def rank(self, graph, personalization, sensitive, *args, **kwargs):
original_ranks = self.ranker(graph,
personalization,
*args,
sensitive=sensitive,
**kwargs)
base_ranks = original_ranks if self.ranker == self.base_ranker else self.base_ranker(
graph, personalization, *args, **kwargs)
training_objective = pg.AM()\
.add(pg.L2(base_ranks), weight=-1.)\
.add(pg.pRule(tf.cast(sensitive.np, tf.float32)), weight=10., max_val=0.8)
with pg.Backend("tensorflow"):
ranks_var = tf.Variable(pg.to_array(original_ranks.np))
optimizer = tf.keras.optimizers.Adam(learning_rate=0.1)
best_loss = float('inf')
best_ranks = None
for epoch in range(2000):
with tf.GradientTape() as tape:
ranks = pg.to_signal(original_ranks, ranks_var)
loss = -training_objective(
ranks) #+ 1.E-5*tf.reduce_sum(ranks_var*ranks_var)
grads = tape.gradient(loss, [ranks_var])
optimizer.apply_gradients(zip(grads, [ranks_var]))
validation_loss = loss
if validation_loss < best_loss:
patience = 100
best_ranks = ranks
best_loss = validation_loss
patience -= 1
if patience == 0:
break
return best_ranks
def _transform(self, ranks: pg.GraphSignal, **kwargs):
if ranks.graph not in self.known_ranks or not self.assume_immutability:
with pg.Backend("numpy"):
A = pg.preprocessor(normalization=self.normalization)(
ranks.graph)
D = pg.degrees(
pg.preprocessor(normalization="none")(ranks.graph))
s = pg.sum(
D)**0.5 / 2 if self.sparsity is None else self.sparsity
D = (D / pg.max(D))**self.beta
S = scipy.sparse.random(
self.dims,
A.shape[0],
density=1. / s,
data_rvs=lambda l: np.random.choice([-1, 1], size=l),
format="csc")
S = S @ scipy.sparse.spdiags(D, 0, *A.shape)
self.embeddigns[ranks.graph] = pg.scipy_sparse_to_backend(S.T)
self.known_ranks[ranks.graph] = [
] # we know that the first term is zero and avoid direct embedding comparison
for _ in range(len(self.weights)):
S = S @ A
self.known_ranks[ranks.graph].append(
pg.scipy_sparse_to_backend(S))
ret = 0
on = pg.conv(ranks.np, self.embeddigns[ranks.graph])
for weight, S in zip(self.weights, self.known_ranks[ranks.graph]):
uv = pg.conv(on, S)
ret = ret + weight * uv
return pg.to_signal(ranks, ret)
def test_appnp_tf():
from tensorflow.keras.layers import Dropout, Dense
from tensorflow.keras.regularizers import L2
class APPNP(tf.keras.Sequential):
def __init__(self, num_inputs, num_outputs, hidden=64):
super().__init__([
Dropout(0.5, input_shape=(num_inputs,)),
Dense(hidden, activation="relu", kernel_regularizer=L2(1.E-5)),
Dropout(0.5),
Dense(num_outputs, activation="relu")])
self.ranker = pg.ParameterTuner(
lambda par: pg.GenericGraphFilter([par[0] ** i for i in range(int(10))],
error_type="iters", max_iters=int(10)),
max_vals=[0.95], min_vals=[0.5], verbose=False,
measure=pg.Mabs, deviation_tol=0.1, tuning_backend="numpy")
def call(self, features, graph, training=False):
predict = super().call(features, training=training)
propagate = self.ranker.propagate(graph, predict, graph_dropout=0.5 if training else 0)
return tf.nn.softmax(propagate, axis=1)
graph, features, labels = pg.load_feature_dataset('synthfeats')
training, test = pg.split(list(range(len(graph))), 0.8)
training, validation = pg.split(training, 1 - 0.2 / 0.8)
model = APPNP(features.shape[1], labels.shape[1])
with pg.Backend('tensorflow'): # pygrank computations in tensorflow backend
graph = pg.preprocessor(renormalize=True, cors=True)(graph) # cors = use in many backends
pg.gnn_train(model, features, graph, labels, training, validation,
optimizer=tf.optimizers.Adam(learning_rate=0.01), verbose=True, epochs=50)
assert float(pg.gnn_accuracy(labels, model(features, graph), test)) == 1. # dataset is super-easy to predict
def test_backend_with():
for backend_name in [
"matvec", "pytorch", "tensorflow", "numpy", "torch_sparse"
]:
with pg.Backend(backend_name) as backend:
assert pg.backend_name() == backend_name
assert backend.backend_name() == backend_name
assert pg.backend_name() == "numpy"
def test_signal_np_auto_conversion():
import tensorflow as tf
import numpy as np
graph = nx.DiGraph([(1, 2), (2, 3)])
signal = pg.to_signal(graph, tf.convert_to_tensor([1., 2., 3.]))
assert isinstance(signal.np, np.ndarray)
with pg.Backend("tensorflow"):
assert pg.backend_name() == "tensorflow"
assert not isinstance(signal.np, np.ndarray)
assert pg.backend_name() == "numpy"
assert isinstance(signal.np, np.ndarray)
def train_model(self, graph, personalization, sensitive, *args, **kwargs):
original_ranks = self.ranker(graph, personalization, *args, **kwargs)
#pretrained_ranks = None if self.pretrainer is None else self.pretrainer(graph, personalization, *args, sensitive=sensitive, **kwargs)
features = tf.concat([
tf.reshape(personalization.np, (-1, 1)),
tf.reshape(original_ranks.np, (-1, 1)),
tf.reshape(sensitive.np, (-1, 1))
],
axis=1)
training_objective = pg.AM()\
.add(pg.L2(tf.cast(original_ranks.np, tf.float32)), weight=1.)\
.add(pg.pRule(tf.cast(sensitive.np, tf.float32)), max_val=0.8, weight=-10.)
