def fit(self, X, y):
"""
Fit data
Inputs:
X (N, M) : A design matrix with N M-dimensional vectors as rows
y (N,) OR (N, D): The output labels. Can be binary vector of 0's and 1's
or True's and False's or a one-hot encoding of D classes
"""
assert is_numpy(X), 'X must be a numpy array'
assert is_numpy(y), 'y must be a numpy array'
assert has_dims(X, 2), 'X must be 2-dimensional'
if has_dims(y, 1):
self.one_class = True
self.y = y.reshape(-1, 1)
elif has_dims(y, 2):
self.one_class = y.shape[1] == 1
self.y = y
else:
raise TypeError('y must be 1 or 2 dimensional')
self.X = X
self.fitted = True
def predict(self, X):
"""
Predict on new design matrix X
Inputs:
X (N, M) : A design matrix with N M-dimensional vectors as rows
"""
assert self.fitted, 'Fit data before predicting'
assert is_numpy(X), 'X must be a numpy array'
assert has_dims(X, 2), 'X must be 2-dimensional'
if self.one_class:
distances = self.pairwise_distances(self.X, X)
sortd = np.argsort(distances, axis=0)[:self.k, :]
labels = self.y[sortd.T]
if self.weights == 'uniform':
guess = np.mean(labels, axis=1) >= 0.5
return guess
elif self.weights == 'distance':
ws = np.nan_to_num(1 / np.take_along_axis(distances, sortd, axis=0).T)
ws = ws / ws.sum(1)[:, None]
guess = (labels * ws).sum(1) >= 0.5
return guess
else:
raise NotImplementedError('this hasnt been done yet... ') #TODO implement one-hot encoding for KNN
raise ValueError('Weights set incorrectly')
def pairwise_distances(self, X1, X2):
"""
Return an (N1, N2) matrix where the element (i, j) is the euclidean
distance between the vectors represented by the ith row of
X1 and the jth row of X2.
Inputs:
X1 (N1, M): A matrix with N1 M-dimensional vectors as rows
X2 (N2, M): A matrix with N2 M-dimensional vectors as rows
Returns:
D (N1, N2): The pairwise euclidean distances
"""
assert is_numpy(X1), 'X1 must be a numpy array'
assert is_numpy(X2), 'X2 must be a numpy array'
assert has_dims(X1, 2), 'X1 must be 2-dimensional'
assert has_dims(X2, 2), 'X2 must be 2-dimensional'
return (- 2 * X1 @ X2.T + (X1 ** 2).sum(1)[:, None] + (X2 ** 2).sum(1)) ** 0.5
def sigmoid(x):
"""
Simple sigmoid
Inputs:
x: float or int or numpy array
Returns:
σ(x)
"""
assert any([is_integer(x), is_float(x),
is_numpy(x)]), 'x must be an integer or a float or numpy array'
return (1 + np.exp(-x))**-1
def fit(self, X, y, method='L-BFGS-B', display_opt=False):
"""
Fit data
Inputs:
X (N, M) : A design matrix with N M-dimensional vectors as rows
y (N,) OR (N, D): The output labels. Can be binary vector of 0's and 1's
or True's and False's or a one-hot encoding of D classes
"""
assert is_numpy(X), 'X must be a numpy array'
assert is_numpy(y), 'y must be a numpy array'
assert has_dims(X, 2), 'X must be 2-dimensional'
assert y.shape[0] == X.shape[
0], 'X and y must have the same 1st dimension length'
if has_dims(y, 1):
self.one_class = True
self.y = y.reshape(-1, 1)
elif has_dims(y, 2):
self.one_class = y.shape[1] == 1
self.y = y
else:
raise TypeError('y must be 1 or 2 dimensional')
self.N, self.M = X.shape
self.z = 2 * self.y - 1
self.X = X
options = {'maxiter': 500, 'disp': display_opt}
result = minimize(self.loss,
np.random.normal(size=self.M),
method=method,
jac=self.grad,
options=options)
self.w = result.x.reshape(-1, 1)