def __do_trial(self, g_init, fdgs, term3, trial): # add and delete edges. gtemp = g_init.copy() seed = (trial + int(time.time())) % (2 ** 32 - 1) rdm_state = np.random.RandomState(seed=seed) # which edges to change. # @todo: should we use just half of the adjacency matrix for undirected graphs? nb_vpairs = nx.number_of_nodes(g_init) * (nx.number_of_nodes(g_init) - 1) # @todo: what if fdgs is bigger than nb_vpairs? idx_change = rdm_state.randint(0, high=nb_vpairs, size=(fdgs if fdgs < nb_vpairs else nb_vpairs)) # print(idx_change) for item in idx_change: node1 = int(item / (nx.number_of_nodes(g_init) - 1)) node2 = (item - node1 * (nx.number_of_nodes(g_init) - 1)) if node2 >= node1: # skip the self pair. node2 += 1 # @todo: is the randomness correct? if not gtemp.has_edge(node1, node2): gtemp.add_edge(node1, node2) else: gtemp.remove_edge(node1, node2) # compute new distances. kernels_to_gtmp, _ = self._graph_kernel.compute(gtemp, self._dataset.graphs, **self._kernel_options) kernel_gtmp, _ = self._graph_kernel.compute(gtemp, gtemp, **self._kernel_options) if self._kernel_options['normalize']: kernels_to_gtmp = [kernels_to_gtmp[i] / np.sqrt(self.__gram_matrix_unnorm[i, i] * kernel_gtmp) for i in range(len(kernels_to_gtmp))] # normalize kernel_gtmp = 1 # @todo: not correct kernel value gram_with_gtmp = np.concatenate((np.array([kernels_to_gtmp]), np.copy(self._graph_kernel.gram_matrix)), axis=0) gram_with_gtmp = np.concatenate((np.array([[kernel_gtmp] + kernels_to_gtmp]).T, gram_with_gtmp), axis=1) dnew = compute_k_dis(0, range(1, 1 + len(self._dataset.graphs)), self.__alphas, gram_with_gtmp, term3=term3, withterm3=True) return gtemp, dnew
def _get_best_graph(Gn, gram_matrix): k_dis_list = [] for idx in range(len(Gn)): k_dis_list.append(compute_k_dis(idx, range(0, len(Gn)), [1 / len(Gn)] * len(Gn), gram_matrix, withterm3=False)) best_index = np.argmin(k_dis_list) best_dis = k_dis_list[best_index] best_graph = Gn[best_index].copy() return best_index, best_dis, best_graph
def __compute_distances_to_true_median(self):
# compute distance in kernel space for set median.
kernels_to_sm, _ = self._graph_kernel.compute(self.__set_median, self._dataset.graphs, **self._kernel_options)
kernel_sm, _ = self._graph_kernel.compute(self.__set_median, self.__set_median, **self._kernel_options)
if self._kernel_options['normalize']:
kernels_to_sm = [kernels_to_sm[i] / np.sqrt(self.__gram_matrix_unnorm[i, i] * kernel_sm) for i in range(len(kernels_to_sm))] # normalize
kernel_sm = 1
# @todo: not correct kernel value
gram_with_sm = np.concatenate((np.array([kernels_to_sm]), np.copy(self._graph_kernel.gram_matrix)), axis=0)
gram_with_sm = np.concatenate((np.array([[kernel_sm] + kernels_to_sm]).T, gram_with_sm), axis=1)
self.__k_dis_set_median = compute_k_dis(0, range(1, 1+len(self._dataset.graphs)),
[1 / len(self._dataset.graphs)] * len(self._dataset.graphs),
gram_with_sm, withterm3=False)
# compute distance in kernel space for generalized median.
kernels_to_gm, _ = self._graph_kernel.compute(self.__gen_median, self._dataset.graphs, **self._kernel_options)
kernel_gm, _ = self._graph_kernel.compute(self.__gen_median, self.__gen_median, **self._kernel_options)
if self._kernel_options['normalize']:
kernels_to_gm = [kernels_to_gm[i] / np.sqrt(self.__gram_matrix_unnorm[i, i] * kernel_gm) for i in range(len(kernels_to_gm))] # normalize
kernel_gm = 1
gram_with_gm = np.concatenate((np.array([kernels_to_gm]), np.copy(self._graph_kernel.gram_matrix)), axis=0)
gram_with_gm = np.concatenate((np.array([[kernel_gm] + kernels_to_gm]).T, gram_with_gm), axis=1)
self.__k_dis_gen_median = compute_k_dis(0, range(1, 1+len(self._dataset.graphs)),
[1 / len(self._dataset.graphs)] * len(self._dataset.graphs),
gram_with_gm, withterm3=False)
