def main():
psi = 256
t = 100
parser = argparse.ArgumentParser()
parser.add_argument('-d', '--dataset', type=str, default='ENZYMES')
parser.add_argument('--crossvalidation',
default=True,
action='store_true',
help='Enable a 10-fold crossvalidation')
parser.add_argument('--h',
type=int,
required=False,
default=5,
help="(Max) number of WL iterations")
args = parser.parse_args()
dataset = args.dataset
h = args.h
data_path = os.path.join('./data', dataset)
output_path = os.path.join('output', dataset)
results_path = os.path.join('results', dataset)
for path in [output_path, results_path]:
if not os.path.exists(path):
os.makedirs(path)
label_file = os.path.join(data_path, 'Labels.txt')
y = np.array(read_labels(label_file))
cv = StratifiedKFold(n_splits=10, shuffle=True)
accuracy_scores = []
label_sequences = compute_wl_embeddings_continuous_by_IK(
data_path, h, t, psi)
label_sequences = np.array(label_sequences)
for train_index, test_index in cv.split(label_sequences, y):
X_train = label_sequences[train_index]
X_test = label_sequences[test_index]
y_train, y_test = y[train_index], y[test_index]
# use liblinear
X_train = average_graph(X_train, t)
X_test = average_graph(X_test, t)
gs = LinearSVC().fit(X_train, y_train)
y_pred = gs.predict(X_test)
accuracy_scores.append(accuracy_score(y_test, y_pred))
print(accuracy_scores)
if not args.crossvalidation:
break
if args.crossvalidation:
print('Mean 10-fold accuracy: {:2.2f} +- {:2.2f} %'.format(
np.mean(accuracy_scores) * 100,
np.std(accuracy_scores) * 100))
else:
print('Final accuracy: {:2.3f} %'.format(
np.mean(accuracy_scores) * 100))
def main(args, logger):
graphs = [ig.read(filename) for filename in args.FILES]
labels = read_labels(args.labels)
# Set the label to be uniform over all graphs in case no labels are
# available. This essentially changes our iteration to degree-based
# checks.
for graph in graphs:
if 'label' not in graph.vs.attributes():
graph.vs['label'] = [0] * len(graph.vs)
logger.info('Read {} graphs and {} labels'.format(len(graphs),
len(labels)))
assert len(graphs) == len(labels)
pwl = PersistentWeisfeilerLehman(
use_cycle_persistence=args.use_cycle_persistence,
use_original_features=args.use_original_features,
use_label_persistence=args.use_persistence_features,
)
if args.use_cycle_persistence:
logger.info('Using cycle persistence')
y = LabelEncoder().fit_transform(labels)
X, num_columns_per_iteration = pwl.transform(graphs, args.num_iterations)
X = StandardScaler().fit_transform(X)
X = MinMaxScaler().fit_transform(X)
logger.info('Finished persistent Weisfeiler-Lehman transformation')
logger.info('Obtained ({} x {}) feature matrix'.format(
X.shape[0], X.shape[1]))
num_classes = len(np.bincount(y))
fig, ax = plt.subplots(nrows=num_classes,
ncols=2,
sharex=True,
sharey=False,
squeeze=False)
for index in range(num_classes):
ax[index][0].matshow(X[y == index], aspect='auto')
ax[index][0].set_title(f'Class {index} (features)')
ax[index][1].matshow(np.mean(X[y == index], axis=0).reshape(1, -1),
aspect='auto')
ax[index][1].set_title(f'Class {index} (mean)')
plt.show()
def main(args, logger):
graphs = [ig.read(filename) for filename in args.FILES]
labels = read_labels(args.labels)
# Set the label to be uniform over all graphs in case no labels are
# available. This essentially changes our iteration to degree-based
# checks.
for graph in graphs:
if 'label' not in graph.vs.attributes():
graph.vs['label'] = [0] * len(graph.vs)
logger.info('Read {} graphs and {} labels'.format(len(graphs),
len(labels)))
assert len(graphs) == len(labels)
pwl = PersistentWeisfeilerLehman(
use_cycle_persistence=args.use_cycle_persistence,
use_original_features=args.use_original_features,
metric=args.metric,
use_label_persistence=True,
)
if args.use_cycle_persistence:
logger.info('Using cycle persistence')
y = LabelEncoder().fit_transform(labels)
# This ignores *all* other feature generation methods and falls back
# to the original Weisfeiler--Lehman subtree kernel.
if args.use_subtree_features:
logger.info('Using original subtree features')
wl_subtree = WeisfeilerLehmanSubtree()
X, num_columns_per_iteration = \
wl_subtree.transform(graphs, args.num_iterations)
else:
X, num_columns_per_iteration = \
pwl.transform(graphs, args.num_iterations)
logger.info('Finished persistent Weisfeiler-Lehman transformation')
logger.info('Obtained ({} x {}) feature matrix'.format(
X.shape[0], X.shape[1]))
np.random.seed(42)
mean_accuracies = []
params = [
'balanced', 'num_iterations', 'filtration', 'use_cycle_persistence',
'use_original_features', 'use_subtree_features', 'metric'
]
cv_results = []
entry = {}
for param in params:
entry[param] = args.__dict__[param]
entry['dataset'] = dirname(args.FILES[0]).split('/')[1]
for i in range(10):
