def test_create_AP_labels(self):
model = Model(set(), collections.Counter(), dict(), set())
complex_parser = Parser("rate_complex")
complex_1 = complex_parser.parse("K(S{i},T{a}).B{o}::cyt").data.children[0]
complex_2 = complex_parser.parse("K(S{a},T{a}).B{o}::cyt").data.children[0]
complex_3 = complex_parser.parse("K(S{a},T{i}).B{o}::cyt").data.children[0]
complex_abstract = complex_parser.parse("K(S{a}).B{_}::cyt").data.children[0]
ordering = (complex_1, complex_2, complex_3)
APs = [Core.Formula.AtomicProposition(complex_abstract, " >= ", "3"),
Core.Formula.AtomicProposition(complex_1, " < ", 2)]
s1 = State(np.array((1, 2, 2)))
s2 = State(np.array((5, 1, 1)))
s3 = State(np.array((2, 4, 3)))
s4 = State(np.array((1, 4, 3)))
states_encoding = {s1: 1, s2: 2, s3: 3, s4: 4}
result_AP_lables = {APs[0]: 'property_0', APs[1]: 'property_1'}
result_state_labels = {1: {'property_0', 'property_1'},
3: {'property_0', 'init'},
4: {'property_0', 'property_1'}}
ts = TS.TransitionSystem.TransitionSystem(ordering)
ts.states_encoding = states_encoding
ts.init = 3
state_labels, AP_lables = model.create_AP_labels(APs, ts, 0)
self.assertEqual(state_labels, result_state_labels)
self.assertEqual(AP_lables, result_AP_lables)
def getDetector(detector, model=None, evaluator=None):
# Get the detector
if detector == None:
assert model != None
model = Model(model, "r")
detector = model.getStr("detector")
model.close()
return importClass(detector, "detector")
def test_create_complex_labels(self):
model = Model(set(), collections.Counter(), dict(), set())
complex_parser = Parser("rate_complex")
complex_1 = complex_parser.parse("K(S{i},T{a}).B{o}::cyt").data.children[0]
complex_2 = complex_parser.parse("K(S{a},T{a}).B{o}::cyt").data.children[0]
complex_3 = complex_parser.parse("K(S{a},T{i}).B{o}::cyt").data.children[0]
complex_abstract = complex_parser.parse("K(S{a}).B{_}::cyt").data.children[0]
ordering = (complex_1, complex_2, complex_3)
complexes = [complex_2, complex_abstract, complex_1]
result_labels = {complex_2: "VAR_1",complex_abstract: "ABSTRACT_VAR_12", complex_1: "VAR_0"}
result_formulas = ['ABSTRACT_VAR_12 = VAR_1+VAR_2; // K(S{a}).B{_}::cyt']
labels, prism_formulas = model.create_complex_labels(complexes, ordering)
self.assertEqual(labels, result_labels)
self.assertEqual(prism_formulas, result_formulas)
def GetTrainingData(self, inputDim, dataPercent):
if len(self.allDataList) == 0:
m = Model()
pares = m.GetAllPar()
totalData = m.GetTotalData()
partTotalDAta = totalData - int(totalData * (1 - dataPercent))
allData = []
for p in pares:
allData = allData + m.GetDataLimited(inputDim, p[0],
partTotalDAta)
print(p[0] + " " + str(len(allData)))
self.allDataList = allData
random.shuffle(self.allDataList)
traingData = self.allDataList[0:int(len(self.allDataList) * 0.7)]
validationData = self.allDataList[(int(len(self.allDataList) * 0.7)) +
1:len(self.allDataList)]
return traingData, validationData
def getDetector(detector, model=None):
# Get the detector
if detector == None:
assert model != None
model = Model(model, "r")
detector = model.getStr("detector")
model.close()
if type(detector) in types.StringTypes:
print >> sys.stderr, "Importing detector", detector
detectorName = detector
if detector.startswith("from"):
exec detector
detector = eval(detector.split(".")[-1])
else:
exec "from " + detector + " import " + detector.split(".")[-1]
detector = eval(detector.split(".")[-1])
else: # assume it is a class
detectorName = detector.__name__
print >> sys.stderr, "Using detector", detectorName
detector = detector
return detector, detectorName
def getDetector(detector, model=None):
# Get the detector
if detector == None:
assert model != None
model = Model(model, "r")
detector = model.getStr("detector")
model.close()
if type(detector) in types.StringTypes:
print >> sys.stderr, "Importing detector", detector
detectorName = detector
if detector.startswith("from"):
exec detector
detector = eval(detector.split(".")[-1])
else:
exec "from " + detector + " import " + detector.split(".")[-1]
detector = eval(detector.split(".")[-1])
else: # assume it is a class
detectorName = detector.__name__
print >> sys.stderr, "Using detector", detectorName
detector = detector
return detector, detectorName
