def build_tree(self):
if self.name not in IMAGE_DS:
raise ValueError('Unknown image dataset: {}'.format(self.name))
ds_class = IMAGE_DS[self.name]
common_params = dict(root=self.data_path, download=True)
if self.name in ['svhn', 'dtd', 'aircraft']:
self.trainset = ds_class(split='train', **common_params)
self.testset = ds_class(split='test', **common_params)
else:
self.trainset = ds_class(train=True, **common_params)
self.testset = ds_class(train=False, **common_params)
self.train_samples, self.train_targets = self._format_ds(self.trainset)
self.test_samples, self.test_targets = self._format_ds(self.testset)
self.height, self.width = self.train_samples.size()[-2:]
taxonomy = TAXONOMY[self.name]
concepts = self._build_tree(taxonomy)
root_concept = ComposedConcept(concepts, id=self.name)
self.tree.add_node(root_concept)
for sub_concept in concepts:
self.tree.add_edge(root_concept, sub_concept)
del self.trainset, self.testset, \
self.train_samples, self.test_samples, \
self.train_targets, self.test_targets
return root_concept
def build_tree(self):
train = MNIST(download=True,
root=os.environ['HOME'] + '/data',
train=True)
test = MNIST(download=True,
root=os.environ['HOME'] + '/data',
train=False)
assert train.class_to_idx == test.class_to_idx
for digit, idx in train.class_to_idx.items():
train_mask = train.targets == idx
train_samples = train.data[train_mask]
test_mask = test.targets == idx
test_samples = test.data[test_mask]
concept = DigitConcept(id=digit,
data=(train_samples, test_samples))
self.tree.add_node(concept)
self.all_nodes.add(concept)
self.leaf_nodes.add(concept)
root_concept = ComposedConcept(self.leaf_concepts, id='all_digits')
for concept in self.leaf_concepts:
self.tree.add_edge(root_concept, concept)
return root_concept
def _build_tree(self, current_level_concepts):
new_concepts = []
if isinstance(current_level_concepts, dict):
# Not yet at the lowest level
for name, lower_concepts in current_level_concepts.items():
concepts = self._build_tree(lower_concepts)
concept_name = '{} {}'.format(self.name, name)
new_concept = ComposedConcept(concepts=concepts,
id=concept_name)
self.tree.add_node(new_concept)
self.all_nodes.add(new_concept)
for c in concepts:
self.tree.add_edge(new_concept, c)
new_concepts.append(new_concept)
elif isinstance(current_level_concepts, list):
# Adding lowest level concepts
for c in current_level_concepts:
samples = self._get_samples(c)
concept_name = '{} {}'.format(self.name, c)
concept = ImageConcept(id=concept_name, samples=samples,
split_rnd=self.split_rnd,
val_size=self.val_size,
val_ratio=self.val_ratio)
self.tree.add_node(concept)
self.all_nodes.add(concept)
self.leaf_nodes.add(concept)
new_concepts.append(concept)
else:
raise NotImplementedError()
return new_concepts
def _generate_samples_from_descr(categories, attributes, n_samples_per_class,
augment):
use_cat_id, attributes = attributes
assert use_cat_id and not attributes, \
"usage of attributes isn't supporte in v1."