# Contains accuracy scores for each cross validation step; the
# means of this list will be used later on.
accuracy_scores = []
cv = StratifiedKFold(n_splits=10, shuffle=True, random_state=i)
for n, indices in enumerate(cv.split(X, y)):
entry_fold = copy.copy(entry)
train_index = indices[0]
test_index = indices[1]
pipeline = Pipeline(
[('fs', FeatureSelector(num_columns_per_iteration)),
('clf',
RandomForestClassifier(
class_weight='balanced' if args.balanced else None,
random_state=42))], )
grid_params = {
'fs__num_iterations': np.arange(0, args.num_iterations + 1),
'clf__n_estimators': [25, 50, 100],
}
clf = GridSearchCV(pipeline,
grid_params,
cv=StratifiedKFold(n_splits=5, shuffle=True),
iid=False,
scoring='accuracy',
n_jobs=4)
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
# TODO: need to discuss whether this is 'allowed' or smart
# to do; this assumes normality of the attributes.
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
scaler = MinMaxScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
acc = accuracy_score(y_test, y_pred)
accuracy_scores.append(acc)
for param, param_val in clf.best_params_.items():
entry_fold[param] = param_val
entry[param] = ''
entry_fold['fold'] = n + 1
entry_fold['it'] = i
entry_fold['acc'] = acc * 100
entry_fold['std'] = 0.0
cv_results.append(entry_fold)
logger.info('Best classifier for this fold:{}'.format(
clf.best_params_))
mean_accuracies.append(np.mean(accuracy_scores))
logger.info(
' - Mean 10-fold accuracy: {:2.2f} [running mean over all folds: {:2.2f}]'
.format(mean_accuracies[-1] * 100,
np.mean(mean_accuracies) * 100))
entry['fold'] = 'all'
entry['it'] = 'all'
entry['acc'] = np.mean(mean_accuracies) * 100
entry['std'] = np.std(mean_accuracies) * 100
cv_results.append(entry)
logger.info('Accuracy: {:2.2f} +- {:2.2f}'.format(
np.mean(mean_accuracies) * 100,
np.std(mean_accuracies) * 100))
if exists(args.result_file):
with open(args.result_file, 'a') as f:
pd.DataFrame(cv_results).to_csv(f, index=False, header=None)
else:
pd.DataFrame(cv_results).to_csv(args.result_file, index=False)
type=str,
help='Labels file',
required=True)
parser.add_argument('-n',
'--num-graphs',
type=int,
required=True,
help='Sample size')
parser.add_argument('-o',
'--out-dir',
type=str,
required=True,
help='Output directory')
args = parser.parse_args()
labels = read_labels(args.labels)
y = LabelEncoder().fit_transform(labels)
n = len(y)
sss = StratifiedShuffleSplit(n_splits=1,
random_state=23,
train_size=args.num_graphs)
for train_index, _ in sss.split(range(n), y):
train_index = sorted(train_index)
files = np.array(args.FILES)
files = files[train_index]
try:
os.makedirs(args.out_dir)
def main(args, logger):
graphs = [ig.read(filename) for filename in args.FILES]
labels = read_labels(args.labels)
# Set the label to be uniform over all graphs in case no labels are
# available. This essentially changes our iteration to degree-based
# checks.
for graph in graphs:
if 'label' not in graph.vs.attributes():
graph.vs['label'] = [0] * len(graph.vs)
logger.info('Read {} graphs and {} labels'.format(len(graphs),
len(labels)))
assert len(graphs) == len(labels)
# Calculate graph kernel
gram_matrix = gk.CalculateEdgeHistKernel(graphs)
y = LabelEncoder().fit_transform(labels)
#np.random.seed(42)
mean_accuracies = []
params = ['balanced']
cv_results = []
entry = {}
for param in params:
entry[param] = args.__dict__[param]
entry['dataset'] = dirname(args.FILES[0]).split('/')[1]
entry['baseline'] = 'edge hist kernel'
for i in range(10):
