def generate(scores, tg_sum, num_graphs):
graphs = []
for i in range(num_graphs):
sparse_mat = utils.graph_from_scores(scores, tg_sum)
g = nx.from_numpy_array(sparse_mat, create_using=nx.Graph())
g.name = 'blah' # filler - renamed later in graph_models.py
graphs.append(g)
return graphs
def train(self,
A_orig,
val_ones,
val_zeros,
max_iters=50000,
stopping=None,
eval_transitions=15e6,
transitions_per_iter=150000,
max_patience=5,
eval_every=500,
plot_every=-1,
save_directory="../snapshots",
model_name=None,
continue_training=False):
"""
Parameters
----------
A_orig: sparse matrix, shape: (N,N)
Adjacency matrix of the original graph to be trained on.
val_ones: np.array, shape (n_val, 2)
The indices of the hold-out set of validation edges
val_zeros: np.array, shape (n_val, 2)
The indices of the hold-out set of validation non-edges
max_iters: int, default: 50,000
The maximum number of training iterations if early stopping does not apply.
stopping: float in (0,1] or None, default: None
The early stopping strategy. None means VAL criterion will be used (i.e. evaluation on the
validation set and stopping after there has not been an improvement for *max_patience* steps.
Set to a value in the interval (0,1] to stop when the edge overlap exceeds this threshold.
eval_transitions: int, default: 15e6
The number of transitions that will be used for evaluating the validation performance, e.g.
if the random walk length is 5, each random walk contains 4 transitions.
transitions_per_iter: int, default: 150000
The number of transitions that will be generated in one batch. Higher means faster
generation, but more RAM usage.
max_patience: int, default: 5
Maximum evaluation steps without improvement of the validation accuracy to tolerate. Only
applies to the VAL criterion.
eval_every: int, default: 500
Evaluate the model every X iterations.
plot_every: int, default: -1
Plot the generator/discriminator losses every X iterations. Set to None or a negative number
to disable plotting.
save_directory: str, default: "../snapshots"
The directory to save model snapshots to.
model_name: str, default: None
Name of the model (will be used for saving the snapshots).
continue_training: bool, default: False
Whether to start training without initializing the weights first. If False, weights will be
initialized.
Returns
-------
log_dict: dict
A dictionary with the following values observed during training:
* The generator and discriminator losses
* The validation performances (ROC and AP)
* The edge overlap values between the generated and original graph
* The sampled graphs for all evaluation steps.
"""
if stopping == None: # use VAL criterion
best_performance = 0.0
patience = max_patience
print("**** Using VAL criterion for early stopping ****")
else: # use EO criterion
assert "float" in str(
type(stopping)) and stopping > 0 and stopping <= 1
print("**** Using EO criterion of {} for early stopping".format(
stopping))
if not os.path.isdir(save_directory):
os.makedirs(save_directory)
if model_name is None:
# Find the file corresponding to the lowest vacant model number to store the snapshots into.
model_number = 0
while os.path.exists("{}/model_best_{}.ckpt".format(
save_directory, model_number)):
model_number += 1
save_file = "{}/model_best_{}.ckpt".format(save_directory,
model_number)
open(save_file, 'a').close() # touch file
else:
save_file = "{}/{}_best.ckpt".format(save_directory, model_name)
print("**** Saving snapshots into {} ****".format(save_file))
if not continue_training:
print("**** Initializing... ****")
self.session.run(self.init_op)
print("**** Done. ****")
else:
print(
"**** Continuing training without initializing weights. ****")
# Validation labels
actual_labels_val = np.append(np.ones(len(val_ones)),
np.zeros(len(val_zeros)))
# Some lists to store data into.
gen_losses = []
disc_losses = []
graphs = []
val_performances = []
eo = []
temperature = self.params['temp_start']
starting_time = time.time()
saver = tf.train.Saver()
transitions_per_walk = self.rw_len - 1
# Sample lots of random walks, used for evaluation of model.
sample_many_count = int(
np.round(transitions_per_iter / transitions_per_walk))
sample_many = self.generate_discrete(sample_many_count, reuse=True)
n_eval_walks = eval_transitions / transitions_per_walk
n_eval_iters = int(np.round(n_eval_walks / sample_many_count))
print("**** Starting training. ****")
for _it in range(max_iters):
if _it > 0 and _it % (2500) == 0:
t = time.time() - starting_time
print(
'{:<7}/{:<8} training iterations, took {} seconds so far...'