def analyseData(self):
workspace = self.pre_form.value()
mdl = Model(workspace['Algorithm'], workspace['Parameters'])
if workspace['Learning Type'] == 'Clustering':
mdl.fitData(workspace['Data'].post_data)
pred = mdl.predictData(workspace['Data'].post_data)
else:
mdl.fitData(workspace['Data'].post_data['Training'].data,
workspace['Data'].post_data['Training'].labels)
pred = mdl.predictData(workspace['Data'].post_data['Testing'].data)
workspace['Model'] = mdl
workspace['Predicted'] = pred
self.post_options.close()
del (self.post_options)
self.post_options = PostAnalysisOptions(self, workspace)
self.rightScroll.setWidget(self.post_options)
self.built = True
def openModel(self, model, mode="r"):
if type(model) in types.StringTypes:
model = Model(model, mode)
self.modelsToClose.append(model)
return model
def setUp(self):
# agents
self.s1 = StructureAgent("X", set())
self.s2 = StructureAgent("Y", set())
self.s3 = StructureAgent("Z", set())
self.c1 = Complex([self.s1], "rep")
self.c2 = Complex([self.s2], "rep")
self.c3 = Complex([self.s3], "rep")
# rules
sequence_1 = (self.s1,)
mid_1 = 1
compartments_1 = ["rep"]
complexes_1 = [(0, 0)]
pairs_1 = [(0, None)]
rate_1 = Rate("k1*[X()::rep]")
self.r1 = Rule(sequence_1, mid_1, compartments_1, complexes_1, pairs_1, rate_1)
sequence_2 = (self.s3, self.s1)
mid_2 = 1
compartments_2 = ["rep"] * 2
complexes_2 = [(0, 0), (1, 1)]
pairs_2 = [(0, 1)]
self.r2 = Rule(sequence_2, mid_2, compartments_2, complexes_2, pairs_2, None)
sequence_3 = (self.s2,)
mid_3 = 0
compartments_3 = ["rep"]
complexes_3 = [(0, 0)]
pairs_3 = [(None, 0)]
rate_3 = Rate("1.0/(1.0+([X()::rep])**4.0)")
self.r3 = Rule(sequence_3, mid_3, compartments_3, complexes_3, pairs_3, rate_3)
# inits
self.inits = collections.Counter({self.c1: 2, self.c2: 1})
# defs
self.defs = {'k1': 0.05, 'k2': 0.12}
self.model = Model({self.r1, self.r2, self.r3}, self.inits, self.defs, set())
# model
self.model_str_1 = """
#! rules
X()::rep => @ k1*[X()::rep]
Z()::rep => X()::rep
=> Y()::rep @ 1/(1+([X()::rep])**4)
#! inits
2 X()::rep
Y()::rep
#! definitions
k1 = 0.05
k2 = 0.12
"""
self.model_parser = Parser("model")
self.model_str_2 = """
#! rules
X(K{i})::rep => X(K{p})::rep @ k1*[X()::rep]
X(T{a})::rep => X(T{o})::rep @ k2*[Z()::rep]
=> Y(P{f})::rep @ 1/(1+([X()::rep])**4)
#! inits
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
Y(P{g}, N{l})::rep
#! definitions
k1 = 0.05
k2 = 0.12
"""
# vectors
ordering = (self.c1, self.c2, self.c3)
self.rate_parser = Parser("rate")
rate_expr = "1/(1+([X()::rep])**4)"
rate_1 = Rate(self.rate_parser.parse(rate_expr).data)
rate_1.vectorize(ordering, dict())
rate_expr = "k1*[X()::rep]"
rate_2 = Rate(self.rate_parser.parse(rate_expr).data)
rate_2.vectorize(ordering, {"k1": 0.05})
init = State(np.array([2, 1, 0]))
vector_reactions = {VectorReaction(State(np.array([0, 0, 0])), State(np.array([0, 1, 0])), rate_1),
VectorReaction(State(np.array([1, 0, 0])), State(np.array([0, 0, 0])), rate_2),
VectorReaction(State(np.array([0, 0, 1])), State(np.array([1, 0, 0])), None)}
self.vm_1 = VectorModel(vector_reactions, init, ordering, None)
# wrong models
self.model_wrong_1 = \
"""#! rules
X(K{i})::rep => X(K{p})::rep @ k1*[X()::rep]
X(T{a})::rep => X(T{o}):;rep @ k2*[Z()::rep]
=> Y(P{f})::rep @ 1/(1+([X()::rep])**4)
#! inits
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
Y(P{g}, N{l})::rep
#! definitions
k1 = 0.05
k2 = 0.12
"""
self.model_wrong_2 = \
"""#! rules
X(K{i})::rep => X(K{p})::rep @ k1*[X()::rep]
X(T{a})::rep = X(T{o})::rep @ k2*[Z()::rep]
=> Y(P{f})::rep @ 1/(1+([X()::rep])**4)
#! inits
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
Y(P{g}, N{l})::rep
#! definitions
k1 = 0.05
k2 = 0.12
"""
self.model_with_comments = """
#! rules
// commenting
X(K{i})::rep => X(K{p})::rep @ k1*[X()::rep] // also here
X(T{a})::rep => X(T{o})::rep @ k2*[Z()::rep]
=> Y(P{f})::rep @ 1/(1+([X()::rep])**4) // ** means power (^)
#! inits
// here
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
Y(P{g}, N{l})::rep // comment just 1 item
#! definitions
// and
k1 = 0.05 // also
k2 = 0.12
"""
self.model_with_complexes = """
#! rules
// commenting
X(T{a}):XX::rep => X(T{o}):XX::rep @ k2*[X().X()::rep]
K{i}:X():XYZ::rep => K{p}:X():XYZ::rep @ k1*[X().Y().Z()::rep] // also here