# Contains accuracy scores for each cross validation step; the
# means of this list will be used later on.
accuracy_scores = []
cv = StratifiedKFold(n_splits=10, shuffle=True, random_state=i)
for n, indices in enumerate(cv.split(graphs, y)):
entry_fold = copy.copy(entry)
train_index = indices[0]
test_index = indices[1]
pipeline = Pipeline(
[('clf',
SVC(class_weight='balanced' if args.balanced else None,
random_state=42,
kernel='precomputed'))], )
grid_params = {'clf__C': [1e-1, 1e0, 1e1]}
X_train, X_test = gram_matrix[
train_index][:,
train_index], gram_matrix[test_index][:,
train_index]
y_train, y_test = y[train_index], y[test_index]
kgscv = KernelGridSearchCV(
pipeline,
param_grid=grid_params,
cv=cv,
)
kgscv.fit(X_train, y_train)
p = kgscv._best_params
sc = kgscv._best_score
clf = kgscv._best_estimator
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
acc = accuracy_score(y_test, y_pred)
accuracy_scores.append(acc)
for param, param_val in kgscv._best_params.items():
entry_fold[param] = param_val
entry[param] = ''
entry_fold['fold'] = n + 1
entry_fold['it'] = i
entry_fold['acc'] = acc * 100
entry_fold['std'] = 0.0
cv_results.append(entry_fold)
logger.info('Best classifier for this fold:{}'.format(
kgscv._best_params))
mean_accuracies.append(np.mean(accuracy_scores))
logger.info(
' - Mean 10-fold accuracy: {:2.2f} [running mean over all folds: {:2.2f}]'
.format(mean_accuracies[-1] * 100,
np.mean(mean_accuracies) * 100))
entry['fold'] = 'all'
entry['it'] = 'all'
entry['acc'] = np.mean(mean_accuracies) * 100
entry['std'] = np.std(mean_accuracies) * 100
cv_results.append(entry)
logger.info('Accuracy: {:2.2f} +- {:2.2f}'.format(
np.mean(mean_accuracies) * 100,
np.std(mean_accuracies) * 100))
if exists(args.result_file):
with open(args.result_file, 'a') as f:
pd.DataFrame(cv_results).to_csv(f, index=False, header=None)
else:
pd.DataFrame(cv_results).to_csv(args.result_file, index=False)
def main(args, logger):
# Read all graphs and labels; there is no direct way of checking
# that the labels are 'correct' for the graphs, but at least the
# code will check that they have the same cardinality.
graphs = [ig.read(filename) for filename in args.FILES]
labels = read_labels(args.labels)
# Simple pre-processing to ensure that all graphs are set up
# equally.
#
# TODO: make this into a shared function?
for graph in graphs:
# Set the label to be uniform over all graphs in case no labels are
# available. This essentially changes our iteration to degree-based
# checks.
if 'label' not in graph.vs.attributes():
graph.vs['label'] = [0] * len(graph.vs)
# Reset edge weights if they already exist
if 'weight' in graph.es.attributes():
graph.es['weight'] = [0] * len(graph.es)
logger.info('Read {} graphs and {} labels'.format(len(graphs),
len(labels)))
assert len(graphs) == len(labels)
# Replace selected metric if necessary; this only applies to the
# uniform metric shortcut.
if args.use_uniform_metric:
args.metric = 'uniform'
pwl = PersistentWeisfeilerLehman(
use_cycle_persistence=args.use_cycle_persistence,
use_original_features=args.use_original_features,
use_label_persistence=True,
metric=args.metric,
p=args.power,
smooth=args.smooth)
if args.use_cycle_persistence:
logger.info('Using cycle persistence')
# Ensures that labels are encoded correctly, regardless of whether
# they are numerical or not.
y = LabelEncoder().fit_transform(labels)
# This ignores *all* other feature generation methods and falls back
# to the original Weisfeiler--Lehman subtree kernel.
if args.use_subtree_features:
logger.info('Using original subtree features')
wl_subtree = WeisfeilerLehmanSubtree()
X, num_columns_per_iteration = \
wl_subtree.transform(graphs, args.num_iterations)
else:
X, num_columns_per_iteration = \
pwl.transform(graphs, args.num_iterations)
logger.info('Finished persistent Weisfeiler-Lehman transformation')
logger.info('Obtained ({} x {}) feature matrix'.format(
X.shape[0], X.shape[1]))
np.random.seed(42)
cv = StratifiedKFold(n_splits=10, shuffle=True)
mean_accuracies = []
for i in range(10):
# Contains accuracy scores for each cross validation step; the
# means of this list will be used later on.
accuracy_scores = []
for train_index, test_index in cv.split(X, y):
rf_clf = RandomForestClassifier(
n_estimators=50,
class_weight='balanced' if args.balanced else None,
random_state=42)
if args.grid_search:
pipeline = Pipeline(
[('fs', FeatureSelector(num_columns_per_iteration)),
('clf', rf_clf)], )
grid_params = {
'fs__num_iterations': np.arange(0,
args.num_iterations + 1),
'clf__n_estimators': [10, 20, 50, 100, 150, 200],
}
clf = GridSearchCV(pipeline,
grid_params,
cv=StratifiedKFold(n_splits=10,
shuffle=True),
iid=False,
scoring='accuracy',
n_jobs=4)
else:
clf = rf_clf
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
scaler = MinMaxScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
accuracy_scores.append(accuracy_score(y_test, y_pred))
logger.debug('Best classifier for this fold: {}'.format(clf))
if args.grid_search:
logger.debug('Best parameters for this fold: {}'.format(
clf.best_params_))
else:
logger.debug('Best parameters for this fold: {}'.format(
clf.get_params()))
mean_accuracies.append(np.mean(accuracy_scores))
logger.info(
''' - Mean 10-fold accuracy: {:2.2f} [running mean over all
folds: {:2.2f}]'''.format(mean_accuracies[-1] * 100,
np.mean(mean_accuracies) * 100))
logger.info('Accuracy: {:2.2f} +- {:2.2f}'.format(
np.mean(mean_accuracies) * 100,
np.std(mean_accuracies) * 100))
action='store_true',
help='Make random forest classifier balanced')
parser.add_argument('-l',
'--labels',
type=str,
help='Labels file',
required=True)
parser.add_argument('-n',
'--num-iterations',
default=3,
type=int,
help='Number of Weisfeiler-Lehman iterations')
args = parser.parse_args()
graphs = [ig.read(filename) for filename in args.FILES]
y = np.array(read_labels(args.labels))
wl = WeisfeilerLehman()
label_dicts = wl.fit_transform(graphs, args.num_iterations)