samples = []
labels = []
for i, cat_concepts in enumerate(categories):
mixture = ComposedConcept(cat_concepts, id=None)
cat_samples = []
cat_labels = []
for s_id, n in enumerate(n_samples_per_class):
split_samples, split_attrs = mixture._get_samples(n,
attributes,
split_id=s_id)
if s_id in augment:
split_samples = augment_samples(split_samples)
split_labels = torch.Tensor().long()
cat_id = torch.tensor([i]).expand(split_samples.shape[0], 1)
split_labels = torch.cat([split_labels, cat_id], dim=1)
cat_samples.append(split_samples)
cat_labels.append(split_labels)
samples.append(cat_samples)
labels.append(cat_labels)
if torch.is_tensor(samples[0][0]):
cat_func = torch.cat
else:
cat_func = np.concatenate
samples = (cat_func(split) for split in zip(*samples))
labels = (torch.cat(split) for split in zip(*labels))
return samples, labels
def _build_tree(self, n_levels, n_children, n_dims, low, high, scale_fact, cov_scale, mean, parent_name):
if isinstance(low, Number):
low = torch.ones(n_dims) * low
if isinstance(high, Number):
high = torch.ones(n_dims) * high
cluster_means = uniform.Uniform(low + mean, high + mean).sample((n_children[0],))
pref = '\t' * (self.n_levels - n_levels)
concepts = []
for i, cluster_mean in enumerate(cluster_means):
logger.debug(pref + 'New cluster centered on {}'.format(cluster_mean))
concept_name = '{}{}'.format(parent_name, i)
if n_levels > 1:
lower_concepts = self._build_tree(n_levels - 1, n_children[1:], n_dims, low=low * scale_fact,
high=high * scale_fact, scale_fact=scale_fact, cov_scale=cov_scale,
mean=cluster_mean, parent_name=concept_name)
concept = ComposedConcept(lower_concepts, cluster_mean=cluster_mean, id=concept_name)
self.tree.add_node(concept)
for c in lower_concepts:
self.tree.add_edge(concept, c)
self.all_nodes.add(concept)
else:
concept = AtomicConcept(cluster_mean, torch.eye(n_dims) * cov_scale, concept_name)
self.tree.add_node(concept)
self.all_nodes.add(concept)
self.leaf_nodes.add(concept)
concepts.append(concept)
return concepts
def build_tree(self):
# if self.name == 'cifar10':
# ds_class = CIFAR10
# elif self.name == 'cifar100':
# ds_class = CIFAR100
# elif self.name == 'mnist':
# ds_class = MNIST
# else:
# raise ValueError('Unknown image dataset: {}'.format(self.name))
#
# self.trainset = ds_class(root=self.data_path, train=True, download=True)
# self.testset = ds_class(root=self.data_path, train=False, download=True)
# self.train_samples, self.train_targets = self._format_ds(self.trainset)
# self.test_samples, self.test_targets = self._format_ds(self.testset)
# self.height, self.width = self.train_samples.size()[2:]
taxonomy = TAXONOMY[self.name]
tasks = self._build_tree(taxonomy)
vdd_task = ComposedConcept(tasks, id=self.name)
self.tree.add_node(vdd_task)
for task in tasks:
self.tree.add_edge(vdd_task, task)
# print(trainset)dd
# del self.trainset, self.testset, \
# self.train_samples, self.test_samples, \
# self.train_targets, self.test_targets
return vdd_task
def build_tree(self):
concepts = self._build_tree(self.n_levels, self.n_children, self.intrinsic_dims, self.low_bound, self.high_bound, self.scale_fact,
self.cov_scale, self.mean, self.name)
concept = ComposedConcept(concepts, cluster_mean=self.mean, id=self.name)
self.tree.add_node(concept)
for c in concepts:
self.tree.add_edge(concept, c)
return concept
def _generate_samples_from_descr(categories, attributes, n_samples_per_class,
augment):
use_cat_id, attributes = attributes
assert use_cat_id or attributes, 'Each task should at least use the ' \
'category id or an attribute as labels'
if not use_cat_id:
all_concepts = np.array(
[concept for cat in categories for concept in cat])
all_attrs = np.array([c.attrs for c in all_concepts])
selected_attr = all_attrs[:, attributes[0]]
categories = [tuple(all_concepts[selected_attr == val])
for val in np.unique(selected_attr)]
if use_cat_id or isinstance(all_concepts[0], AtomicConcept):
samples = []
labels = []
for i, cat_concepts in enumerate(categories):
mixture = ComposedConcept(cat_concepts, id=None)
cat_samples = []
cat_labels = []
for s_id, n in enumerate(n_samples_per_class):
split_samples, split_attrs = mixture._get_samples(n, attributes,
split_id=s_id)
if s_id in augment:
split_samples = augment_samples(split_samples)
split_labels = torch.Tensor().long()
if use_cat_id:
cat_id = torch.tensor([i]).expand(split_samples.shape[0], 1)
split_labels = torch.cat([split_labels, cat_id], dim=1)
if attributes:
raise NotImplementedError('Attrs aren\'t supported '
'anymore')
split_labels = torch.cat([split_labels, split_attrs], dim=1)
cat_samples.append(split_samples)
cat_labels.append(split_labels)
samples.append(cat_samples)
labels.append(cat_labels)
if torch.is_tensor(samples[0][0]):
cat_func = torch.cat
else:
cat_func = np.concatenate
samples = (cat_func(split) for split in zip(*samples))
labels = (torch.cat(split) for split in zip(*labels))
else:
# Grouping the concepts by attribute value to create the categories
all_concepts = np.array(
[concept for cat in categories for concept in cat])
samples, labels = get_samples_using_attrs(all_concepts, attributes,
n_samples_per_class)
return samples, labels
def build_tree(self):
child_concepts = [child.root_node for child in self.children]
root_concept = ComposedConcept(child_concepts, id=self.name)
self.tree.add_node(root_concept)
for child in self.children:
self.tree = nx.compose(self.tree, child.tree)
self.tree.add_edge(root_concept, child.root_node)
self.leaf_nodes = self.leaf_nodes.union(child.leaf_nodes)
self.all_nodes = self.all_nodes.union(child.all_nodes)
return root_concept
def _build_tree(self, current_level_concepts):
all_task_concepts = []
for task, classes in current_level_concepts.items():
task_concepts = self.add_task(task)
new_concept = ComposedConcept(concepts=task_concepts, id=task)
self.tree.add_node(new_concept)
self.all_nodes.add(new_concept)
for c in task_concepts:
self.tree.add_edge(new_concept, c)
all_task_concepts.append(new_concept)
return all_task_concepts
def _generate_samples_from_descr(categories, attributes, n_samples_per_class,
augment, rnd):
use_cat_id, attributes = attributes
assert use_cat_id and not attributes, \
"usage of attributes isn't supporte in v1."
samples = []
labels = []
for i, cat_concepts in enumerate(categories):
mixture = ComposedConcept(cat_concepts, id=None)
cat_samples = []
cat_labels = []
for s_id, n in enumerate(n_samples_per_class):
if isinstance(n, list):
n_samples, n_labeled = n[0], n[1]
else:
n_samples, n_labeled = n, n
split_samples, split_attrs = mixture._get_samples(n_samples,
attributes,
split_id=s_id,
rng=rnd)
if s_id in augment:
split_samples = augment_samples(split_samples)
split_labels = torch.Tensor().long()
cat_id = torch.tensor([i]).expand(split_samples.shape[0], 1)
split_labels = torch.cat([split_labels, cat_id], dim=1)
#Create unlabeled samples
n_unlabeled = len(split_samples) - n_labeled
if n_unlabeled > 0:
idx_unlabeled = rnd.choice(list(range(len(split_samples))),
n_unlabeled,
replace=False)
split_labels[
idx_unlabeled] = -1 #mark unlabeled samples with -1
cat_samples.append(split_samples)
cat_labels.append(split_labels)
samples.append(cat_samples)
labels.append(cat_labels)
if torch.is_tensor(samples[0][0]):
cat_func = torch.cat
else:
cat_func = np.concatenate
samples = (cat_func(split) for split in zip(*samples))
labels = (torch.cat(split) for split in zip(*labels))
return samples, labels
def build_tree(self):
# if self.name == 'cifar10':
# ds_class = CIFAR10
# elif self.name == 'cifar100':
# ds_class = CIFAR100
# elif self.name == 'mnist':
# ds_class = MNIST
# elif self.name == 'fashion-mnist':
# ds_class = FashionMNIST
# elif self.name == 'svhn':
# ds_class = SVHN
# else:
# raise ValueError('Unknown image dataset: {}'.format(self.name))
# common_params = dict(root=self.data_path, download=True)
# if self.name == 'svhn':
# self.trainset = ds_class(split='train', **common_params)
# self.testset = ds_class(split='test', **common_params)
# else:
# self.trainset = ds_class(train=True, **common_params)
# self.testset = ds_class(train=False, **common_params)
# self.train_samples, self.train_targets = self._format_ds(self.trainset)
# self.test_samples, self.test_targets = self._format_ds(self.testset)
# self.height, self.width = self.train_samples.size()[2:]
#
# taxonomy = TAXONOMY[self.name]
#
# concepts = self._build_tree(taxonomy)
child_concepts = [child.root_node for child in self.children]
root_concept = ComposedConcept(child_concepts, id=self.name)
self.tree.add_node(root_concept)
for child in self.children:
self.tree = nx.compose(self.tree, child.tree)
self.tree.add_edge(root_concept, child.root_node)
self.leaf_nodes = self.leaf_nodes.union(child.leaf_nodes)
self.all_nodes = self.all_nodes.union(child.all_nodes)
# print(trainset)dd
# del self.trainset, self.testset, \
# self.train_samples, self.test_samples, \
# self.train_targets, self.test_targets
return root_concept