=> P{f}:XP::rep @ 1/(1+([X().P{_}::rep])**4) // ** means power (^)
#! inits
// here
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
Y(P{g}, N{l})::rep // comment just 1 item
#! definitions
// and
k1 = 0.05 // also
k2 = 0.12
#! complexes
XYZ = X().Y().Z() // a big complex
XX = X().X()
XP = X().P{_}
"""
self.model_without_complexes = """
#! rules
// commenting
X(T{a}).X()::rep => X(T{o}).X()::rep @ k2*[X().X()::rep]
X(K{i}).Y().Z()::rep => X(K{p}).Y().Z()::rep @ k1*[X().Y().Z()::rep] // also here
=> X().P{f}::rep @ 1/(1+([X().P{_}::rep])**4) // ** means power (^)
#! inits
// here
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
Y(P{g}, N{l})::rep // comment just 1 item
#! definitions
// and
k1 = 0.05 // also
k2 = 0.12
"""
self.model_with_variable = """
#! rules
// commenting
T{a}:X():?::rep => T{o}:X():?::rep @ k2*[X().X()::rep] ; ? = { XX, XY }
K{i}:X():XY::rep => K{p}:X():XY::rep @ k1*[X().Y().Z().X()::rep] // also here
#! inits
// here
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
#! definitions
// and
k1 = 0.05 // also
k2 = 0.12
#! complexes
XX = X().X()
XY = X().Y()
"""
self.model_without_variable = """
#! rules
// commenting
X(K{i}).Y()::rep => X(K{p}).Y()::rep @ k1*[X().Y().Z().X()::rep]
X(T{a}).X()::rep => X(T{o}).X()::rep @ k2*[X().X()::rep]
X(T{a}).Y()::rep => X(T{o}).Y()::rep @ k2*[X().X()::rep]
#! inits
// here
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
#! definitions
// and
k1 = 0.05 // also
k2 = 0.12
"""
self.model_with_redundant = """
#! rules
K(S{u}).B()::cyt => K(S{p})::cyt + B()::cyt + D(A{_})::cell @ 3*[K().B()::cyt]/2*v_1
K().B()::cyt => K()::cyt + B()::cyt + D(A{_})::cell @ 3*[K().B()::cyt]/2*v_1
K().K()::cyt => K()::cyt + K()::cyt
K(S{i}).K()::cyt => K(S{a})::cyt + K()::cyt
K(S{i}, T{p}).K()::cyt => K(S{a}, T{p})::cyt + K()::cyt
#! inits
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
#! definitions
v_1 = 0.05
k2 = 0.12
"""
self.model_without_redundant = """
#! rules
K().B()::cyt => K()::cyt + B()::cyt + D(A{_})::cell @ 3*[K().B()::cyt]/2*v_1
K().K()::cyt => K()::cyt + K()::cyt
#! inits
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
#! definitions
v_1 = 0.05
k2 = 0.12
"""
self.model_with_context = """
#! rules
K(S{i}).B(T{a})::cyt => K(S{i})::cyt + B(T{a})::cyt @ 3*[K(S{i}).B(T{a})::cyt]/2*v_1
A{p}.K(S{i},T{i})::cyt => A{i}::cyt + K(S{a},T{a})::cyt
K(S{i},T{i})::cyt => K(S{a},T{i})::cyt
#! inits
2 K(S{i}).B(T{a})::cyt
1 A{p}.K(S{i},T{i})::cyt
#! definitions
v_1 = 0.05
k2 = 0.12
"""
self.model_without_context = """
#! rules
K().B()::cyt => K()::cyt + B()::cyt @ 3*[K().B()::cyt]/2*v_1
A{_}.K()::cyt => A{_}::cyt + K()::cyt
#! inits
2 K().B()::cyt
1 A{_}.K()::cyt
#! definitions
v_1 = 0.05
k2 = 0.12
"""
self.model_reachable = """
#! rules
K(S{i}).B()::cyt => K(S{a})::cyt + B()::cyt @ 3*[K(S{i}).B()::cyt]/2*v_1
K(S{a})::cyt + A{i}::cyt => K(S{a}).A{i}::cyt
K().A{i}::cyt => K().A{a}::cyt
#! inits
2 K(S{i}).B()::cyt
1 A{i}::cyt
#! definitions
v_1 = 0.05
k2 = 0.12
"""
self.model_nonreachable = """
#! rules
K(S{i}).B()::cyt => K(S{a})::cyt + B()::cyt @ 3*[K(S{i}).B()::cyt]/2*v_1
K(S{a})::cyt + A{i}::cyt => K(S{a}).A{i}::cyt
#! inits
2 K(S{i}).B()::cyt
1 A{i}::cyt
#! definitions
v_1 = 0.05
k2 = 0.12
"""
self.model_parametrised = """
#! rules
// commenting
X(K{i})::rep => X(K{p})::rep @ k1*[X()::rep] // also here
X(T{a})::rep => X(T{o})::rep @ k2*[Z()::rep]
=> Y(P{f})::rep @ 1/(v_3+([X()::rep])**4) // ** means power (^)
#! inits
2 X(K{c}, T{e}).X(K{c}, T{j})::rep
Y(P{g}, N{l})::rep // comment just 1 item
#! definitions
k1 = 0.05
"""
self.miyoshi = """
def train(output, task=None, detector=None, inputFiles=None, models=None, parse=None,
processUnmerging=None, processModifiers=None, isSingleStage=False,
bioNLPSTParams=None, preprocessorParams=None, exampleStyles=None,
classifierParams=None, doFullGrid=False, deleteOutput=False, copyFrom=None,
log="log.txt", step=None, omitSteps=None, debug=False, connection=None):
"""
Train a new model for event or relation detection.
@param output: A directory where output files will appear.
@param task: If defined, overridable default settings are used for many of the training parameters. Must be one of the supported TEES tasks.
@param detector: a Detector object, or a string defining one to be imported