# compute distance in kernel space for each graph in median set.
k_dis_median_set = []
for idx in range(len(self._dataset.graphs)):
k_dis_median_set.append(compute_k_dis(idx+1, range(1, 1+len(self._dataset.graphs)),
[1 / len(self._dataset.graphs)] * len(self._dataset.graphs),
gram_with_gm, withterm3=False))
idx_k_dis_median_set_min = np.argmin(k_dis_median_set)
self.__k_dis_dataset = k_dis_median_set[idx_k_dis_median_set_min]
self.__best_from_dataset = self._dataset.graphs[idx_k_dis_median_set_min].copy()
if self._verbose >= 2:
print()
print('distance in kernel space for set median:', self.__k_dis_set_median)
print('distance in kernel space for generalized median:', self.__k_dis_gen_median)
print('minimum distance in kernel space for each graph in median set:', self.__k_dis_dataset)
print('distance in kernel space for each graph in median set:', k_dis_median_set)
def __kernel_knn_cv_best_ds(dataset_all, ds_name, knn_options, kernel_options,
gram_matrix_unnorm, time_precompute_gm,
train_examples, save_results, dir_save,
fn_output_detail, fn_output_summary):
Gn = dataset_all.graphs
y_all = dataset_all.targets
n_neighbors, n_splits, test_size = knn_options['n_neighbors'], knn_options[
'n_splits'], knn_options['test_size']
# get shuffles.
train_indices, test_indices, train_nums, y_app = __get_shuffles(
y_all, n_splits, test_size)
accuracies = []
for trial in range(len(train_indices)):
print('\ntrial =', trial)
train_index = train_indices[trial]
test_index = test_indices[trial]
G_app = [Gn[i] for i in train_index]
G_test = [Gn[i] for i in test_index]
y_test = [y_all[i] for i in test_index]
gm_unnorm_trial = gram_matrix_unnorm[
train_index, :][:, train_index].copy()
# get best graph from trainset according to distance in kernel space for each class.
best_graphs = []
train_nums_tmp = [0] + train_nums
print('\ngetting best graph from trainset for each class...')
for i_class in range(len(train_nums_tmp) - 1):
print(i_class + 1, 'of', len(train_nums_tmp) - 1, 'classes.')
i_start = int(np.sum(train_nums_tmp[0:i_class + 1]))
i_end = i_start + train_nums_tmp[i_class + 1]
G_class = G_app[i_start:i_end]
gm_unnorm_class = gm_unnorm_trial[i_start:i_end, i_start:i_end]
gm_class = normalize_gram_matrix(gm_unnorm_class.copy())
k_dis_list = []
for idx in range(len(G_class)):
k_dis_list.append(
compute_k_dis(idx,
range(0, len(G_class)),
[1 / len(G_class)] * len(G_class),
gm_class,
withterm3=False))
idx_k_dis_min = np.argmin(k_dis_list)
best_graphs.append(G_class[idx_k_dis_min].copy())
# perform k-nn.
print('\nperforming k-nn...')
# compute dis_mat between medians.
dataset = dataset_all.copy()
dataset.load_graphs([g.copy() for g in best_graphs], targets=None)
gm_app_unnorm, _ = __compute_gram_matrix_unnorm(
dataset, kernel_options.copy())
