def _get_coords(filename):
alb = file(filename)
start_line = None
end_line = None
for line in alb:
if line.startswith("["):
if not start_line:
start_line = line # rstrip not needed
else:
end_line = line
if end_line is None: # sequence is too short
return [(0, 0), (0, 0)]
return list(zip(*map(_alb_line2coords, [start_line, end_line]))) # returns [(start0, end0), (start1, end1)]
def _get_coords(filename):
alb = file(filename)
start_line = None
end_line = None
for line in alb:
if line.startswith("["):
if not start_line:
start_line = line # rstrip not needed
else:
end_line = line
if end_line is None: # sequence is too short
return [(0, 0), (0, 0)]
return list(zip(*map(_alb_line2coords, [start_line, end_line]))
) # returns [(start0, end0), (start1, end1)]
def __eq__(self, other):
if len(self.data) != len(other.data):
return 0
ok = reduce(lambda x, y: x and y,
map(lambda x, y: x == y, self.data, other.data))
return ok
def __eq__(self, other):
if len(self.data) != len(other.data):
return 0
ok = reduce(lambda x, y: x and y, map(lambda x, y: x == y, self.data, other.data))
return ok
def train(training_set, results, feature_fns, update_fn=None,
max_iis_iterations=10000, iis_converge=1.0e-5,
max_newton_iterations=100, newton_converge=1.0e-10):
"""Train a maximum entropy classifier, returns MaxEntropy object.
Train a maximum entropy classifier on a training set.
training_set is a list of observations. results is a list of the
class assignments for each observation. feature_fns is a list of
the features. These are callback functions that take an
observation and class and return a 1 or 0. update_fn is a
callback function that is called at each training iteration. It is
passed a MaxEntropy object that encapsulates the current state of
the training.
The maximum number of iterations and the convergence criterion for IIS
are given by max_iis_iterations and iis_converge, respectively, while
max_newton_iterations and newton_converge are the maximum number
of iterations and the convergence criterion for Newton's method.
"""
if not training_set:
raise ValueError("No data in the training set.")
if len(training_set) != len(results):
raise ValueError("training_set and results should be parallel lists.")
# Rename variables for convenience.
xs, ys = training_set, results
# Get a list of all the classes that need to be trained.
classes = sorted(set(results))
# Cache values for all features.
features = [_eval_feature_fn(fn, training_set, classes)
for fn in feature_fns]
# Cache values for f#.
f_sharp = _calc_f_sharp(len(training_set), len(classes), features)
# Pre-calculate the empirical expectations of the features.
e_empirical = _calc_empirical_expects(xs, ys, classes, features)
# Now train the alpha parameters to weigh each feature.
alphas = [0.0] * len(features)
iters = 0
while iters < max_iis_iterations:
nalphas = _train_iis(xs, classes, features, f_sharp,
alphas, e_empirical,
max_newton_iterations, newton_converge)
diff = map(lambda x, y: numpy.fabs(x-y), alphas, nalphas)
diff = reduce(lambda x, y: x+y, diff, 0)
alphas = nalphas
me = MaxEntropy()
me.alphas, me.classes, me.feature_fns = alphas, classes, feature_fns
if update_fn is not None:
update_fn(me)
if diff < iis_converge: # converged
break
else:
raise RuntimeError("IIS did not converge")
return me
