def __init__(self, module, dataset=None, learningrate=0.01, lrdecay=1.0,
momentum=0., verbose=False, batchlearning=False,
weightdecay=0.):
"""Create a BackpropTrainer to train the specified `module` on the
specified `dataset`.
The learning rate gives the ratio of which parameters are changed into
the direction of the gradient. The learning rate decreases by `lrdecay`,
which is used to to multiply the learning rate after each training
step. The parameters are also adjusted with respect to `momentum`, which
is the ratio by which the gradient of the last timestep is used.
If `batchlearning` is set, the parameters are updated only at the end of
each epoch. Default is False.
`weightdecay` corresponds to the weightdecay rate, where 0 is no weight
decay at all.
"""
Trainer.__init__(self, module)
self.setData(dataset)
self.verbose = verbose
self.batchlearning = batchlearning
self.weightdecay = weightdecay
self.epoch = 0
self.totalepochs = 0
# set up gradient descender
self.descent = GradientDescent()
self.descent.alpha = learningrate
self.descent.momentum = momentum
self.descent.alphadecay = lrdecay
self.descent.init(module.params)
def __init__(self, module, dataset=None, learningrate=0.01, lrdecay=1.0,
momentum=0., verbose=False, batchlearning=False,
weightdecay=0.):
"""Create a BackpropTrainer to train the specified `module` on the
specified `dataset`.
The learning rate gives the ratio of which parameters are changed into
the direction of the gradient. The learning rate decreases by `lrdecay`,
which is used to to multiply the learning rate after each training
step. The parameters are also adjusted with respect to `momentum`, which
is the ratio by which the gradient of the last timestep is used.
If `batchlearning` is set, the parameters are updated only at the end of
each epoch. Default is False.
`weightdecay` corresponds to the weightdecay rate, where 0 is no weight
decay at all.
"""
Trainer.__init__(self, module)
self.setData(dataset)
self.verbose = verbose
self.batchlearning = batchlearning
self.weightdecay = weightdecay
self.epoch = 0
self.totalepochs = 0
# set up gradient descender
self.descent = GradientDescent()
self.descent.alpha = learningrate
self.descent.momentum = momentum
self.descent.alphadecay = lrdecay
self.descent.init(module.params)
def __init__(self,model,dataset,**kwargs):
ap = KWArgsProcessor(self, kwargs)
Trainer.__init__(self, model)
ap = KWArgsProcessor(self, kwargs)
# misc
ap.add('verbosity', default=0)
# population
ap.add('subPopulationSize', private=True, default=8)
ap.add('nCombinations', private=True, default=4)
ap.add('nParents', private=True, default=None)
ap.add('initialWeightRange', private=True, default=(-0.1, 0.1))
ap.add('weightInitializer', private=True, default=Randomization(self._initialWeightRange[0], self._initialWeightRange[1]))
# mutation
ap.add('mutationAlpha', private=True, default=0.01)
ap.add('mutationVariate', private=True, default=CauchyVariate(0, self._mutationAlpha))
# evaluation
ap.add('wtRatio', private=True, default=(1, 3))
# burst mutation
ap.add('nBurstMutationEpochs', default=Infinity)
# network
ap.add('backprojectionFactor', private=True, default=float(model.backprojectionFactor))
model.backprojectionFactor = self._backprojectionFactor
# aggregated objects
ap.add('selection', default=EvolinoSelection())
ap.add('reproduction', default=EvolinoReproduction(mutationVariate=self.mutationVariate))
ap.add('burstMutation', default=EvolinoBurstMutation())
ap.add('evaluation', default=EvolinoEvaluation(model, dataset, **kwargs))
self.model=model
self.dataset=dataset
genome = self.model.getGenome()
self.population = EvolinoPopulation(EvolinoSubIndividual(genome),self.subPopulationSize,self.nCombinations,self.weightInitializer)
filters = []
filters.append(self.evaluation)
filters.append(self.selection)
filters.append(self.reproduction)
self._filters = filters
self.totalepochs = 0
self._max_fitness = self.evaluation.max_fitness
self._max_fitness_epoch = self.totalepochs
def __init__(self, module, dataset, RHO=0.0001, SIG=0.5, reEvaluate=0.1,
extrapolate=3.0, maxEvaluate=20, maxSlope=100.0):
""" Initialization method. Unlike the normal BackpropTrainer, the
training dataset must be set during the initialization procedure when
using conjugate gradient descent because the dataset will be used in
the cost function and can not be passed as a variable explicitly.
"""
assert dataset is not None, "The training dataset can not be empty."
