
doer.load()
tstReport = doer.test(lTst)
traceln(tstReport)
if lRun:
if options.storeX or options.applyY:
try:
doer.load()
except:
pass #we only need the transformer
lsOutputFilename = doer.runForExternalMLMethod(
lRun, options.storeX, options.applyY)
else:
doer.load()
lsOutputFilename = doer.predict(lRun)
traceln("Done, see in:\n %s" % lsOutputFilename)
def main_command_line(TableSkewedRowCut_CLASS):
version = "v.01"
usage, description, parser = DU_CRF_Task.getBasicTrnTstRunOptionParser(sys.argv[0], version)
# parser.add_option("--annotate", dest='bAnnotate', action="store_true",default=False, help="Annotate the textlines with BIES labels")
#FOR GCN
parser.add_option("--revertEdges", dest='bRevertEdges', action="store_true", help="Revert the direction of the edges")
parser.add_option("--detail", dest='bDetailedReport', action="store_true", default=False,help="Display detailed reporting (score per document)")
parser.add_option("--baseline", dest='bBaseline', action="store_true", default=False, help="report baseline method")
parser.add_option("--line_see_line", dest='iLineVisibility', action="store",
type=int, default=GraphSkewedCut.iLineVisibility,
help="seeline2line: how far in pixel can a line see another cut line?")
parser.add_option("--block_see_line", dest='iBlockVisibility', action="store",
type=int, default=GraphSkewedCut.iBlockVisibility,
help="seeblock2line: how far in pixel can a block see a cut line?")
parser.add_option("--height", dest="fCutHeight", default=GraphSkewedCut.fCutHeight
, action="store", type=float, help="Minimal height of a cut")
parser.add_option("--cut-above", dest='bCutAbove', action="store_true", default=False
,help="Each object defines one or several cuts above it (instead of below as by default)")
parser.add_option("--angle", dest='lsAngle'
, action="store", type="string", default="-1,0,+1"
,help="Allowed cutting angles, in degree, comma-separated")
parser.add_option("--graph", dest='bGraph', action="store_true", help="Store the graph in the XML for displaying it")
# ---
#parse the command line
(options, args) = parser.parse_args()
if options.bGraph:
import os.path
# hack
TableSkewedRowCut_CLASS.bCutAbove = options.bCutAbove
traceln("\t%s.bCutAbove=" % TableSkewedRowCut_CLASS.__name__, TableSkewedRowCut_CLASS.bCutAbove)
TableSkewedRowCut_CLASS.lRadAngle = [math.radians(v) for v in [float(s) for s in options.lsAngle.split(",")]]
traceln("\t%s.lRadAngle=" % TableSkewedRowCut_CLASS.__name__, TableSkewedRowCut_CLASS.lRadAngle)
for sInputFilename in args:
sp, sf = os.path.split(sInputFilename)
sOutFilename = os.path.join(sp, "graph-" + sf)
doer = TableSkewedRowCut_CLASS("debug", "."
, iBlockVisibility=options.iBlockVisibility
, iLineVisibility=options.iLineVisibility
, fCutHeight=options.fCutHeight
, bCutAbove=options.bCutAbove
, lRadAngle=[math.radians(float(s)) for s in options.lsAngle.split(",")])
o = doer.cGraphClass()
o.parseDocFile(sInputFilename, 9)
o.parseDocLabels()
o.addParsedLabelToDom()
o.addEdgaddEdgeToDoc print('Graph edges added to %s'%sOutFilename)
o.doc.write(sOutFilename, encoding='utf-8',pretty_print=True,xml_declaration=True)
SkewedCutAnnotator.gtStatReport()
exit(0)
# ---
try:
sModelDir, sModelName = args
except Exception as e:
traceln("Specify a model folder and a model name!")
_exit(usage, 1, e)
main(TableSkewedRowCut_CLASS, sModelDir, sModelName, options)
