def test_kernel_computation(self):
""" Testing computation of the lp distance. """
self.report('Testing computed kernel values for specific cnns.')
betas = [0.0001]
scale = 2.1
struct_coeffs = 1.0
mislabel_coeffs = 1.0
tp_comp = nn_comparators.OTMANNDistanceComputer(
self.all_layer_labels,
self.label_mismatch_penalty,
self.non_assignment_penalty,
nn_comparators.CNN_STRUCTURAL_PENALTY_GROUPS,
nn_comparators.PATH_LENGTH_TYPES,
dflt_mislabel_coeffs=mislabel_coeffs,
dflt_struct_coeffs=struct_coeffs)
tp_kernel = nn_comparators.generate_otmann_kernel_from_params(
'prod',
self.all_layer_labels,
self.label_mismatch_penalty,
self.non_assignment_penalty,
nn_comparators.CNN_STRUCTURAL_PENALTY_GROUPS,
nn_comparators.PATH_LENGTH_TYPES,
mislabel_coeffs,
struct_coeffs,
betas,
scale,
dist_type='lp')
cnn_dists = tp_comp(self.cnns, self.cnns)
cnn_kernel = tp_kernel(self.cnns)
diff = np.linalg.norm(scale * np.exp(-cnn_dists[0] * betas[0]) -
cnn_kernel)
assert diff < np.linalg.norm(cnn_kernel) * 1e-6
def test_sum_product_equivalence(self):
""" Unit-test for testing that both kernels compute the same thing in certain cases.
"""
dist_type_vals = ['lp', 'emd']
for dist_type in dist_type_vals:
if dist_type == 'lp':
betas = self.lp_betas
scales = [1]
elif dist_type == 'emd':
betas = self.emd_betas
scales = [1]
sum_kernel = nn_comparators.generate_otmann_kernel_from_params(
'sum',
self.all_layer_labels,
self.label_mismatch_penalty,
self.non_assignment_penalty,
nn_comparators.MLP_STRUCTURAL_PENALTY_GROUPS,
nn_comparators.PATH_LENGTH_TYPES,
self.mislabel_coeffs,
self.struct_coeffs,
betas,
scales,
dist_type=dist_type)
prod_kernel = nn_comparators.generate_otmann_kernel_from_params(
'prod',
self.all_layer_labels,
self.label_mismatch_penalty,
self.non_assignment_penalty,
nn_comparators.MLP_STRUCTURAL_PENALTY_GROUPS,
nn_comparators.PATH_LENGTH_TYPES,
self.mislabel_coeffs,
self.struct_coeffs,
betas,
scales,
dist_type=dist_type)
sum_nn_K = sum_kernel(self.mlps)
prod_nn_K = prod_kernel(self.mlps)
assert np.linalg.norm(sum_nn_K - prod_nn_K) < _TOL
def test_instantiation_and_computation(self):
""" Testing instantiation. """
self.report(
'Testing instantiation of DistSumNNKernelTestCase and computation '
+ 'for specific networks.')
dist_type_vals = ['lp', 'emd', 'lp-emd']
all_kernels = []
for dist_type in dist_type_vals:
if dist_type == 'lp':
betas = self.lp_betas
scales = [1]
elif dist_type == 'emd':
betas = self.emd_betas
scales = [1]
else:
betas = [
j for i in zip(self.lp_betas, self.emd_betas) for j in i
]
scales = [1, 1]
tp_kernel = nn_comparators.generate_otmann_kernel_from_params(
'sum',
self.all_layer_labels,
self.label_mismatch_penalty,
self.non_assignment_penalty,
nn_comparators.MLP_STRUCTURAL_PENALTY_GROUPS,
nn_comparators.PATH_LENGTH_TYPES,
self.mislabel_coeffs,
self.struct_coeffs,
betas,
scales,
dist_type=dist_type)
nn_K = tp_kernel(self.mlps)
nn_eig_vals, _ = np.linalg.eig(nn_K)
self.report(
'dist-type: %s, eigvals: %s.' %
(dist_type, get_list_of_floats_as_str(sorted(nn_eig_vals))))
assert nn_K.shape == (len(self.mlps), len(self.mlps))
self.report(
'%s transport kernel:\n%s' %
(dist_type, str(np.round(nn_K, 3))), 'test_result')
assert np.all(np.diag(nn_K) == sum(scales))
all_kernels.append(nn_K)
# Check if it is in fact the sum
if 'lp' in dist_type_vals and 'emd' in dist_type_vals and 'lp-emd' in dist_type_vals:
lp_kernel = all_kernels[dist_type_vals.index('lp')]
emd_kernel = all_kernels[dist_type_vals.index('emd')]
lpemd_kernel = all_kernels[dist_type_vals.index('lp-emd')]
assert np.linalg.norm(lpemd_kernel - lp_kernel - emd_kernel) < _TOL