# Each entry in the list represents the label sequence of a single
# graph. The label sequence contains the vertices in its rows, and
# the individual iterations in its columns.
#
# Hence, (i, j) will contain the label of vertex i at iteration j.
label_sequences = [
np.full((len(graph.vs), args.num_iterations + 1), np.nan)
for graph in graphs
]
for iteration in sorted(label_dicts.keys()):
def main(args, logger):
graphs = [ig.read(filename) for filename in args.FILES]
labels = read_labels(args.labels)
# Stores *all* vertex labels of the given graph in order to
# determine the conversion factor for persistence diagrams.
vertex_labels = set()
# Set the label to be uniform over all graphs in case no labels are
# available. This essentially changes our iteration to degree-based
# checks.
for graph in graphs:
if 'label' not in graph.vs.attributes():
graph.vs['label'] = [0] * len(graph.vs)
vertex_labels.update(graph.vs['label'])
logger.info('Read {} graphs and {} labels'.format(len(graphs),
len(labels)))
assert len(graphs) == len(labels)
pwl = PersistentWeisfeilerLehman(
use_cycle_persistence=args.use_cycle_persistence,
use_original_features=args.use_original_features,
use_label_persistence=True,
store_persistence_diagrams=True,
)
if args.use_cycle_persistence:
logger.info('Using cycle persistence')
y = LabelEncoder().fit_transform(labels)
X, num_columns_per_iteration = pwl.transform(graphs, args.num_iterations)
persistence_diagrams = pwl._persistence_diagrams
fig, ax = plt.subplots(args.num_iterations + 1)
for iteration in persistence_diagrams.keys():
M = collections.defaultdict(list)
for index, pd in enumerate(persistence_diagrams[iteration]):
label = y[index]
for _, d, _ in pd:
M[label].append(d)
d_min = sys.float_info.max
d_max = -d_min
for hist in M.values():
d_min = min(d_min, min(hist))
d_max = max(d_max, max(hist))
bins = np.linspace(d_min, d_max, 10)
for label, hist in M.items():
sns.distplot(hist,
bins=bins,
rug=True,
kde=True,
hist=False,
ax=ax[iteration])
plt.show()
L = len(vertex_labels)
assert L > 0
original_labels = pwl._original_labels
# Will store *all* persistence diagrams in the form of a probability
# distribution.
M = np.zeros((len(graphs), (args.num_iterations + 1) * L))
# Will store *all* pairwise distances according to the
# Jensen--Shannon divergence (JS), or, alternatively,
# the Kullback--Leibler divergence (KL).
D_KL = np.zeros((len(graphs), len(graphs)))
D_JS = np.zeros((len(graphs), len(graphs)))
D = np.zeros((len(graphs), len(graphs)))
for iteration in persistence_diagrams.keys():
M, D_KL, D_JS = make_kernel_matrices(
persistence_diagrams[iteration],
original_labels, # notice that they do *not* change
L)
D += D_JS
D = -D
fig, ax = plt.subplots(len(set(y)))
for label in sorted(set(y)):
ax[label].matshow(M[y == label], aspect='auto', vmin=0, vmax=1)
plt.show()
logger.info('Finished persistent Weisfeiler-Lehman transformation')
logger.info('Obtained ({} x {}) feature matrix'.format(
X.shape[0], X.shape[1]))
np.random.seed(42)
cv = StratifiedKFold(n_splits=10, shuffle=True)
mean_accuracies = []
for i in range(10):
# Contains accuracy scores for each cross validation step; the
# means of this list will be used later on.
accuracy_scores = []
for train_index, test_index in cv.split(X, y):
rf_clf = RandomForestClassifier(
n_estimators=50,
class_weight='balanced' if args.balanced else None)
if args.grid_search:
pipeline = Pipeline(
[('fs', FeatureSelector(num_columns_per_iteration)),
('clf', rf_clf)], )
grid_params = {
'fs__num_iterations': np.arange(0,
args.num_iterations + 1),
'clf__n_estimators': [10, 20, 50, 100, 150, 200],
}
clf = GridSearchCV(pipeline,
grid_params,
cv=StratifiedKFold(n_splits=10,
shuffle=True),
iid=False,
scoring='accuracy',
n_jobs=4)
else:
clf = rf_clf
clf = SVC(kernel='precomputed')
clf.fit(D, y)
y_test = y
y_pred = clf.predict(D)
#X_train, X_test = X[train_index], X[test_index]
#y_train, y_test = y[train_index], y[test_index]