.format(_it, max_iters, int(t)))
# Generator training iteration
gen_loss, _ = self.session.run([self.gen_cost, self.gen_train_op],
feed_dict={self.tau: temperature})
_disc_l = []
# Multiple discriminator training iterations.
for _ in range(self.params['disc_iters']):
disc_loss, _ = self.session.run(
[self.disc_cost, self.disc_train_op],
feed_dict={self.tau: temperature})
_disc_l.append(disc_loss)
gen_losses.append(gen_loss)
disc_losses.append(np.mean(_disc_l))
# Evaluate the model's progress.
if _it > 0 and _it % eval_every == 0:
# Sample lots of random walks.
smpls = []
for _ in range(n_eval_iters):
smpls.append(self.session.run(sample_many,
{self.tau: 0.5}))
# Compute score matrix
gr = utils.score_matrix_from_random_walks(
np.array(smpls).reshape([-1, self.rw_len]), self.N)
gr = gr.tocsr()
# Assemble a graph from the score matrix
_graph = utils.graph_from_scores(gr, A_orig.sum())
# Compute edge overlap
edge_overlap = utils.edge_overlap(A_orig.toarray(), _graph)
graphs.append(_graph)
eo.append(edge_overlap)
edge_scores = np.append(gr[tuple(val_ones.T)].A1,
gr[tuple(val_zeros.T)].A1)
# Compute Validation ROC-AUC and average precision scores.
val_performances.append(
(roc_auc_score(actual_labels_val, edge_scores),
average_precision_score(actual_labels_val, edge_scores)))
# Update Gumbel temperature
temperature = np.maximum(
self.params['temp_start'] *
np.exp(-(1 - self.params['temperature_decay']) * _it),
self.params['min_temperature'])
print(
"**** Iter {:<6} Val ROC {:.3f}, AP: {:.3f}, EO {:.3f} ****"
.format(_it, val_performances[-1][0],
val_performances[-1][1],
edge_overlap / A_orig.sum()))
if stopping is None: # Evaluate VAL criterion
if np.sum(val_performances[-1]) > best_performance:
# New "best" model
best_performance = np.sum(val_performances[-1])
patience = max_patience
_ = saver.save(self.session, save_file)
else:
patience -= 1
if patience == 0:
print("**** EARLY STOPPING AFTER {} ITERATIONS ****".
format(_it))
break
elif edge_overlap / A_orig.sum(
) >= stopping: # Evaluate EO criterion
print(
"**** EARLY STOPPING AFTER {} ITERATIONS ****".format(
_it))
break
if plot_every > 0 and (_it + 1) % plot_every == 0:
if len(disc_losses) > 10:
plt.plot(disc_losses[9::], label="Critic loss")
plt.plot(gen_losses[9::], label="Generator loss")
else:
plt.plot(disc_losses, label="Critic loss")
plt.plot(gen_losses, label="Generator loss")
plt.legend()
plt.show()
print("**** Training completed after {} iterations. ****".format(_it))
plt.plot(disc_losses[9::], label="Critic loss")
plt.plot(gen_losses[9::], label="Generator loss")
plt.legend()
plt.show()
if stopping is None:
saver.restore(self.session, save_file)
#### Training completed.
log_dict = {
"disc_losses": disc_losses,
'gen_losses': gen_losses,
'val_performances': val_performances,
'edge_overlaps': eo,
'generated_graphs': graphs
}
return log_dict
def gen(scores, tg_sum):
return utils.graph_from_scores(scores, tg_sum)
test_labels = np.concatenate(
(np.ones(len(test_ones)), np.zeros(len(test_zeros))))
test_scores = np.concatenate((scores_matrix[tuple(test_ones.T)].A1,
scores_matrix[tuple(test_zeros.T)].A1))
# In[37]:
print(roc_auc_score(test_labels, test_scores))
# In[38]:
print(average_precision_score(test_labels, test_scores))
A_select = train_graph
print(A_select.sum())
sampled_graph = utils.graph_from_scores(scores_matrix, A_select.sum())
np.savetxt('netgan/plots/sampled_graph.txt', sampled_graph)
stats = utils.compute_graph_statistics(sampled_graph)
f = open('netgan/plots/stats.txt', "w")
f.write(str(stats))
f.close()
sampled_graph_from_walk = utils.graph_from_transitions(
transition_tensor, edges, A_select.sum(), _N)
print(type(sampled_graph_from_walk))
print(type(sampled_graph))
np.savetxt('netgan/plots/sampled_graph_from_walk.txt',
sampled_graph_from_walk)
stats = utils.compute_graph_statistics(sampled_graph_from_walk)