# compute the entire Gram matrix.
graph_kernel = __get_graph_kernel(dataset.copy(),
kernel_options.copy())
kernels_to_best_graphs = []
for g in best_graphs:
kernels_to_best_graph, _ = graph_kernel.compute(
g, G_test, **kernel_options.copy())
kernels_to_best_graphs.append(kernels_to_best_graph)
kernels_to_best_graphs = np.array(kernels_to_best_graphs)
gm_all = np.concatenate((gm_app_unnorm, kernels_to_best_graphs),
axis=1)
gm_all = np.concatenate(
(gm_all,
np.concatenate(
(kernels_to_best_graphs.T,
gram_matrix_unnorm[test_index, :][:, test_index].copy()),
axis=1)),
axis=0)
gm_all = normalize_gram_matrix(gm_all.copy())
dis_mat, _, _, _ = compute_distance_matrix(gm_all)
N = len(best_graphs)
d_app = dis_mat[range(N), :][:, range(N)].copy()
d_test = np.zeros((N, len(test_index)))
for i in range(N):
for j in range(len(test_index)):
d_test[i, j] = dis_mat[i, j]
accuracies.append(
knn_classification(d_app,
d_test,
y_app,
y_test,
n_neighbors,
verbose=True,
text=train_examples))
# write result detail.
if save_results:
f_detail = open(dir_save + fn_output_detail, 'a')
print('writing results to files...')
csv.writer(f_detail).writerow([
ds_name, kernel_options['name'], train_examples, trial,
knn_options['n_neighbors'],
len(gm_all), knn_options['test_size'], accuracies[-1][0],
accuracies[-1][1]
])
f_detail.close()
results = {}
results['ave_perf_train'] = np.mean([i[0] for i in accuracies], axis=0)
results['std_perf_train'] = np.std([i[0] for i in accuracies],
axis=0,
ddof=1)
results['ave_perf_test'] = np.mean([i[1] for i in accuracies], axis=0)
results['std_perf_test'] = np.std([i[1] for i in accuracies],
axis=0,
ddof=1)
# write result summary for each letter.
if save_results:
f_summary = open(dir_save + fn_output_summary, 'a')
csv.writer(f_summary).writerow([
ds_name, kernel_options['name'], train_examples,
knn_options['n_neighbors'], knn_options['test_size'],
results['ave_perf_train'], results['ave_perf_test'],
results['std_perf_train'], results['std_perf_test'],
time_precompute_gm
])
f_summary.close()
def run(self):
self._graph_kernel = get_graph_kernel_by_name(self._kernel_options['name'],
node_labels=self._dataset.node_labels,
edge_labels=self._dataset.edge_labels,
node_attrs=self._dataset.node_attrs,
edge_attrs=self._dataset.edge_attrs,
ds_infos=self._dataset.get_dataset_infos(keys=['directed']),
kernel_options=self._kernel_options)
# record start time.
start = time.time()
# 1. precompute gram matrix.
if self.__gram_matrix_unnorm is None:
gram_matrix, run_time = self._graph_kernel.compute(self._dataset.graphs, **self._kernel_options)
self.__gram_matrix_unnorm = self._graph_kernel.gram_matrix_unnorm
end_precompute_gm = time.time()
self.__runtime_precompute_gm = end_precompute_gm - start
else:
if self.__runtime_precompute_gm is None:
raise Exception('Parameter "runtime_precompute_gm" must be given when using pre-computed Gram matrix.')
self._graph_kernel.gram_matrix_unnorm = self.__gram_matrix_unnorm
if self._kernel_options['normalize']:
self._graph_kernel.gram_matrix = self._graph_kernel.normalize_gm(np.copy(self.__gram_matrix_unnorm))
else:
self._graph_kernel.gram_matrix = np.copy(self.__gram_matrix_unnorm)
end_precompute_gm = time.time()
start -= self.__runtime_precompute_gm
# 2. compute k nearest neighbors of phi in D_N.
if self._verbose >= 2:
print('\nstart computing k nearest neighbors of phi in D_N...\n')
D_N = self._dataset.graphs
if self.__alphas is None:
self.__alphas = [1 / len(D_N)] * len(D_N)
k_dis_list = [] # distance between g_star and each graph.
term3 = 0
for i1, a1 in enumerate(self.__alphas):
for i2, a2 in enumerate(self.__alphas):
term3 += a1 * a2 * self._graph_kernel.gram_matrix[i1, i2]
for idx in range(len(D_N)):
k_dis_list.append(compute_k_dis(idx, range(0, len(D_N)), self.__alphas, self._graph_kernel.gram_matrix, term3=term3, withterm3=True))
# sort.
sort_idx = np.argsort(k_dis_list)
dis_gs = [k_dis_list[idis] for idis in sort_idx[0:self.__k]] # the k shortest distances.
