def iterate(self, context):
evidence, logLikTrain = context.trainSample.evidence(
context.model.sequenceModel, self.shallUseMaximumApproximation)
print(('LL train: %s (before)' % logLikTrain), file=context.log)
context.logLikTrain.append(logLikTrain)
if (not context.develSample) and (context.iteration >
self.minIterations):
context.registerNewModel(context.model, logLikTrain)
order = evidence.maximumHistoryLength()
print(' evidence order: %s' % order, file=context.log)
if context.order is not None and order != context.order:
print(' warning: evidence order changed from %d to %d!' %
(context.order, order),
file=context.log)
context.order = order
print(' evidence types: %s' % evidence.size(), file=context.log)
print(' evidence total / max: %s / %s' %
(evidence.total(), evidence.maximum()),
file=context.log)
self.showMostEvident(context.log, evidence, 10) ### TESTING
newModel = Model(self.sequitur)
newModel.discount = context.discountAdjuster.adjust(
context, evidence, order)
newModel.sequenceModel = self.sequenceModel(evidence,
newModel.discount)
print(' model size: %s' % newModel.sequenceModel.size(),
file=context.log)
if context.develSample:
logLikDevel = context.develSample.logLik(
newModel.sequenceModel, self.shallUseMaximumApproximation)
print('LL devel: %s' % logLikDevel, file=context.log)
context.logLikDevel.append(logLikDevel)
for observer in self.observers:
observer(context.log, context, newModel)
if (context.develSample) and (context.iteration >= self.minIterations):
context.registerNewModel(newModel, logLikDevel)
shouldStop = False
if context.bestModel:
if context.develSample:
crit = context.logLikDevel
else:
crit = context.logLikTrain
crit = [-ll for ll in crit[-self.convergenceWindow:]]
if not Minimization.hasSignificantDecrease(crit):
print('iteration converged.', file=context.log)
shouldStop = True
context.model = newModel
return shouldStop
def iterate(self, context):
evidence, logLikTrain = context.trainSample.evidence(
context.model.sequenceModel,
self.shallUseMaximumApproximation)
print >> context.log, 'LL train: %s (before)' % logLikTrain
context.logLikTrain.append(logLikTrain)
if (not context.develSample) and (context.iteration > self.minIterations):
context.registerNewModel(context.model, logLikTrain)
order = evidence.maximumHistoryLength()
print >> context.log, ' evidence order: %s' % order
if context.order is not None and order != context.order:
print >> context.log, ' warning: evidence order changed from %d to %d!' % (context.order, order)
context.order = order
print >> context.log, ' evidence types: %s' % evidence.size()
print >> context.log, ' evidence total / max: %s / %s' % (evidence.total(), evidence.maximum())
self.showMostEvident(context.log, evidence, 10) ### TESTING
newModel = Model(self.sequitur)
newModel.discount = context.discountAdjuster.adjust(context, evidence, order)
newModel.sequenceModel = self.sequenceModel(evidence, newModel.discount)
print >> context.log, ' model size: %s' % newModel.sequenceModel.size()
if context.develSample:
logLikDevel = context.develSample.logLik(newModel.sequenceModel, self.shallUseMaximumApproximation)
print >> context.log, 'LL devel: %s' % logLikDevel
context.logLikDevel.append(logLikDevel)
for observer in self.observers:
observer(context.log, context, newModel)
if (context.develSample) and (context.iteration >= self.minIterations):
context.registerNewModel(newModel, logLikDevel)
shouldStop = False
if context.bestModel:
if context.develSample:
crit = context.logLikDevel
else:
crit = context.logLikTrain
crit = [ -ll for ll in crit[-self.convergenceWindow:] ]
if not Minimization.hasSignificantDecrease(crit):
print >> context.log, 'iteration converged.'