## TODO: need to discuss whether this is 'allowed' or smart
## to do; this assumes normality of the attributes.
#scaler = StandardScaler()
#X_train = scaler.fit_transform(X_train)
#X_test = scaler.transform(X_test)
#scaler = MinMaxScaler()
#X_train = scaler.fit_transform(X_train)
#X_test = scaler.transform(X_test)
#clf.fit(X_train, y_train)
#y_pred = clf.predict(X_test)
accuracy_scores.append(accuracy_score(y_test, y_pred))
logger.debug('Best classifier for this fold: {}'.format(clf))
if args.grid_search:
logger.debug('Best parameters for this fold: {}'.format(
clf.best_params_))
else:
logger.debug('Best parameters for this fold: {}'.format(
clf.get_params()))
mean_accuracies.append(np.mean(accuracy_scores))
logger.info(
' - Mean 10-fold accuracy: {:2.2f} [running mean over all folds: {:2.2f}]'
.format(mean_accuracies[-1] * 100,
np.mean(mean_accuracies) * 100))
logger.info('Accuracy: {:2.2f} +- {:2.2f}'.format(
np.mean(mean_accuracies) * 100,
np.std(mean_accuracies) * 100))
import utilities
# Decide read/write mode based on python version
read_mode, write_mode = ('r', 'w') if six.PY2 else ('rt', 'wt')
# Set the path to your consolidated files
path = '/Users/chrysovalantis/Documents/UCY/EPL451/Project'
os.chdir(path)
# File names
ftrain = 'train_consolidation.txt'
ftest = 'test_consolidation.txt'
flabel = 'trainLabels.csv'
fsubmission = 'submission.csv'
labels = utilities.read_labels(flabel)
# Dimensions for train set
ntrain = 10868
nfeature = 16 ** 2 + 1 + 1 # For two_byte_codes, no_que_marks, label
train = utilities.read_train(ntrain, nfeature, labels, ftrain)
X = train[:, :-1]
y = train[:, -1]
del labels
del train
# Parameters for trees
random_state = 5342
n_jobs = 8
def main(args, logger):
graphs = [ig.read(filename) for filename in args.FILES]
labels = read_labels(args.labels)
# Set the label to be uniform over all graphs in case no labels are
# available. This essentially changes our iteration to degree-based
# checks.
for graph in graphs:
if 'label' not in graph.vs.attributes():
graph.vs['label'] = [0] * len(graph.vs)
logger.info('Read {} graphs and {} labels'.format(len(graphs), len(labels)))
assert len(graphs) == len(labels)
pwl = PersistentWeisfeilerLehman(
use_label_persistence=True,
store_persistence_diagrams=False, # TODO: might need this later on?
)
y = LabelEncoder().fit_transform(labels)
X, num_columns_per_iteration = pwl.transform(graphs, args.num_iterations)
logger.info('Finished persistent Weisfeiler-Lehman transformation')
logger.info('Obtained ({} x {}) feature matrix'.format(X.shape[0], X.shape[1]))
X = to_probability_distribution(X, num_columns_per_iteration)
np.random.seed(42)
cv = StratifiedKFold(n_splits=3, shuffle=True)
mean_accuracies = []
def product_kernel(X, Y, k):
# Index for the current iteration; indicates which column is
# used as the *first* one.
start_index = 0
K = np.zeros((X.shape[0], Y.shape[0]))
for iteration in sorted(num_columns_per_iteration.keys()):
end_index = num_columns_per_iteration[iteration]
P = X[:, start_index:end_index]
Q = Y[:, start_index:end_index]
# TODO: can this be made more efficient?
K_iteration = np.array(
[k(p, q) for p, q in itertools.product(P, Q)]).reshape(
X.shape[0], Y.shape[0]
)
K += K_iteration
start_index += end_index
return K
def jensen_shannon_kernel(X, Y):
return product_kernel(X, Y, jensen_shannon)
def kullback_leibler_kernel(X, Y):
return -product_kernel(X, Y, kullback_leibler)
for i in range(3):
# Contains accuracy scores for each cross validation step; the
# means of this list will be used later on.
accuracy_scores = []
for train_index, test_index in cv.split(X, y):
clf = SVC(
kernel=jensen_shannon_kernel,
)
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
# TODO: need to discuss whether this is 'allowed' or smart
# to do; this assumes normality of the attributes.
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
scaler = MinMaxScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
accuracy_scores.append(accuracy_score(y_test, y_pred))
logger.debug('Best classifier for this fold: {}'.format(clf))
logger.debug('Best parameters for this fold: {}'.format(clf.get_params()))
mean_accuracies.append(np.mean(accuracy_scores))
logger.info(' - Mean 10-fold accuracy: {:2.2f} [running mean over all folds: {:2.2f}]'.format(mean_accuracies[-1] * 100, np.mean(mean_accuracies) * 100))
logger.info('Accuracy: {:2.2f} +- {:2.2f}'.format(np.mean(mean_accuracies) * 100, np.std(mean_accuracies) * 100))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-d', '--dataset', type=str, help='Provide the dataset name (MUTAG or Enzymes)',
choices=['MUTAG', 'ENZYMES'])
parser.add_argument('--crossvalidation', default=False, action='store_true', help='Enable a 10-fold crossvalidation')
parser.add_argument('--gridsearch', default=False, action='store_true', help='Enable grid search')
parser.add_argument('--sinkhorn', default=False, action='store_true', help='Use sinkhorn approximation')
parser.add_argument('--h', type = int, required=False, default=2, help = "(Max) number of WL iterations")
args = parser.parse_args()
dataset = args.dataset
h = args.h
sinkhorn = args.sinkhorn
print(f'Generating results for {dataset}...')