model = self.model()
with pg.Backend("tensorflow"):
best_loss = float('inf')
best_ranks = None
optimizer = tf.keras.optimizers.Adam(learning_rate=0.01)
#degrade = 1
for epoch in range(5000):
with tf.GradientTape() as tape:
personalization = pg.to_signal(personalization,
model(features))
#personalization.np = tf.nn.relu(personalization.np*2-1)
ranks = self.ranker(graph, personalization, *args,
**kwargs)
loss = training_objective(ranks)
for var in model.trainable_variables:
loss = loss + 1.E-5 * tf.reduce_sum(var * var)
#loss = loss * degrade
grads = tape.gradient(loss, model.trainable_variables)
#degrade *= 0.9
optimizer.apply_gradients(zip(grads,
model.trainable_variables))
validation_loss = training_objective(ranks)
if validation_loss < best_loss:
patience = 10
best_ranks = ranks
best_loss = validation_loss
print("epoch", epoch, "loss", validation_loss, "prule",
pg.pRule(tf.cast(sensitive.np, tf.float32))(ranks))
patience -= 1
if patience == 0:
break
return best_ranks
def test_appnp_torch():
graph, features, labels = pg.load_feature_dataset('synthfeats')
training, test = pg.split(list(range(len(graph))), 0.8)
training, validation = pg.split(training, 1 - 0.2 / 0.8)
class AutotuneAPPNP(torch.nn.Module):
def __init__(self, num_inputs, num_outputs, hidden=64):
super().__init__()
self.layer1 = torch.nn.Linear(num_inputs, hidden)
self.layer2 = torch.nn.Linear(hidden, num_outputs)
self.activation = torch.nn.ReLU()
self.dropout = torch.nn.Dropout(0.5)
self.num_outputs = num_outputs
self.ranker = pg.ParameterTuner(
lambda par: pg.GenericGraphFilter([par[0] ** i for i in range(int(10))],
error_type="iters", max_iters=int(10)),
max_vals=[0.95], min_vals=[0.5], verbose=False,
measure=pg.Mabs, deviation_tol=0.1, tuning_backend="numpy")
def forward(self, features, graph, training=False):
predict = self.dropout(torch.FloatTensor(features))
predict = self.dropout(self.activation(self.layer1(predict)))
predict = self.activation(self.layer2(predict))
predict = self.ranker.propagate(graph, predict, graph_dropout=0.5 if training else 0)
ret = torch.nn.functional.softmax(predict, dim=1)
self.loss = 0
for param in self.layer1.parameters():
self.loss = self.loss + 1E-5*torch.norm(param)
return ret
def init_weights(m):
if isinstance(m, torch.nn.Linear):
torch.nn.init.xavier_uniform_(m.weight)
m.bias.data.fill_(0.01)
model = AutotuneAPPNP(features.shape[1], labels.shape[1])
graph = pg.preprocessor(renormalize=True, cors=True)(graph)
model.apply(init_weights)
with pg.Backend('pytorch'):
pg.gnn_train(model, features, graph, labels, training, validation, epochs=50)
assume_immutability=True,
use_quotient=False,
error_type="iters",
max_iters=10) # 10 iterations
def call(self,
features,
graph,
training=False): # can call with tensor graph
predict = super().call(features, training=training)
propagate = self.ranker.propagate(graph,
predict,
graph_dropout=0.5 * training)
return tf.nn.softmax(propagate, axis=1)
graph, features, labels = pg.load_feature_dataset('citeseer')
training, test = pg.split(list(range(len(graph))), 0.8, seed=5) # seeded split
training, validation = pg.split(training, 1 - 0.2 / 0.8)
model = APPNP(features.shape[1], labels.shape[1])
with pg.Backend('tensorflow'): # pygrank with tensorflow backend
pg.gnn_train(model,
features,
graph,
labels,
training,
validation,
optimizer=tf.optimizers.Adam(learning_rate=0.01),
verbose=True)
print("Accuracy", pg.gnn_accuracy(labels, model(features, graph), test))
max_vals=[1],
min_vals=[0.5],
verbose=False,
measure=pg.Mabs,
deviation_tol=0.01,
tuning_backend="numpy")
def call(self, features, graph, training=False):
predict = super().call(features, training=training)
propagate = self.ranker.propagate(graph,
predict,
graph_dropout=0.5 * training)
return tf.nn.softmax(propagate, axis=1)
graph, features, labels = pg.load_feature_dataset('citeseer')
training, test = pg.split(list(range(len(graph))), 0.8, seed=5)
training, validation = pg.split(training, 1 - 0.2 / 0.8)
model = APPNP(features.shape[1], labels.shape[1])
with pg.Backend('tensorflow'): # pygrank computations in tensorflow backend
pg.gnn_train(model,
features,
graph,
labels,
training,
validation,
optimizer=tf.optimizers.Adam(learning_rate=0.01),
verbose=True,
test=test)
print("Accuracy", pg.gnn_accuracy(labels, model(features, graph), test))