def __init__(self,
adj,
x,
labels,
idx_train=None,
idx_val=None,
seed=None,
name=None,
device='CPU:0',
surrogate=None,
surrogate_args={},
**kwargs):
super().__init__(adj,
x=x,
labels=labels,
seed=seed,
name=name,
device=device,
**kwargs)
if surrogate is None:
surrogate = train_a_surrogate(self, 'GCN', idx_train, idx_val,
**kwargs)
elif not isinstance(surrogate, SemiSupervisedModel):
raise RuntimeError(
"surrogate model should be instance of `graphgallery.nn.SemiSupervisedModel`."
)
with tf.device(self.device):
self.surrogate = surrogate
self.loss_fn = sparse_categorical_crossentropy
self.tf_x = astensor(self.x)
def __init__(self,
adj,
x,
labels,
idx_train=None,
seed=None,
name=None,
device='CPU:0',
surrogate=None,
surrogate_args={},
**kwargs):
super().__init__(adj,
x=x,
labels=labels,
seed=seed,
name=name,
device=device,
**kwargs)
if surrogate is None:
surrogate = train_a_surrogate(self, 'DenseGCN', idx_train,
**surrogate_args)
elif not isinstance(surrogate, DenseGCN):
raise RuntimeError(
"surrogate model should be the instance of `graphgallery.nn.DenseGCN`."
)
self.allow_feature_attack = True
with tf.device(self.device):
self.surrogate = surrogate
self.loss_fn = sparse_categorical_crossentropy
def __init__(self,
adj,
x,
labels,
idx_train,
surrogate=None,
surrogate_args={},
seed=None,
name=None,
device='CPU:0',
**kwargs):
super().__init__(adj=adj,
x=x,
labels=labels,
seed=seed,
name=name,
device=device,
**kwargs)
adj, x = self.adj, self.x
if surrogate is None:
surrogate = train_a_surrogate(self, 'DenseGCN', idx_train,
**surrogate_args)
elif not isinstance(surrogate, DenseGCN):
raise RuntimeError(
"surrogate model should be the instance of `graphgallery.nn.DenseGCN`."
)
idx_attack = asintarr(idx_train)
with tf.device(self.device):
self.idx_attack = astensor(idx_attack)
self.labels_attack = astensor(labels[idx_attack])
self.tf_adj = astensor(self.adj.A)
self.tf_x = astensor(x)
self.complementary = tf.ones_like(self.tf_adj) - tf.eye(
self.n_nodes) - 2. * self.tf_adj
self.loss_fn = sparse_categorical_crossentropy
self.adj_changes = tf.Variable(
tf.zeros(adj.shape, dtype=self.floatx))
self.surrogate = surrogate
# used for CW_loss=True
self.label_matrix = tf.gather(tf.eye(self.n_classes),
self.labels_attack)
def __init__(self,
adj,
x,
labels,
idx_train=None,
surrogate=None,
surrogate_args={},
seed=None,
name=None,
device='CPU:0',
**kwargs):
super().__init__(adj=adj,
x=x,
labels=labels,
seed=seed,
name=name,
device=device,
**kwargs)
adj, x = self.adj, self.x
# nettack can conduct feature attack
self.allow_feature_attack = True
if surrogate is None:
# By default, Nettack uses a linear GCN model.
if not 'activations' in surrogate_args:
surrogate_args['activations'] = None
surrogate = train_a_surrogate(self, 'GCN', idx_train,
**surrogate_args)
elif not isinstance(surrogate, GCN):
raise RuntimeError(
'surrogate model should be the instance of `graphgallery.nn.models.GCN`.'
)
self.sparse_x = sp.csr_matrix(x)
surrogate_weights = surrogate.get_weights()
assert len(
surrogate_weights
) == 2, 'surrogate weights are the weights of two layer GCN without bias, i.e., W1 and W2.'
# GCN weight matrices
W1, W2 = surrogate_weights
self.W = W1 @ W2
self.cooc_matrix = self.sparse_x.T @ self.sparse_x
def __init__(self,
adj,
x,
labels,
idx_train=None,
hops=2,
seed=None,
name=None,
device='CPU:0',
surrogate=None,
surrogate_args={},
**kwargs):
super().__init__(adj,
x=x,
labels=labels,
seed=seed,
name=name,
device=device,
**kwargs)
if surrogate is None:
surrogate = train_a_surrogate(self, 'SGC', idx_train,
**surrogate_args)
elif not isinstance(surrogate, SGC):
raise RuntimeError(
"surrogate model should be the instance of `graphgallery.nn.SGC`."
)
self.hops = hops
# nodes with the same class labels
self.similar_nodes = [
np.where(labels == c)[0] for c in range(self.n_classes)
]
with tf.device(self.device):
W, b = surrogate.weights
X = astensor(x)
self.b = b
self.XW = X @ W
self.surrogate = surrogate
self.SGC = SGConvolution(hops)
self.loss_fn = sparse_categorical_crossentropy
def __init__(self,
adj,
x,
labels,
idx_train,
idx_unlabeled=None,
surrogate=None,
surrogate_args={},
seed=None,
name=None,
device='CPU:0',
**kwargs):
super().__init__(adj=adj,
x=x,
labels=labels,
seed=seed,
name=name,
device=device,
**kwargs)
adj, x = self.adj, self.x
if surrogate is None:
surrogate = train_a_surrogate(self, 'DenseGCN', idx_train,
**surrogate_args)
elif not isinstance(surrogate, DenseGCN):
raise RuntimeError(
"surrogate model should be the instance of `graphgallery.nn.DenseGCN`."
)
# poisoning attack in DeepRobust
if idx_unlabeled is None:
idx_attack = idx_train
labels_attack = labels[idx_train]
else: # Evasion attack in original paper
idx_unlabeled = asintarr(idx_unlabeled)
self_training_labels = self.estimate_self_training_labels(
surrogate, idx_unlabeled)
idx_attack = np.hstack([idx_train, idx_unlabeled])
labels_attack = np.hstack(
[labels[idx_train], self_training_labels])
with tf.device(self.device):
self.idx_attack = astensor(idx_attack)
self.labels_attack = astensor(labels_attack)
self.tf_adj = astensor(self.adj.A)
self.tf_x = astensor(x)
self.complementary = tf.ones_like(self.tf_adj) - tf.eye(
self.n_nodes) - 2. * self.tf_adj
self.loss_fn = SparseCategoricalCrossentropy()
self.adj_changes = tf.Variable(
tf.zeros(adj.shape, dtype=self.floatx))
self.surrogate = surrogate
# used for `CW_loss=True`
self.label_matrix = tf.gather(tf.eye(self.n_classes),
self.labels_attack)
self.range_idx = tf.range(idx_attack.size, dtype=self.intx)
self.indices_real = tf.stack([self.range_idx, self.labels_attack],
axis=1)