nb_best = len(np.argwhere(dis_gs == dis_gs[0]).flatten().tolist())
g0hat_list = [D_N[idx].copy() for idx in sort_idx[0:nb_best]] # the nearest neighbors of phi in D_N
self.__best_from_dataset = g0hat_list[0] # get the first best graph if there are muitlple.
self.__k_dis_dataset = dis_gs[0]
if self.__k_dis_dataset == 0: # get the exact pre-image.
end_generate_preimage = time.time()
self.__runtime_generate_preimage = end_generate_preimage - end_precompute_gm
self.__runtime_total = end_generate_preimage - start
self.__preimage = self.__best_from_dataset.copy()
self.__k_dis_preimage = self.__k_dis_dataset
if self._verbose:
print()
print('=============================================================================')
print('The exact pre-image is found from the input dataset.')
print('-----------------------------------------------------------------------------')
print('Distance in kernel space for the best graph from dataset and for preimage:', self.__k_dis_dataset)
print('Time to pre-compute Gram matrix:', self.__runtime_precompute_gm)
print('Time to generate pre-images:', self.__runtime_generate_preimage)
print('Total time:', self.__runtime_total)
print('=============================================================================')
print()
return
dhat = dis_gs[0] # the nearest distance
Gk = [D_N[ig].copy() for ig in sort_idx[0:self.__k]] # the k nearest neighbors
Gs_nearest = [nx.convert_node_labels_to_integers(g) for g in Gk] # [g.copy() for g in Gk]
# 3. start iterations.
if self._verbose >= 2:
print('starting iterations...')
gihat_list = []
dihat_list = []
r = 0
dis_of_each_itr = [dhat]
if self.__parallel:
self._kernel_options['parallel'] = None
self.__itrs = 0
self.__num_updates = 0
timer = Timer(self.__time_limit_in_sec)
while not self.__termination_criterion_met(timer, self.__itrs, r):
print('\n- r =', r)
found = False
dis_bests = dis_gs + dihat_list
# compute numbers of edges to be inserted/deleted.
# @todo what if the log is negetive? how to choose alpha (scalar)?
fdgs_list = np.array(dis_bests)
if np.min(fdgs_list) < 1: # in case the log is negetive.
fdgs_list /= np.min(fdgs_list)
fdgs_list = [int(item) for item in np.ceil(np.log(fdgs_list))]
if np.min(fdgs_list) < 1: # in case the log is smaller than 1.
fdgs_list = np.array(fdgs_list) + 1
# expand the number of modifications to increase the possiblity.
nb_vpairs_list = [nx.number_of_nodes(g) * (nx.number_of_nodes(g) - 1) for g in (Gs_nearest + gihat_list)]
nb_vpairs_min = np.min(nb_vpairs_list)
idx_fdgs_max = np.argmax(fdgs_list)
fdgs_max_old = fdgs_list[idx_fdgs_max]
fdgs_max = fdgs_max_old
nb_modif = 1
for idx, nb in enumerate(range(nb_vpairs_min, nb_vpairs_min - fdgs_max, -1)):
nb_modif *= nb / (fdgs_max - idx)
while fdgs_max < nb_vpairs_min and nb_modif < self.__l:
fdgs_max += 1
nb_modif *= (nb_vpairs_min - fdgs_max + 1) / fdgs_max
nb_increase = int(fdgs_max - fdgs_max_old)
if nb_increase > 0:
fdgs_list += 1
for ig, gs in enumerate(Gs_nearest + gihat_list):
if self._verbose >= 2:
print('-- computing', ig + 1, 'graphs out of', len(Gs_nearest) + len(gihat_list))
gnew, dhat, found = self.__generate_l_graphs(gs, fdgs_list[ig], dhat, ig, found, term3)
if found:
r = 0
gihat_list = [gnew]
dihat_list = [dhat]
else:
r += 1
dis_of_each_itr.append(dhat)
self.__itrs += 1
if self._verbose >= 2:
print('Total number of iterations is', self.__itrs, '.')
print('The preimage is updated', self.__num_updates, 'times.')
print('The shortest distances for previous iterations are', dis_of_each_itr, '.')