@param inputFiles: A dictionary of file names, with keys "train", "devel" and, "test"
@param models: A dictionary of file names defining the place for the new models, with keys "devel" and, "test"
@param parse: The parse element name in the training interaction XML
@param processUnmerging: Use the unmerging step of EventDetector. True, False or None for task default.
@param processModifiers: Use the modifier detection step of EventDetector. True, False or None for task default.
@param isSingleStage: False for EventDetector, True for a single stage detector.
@param bioNLPSTParams: Parameters controlling BioNLP ST format output.
@param preprocessorParams: Parameters controlling the preprocessor. Not used for training, but saved to the model for use when classifying.
@param exampleStyles: A parameter set for controlling example builders.
@param classifierParams: A parameter set for controlling classifiers.
@param doFullGrid: Whether all parameters, as opposed to just recall adjustment, are tested in the EventDetector grid search.
@param deleteOutput: Remove an existing output directory
@param copyFrom: Copy an existing output directory for use as a template
@param log: An optional alternative name for the log file. None is for no logging.
@param step: A step=substep pair, where the steps are "TRAIN", "DEVEL", "EMPTY" and "TEST"
@param omitSteps: step=substep parameters, where multiple substeps can be defined.
@param debug: In debug mode, more output is shown, and some temporary intermediate files are saved
@param connection: A parameter set defining a local or remote connection for training the classifier
"""
# Insert default arguments where needed
inputFiles = Parameters.get(inputFiles, {"train":None, "devel":None, "test":None})
models = Parameters.get(models, {"devel":None, "test":None})
exampleStyles = Parameters.get(exampleStyles, {"examples":None, "trigger":None, "edge":None, "unmerging":None, "modifiers":None})
classifierParams = Parameters.get(classifierParams, {"examples":None, "trigger":None, "recall":None, "edge":None, "unmerging":None, "modifiers":None})
processUnmerging = getDefinedBool(processUnmerging)
processModifiers = getDefinedBool(processModifiers)
# Initialize working directory
workdir(output, deleteOutput, copyFrom, log)
# Get task specific parameters
detector, processUnmerging, processModifiers, isSingleStage, bioNLPSTParams, preprocessorParams, exampleStyles, classifierParams, removeNamesFromEmpty = getTaskSettings(task,
detector, processUnmerging, processModifiers, isSingleStage, bioNLPSTParams, preprocessorParams, inputFiles, exampleStyles, classifierParams)
if task != None: task = task.replace("-MINI", "").replace("-FULL", "")
# Define processing steps
selector, detectorSteps, omitDetectorSteps = getSteps(step, omitSteps, ["TRAIN", "DEVEL", "EMPTY", "TEST"])
# Initialize the detector
detector, detectorName = getDetector(detector)
detector = detector() # initialize object
detector.debug = debug
detector.bioNLPSTParams = detector.getBioNLPSharedTaskParams(bioNLPSTParams)
#detector.useBioNLPSTFormat = useBioNLPSTFormat # classify-output and grid evaluation in ST-format
#detector.stWriteScores = True # write confidence scores into additional st-format files
connection = getConnection(connection)
detector.setConnection(connection)
connection.debug = debug
if deleteOutput:
connection.clearWorkDir()
# Train
if selector.check("TRAIN"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------------ Train Detector ------------------"
print >> sys.stderr, "----------------------------------------------------"
if isSingleStage:
detector.train(inputFiles["train"], inputFiles["devel"], models["devel"], models["test"],
exampleStyles["examples"], classifierParams["examples"], parse, None, task,
fromStep=detectorSteps["TRAIN"], workDir="training")
else:
detector.train(inputFiles["train"], inputFiles["devel"], models["devel"], models["test"],
exampleStyles["trigger"], exampleStyles["edge"], exampleStyles["unmerging"], exampleStyles["modifiers"],
classifierParams["trigger"], classifierParams["edge"], classifierParams["unmerging"], classifierParams["modifiers"],
classifierParams["recall"], processUnmerging, processModifiers,