#---------------------------------
# Setup
#---------------------------------
# Start by making directories for intermediate and final files
data_path = os.path.join('../data', dataset)
output_path = os.path.join('output', dataset)
results_path = os.path.join('results', dataset)
for path in [output_path, results_path]:
if not os.path.exists(path):
os.makedirs(path)
#---------------------------------
# Embeddings
#---------------------------------
# Load the data and generate the embeddings
embedding_type = 'continuous' if dataset == 'ENZYMES' else 'discrete'
print(f'Generating {embedding_type} embeddings for {dataset}.')
if dataset == 'ENZYMES':
label_sequences = compute_wl_embeddings_continuous(data_path, h)
else:
label_sequences = compute_wl_embeddings_discrete(data_path, h)
# Save embeddings to output folder
out_name = f'{dataset}_wl_{embedding_type}_embeddings_h{h}.npy'
np.save(os.path.join(output_path, out_name), label_sequences)
print(f'Embeddings for {dataset} computed, saved to {os.path.join(output_path, out_name)}.')
print()
#---------------------------------
# Wasserstein & Kernel computations
#---------------------------------
# Run Wasserstein distance computation
print('Computing the Wasserstein distances...')
wasserstein_distances = compute_wasserstein_distance(label_sequences, h, sinkhorn=sinkhorn,
discrete=dataset=='MUTAG')
# Save Wasserstein distance matrices
for i, D_w in enumerate(wasserstein_distances):
filext = 'wasserstein_distance_matrix'
if sinkhorn:
filext += '_sinkhorn'
filext += f'_it{i}.npy'
np.save(os.path.join(output_path,filext), D_w)
print('Wasserstein distances computation done. Saved to file.')
print()
# Transform to Kernel
# Here the flags come into play
if args.gridsearch:
# Gammas in eps(-gamma*M):
gammas = np.logspace(-4,1,num=6)
# iterate over the iterations too
hs = range(h)
param_grid = [
{'C': np.logspace(-3,3,num=7)}
]
else:
gammas = [0.001]
hs = [h]
kernel_matrices = []
kernel_params = []
for i, current_h in enumerate(hs):
# Generate the full list of kernel matrices from which to select
M = wasserstein_distances[current_h]
for g in gammas:
K = np.exp(-g*M)
kernel_matrices.append(K)
kernel_params.append((current_h, g))
# Check for no hyperparam:
if not args.gridsearch:
assert len(kernel_matrices) == 1
print('Kernel matrices computed.')
print()
#---------------------------------
# Classification
#---------------------------------
# Run hyperparameter search if needed
print(f'Running SVMs, crossvalidation: {args.crossvalidation}, gridsearch: {args.gridsearch}.')
# Load labels
label_file = os.path.join(data_path, 'Labels.txt')
y = np.array(read_labels(label_file))
# Contains accuracy scores for each cross validation step; the
# means of this list will be used later on.
accuracy_scores = []
np.random.seed(42)
cv = StratifiedKFold(n_splits=10, shuffle=True)
# Hyperparam logging
best_C = []
best_h = []
best_gamma = []
for train_index, test_index in cv.split(kernel_matrices[0], y):
K_train = [K[train_index][:, train_index] for K in kernel_matrices]
K_test = [K[test_index][:, train_index] for K in kernel_matrices]
y_train, y_test = y[train_index], y[test_index]
# Gridsearch
if args.gridsearch:
gs, best_params = custom_grid_search_cv(SVC(kernel='precomputed'),
param_grid, K_train, y_train, cv=5)
# Store best params
C_ = best_params['params']['C']
h_, gamma_ = kernel_params[best_params['K_idx']]
y_pred = gs.predict(K_test[best_params['K_idx']])
else:
gs = SVC(C=100, kernel='precomputed').fit(K_train[0], y_train)
y_pred = gs.predict(K_test[0])
h_, gamma_, C_ = h, gammas[0], 100
best_C.append(C_)
best_h.append(h_)
best_gamma.append(gamma_)
accuracy_scores.append(accuracy_score(y_test, y_pred))
if not args.crossvalidation:
break
#---------------------------------
# Printing and logging
#---------------------------------
if args.crossvalidation:
print('Mean 10-fold accuracy: {:2.2f} +- {:2.2f} %'.format(
np.mean(accuracy_scores) * 100,
np.std(accuracy_scores) * 100))
else:
print('Final accuracy: {:2.3f} %'.format(np.mean(accuracy_scores)*100))
# Save to file
if args.crossvalidation or args.gridsearch:
extension = ''
if args.crossvalidation:
extension += '_crossvalidation'
if args.gridsearch:
extension += '_gridsearch'
results_filename = os.path.join(results_path, f'results_{dataset}'+extension+'.csv')
n_splits = 10 if args.crossvalidation else 1
pd.DataFrame(np.array([best_h, best_C, best_gamma, accuracy_scores]).T,
columns=[['h', 'C', 'gamma', 'accuracy']],
index=['fold_id{}'.format(i) for i in range(n_splits)]).to_csv(results_filename)
print(f'Results saved in {results_filename}.')