# get results and print.
end_generate_preimage = time.time()
self.__runtime_generate_preimage = end_generate_preimage - end_precompute_gm
self.__runtime_total = end_generate_preimage - start
self.__preimage = (g0hat_list[0] if len(gihat_list) == 0 else gihat_list[0])
self.__k_dis_preimage = dhat
if self._verbose:
print()
print('=============================================================================')
print('Finished generation of preimages.')
print('-----------------------------------------------------------------------------')
print('Distance in kernel space for the best graph from dataset:', self.__k_dis_dataset)
print('Distance in kernel space for the preimage:', self.__k_dis_preimage)
print('Total number of iterations for optimizing:', self.__itrs)
print('Total number of updating preimage:', self.__num_updates)
print('Time to pre-compute Gram matrix:', self.__runtime_precompute_gm)
print('Time to generate pre-images:', self.__runtime_generate_preimage)
print('Total time:', self.__runtime_total)
print('=============================================================================')
print()
def xp_simple_preimage():
import numpy as np
"""**1. Get dataset.**"""
from gklearn.utils import Dataset, split_dataset_by_target
# Predefined dataset name, use dataset "MAO".
ds_name = 'MAO'
# The node/edge labels that will not be used in the computation.
irrelevant_labels = {
'node_attrs': ['x', 'y', 'z'],
'edge_labels': ['bond_stereo']
}
# Initialize a Dataset.
dataset_all = Dataset()
# Load predefined dataset "MAO".
dataset_all.load_predefined_dataset(ds_name)
# Remove irrelevant labels.
dataset_all.remove_labels(**irrelevant_labels)
# Split the whole dataset according to the classification targets.
datasets = split_dataset_by_target(dataset_all)
# Get the first class of graphs, whose median preimage will be computed.
dataset = datasets[0]
len(dataset.graphs)
"""**2. Set parameters.**"""
import multiprocessing
# Parameters for MedianPreimageGenerator (our method).
mpg_options = {
'fit_method':
'k-graphs', # how to fit edit costs. "k-graphs" means use all graphs in median set when fitting.
'init_ecc': [4, 4, 2, 1, 1, 1], # initial edit costs.
'ds_name': ds_name, # name of the dataset.
'parallel': True, # whether the parallel scheme is to be used.
'time_limit_in_sec':
0, # maximum time limit to compute the preimage. If set to 0 then no limit.
'max_itrs':
10, # maximum iteration limit to optimize edit costs. If set to 0 then no limit.
'max_itrs_without_update':
3, # If the times that edit costs is not update is more than this number, then the optimization stops.
'epsilon_residual':
0.01, # In optimization, the residual is only considered changed if the change is bigger than this number.
'epsilon_ec':
0.1, # In optimization, the edit costs are only considered changed if the changes are bigger than this number.
'verbose': 2 # whether to print out results.
}
# Parameters for graph kernel computation.
kernel_options = {
'name': 'PathUpToH', # use path kernel up to length h.
'depth': 9,
'k_func': 'MinMax',
'compute_method': 'trie',
'parallel': 'imap_unordered', # or None
'n_jobs': multiprocessing.cpu_count(),
'normalize':
True, # whether to use normalized Gram matrix to optimize edit costs.
'verbose': 2 # whether to print out results.
}
# Parameters for GED computation.
ged_options = {
'method': 'IPFP', # use IPFP huristic.
'initialization_method': 'RANDOM', # or 'NODE', etc.
'initial_solutions':
10, # when bigger than 1, then the method is considered mIPFP.
'edit_cost': 'CONSTANT', # use CONSTANT cost.
'attr_distance':
'euclidean', # the distance between non-symbolic node/edge labels is computed by euclidean distance.
'ratio_runs_from_initial_solutions': 1,
'threads': multiprocessing.cpu_count(
), # parallel threads. Do not work if mpg_options['parallel'] = False.
'init_option': 'EAGER_WITHOUT_SHUFFLED_COPIES'
}
# Parameters for MedianGraphEstimator (Boria's method).
mge_options = {
'init_type':
'MEDOID', # how to initial median (compute set-median). "MEDOID" is to use the graph with smallest SOD.
'random_inits':
10, # number of random initialization when 'init_type' = 'RANDOM'.
'time_limit':
600, # maximum time limit to compute the generalized median. If set to 0 then no limit.
'verbose': 2, # whether to print out results.
'refine': False # whether to refine the final SODs or not.