doFullGrid, task, parse, None,
fromStep=detectorSteps["TRAIN"], workDir="training")
# Save the detector type
for model in [models["devel"], models["test"]]:
if os.path.exists(model):
model = Model(model, "a")
model.addStr("detector", detectorName)
if preprocessorParams != None:
preprocessor = Preprocessor()
model.addStr("preprocessorParams", Parameters.toString(preprocessor.getParameters(preprocessorParams)))
model.save()
model.close()
if selector.check("DEVEL"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------ Check devel classification ------------"
print >> sys.stderr, "----------------------------------------------------"
detector.classify(inputFiles["devel"], models["devel"], "classification-devel/devel", goldData=inputFiles["devel"], fromStep=detectorSteps["DEVEL"], workDir="classification-devel")
if selector.check("EMPTY"):
# By passing an emptied devel set through the prediction system, we can check that we get the same predictions
# as in the DEVEL step, ensuring the model does not use leaked information.
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------ Empty devel classification ------------"
print >> sys.stderr, "----------------------------------------------------"
detector.classify(getEmptyCorpus(inputFiles["devel"], removeNames=removeNamesFromEmpty), models["devel"], "classification-empty/devel-empty", fromStep=detectorSteps["EMPTY"], workDir="classification-empty")
if selector.check("TEST"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------- Test set classification --------------"
print >> sys.stderr, "----------------------------------------------------"
if inputFiles["test"] == None or not os.path.exists(inputFiles["test"]):
print >> sys.stderr, "Skipping, test file", inputFiles["test"], "does not exist"
else:
detector.bioNLPSTParams["scores"] = False # the evaluation server doesn't like additional files
detector.classify(inputFiles["test"], models["test"], "classification-test/test", fromStep=detectorSteps["TEST"], workDir="classification-test")
if detector.bioNLPSTParams["convert"]:
Utils.STFormat.Compare.compare("classification-test/test-events.tar.gz", "classification-devel/devel-events.tar.gz", "a2")
def train(output,
task=None,
detector=None,
inputFiles=None,
models=None,
parse=None,
processUnmerging=None,
processModifiers=None,
bioNLPSTParams=None,
preprocessorParams=None,
exampleStyles=None,
classifierParams=None,
doFullGrid=False,
deleteOutput=False,
copyFrom=None,
log="log.txt",
step=None,
omitSteps=None,
debug=False,
connection=None,
subset=None,
folds=None):
"""
Train a new model for event or relation detection.
@param output: A directory where output files will appear.
@param task: If defined, overridable default settings are used for many of the training parameters. Must be one of the supported TEES tasks.
@param detector: a Detector object, or a string defining one to be imported
@param inputFiles: A dictionary of file names, with keys "train", "devel" and, "test"
@param models: A dictionary of file names defining the place for the new models, with keys "devel" and, "test"
@param parse: The parse element name in the training interaction XML
@param processUnmerging: Use the unmerging step of EventDetector. True, False or None for task default.
@param processModifiers: Use the modifier detection step of EventDetector. True, False or None for task default.
@param bioNLPSTParams: Parameters controlling BioNLP ST format output.
@param preprocessorParams: Parameters controlling the preprocessor. Not used for training, but saved to the model for use when classifying.
@param exampleStyles: A parameter set for controlling example builders.
@param classifierParams: A parameter set for controlling classifiers.
@param doFullGrid: Whether all parameters, as opposed to just recall adjustment, are tested in the EventDetector grid search.
@param deleteOutput: Remove an existing output directory
@param copyFrom: Copy an existing output directory for use as a template
@param log: An optional alternative name for the log file. None is for no logging.
@param step: A step=substep pair, where the steps are "TRAIN", "DEVEL", "EMPTY" and "TEST"
@param omitSteps: step=substep parameters, where multiple substeps can be defined.