else:
print('No results saved to file as --crossvalidation or --gridsearch were not selected.')
def main(args, logger):
labels = read_labels(args.labels)
# Load matrices
matrices = np.load(args.MATRICES)
print(
f"Loaded {len(list(matrices.keys()))} matrices, with shape {matrices['0'].shape}"
)
matrix_dict = {}
for h in matrices.keys():
matrix_dict[int(h)] = {'gram': matrices[h]}
y = LabelEncoder().fit_transform(labels)
np.random.seed(42)
mean_accuracies = []
kernel_params = np.array(basename(args.MATRICES)[:-4].split('_'))[1:]
params = ['dataset', 'max_h', 'sigma']
cv_results = []
entry = {}
for i, param in enumerate(params):
entry[param] = kernel_params[i]
for i in range(10):
# Contains accuracy scores for each cross validation step; the
# means of this list will be used later on.
accuracy_scores = []
cv = StratifiedKFold(n_splits=10, shuffle=True, random_state=i)
for n, indices in enumerate(cv.split(matrix_dict[0]['gram'], y)):
entry_fold = copy.copy(entry)
train_index = indices[0]
test_index = indices[1]
y_train = y[train_index]
y_test = y[test_index]
# Override current full matrices
for h, m_dict in matrix_dict.items():
X_train = m_dict['gram'][train_index][:, train_index]
X_test = m_dict['gram'][test_index][:, train_index]
m_dict['X_train'] = X_train
m_dict['X_test'] = X_test
pipeline = Pipeline(
[('clf',
SVC(class_weight='balanced' if args.balanced else None,
random_state=42,
kernel='precomputed'))], )
grid_params = {
'clf__C': [1e-1, 1e0, 1e1],
}
clf, best_params = custom_grid_search_cv(pipeline, grid_params,
matrix_dict, y_train)
X_test = np.zeros(shape=matrix_dict[0]['X_test'].shape)
counter = 0
for h in range(best_params['h'] + 1):
X_test += matrix_dict[h]['X_test']
counter += 1
y_pred = clf.predict(X_test)
acc = accuracy_score(y_test, y_pred)
accuracy_scores.append(acc)
best_params['params']['h'] = best_params['h']
for param, param_val in best_params['params'].items():
entry_fold[param] = param_val
entry[param] = ''
entry_fold['fold'] = n + 1
entry_fold['it'] = i
entry_fold['acc'] = acc * 100
entry_fold['std'] = 0.0
cv_results.append(entry_fold)
print('Best classifier for this fold:{}'.format(
best_params['params']))
mean_accuracies.append(np.mean(accuracy_scores))
print(
' - Mean 10-fold accuracy: {:2.2f} [running mean over all folds: {:2.2f}]'
.format(mean_accuracies[-1] * 100,
np.mean(mean_accuracies) * 100))
entry['fold'] = 'all'
entry['it'] = 'all'
entry['acc'] = np.mean(mean_accuracies) * 100
entry['std'] = np.std(mean_accuracies) * 100
cv_results.append(entry)
logger.info('Accuracy: {:2.2f} +- {:2.2f}'.format(
np.mean(mean_accuracies) * 100,
np.std(mean_accuracies) * 100))
if exists(args.result_file):
with open(args.result_file, 'a') as f:
pd.DataFrame(cv_results).to_csv(f, index=False, header=None)
else:
pd.DataFrame(cv_results).to_csv(args.result_file, index=False)
def main(args, logger):
graphs = [ig.read(filename) for filename in args.FILES]
labels = read_labels(args.labels)
for graph in graphs:
# Make sure that no label information exists as a graph
# attribute already.
assert 'label' not in graph.vs.attributes()
graph.vs['degree'] = graph.vs.degree()
logger.info('Read {} graphs and {} labels'.format(len(graphs), len(labels)))
assert len(graphs) == len(labels)
prop = WeisfeilerLehmanAttributePropagation()
attributes_per_iteration = prop.transform(
graphs,
'degree',
args.num_iterations
)
use_vertex_weights = args.vertex_weights
# Stores *all* persistence diagrams because they will be used to
# represent the data set later on.
persistence_diagrams_per_iteration = collections.defaultdict(list)
for iteration in sorted(attributes_per_iteration.keys()):
