def fit(self, x_train, w_train):
# Train all committee classifiers with their corresponding views.
for (classifier, view) in self.committee:
classifier.fit(x_train[:, view], w_train)
# Estimate a prior probabilities p(wi) for each class wi.
self.p_w = DataLoader.compute_a_priori(w_train)
return self
def fit(self, x_train, w_train):
"""
Stores training points *x_train* and their correponsindg labels *w_train*,
and estimates the a prior probabilities p(w_i) for each class w_i.
"""
# Store examples.
self.x_train = x_train
self.w_train = w_train
# Estimate a prior probabilities p(wi) for each class wi.
self.p_w = DataLoader.compute_a_priori(w_train)
self.num_classes = len(self.p_w)
return self
def fit(self, x_train, w_train):
"""
Estimates a prior probabilities p(w_i), the mean i and the variance i for
each class in the training dataset. That is, from (x_train, w_train).
Once trained, you can call classify() to predict the class/label
for a given feature vector.
Note: all classes must have at least one sample.
"""
# Break down dataset into smaller groups sharing the same label.
x_groups = DataLoader.group_by_label(x_train, w_train)
# Estimate a prior probabilities p(wi) for each class wi.
self.p_w = DataLoader.compute_a_priori(w_train)
# Estimate mean and [diagonal] variances for each class w_i.
# Pattern Classification (Second Edition), Section 3.2.3.
self.mu = np.array(
list(map(
lambda x_train_i: np.mean(x_train_i, axis=0),
x_groups,
)))
self.sigma = np.array(
list(
map(
lambda i: np.mean((x_groups[i] - self.mu[i])**2, axis=0),
range(len(x_groups)),
)))
# For the sake of optimization, we may precompute some constants in the
# gaussian pdf equations - the amplitudes and the inverse of sigma.
n = len(self.mu)
pi_const = np.power(2 * np.pi, n / 2.0)
det_sigma = np.abs(np.product(self.sigma, axis=1))
self.inv_sigma = 1.0 / (self.sigma + epsilon)
self.amplitudes = 1.0 / (pi_const * np.sqrt(det_sigma + epsilon))
return self