}
print('done.')
"""**3. Compute the Gram matrix and distance matrix.**"""
from gklearn.utils.utils import get_graph_kernel_by_name
# Get a graph kernel instance.
graph_kernel = get_graph_kernel_by_name(
kernel_options['name'],
node_labels=dataset.node_labels,
edge_labels=dataset.edge_labels,
node_attrs=dataset.node_attrs,
edge_attrs=dataset.edge_attrs,
ds_infos=dataset.get_dataset_infos(keys=['directed']),
kernel_options=kernel_options)
# Compute Gram matrix.
gram_matrix, run_time = graph_kernel.compute(dataset.graphs,
**kernel_options)
# Compute distance matrix.
from gklearn.utils import compute_distance_matrix
dis_mat, _, _, _ = compute_distance_matrix(gram_matrix)
print('done.')
"""**4. Find the candidate graph.**"""
from gklearn.preimage.utils import compute_k_dis
# Number of the nearest neighbors.
k_neighbors = 10
# For each graph G in dataset, compute the distance between its image \Phi(G) and the mean of its neighbors' images.
dis_min = np.inf # the minimum distance between possible \Phi(G) and the mean of its neighbors.
for idx, G in enumerate(dataset.graphs):
# Find the k nearest neighbors of G.
dis_list = dis_mat[
idx] # distance between \Phi(G) and image of each graphs.
idx_sort = np.argsort(
dis_list) # sort distances and get the sorted indices.
idx_nearest = idx_sort[1:k_neighbors +
1] # indices of the k-nearest neighbors.
dis_k_nearest = [dis_list[i] for i in idx_nearest
] # k-nearest distances, except the 0.
G_k_nearest = [dataset.graphs[i]
for i in idx_nearest] # k-nearest neighbors.
# Compute the distance between \Phi(G) and the mean of its neighbors.
dis_tmp = compute_k_dis(
idx, # the index of G in Gram matrix.
idx_nearest, # the indices of the neighbors
[1 / k_neighbors] * k_neighbors, # coefficients for neighbors.
gram_matrix,
withterm3=False)
# Check if the new distance is smallers.
if dis_tmp < dis_min:
dis_min = dis_tmp
G_cand = G
G_neighbors = G_k_nearest
print('The minimum distance is', dis_min)
"""**5. Run median preimage generator.**"""
from gklearn.preimage import MedianPreimageGenerator
# Set the dataset as the k-nearest neighbors.
dataset.load_graphs(G_neighbors)
# Create median preimage generator instance.
mpg = MedianPreimageGenerator()
# Add dataset.
mpg.dataset = dataset
# Set parameters.
mpg.set_options(**mpg_options.copy())
mpg.kernel_options = kernel_options.copy()
mpg.ged_options = ged_options.copy()
mpg.mge_options = mge_options.copy()
# Run.
mpg.run()
"""**4. Get results.**"""
# Get results.
import pprint
pp = pprint.PrettyPrinter(indent=4) # pretty print
results = mpg.get_results()
pp.pprint(results)
draw_graph(mpg.set_median)
draw_graph(mpg.gen_median)
draw_graph(G_cand)
# Compute Gram matrix.
gram_matrix, run_time = graph_kernel.compute(
dataset.graphs,
parallel='imap_unordered', # or None.
n_jobs=multiprocessing.cpu_count(), # number of parallel jobs.
normalize=True, # whether to return normalized Gram matrix.
verbose=2 # whether to print out results.
)
"""**3. Compute distance in kernel space.**
Given a dataset $\mathcal{G}_N$, compute the distance in kernel space between the image of $G_1 \in \mathcal{G}_N$ and the mean of images of $\mathcal{G}_k \subset \mathcal{G}_N$.
"""
from gklearn.preimage.utils import compute_k_dis
# Index of $G_1$.
idx_1 = 10
# Indices of graphs in $\mathcal{G}_k$.
idx_graphs = range(0, 10)
# Compute the distance in kernel space.
dis_k = compute_k_dis(
idx_1,
idx_graphs,
[1 / len(idx_graphs)] * len(
idx_graphs
), # weights for images of graphs in $\mathcal{G}_k$; all equal when computing the mean.
gram_matrix, # gram matrix of al graphs.
withterm3=False)
print(dis_k)
