def _weights(self, index, dists, sigma_sq, y_desired):
dists = [dist(self.fmodel.dataset.get_y(idx), y_desired)
for idx in index] # could be optimized
w = np.fromiter((gaussian_kernel(d/self.fmodel.dim_y, sigma_sq)
for d in dists), np.float)
# We eliminate the outliers # TODO : actually reduce w and index
wsum = w.sum()
if wsum == 0:
return 1.0/len(dists)*np.ones((len(dists),))
else:
eps = wsum * 1e-15 / self.fmodel.dim_y
return np.fromiter((w_i/wsum if w_i > eps else 0.0 for w_i in w), np.float)
def _weights(self, index, dists, sigma_sq, y_desired):
dists = [
dist(self.fmodel.dataset.get_y(idx), y_desired) for idx in index
] # could be optimized
w = np.fromiter((gaussian_kernel(d / self.fmodel.dim_y, sigma_sq)
for d in dists), np.float)
# We eliminate the outliers # TODO : actually reduce w and index
wsum = w.sum()
if wsum == 0:
return 1.0 / len(dists) * np.ones((len(dists), ))
else:
eps = wsum * 1e-15 / self.fmodel.dim_y
return np.fromiter((w_i / wsum if w_i > eps else 0.0 for w_i in w),
np.float)
def _guess_x(self, y_desired, **kwargs):
"""Choose the relevant neighborhood to feed the inverse model, based
on the minimum spread of the corresponding neighborhood in S.
for each (x, y) with y neighbor of y_desired,
1. find the neighborhood of x, (xi, yi)_k.
2. compute the standart deviation of the error between yi and y_desired.
3. select the neighborhood of minimum standart deviation
TODO : Implement another method taking the spread in M too.
"""
k = kwargs.get('k', self.k)
_, indexes = self.fmodel.dataset.nn_y(y_desired, k = k)
min_std, min_xi = float('inf'), None
for i in indexes:
xi = self.fmodel.dataset.get_x(i)
_, indexes_xi = self.fmodel.dataset.nn_x(xi, k = k)
std_xi = np.std([dist(self.fmodel.dataset.get_y(j), y_desired) for j in indexes_xi])
if std_xi < min_std:
min_std, min_xi = std_xi, xi
#print(std_xi, tuple(yi))
return [min_xi]
def _guess_x(self, y_desired, **kwargs):
"""Choose the relevant neighborhood to feed the inverse model, based
on the minimum spread of the corresponding neighborhood in S.
for each (x, y) with y neighbor of y_desired,
1. find the neighborhood of x, (xi, yi)_k.
2. compute the standart deviation of the error between yi and y_desired.
3. select the neighborhood of minimum standart deviation
TODO : Implement another method taking the spread in M too.
"""
k = kwargs.get('k', self.k)
_, indexes = self.fmodel.dataset.nn_y(y_desired, k=k)
min_std, min_xi = float('inf'), None
for i in indexes:
xi = self.fmodel.dataset.get_x(i)
_, indexes_xi = self.fmodel.dataset.nn_x(xi, k=k)
std_xi = np.std([
dist(self.fmodel.dataset.get_y(j), y_desired)
for j in indexes_xi
])
if std_xi < min_std:
min_std, min_xi = std_xi, xi
#print(std_xi, tuple(yi))
return [min_xi]
