class FA(object):
""" This is the base class for general function approximators. """
parametric = True
def __init__(self, indim, outdim):
""" initialize function approximator with input and output dimension. """
self.indim = indim
self.outdim = outdim
self.dataset = Dataset(indim, outdim)
self.reset()
def predict(self, inp):
""" predict the output for the given input. """
abstractMethod()
def update(self, inp, tgt):
""" update the function approximator to return something closer
to target when queried for input next time. Some function
approximators only collect the input/target tuples here and
learn only on a call to learn(). """
self.dataset.append(inp, tgt)
def reset(self):
""" this initializes the function approximator to an initial state,
forgetting everything it has learned before. """
self.dataset.clear()
def train(self):
""" some function approximators learn offline after collecting all
samples. this function executes one such learning step. """
pass
def _getParameters(self):
""" getter method for parameters. """
pass
def _setParameters(self, parameters):
""" setter method for parameters. """
pass
def _asFlatArray(self, value):
return np.asarray(value).flatten()
# overwrite the two above functions and re-declare this property
parameters = property(_getParameters, _setParameters)
class Classifier(object):
""" This is the base class for general classifiers. """
parametric = False
def __init__(self, indim, nclasses):
""" initialize function approximator with input and output dimension. """
self.indim = indim
self.nclasses = nclasses
self.dataset = Dataset(indim, 1)
self.reset()
def classify(self, inp):
""" predict the output for the given input. """
abstractMethod()
def update(self, inp, tgt, imp=1.):
""" update the function approximator to return something closer
to target when queried for input next time. Some function
approximators only collect the input/target tuples here and
learn only on a call to learn().
tgt is an integer for the class number. conversion to one-of-k
coding is done internally. imp is the importance [0., 1.] of
the sample.
"""
self.dataset.append(inp, tgt, imp)
def reset(self):
""" this initializes the function approximator to an initial state,
forgetting everything it has learned before. """
self.dataset.clear()
def train(self):
""" some function approximators learn offline after collecting all
samples. this function executes one such learning step. """
pass
def _getParameters(self):
""" getter method for parameters. """
pass
def _setParameters(self, parameters):
""" setter method for parameters. """
pass
# overwrite the two above functions and re-declare this property
parameters = property(_getParameters, _setParameters)
def _asFlatArray(self, value):
return np.asarray(value).flatten()
def _asOneOfK(self, n, twodim=False):
""" returns a k-dimensional vector with all zeros except the n-th
element which is 1. if twodim is set to True, a (1,k)-dimensional
row vector is returned instead.
"""
k = self.nclasses
ret = np.zeros(k)
ret[n] = 1.
if twodim:
ret = ret.reshape(1,k)
return ret