Trainer.__init__(self, module)
self.setData(dataset)
self.epoch = 0
self.descent = PRConjugateGradientDescent(RHO=RHO, SIG=SIG, reEvaluate=reEvaluate,
extrapolate=extrapolate, maxEvaluate=maxEvaluate, maxSlope=maxSlope)
def __init__(self,
module,
dataset=None,
learningrate=0.01,
lrdecay=1.0,
momentum=0.,
verbose=False,
batchlearning=False,
weightdecay=0.,
errfun=None):
"""Create a BackpropTrainer to train the specified `module` on the
specified `dataset`.
The learning rate gives the ratio of which parameters are changed into
the direction of the gradient. The learning rate decreases by
`lrdecay`, which is used to to multiply the learning rate after each
training step. The parameters are also adjusted with respect to
`momentum`, which is the ratio by which the gradient of the last
timestep is used.
If `batchlearning` is set, the parameters are updated only at the end
of each epoch. Default is False.
`weightdecay` corresponds to the weightdecay rate, where 0 is no weight
decay at all.
Arguments:
errfun (func): Function that takes 2 positional arguments,
the target (true) and predicted (estimated) output vectors, and
returns an estimate of the signed distance to the target (true)
output. default = lambda targ, est: (targ - est))
"""
Trainer.__init__(self, module)
self.setData(dataset)
self.verbose = verbose
self.batchlearning = batchlearning
self.weightdecay = weightdecay
self.epoch = 0
self.totalepochs = 0
# set up gradient descender
self.descent = GradientDescent()
self.descent.alpha = learningrate
self.descent.momentum = momentum
self.descent.alphadecay = lrdecay
self.descent.init(module.params)
self.errfun = errfun or abs_error
def __init__(self, module, dataset=None, learningrate=0.01, lrdecay=1.0,
momentum=0., verbose=False, batchlearning=False,
weightdecay=0., errfun=None):
"""Create a BackpropTrainer to train the specified `module` on the
specified `dataset`.
The learning rate gives the ratio of which parameters are changed into
the direction of the gradient. The learning rate decreases by
`lrdecay`, which is used to to multiply the learning rate after each
training step. The parameters are also adjusted with respect to
`momentum`, which is the ratio by which the gradient of the last
timestep is used.
If `batchlearning` is set, the parameters are updated only at the end
of each epoch. Default is False.
`weightdecay` corresponds to the weightdecay rate, where 0 is no weight
decay at all.
Arguments:
errfun (func): Function that takes 2 positional arguments,
the target (true) and predicted (estimated) output vectors, and
returns an estimate of the signed distance to the target (true)
output. default = lambda targ, est: (targ - est))
"""
Trainer.__init__(self, module)
self.setData(dataset)
self.verbose = verbose
self.batchlearning = batchlearning
self.weightdecay = weightdecay
self.epoch = 0
self.totalepochs = 0
# set up gradient descender
self.descent = GradientDescent()
self.descent.alpha = learningrate
self.descent.momentum = momentum
self.descent.alphadecay = lrdecay
self.descent.init(module.params)
self.errfun = errfun or abs_error
def __init__(self, evolino_network, dataset, **kwargs):
"""
:key subPopulationSize: Size of the subpopulations.
:key nCombinations: Number of times each chromosome is built into an individual. default=1
:key nParents: Number of individuals left in a subpopulation after selection.
:key initialWeightRange: Range of the weights of the RNN after initialization. default=(-0.1,0.1)
:key weightInitializer: Initializer object for the weights of the RNN. default=Randomization(...)
:key mutationAlpha: The mutation's intensity. default=0.01
:key mutationVariate: The variate used for mutation. default=CauchyVariate(...)
:key wtRatio: The quotient: washout-time/training-time. Needed to
split the sequences into washout phase and training phase.
:key nBurstMutationEpochs: Number of epochs without increase of fitness in a row,
before burstmutation is applied. default=Infinity
:key backprojectionFactor: Weight of the backprojection. Usually
supplied through evolino_network.