# Determine maximum attribute value over *all* graphs and their
# respective filtrations.
max_attribute = max([np.max(attributes_per_iteration[iteration][index]) for index, _ in enumerate(graphs)])
unpaired_value = 2 * max_attribute
if use_vertex_weights:
pdc = PersistenceDiagramCalculator(
unpaired_value=unpaired_value,
vertex_attribute='degree',
)
else:
pdc = PersistenceDiagramCalculator(
unpaired_value=unpaired_value
)
for index, graph in enumerate(graphs):
attributes = attributes_per_iteration[iteration][index]
graph.vs['degree'] = attributes
weighted_graph = assign_filtration_values(
graph,
attributes,
normalize=args.normalize
)
pd, edge_indices_cycles = pdc.fit_transform(graph)
# Store the persistence diagram as a 2D array in order to
# facilitate the subsequent kernel calculations.
persistence_diagrams_per_iteration[iteration].append(
np.array([(c, d) for c, d, _ in pd])
)
np.savetxt(
'/tmp/{:04d}_d0_h{:d}.txt'.format(index, iteration),
np.array([(c, d) for c, d, _ in pd]),
fmt='%.f'
)
def main(args, logger):
graphs = [ig.read(filename) for filename in args.FILES]
labels = read_labels(args.labels)
# Set the label to be uniform over all graphs in case no labels are
# available. This essentially changes our iteration to degree-based
# checks.
for graph in graphs:
if 'label' not in graph.vs.attributes():
graph.vs['label'] = [0] * len(graph.vs)
logger.info('Read {} graphs and {} labels'.format(len(graphs),
len(labels)))
assert len(graphs) == len(labels)
pwl_list = []
for p in [1, 2]:
pwl = PersistentWeisfeilerLehman(
use_cycle_persistence=args.use_cycle_persistence,
use_original_features=args.use_original_features,
metric=args.metric,
use_label_persistence=True,
p=p)
X, num_columns_per_iteration = pwl.transform(graphs,
args.num_iterations)
pwl_list.append({'p': p, 'X': X})
logger.info(
f'Finished persistent Weisfeiler-Lehman transformation for \
p={p}')
logger.info('Obtained ({} x {}) feature matrix'.format(
X.shape[0], X.shape[1]))
if args.use_cycle_persistence:
logger.info('Using cycle persistence')
y = LabelEncoder().fit_transform(labels)
np.random.seed(42)
mean_accuracies = []
params = [
'balanced', 'num_iterations', 'filtration', 'use_cycle_persistence',
'use_original_features', 'metric'
]
cv_results = []
entry = {}
for param in params:
entry[param] = args.__dict__[param]
entry['dataset'] = dirname(args.FILES[0]).split('/')[1]
for i in range(10):
# Contains accuracy scores for each cross validation step; the
# means of this list will be used later on.
accuracy_scores = []
cv = StratifiedKFold(n_splits=10, shuffle=True, random_state=i)
for n, indices in enumerate(cv.split(X, y)):
entry_fold = copy.copy(entry)
train_index = indices[0]
test_index = indices[1]
y_train = y[train_index]
y_test = y[test_index]
# Override current full matrices
for pwl_dict in pwl_list:
scaler = StandardScaler()
X_train = scaler.fit_transform(pwl_dict['X'][train_index])
X_test = scaler.transform(pwl_dict['X'][test_index])
scaler = MinMaxScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
pwl_dict['X_train'] = X_train
pwl_dict['X_test'] = X_test
pipeline = Pipeline(
[('fs', FeatureSelector(num_columns_per_iteration)),
('clf',
RandomForestClassifier(
class_weight='balanced' if args.balanced else None,
random_state=42,
n_jobs=4))], )
grid_params = {
'fs__num_iterations': np.arange(0, args.num_iterations + 1),
'clf__n_estimators': [25, 50, 100],
}
clf, best_params = custom_grid_search_cv(pipeline, grid_params,
pwl_list, y_train)
X_test = pwl_list[best_params['pwl_idx']]['X_test']
y_pred = clf.predict(X_test)
acc = accuracy_score(y_test, y_pred)
accuracy_scores.append(acc)
best_params['params']['p'] = best_params['pwl_idx'] + 1
for param, param_val in best_params['params'].items():
entry_fold[param] = param_val
entry[param] = ''
entry_fold['fold'] = n + 1
entry_fold['it'] = i
entry_fold['acc'] = acc * 100
entry_fold['std'] = 0.0
cv_results.append(entry_fold)
logger.info('Best classifier for this fold:{}'.format(
best_params['params']))
mean_accuracies.append(np.mean(accuracy_scores))
logger.info(
' - Mean 10-fold accuracy: {:2.2f} [running mean over all folds: {:2.2f}]'
.format(mean_accuracies[-1] * 100,
np.mean(mean_accuracies) * 100))
entry['fold'] = 'all'
entry['it'] = 'all'
entry['acc'] = np.mean(mean_accuracies) * 100
entry['std'] = np.std(mean_accuracies) * 100
cv_results.append(entry)
logger.info('Accuracy: {:2.2f} +- {:2.2f}'.format(
np.mean(mean_accuracies) * 100,
np.std(mean_accuracies) * 100))
if exists(args.result_file):
with open(args.result_file, 'a') as f:
pd.DataFrame(cv_results).to_csv(f, index=False, header=None)
else:
pd.DataFrame(cv_results).to_csv(args.result_file, index=False)
