def _split_weighted_sample(self, X, y, sample_weight, is_stratified=False):
if is_stratified:
kfold_model = StratifiedKFold(n_splits=self.n_splits,
shuffle=self.shuffle,
random_state=self.random_state)
else:
kfold_model = KFold(n_splits=self.n_splits,
shuffle=self.shuffle,
random_state=self.random_state)
if sample_weight is None:
return kfold_model.split(X, y)
weights_sum = np.sum(sample_weight)
max_deviations = []
all_splits = []
for i in range(self.n_trials + 1):
splits = [test for (train, test) in list(kfold_model.split(X, y))]
weight_fracs = np.array(
[np.sum(sample_weight[split]) / weights_sum for split in splits])
if np.all(weight_fracs > .95 / self.n_splits):
# Found a good split, return.
return self._get_folds_from_splits(splits, X.shape[0])
# Record all splits in case the stratification by weight yeilds a worse partition
all_splits.append(splits)
max_deviation = np.max(np.abs(weight_fracs - 1 / self.n_splits))
max_deviations.append(max_deviation)
# Reseed random generator and try again
kfold_model.shuffle = True
kfold_model.random_state = None
# If KFold fails after n_trials, we try the next best thing: stratifying by weight groups
warnings.warn(
"The KFold algorithm failed to find a weight-balanced partition after "
+
"{n_trials} trials. Falling back on a weight stratification algorithm."
.format(n_trials=self.n_trials), UserWarning)
if is_stratified:
stratified_weight_splits = [[]] * self.n_splits
for y_unique in np.unique(y.flatten()):
class_inds = np.argwhere(y == y_unique).flatten()
class_splits = self._get_splits_from_weight_stratification(
sample_weight[class_inds])
stratified_weight_splits = [
split + list(class_inds[class_split]) for split, class_split in
zip(stratified_weight_splits, class_splits)
]
else:
stratified_weight_splits = self._get_splits_from_weight_stratification(
sample_weight)
weight_fracs = np.array([
np.sum(sample_weight[split]) / weights_sum
for split in stratified_weight_splits
])
if np.all(weight_fracs > .95 / self.n_splits):
# Found a good split, return.
return self._get_folds_from_splits(stratified_weight_splits,
X.shape[0])
else:
# Did not find a good split
# Record the devaiation for the weight-stratified split to compare with KFold splits
all_splits.append(stratified_weight_splits)
max_deviation = np.max(np.abs(weight_fracs - 1 / self.n_splits))
max_deviations.append(max_deviation)
# Return most weight-balanced partition
min_deviation_index = np.argmin(max_deviations)
return self._get_folds_from_splits(all_splits[min_deviation_index],
X.shape[0])
def main():
global args
args = parser.parse_args()
# Check if CUDA is enabled
args.cuda = not args.no_cuda and torch.cuda.is_available()
unlabeled_datasets = [
"IMDB-BINARY", "IMDB-MULTI", "REDDIT-BINARY", "REDDIT-MULTI-5K",
"COLLAB", "SYNTHETIC", "raw-gitgraph"
]
if args.dataset in unlabeled_datasets:
use_node_labels = False
from graph_kernels import sp_kernel, wl_kernel
else:
use_node_labels = True
from graph_kernels_labeled import sp_kernel, wl_kernel
kernels = [wl_kernel]
n_kernels = len(kernels)
print('Computing graph maps')
Q, subgraphs, labels, shapes = compute_nystrom(args.dataset,
use_node_labels, args.d,
args.community_detection,
kernels)
M = np.zeros((shapes[0], shapes[1], n_kernels))
for idx, k in enumerate(kernels):
M[:, :, idx] = Q[idx]
Q = M
# Binarize labels
le = LabelEncoder()
y = le.fit_transform(labels)
# Build vocabulary
max_n_communities = max([len(x.split(" ")) for x in subgraphs])
x = np.zeros((len(subgraphs), max_n_communities), dtype=np.int32)
for i in range(len(subgraphs)):
communities = subgraphs[i].split()
for j in range(len(communities)):
x[i, j] = int(communities[j])
print(x[0, :])
kf = StratifiedKFold(n_splits=10, random_state=None)
kf.shuffle = True
accs = []
it = 0
print('Starting cross-validation')
for train_index, test_index in kf.split(x, y):
it += 1
best_acc1 = 0
x_train, x_test = x[train_index], x[test_index]
y_train, y_test = y[train_index], y[test_index]
x_train, x_val, y_train, y_val = train_test_split(x_train,
y_train,
test_size=0.1)
train_loader, val_loader, test_loader = create_train_val_test_loaders(
Q, x_train, x_val, x_test, y_train, y_val, y_test, args.batch_size)
print('\tCreate model')
model = CNN(input_size=args.n_filters,
hidden_size=args.hidden_size,
n_classes=np.unique(y).size,
d=args.d,
n_kernels=n_kernels,
max_n_communities=max_n_communities)
print('Optimizer')
optimizer = optim.Adam(model.parameters(), lr=args.lr)
criterion = nn.CrossEntropyLoss()
evaluation = lambda output, target: torch.sum(output.eq(target)
) / target.size()[0]
lr = args.lr
lr_step = (args.lr - args.lr * args.lr_decay) / (
args.epochs * args.schedule[1] - args.epochs * args.schedule[0])
if os.path.isdir(args.checkpoint_dir):
shutil.rmtree(args.checkpoint_dir)
os.makedirs(args.checkpoint_dir)
print('Check cuda')
if args.cuda:
print('\t* Cuda')
model = model.cuda()
criterion = criterion.cuda()
# Epoch for loop
for epoch in range(0, args.epochs):
if epoch > args.epochs * args.schedule[
0] and epoch < args.epochs * args.schedule[1]:
lr -= lr_step
for param_group in optimizer.param_groups:
param_group['lr'] = lr
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, evaluation)
# evaluate on test set
acc1 = validate(val_loader, model, criterion, evaluation)
is_best = acc1 > best_acc1
best_acc1 = max(acc1, best_acc1)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
},
is_best=is_best,
directory=args.checkpoint_dir)
# get the best checkpoint and test it with test set
best_model_file = os.path.join(args.checkpoint_dir, 'model_best.pth')
if not os.path.isdir(args.checkpoint_dir):
os.makedirs(args.checkpoint_dir)
if os.path.isfile(best_model_file):
print("=> loading best model '{}'".format(best_model_file))
checkpoint = torch.load(best_model_file)
args.start_epoch = checkpoint['epoch']
best_acc1 = checkpoint['best_acc1']
model.load_state_dict(checkpoint['state_dict'])
if args.cuda:
model.cuda()
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded best model '{}' (epoch {})".format(
best_model_file, checkpoint['epoch']))
else:
print("=> no best model found at '{}'".format(best_model_file))
# For testing
acc = validate(test_loader, model, criterion, evaluation)
print("Accuracy at iteration " + str(it) + ": " + str(acc))
accs.append(acc)
print("Average accuracy: ", np.mean(accs))
print("std: ", np.std(accs))
