def test_matrix_output(self, y_values, epsilon, metric, metric_params):
"""
Test that we are returning the correct kernel values.
"""
# Setup true values to test again.
if y_values is None:
y_values_ref = x_values
else:
y_values_ref = y_values
if metric == 'minkowski':
pw_distance = cdist(y_values_ref,
x_values,
metric='minkowski',
p=metric_params['p'])
else:
pw_distance = cdist(y_values_ref, x_values, metric=metric)
true_values = np.exp(-1. * pw_distance**2 / epsilon)
# Construct the kernel and fit to data.
mykernel = kernel.Kernel(kernel_type='gaussian',
metric=metric,
metric_params=metric_params,
epsilon=epsilon,
k=len(x_values))
mykernel.fit(x_values)
K_matrix = mykernel.compute(y_values).toarray()
# Check that error values are beneath tolerance.
error_values = (K_matrix - true_values).ravel()
total_error = np.linalg.norm(error_values)
assert (total_error < 1E-8)
def test_auto_epsilon_selection(self):
X = np.arange(100).reshape(-1, 1)
mykernel = kernel.Kernel(kernel_type='gaussian',
metric='euclidean',
epsilon='bgh',
k=10)
mykernel.fit(X)
assert (mykernel.epsilon_fitted == 1.0)
assert (mykernel.d == 1.0)
def test_auto_epsilon_selection(self, eps_method):
X = np.arange(100).reshape(-1, 1)
mykernel = kernel.Kernel(kernel_type='gaussian',
metric='euclidean',
epsilon=eps_method,
k=10)
mykernel.fit(X)
if eps_method == 'bgh':
assert (mykernel.epsilon_fitted == 0.25)
else:
assert (mykernel.epsilon_fitted == 0.50)
assert (mykernel.d == 1.0)
def test_sparse_input(self, x_values, y_values, metric, metric_params,
use_sparse):
"""
Test that we are returning the correct kernel values.
"""
# Setup true values to test again.
epsilon = 10.
bandwidth_fxn = None
if y_values is None:
y_values_ref = x_values
else:
y_values_ref = y_values
if metric == 'minkowski':
pw_distance = cdist(y_values_ref,
x_values,
metric='minkowski',
p=metric_params['p'])
else:
pw_distance = cdist(y_values_ref, x_values, metric=metric)
if bandwidth_fxn is None:
ref_bandwidth_fxn = lambda x: np.ones(x.shape[0])
else:
ref_bandwidth_fxn = bandwidth_fxn
if use_sparse:
x_values = sps.csr_matrix(x_values)
y_values_ref = sps.csr_matrix(y_values_ref)
x_bandwidth = ref_bandwidth_fxn(x_values)
y_bandwidth = ref_bandwidth_fxn(y_values_ref).reshape(-1, 1)
scaled_sq_dists = pw_distance**2 / (x_bandwidth * y_bandwidth)
true_values = np.exp(-1. * scaled_sq_dists / (4. * epsilon))
# Construct the kernel and fit to data.
mykernel = kernel.Kernel(kernel_type='gaussian',
metric=metric,
metric_params=metric_params,
epsilon=epsilon,
k=x_values.shape[0],
bandwidth_type=bandwidth_fxn)
mykernel.fit(x_values)
K_matrix = mykernel.compute(y_values).toarray()
# Check that error values are beneath tolerance.
error_values = (K_matrix - true_values).ravel()
total_error = np.linalg.norm(error_values)
assert (total_error < 1E-8)
def test_neighborlists(self, x_values, k, neighbor_params):
"""
Test that neighborlisting gives the right number of elements.
"""
# Correct number of nearest neighbors.
k0 = min(k, x_values.shape[0])
# Construct kernel matrix.
mykernel = kernel.Kernel(kernel_type='gaussian',
metric='euclidean',
epsilon=1.,
k=k0,
neighbor_params=neighbor_params)
mykernel.fit(x_values)
K_matrix = mykernel.compute(x_values)
# Check if each row has correct number of elements
row_has_k_elements = (K_matrix.nnz == k0 * x_values.shape[0])
assert (row_has_k_elements)