@param debug: In debug mode, more output is shown, and some temporary intermediate files are saved
@param connection: A parameter set defining a local or remote connection for training the classifier
@param subset: A parameter set for making subsets of input files
"""
# Insert default arguments where needed
inputFiles = setDictDefaults(inputFiles, {
"train": None,
"devel": None,
"test": None
})
models = setDictDefaults(models, {"devel": None, "test": None})
exampleStyles = setDictDefaults(
exampleStyles, {
"examples": None,
"trigger": None,
"edge": None,
"unmerging": None,
"modifiers": None
})
classifierParams = setDictDefaults(
classifierParams, {
"examples": None,
"trigger": None,
"recall": None,
"edge": None,
"unmerging": None,
"modifiers": None
})
subset = setDictDefaults(Parameters.get(subset), {
"train": None,
"devel": None,
"test": None,
"seed": 0,
"all": None
})
folds = setDictDefaults(folds, {
"train": None,
"devel": None,
"test": None
})
processUnmerging = getDefinedBool(processUnmerging)
processModifiers = getDefinedBool(processModifiers)
# Initialize working directory
workdir(output, deleteOutput, copyFrom, log)
# Get task specific parameters
detector, bioNLPSTParams, preprocessorParams = getTaskSettings(
task, detector, bioNLPSTParams, preprocessorParams, inputFiles,
exampleStyles, classifierParams)
# Learn training settings from input files
detector = learnSettings(inputFiles, detector, classifierParams)
# Get corpus subsets
getFolds(inputFiles, folds)
getSubsets(inputFiles, subset)
if task != None:
task = task.replace("-FULL", "")
# Define processing steps
selector, detectorSteps, omitDetectorSteps = getSteps(
step, omitSteps, ["TRAIN", "DEVEL", "EMPTY", "TEST"])
# Initialize the detector
detector, detectorName = getDetector(detector)
detector = detector() # initialize object
detector.debug = debug
detector.bioNLPSTParams = detector.getBioNLPSharedTaskParams(
bioNLPSTParams)
#detector.useBioNLPSTFormat = useBioNLPSTFormat # classify-output and grid evaluation in ST-format
#detector.stWriteScores = True # write confidence scores into additional st-format files
connection = getConnection(connection)
detector.setConnection(connection)
connection.debug = debug
if deleteOutput:
connection.clearWorkDir()
# Train
if selector.check("TRAIN"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------------ Train Detector ------------------"
print >> sys.stderr, "----------------------------------------------------"
if isinstance(detector, SingleStageDetector):
detector.train(inputFiles["train"],
inputFiles["devel"],
models["devel"],
models["test"],
exampleStyles["examples"],
classifierParams["examples"],
parse,
None,
task,
fromStep=detectorSteps["TRAIN"],
workDir="training")
else:
detector.train(inputFiles["train"],
inputFiles["devel"],
models["devel"],
models["test"],
exampleStyles["trigger"],
exampleStyles["edge"],
exampleStyles["unmerging"],
exampleStyles["modifiers"],
classifierParams["trigger"],
classifierParams["edge"],
classifierParams["unmerging"],
classifierParams["modifiers"],
classifierParams["recall"],
processUnmerging,
processModifiers,
doFullGrid,
task,
parse,
None,
fromStep=detectorSteps["TRAIN"],
workDir="training")
# Save the detector type
for model in [models["devel"], models["test"]]:
if model != None and os.path.exists(model):
model = Model(model, "a")
model.addStr("detector", detectorName)
if preprocessorParams != None:
preprocessor = Preprocessor()
model.addStr(
"preprocessorParams",
Parameters.toString(
preprocessor.getParameters(preprocessorParams)))
model.save()
model.close()
if selector.check("DEVEL"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------ Check devel classification ------------"
print >> sys.stderr, "----------------------------------------------------"
#detector.bioNLPSTParams["scores"] = False # the evaluation server doesn't like additional files
detector.classify(inputFiles["devel"],
models["devel"],
"classification-devel/devel",
goldData=inputFiles["devel"],
fromStep=detectorSteps["DEVEL"],
workDir="classification-devel")
if selector.check("EMPTY"):
# By passing an emptied devel set through the prediction system, we can check that we get the same predictions
# as in the DEVEL step, ensuring the model does not use leaked information.
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------ Empty devel classification ------------"
print >> sys.stderr, "----------------------------------------------------"
#detector.bioNLPSTParams["scores"] = False # the evaluation server doesn't like additional files
detector.classify(getEmptyCorpus(
inputFiles["devel"],
removeNames=("names" in str(exampleStyles["examples"])
or "names" in str(exampleStyles["trigger"]))),
models["devel"],
"classification-empty/devel-empty",
fromStep=detectorSteps["EMPTY"],
workDir="classification-empty")
if selector.check("TEST"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------- Test set classification --------------"
print >> sys.stderr, "----------------------------------------------------"
if inputFiles["test"] == None or not os.path.exists(
inputFiles["test"]):
print >> sys.stderr, "Skipping, test file", inputFiles[
"test"], "does not exist"
else:
#detector.bioNLPSTParams["scores"] = False # the evaluation server doesn't like additional files
detector.classify(inputFiles["test"],
models["test"],
"classification-test/test",
fromStep=detectorSteps["TEST"],
workDir="classification-test")
if detector.bioNLPSTParams["convert"]:
Utils.STFormat.Compare.compare(
"classification-test/test-events.tar.gz",
"classification-devel/devel-events.tar.gz", "a2")
def train(output, task=None, detector=None, inputFiles=None, models=None, parse=None,
processUnmerging=None, processModifiers=None,
bioNLPSTParams=None, preprocessorParams=None, exampleStyles=None,
classifierParams=None, doFullGrid=False, deleteOutput=False, copyFrom=None,
log="log.txt", step=None, omitSteps=None, debug=False, connection=None, subset=None,
folds=None, corpusDir=None, corpusPreprocessing=None, evaluator=None):
"""
Train a new model for event or relation detection.