:key selection: Selection object for evolino
:key reproduction: Reproduction object for evolino
:key burstMutation: BurstMutation object for evolino
:key evaluation: Evaluation object for evolino
:key verbosity: verbosity level
"""
Trainer.__init__(self, evolino_network)
self.network = evolino_network
self.setData(dataset)
ap = KWArgsProcessor(self, kwargs)
# misc
ap.add('verbosity', default=0)
# population
ap.add('subPopulationSize', private=True, default=8)
ap.add('nCombinations', private=True, default=4)
ap.add('nParents', private=True, default=None)
ap.add('initialWeightRange', private=True, default=(-0.1, 0.1))
ap.add('weightInitializer', private=True, default=Randomization(self._initialWeightRange[0], self._initialWeightRange[1]))
# mutation
ap.add('mutationAlpha', private=True, default=0.01)
ap.add('mutationVariate', private=True, default=CauchyVariate(0, self._mutationAlpha))
# evaluation
ap.add('wtRatio', private=True, default=(1, 3))
# burst mutation
ap.add('nBurstMutationEpochs', default=Infinity)
# network
ap.add('backprojectionFactor', private=True, default=float(evolino_network.backprojectionFactor))
evolino_network.backprojectionFactor = self._backprojectionFactor
# aggregated objects
ap.add('selection', default=EvolinoSelection())
ap.add('reproduction', default=EvolinoReproduction(mutationVariate=self.mutationVariate))
ap.add('burstMutation', default=EvolinoBurstMutation())
ap.add('evaluation', default=EvolinoEvaluation(evolino_network, self.ds, **kwargs))
self.selection.nParents = self.nParents
self._population = EvolinoPopulation(
EvolinoSubIndividual(evolino_network.getGenome()),
self._subPopulationSize,
self._nCombinations,
self._weightInitializer
)
filters = []
filters.append(self.evaluation)
filters.append(self.selection)
filters.append(self.reproduction)
self._filters = filters
self.totalepochs = 0
self._max_fitness = self.evaluation.max_fitness
self._max_fitness_epoch = self.totalepochs
def __init__(self, evolino_network, dataset, **kwargs):
"""
:key subPopulationSize: Size of the subpopulations.
:key nCombinations: Number of times each chromosome is built into an individual. default=1
:key nParents: Number of individuals left in a subpopulation after selection.
:key initialWeightRange: Range of the weights of the RNN after initialization. default=(-0.1,0.1)
:key weightInitializer: Initializer object for the weights of the RNN. default=Randomization(...)
:key mutationAlpha: The mutation's intensity. default=0.01
:key mutationVariate: The variate used for mutation. default=CauchyVariate(...)
:key wtRatio: The quotient: washout-time/training-time. Needed to
split the sequences into washout phase and training phase.
:key nBurstMutationEpochs: Number of epochs without increase of fitness in a row,
before burstmutation is applied. default=Infinity
:key backprojectionFactor: Weight of the backprojection. Usually
supplied through evolino_network.
:key selection: Selection object for evolino
:key reproduction: Reproduction object for evolino
:key burstMutation: BurstMutation object for evolino
:key evaluation: Evaluation object for evolino
:key verbosity: verbosity level
"""
Trainer.__init__(self, evolino_network)
self.network = evolino_network
self.setData(dataset)
ap = KWArgsProcessor(self, kwargs)
# misc
ap.add('verbosity', default=0)
# population
ap.add('subPopulationSize', private=True, default=8)
ap.add('nCombinations', private=True, default=4)
ap.add('nParents', private=True, default=None)
ap.add('initialWeightRange', private=True, default=(-0.1, 0.1))
ap.add('weightInitializer',
private=True,
default=Randomization(self._initialWeightRange[0],
self._initialWeightRange[1]))
# mutation
ap.add('mutationAlpha', private=True, default=0.01)
ap.add('mutationVariate',
private=True,
default=CauchyVariate(0, self._mutationAlpha))
# evaluation
ap.add('wtRatio', private=True, default=(1, 3))
# burst mutation
ap.add('nBurstMutationEpochs', default=Infinity)
# network
ap.add('backprojectionFactor',
private=True,
default=float(evolino_network.backprojectionFactor))
evolino_network.backprojectionFactor = self._backprojectionFactor
# aggregated objects
ap.add('selection', default=EvolinoSelection())
ap.add(
'reproduction',
default=EvolinoReproduction(mutationVariate=self.mutationVariate))
ap.add('burstMutation', default=EvolinoBurstMutation())
ap.add('evaluation',
default=EvolinoEvaluation(evolino_network, self.ds, **kwargs))
self.selection.nParents = self.nParents
self._population = EvolinoPopulation(
EvolinoSubIndividual(evolino_network.getGenome()),
self._subPopulationSize, self._nCombinations,
self._weightInitializer)
filters = []
filters.append(self.evaluation)
filters.append(self.selection)
filters.append(self.reproduction)
self._filters = filters
self.totalepochs = 0
self._max_fitness = self.evaluation.max_fitness
self._max_fitness_epoch = self.totalepochs