shouldStop = True
context.model = newModel
return shouldStop
def adjustHigherOrder(self, evidence, order, maximumDiscount):
def criterion(discount):
disc = tuple(num.maximum(0.0, discount))
sm = self.modelFactory.sequenceModel(evidence, disc)
ll = self.develSample.logLik(sm, self.shallUseMaximumApproximation)
crit = - ll \
- sum(num.minimum(discount, 0)) \
+ sum(num.maximum(discount - maximumDiscount, 0))
print discount, ll, crit # TESTING
return crit
initialGuess = self.discounts[-1]
firstDirection = None
if initialGuess is None:
initialGuess = 0.1 * num.arange(1, order + 2, dtype=num.float64)
elif len(initialGuess) < order + 1:
oldGuess = initialGuess
oldSize = len(initialGuess)
initialGuess = num.zeros(order + 1, dtype=num.float64)
initialGuess[:oldSize] = oldGuess
initialGuess[oldSize:] = oldGuess[-1]
elif len(initialGuess) > order + 1:
initialGuess = initialGuess[:order + 1]
else:
previous = self.discounts[-2]
if previous is not None and len(previous) == order + 1:
firstDirection = initialGuess - previous
if not num.sometrue(num.abs(firstDirection) > 1e-4):
firstDirection = None
directions = num.identity(order + 1, dtype=num.float64)
directions = directions[::-1]
if firstDirection is not None:
directions = num.concatenate(
(firstDirection[num.newaxis, :], directions))
directions *= 0.1
print directions # TESTING
discount, ll = Minimization.directionSetMinimization(criterion,
initialGuess,
directions,
tolerance=1e-4)
discount = num.maximum(0.0, discount)
return discount, -ll
def adjustHigherOrder(self, evidence, order, maximumDiscount):
def criterion(discount):
disc = tuple(num.maximum(0.0, discount))
sm = self.modelFactory.sequenceModel(evidence, disc)
ll = self.develSample.logLik(sm, self.shallUseMaximumApproximation)
crit = - ll \
- sum(num.minimum(discount, 0)) \
+ sum(num.maximum(discount - maximumDiscount, 0))
print discount, ll, crit # TESTING
return crit
initialGuess = self.discounts[-1]
firstDirection = None
if initialGuess is None:
initialGuess = 0.1 * num.arange(1, order+2, dtype=num.float64)
elif len(initialGuess) < order+1:
oldGuess = initialGuess
oldSize = len(initialGuess)
initialGuess = num.zeros(order+1, dtype=num.float64)
initialGuess[:oldSize] = oldGuess
initialGuess[oldSize:] = oldGuess[-1]
elif len(initialGuess) > order+1:
initialGuess = initialGuess[:order+1]
else:
previous = self.discounts[-2]
if previous is not None and len(previous) == order+1:
firstDirection = initialGuess - previous
if not num.sometrue(num.abs(firstDirection) > 1e-4):
firstDirection = None
directions = num.identity(order+1, dtype=num.float64)
directions = directions[::-1]
if firstDirection is not None:
directions = num.concatenate((firstDirection[num.newaxis,:], directions))
directions *= 0.1
print directions # TESTING
discount, ll = Minimization.directionSetMinimization(
criterion, initialGuess, directions, tolerance=1e-4)
discount = num.maximum(0.0, discount)
return discount, -ll
def adjustOrderZero(self, evidence, maximumDiscount):
def criterion(discount):
sm = self.modelFactory.sequenceModel(evidence, [max(0.0, discount)])
ll = self.develSample.logLik(sm, self.shallUseMaximumApproximation)
crit = - ll - min(discount, 0) + max(discount - maximumDiscount, 0)
print discount, ll, crit # TESTING
return crit
initialGuess = self.discounts[-1]
previous = self.discounts[-2]
if initialGuess is None:
initialGuess = 0.1
else:
initialGuess = initialGuess[0]
discount, ll = Minimization.linearMinimization(
criterion, initialGuess, tolerance=1e-4)
discount = max(0.0, discount)
discount = num.array([discount])
return discount, -ll