@param output: A directory where output files will appear.
@param task: If defined, overridable default settings are used for many of the training parameters. Must be one of the supported TEES tasks.
@param detector: a Detector object, or a string defining one to be imported
@param inputFiles: A dictionary of file names, with keys "train", "devel" and, "test"
@param models: A dictionary of file names defining the place for the new models, with keys "devel" and, "test"
@param parse: The parse element name in the training interaction XML
@param processUnmerging: Use the unmerging step of EventDetector. True, False or None for task default.
@param processModifiers: Use the modifier detection step of EventDetector. True, False or None for task default.
@param bioNLPSTParams: Parameters controlling BioNLP ST format output.
@param preprocessorParams: Parameters controlling the preprocessor. Not used for training, but saved to the model for use when classifying.
@param exampleStyles: A parameter set for controlling example builders.
@param classifierParams: A parameter set for controlling classifiers.
@param doFullGrid: Whether all parameters, as opposed to just recall adjustment, are tested in the EventDetector grid search.
@param deleteOutput: Remove an existing output directory
@param copyFrom: Copy an existing output directory for use as a template
@param log: An optional alternative name for the log file. None is for no logging.
@param step: A step=substep pair, where the steps are "TRAIN", "DEVEL", "EMPTY" and "TEST"
@param omitSteps: step=substep parameters, where multiple substeps can be defined.
@param debug: In debug mode, more output is shown, and some temporary intermediate files are saved
@param connection: A parameter set defining a local or remote connection for training the classifier
@param subset: A parameter set for making subsets of input files
"""
# Insert default arguments where needed
inputFiles = setDictDefaults(inputFiles, {"train":None, "devel":None, "test":None})
models = setDictDefaults(models, {"devel":"model-devel", "test":"model-test"})
exampleStyles = setDictDefaults(exampleStyles, {"examples":None, "trigger":None, "edge":None, "unmerging":None, "modifiers":None})
classifierParams = setDictDefaults(classifierParams, {"examples":None, "trigger":None, "recall":None, "edge":None, "unmerging":None, "modifiers":None})
subset = setDictDefaults(Parameters.get(subset), {"train":None, "devel":None, "test":None, "seed":0, "all":None})
folds = setDictDefaults(folds, {"train":None, "devel":None, "test":None})
processUnmerging = getDefinedBool(processUnmerging)
processModifiers = getDefinedBool(processModifiers)
# Initialize working directory
workdir(output, deleteOutput, copyFrom, log)
# Get task specific parameters
useKerasDetector = False
if detector != None and "keras" in detector.lower():
print >> sys.stderr, "Using a Keras Detector"
useKerasDetector = True
if detector.lower() == "keras":
detector = None
detector, bioNLPSTParams, preprocessorParams, folds = getTaskSettings(task, detector,
bioNLPSTParams, preprocessorParams, inputFiles, exampleStyles, classifierParams, folds, corpusDir=corpusDir, useKerasDetector=useKerasDetector)
# Learn training settings from input files
detector = learnSettings(inputFiles, detector, classifierParams, task, exampleStyles, useKerasDetector=useKerasDetector)
# Get corpus subsets
getFolds(inputFiles, folds)
getSubsets(inputFiles, subset)
if task != None:
task = task.replace("-FULL", "")
if "." in task:
_, subTask = getSubTask(task)
if subTask != 3:
processModifiers = False
# Preprocess the corpus if required
if corpusPreprocessing != None:
preprocessor = Preprocessor(steps=corpusPreprocessing)
assert preprocessor.steps[0].name == "MERGE_SETS"
assert preprocessor.steps[-1].name == "DIVIDE_SETS"
preprocessedCorpusDir = os.path.join(output, "corpus")
#outputFiles = {x:os.path.join(preprocessedCorpusDir, os.path.basename(inputFiles[x])) for x in inputFiles}
preprocessor.process(inputFiles, os.path.join(preprocessedCorpusDir, task))
#inputFiles = outputFiles
for setName in inputFiles.keys():
if inputFiles[setName] != None:
inputFiles[setName] = os.path.join(preprocessedCorpusDir, task + "-" + setName + ".xml")
# Define processing steps
selector, detectorSteps, omitDetectorSteps = getSteps(step, omitSteps, ["TRAIN", "DEVEL", "EMPTY", "TEST"])
# Initialize the detector
detector, detectorName = getDetector(detector, evaluator=evaluator)
evaluator, evaluatorName = importClass(evaluator, "evaluator")
detector = detector() # initialize object
if evaluator != None:
print >> sys.stderr, "Using evaluator", evaluator.__name__
detector.evaluator = evaluator
detector.debug = debug
detector.bioNLPSTParams = detector.getBioNLPSharedTaskParams(bioNLPSTParams)
#detector.useBioNLPSTFormat = useBioNLPSTFormat # classify-output and grid evaluation in ST-format
#detector.stWriteScores = True # write confidence scores into additional st-format files
connection = getConnection(connection)
detector.setConnection(connection)
connection.debug = debug
if deleteOutput:
connection.clearWorkDir()
# Train
if selector.check("TRAIN"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------------ Train Detector ------------------"
print >> sys.stderr, "----------------------------------------------------"
if not isinstance(detector, EventDetector):
detector.train(inputFiles["train"], inputFiles["devel"], models["devel"], models["test"],
exampleStyles["examples"], classifierParams["examples"], parse, None, task,
fromStep=detectorSteps["TRAIN"], workDir="training", testData=inputFiles["test"])
else:
detector.train(inputFiles["train"], inputFiles["devel"], models["devel"], models["test"],
exampleStyles["trigger"], exampleStyles["edge"], exampleStyles["unmerging"], exampleStyles["modifiers"],
classifierParams["trigger"], classifierParams["edge"], classifierParams["unmerging"], classifierParams["modifiers"],
classifierParams["recall"], processUnmerging, processModifiers,
doFullGrid, task, parse, None,
fromStep=detectorSteps["TRAIN"], workDir="training", testData=inputFiles["test"])
# Save the detector type
for model in [models["devel"], models["test"]]:
if model != None and os.path.exists(model):
model = Model(model, "a")
model.addStr("detector", detectorName)
if evaluatorName != None:
model.addStr("detector", evaluatorName)
if preprocessorParams != None:
preprocessor = Preprocessor()
model.addStr("preprocessorParams", Parameters.toString(preprocessor.getParameters(preprocessorParams)))
model.save()
model.close()
if selector.check("DEVEL"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------ Check devel classification ------------"
print >> sys.stderr, "----------------------------------------------------"
#detector.bioNLPSTParams["scores"] = False # the evaluation server doesn't like additional files
detector.classify(inputFiles["devel"], models["devel"], "classification-devel/devel", goldData=inputFiles["devel"], fromStep=detectorSteps["DEVEL"], workDir="classification-devel")
if selector.check("EMPTY"):
# By passing an emptied devel set through the prediction system, we can check that we get the same predictions
# as in the DEVEL step, ensuring the model does not use leaked information.
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------ Empty devel classification ------------"
print >> sys.stderr, "----------------------------------------------------"
#detector.bioNLPSTParams["scores"] = False # the evaluation server doesn't like additional files
removalScope = "non-given"
if "names" in str(exampleStyles["examples"]) or "names" in str(exampleStyles["trigger"]):
removalScope = "all"
elif "Edge" in detector.__class__.__name__:
removalScope = "interactions"
detector.classify(getEmptyCorpus(inputFiles["devel"], scope=removalScope), models["devel"], "classification-empty/devel-empty", fromStep=detectorSteps["EMPTY"], workDir="classification-empty")
print >> sys.stderr, "*** Evaluate empty devel classification ***"
if os.path.exists("classification-empty/devel-empty-pred.xml.gz"):
EvaluateInteractionXML.run(detector.evaluator, "classification-empty/devel-empty-pred.xml.gz", inputFiles["devel"], parse)
else:
print >> sys.stderr, "No output file for evaluation"
if selector.check("TEST"):
print >> sys.stderr, "----------------------------------------------------"
print >> sys.stderr, "------------- Test set classification --------------"
print >> sys.stderr, "----------------------------------------------------"
if inputFiles["test"] == None or not os.path.exists(inputFiles["test"]):
print >> sys.stderr, "Skipping, test file", inputFiles["test"], "does not exist"
else:
#detector.bioNLPSTParams["scores"] = False # the evaluation server doesn't like additional files
detector.classify(inputFiles["test"], models["test"] if models["test"] != None else models["devel"], "classification-test/test", fromStep=detectorSteps["TEST"], workDir="classification-test")
if detector.bioNLPSTParams["convert"]:
extension = ".zip" if (detector.bioNLPSTParams["convert"] == "zip") else ".tar.gz"
Utils.STFormat.Compare.compare("classification-test/test-events" + extension, "classification-devel/devel-events" + extension, "a2")
# Stop logging
if log != None:
Stream.closeLog(log)