def test_gae():
model = GAE(encoder=lambda x: x)
model.reset_parameters()
x = torch.Tensor([[1, -1], [1, 2], [2, 1]])
z = model.encode(x)
assert z.tolist() == x.tolist()
adj = model.decoder.forward_all(z)
assert adj.tolist() == torch.sigmoid(
torch.Tensor([[+2, -1, +1], [-1, +5, +4], [+1, +4, +5]])).tolist()
edge_index = torch.tensor([[0, 1], [1, 2]])
value = model.decode(z, edge_index)
assert value.tolist() == torch.sigmoid(torch.Tensor([-1, 4])).tolist()
edge_index = torch.tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]])
data = Data(edge_index=edge_index)
data.num_nodes = edge_index.max().item() + 1
data = train_test_split_edges(data, val_ratio=0.2, test_ratio=0.3)
z = torch.randn(11, 16)
loss = model.recon_loss(z, data.train_pos_edge_index)
assert loss.item() > 0
auc, ap = model.test(z, data.val_pos_edge_index, data.val_neg_edge_index)
assert auc >= 0 and auc <= 1 and ap >= 0 and ap <= 1
def test_gae():
model = GAE(encoder=lambda x: x)
model.reset_parameters()
x = torch.Tensor([[1, -1], [1, 2], [2, 1]])
z = model.encode(x)
assert z.tolist() == x.tolist()
adj = model.decode(z)
assert adj.tolist() == torch.sigmoid(
torch.Tensor([[+2, -1, +1], [-1, +5, +4], [+1, +4, +5]])).tolist()
edge_index = torch.tensor([[0, 1], [1, 2]])
value = model.decode_indices(z, edge_index)
assert value.tolist() == torch.sigmoid(torch.Tensor([-1, 4])).tolist()
edge_index = torch.tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]])
data = Data(edge_index=edge_index)
data = model.split_edges(data, val_ratio=0.2, test_ratio=0.3)
assert data.val_pos_edge_index.size() == (2, 2)
assert data.val_neg_edge_index.size() == (2, 2)
assert data.test_pos_edge_index.size() == (2, 3)
assert data.test_neg_edge_index.size() == (2, 3)
assert data.train_pos_edge_index.size() == (2, 5)
assert data.train_neg_adj_mask.size() == (11, 11)
assert data.train_neg_adj_mask.sum().item() == (11**2 - 11) / 2 - 4 - 6 - 5
z = torch.randn(11, 16)
loss = model.recon_loss(z, data.train_pos_edge_index)
assert loss.item() > 0
auc, ap = model.test(z, data.val_pos_edge_index, data.val_neg_edge_index)
assert auc >= 0 and auc <= 1 and ap >= 0 and ap <= 1
def test_gae():
model = GAE(encoder=lambda x: x)
model.reset_parameters()
x = torch.Tensor([[1, -1], [1, 2], [2, 1]])
z = model.encode(x)
assert z.tolist() == x.tolist()
adj = model.decoder.forward_all(z)
assert adj.tolist() == torch.sigmoid(
torch.Tensor([[+2, -1, +1], [-1, +5, +4], [+1, +4, +5]])).tolist()
edge_index = torch.tensor([[0, 1], [1, 2]])
value = model.decode(z, edge_index)
assert value.tolist() == torch.sigmoid(torch.Tensor([-1, 4])).tolist()
edge_index = torch.tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]])
data = Data(edge_index=edge_index, num_nodes=11)
transform = RandomLinkSplit(split_labels=True,
add_negative_train_samples=False)
train_data, val_data, test_data = transform(data)
z = torch.randn(11, 16)
loss = model.recon_loss(z, train_data.pos_edge_label_index)
assert loss.item() > 0
auc, ap = model.test(z, val_data.pos_edge_label_index,
val_data.neg_edge_label_index)
assert auc >= 0 and auc <= 1 and ap >= 0 and ap <= 1
def train(dataset, args, writer = None):
task = args.task
test_loader = loader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True)
if task == 'link':
model = GAE(models.GNNStack(dataset.num_node_features, args.hidden_dim, int(dataset.num_classes),
args))
elif task == 'node':
model = models.GNNStack(dataset.num_node_features, args.hidden_dim, int(dataset.num_classes),
args)
else:
raise RuntimeError("Unknown task.")
metrics_for_labels = True if args.metrics_for_labels == 'True' else False
scheduler, opt = build_optimizer(args, model.parameters())
print("Training \nModel: {}, Data representation: {}. Dataset: {}, Task type: {}". format(args.model_name, args.graph_type, args.dataset, args.task))
metric_text = 'test accuracy' if task == 'node' else 'test precision'
for epoch in range(args.epochs):
total_loss = 0
model.train()
for batch in loader:
opt.zero_grad()
if task == 'node':
pred = model(batch)
label = batch.y
pred = pred[batch.train_mask]
label = label[batch.train_mask]
loss = model.loss(pred, label)
else:
train_pos_edge_index = batch.train_pos_edge_index
z = model.encode(batch)
loss = model.recon_loss(z, train_pos_edge_index)
loss.backward()
opt.step()
total_loss += loss.item() * batch.num_graphs
total_loss /= len(loader.dataset)
if writer == None:
print(total_loss)
else:
writer.add_scalar("loss", total_loss, epoch)
if epoch % 10 == 0:
test_metric, _ = test(loader, model, task = task)
if writer == None:
print(test_metric, metric_text)
else:
writer.add_scalar(metric_text, test_metric, epoch)
if metrics_for_labels == True and epoch == args.epochs -1:
_, labels_metrics = test(loader, model, task = task, metrics_for_labels=metrics_for_labels)
print('{} for labels:\n {}'.format(metric_text, labels_metrics))
def run_experiment(args):
"""
Performing experiment for the given arguments
"""
dataset, data = load_data(args.dataset)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Define Model
encoder = create_encoder(args.model, dataset.num_features,
args.latent_dim).to(device)
decoder = create_decoder(args.decoder).to(device)
if args.model == 'GAE':
model = GAE(encoder=encoder, decoder=decoder).to(device)
else:
model = VGAE(encoder=encoder, decoder=decoder).to(device)
# Split edges of a torch_geometric.data.Data object into pos negative train/val/test edges
# default ratios of positive edges: val_ratio=0.05, test_ratio=0.1
print("Data.edge_index.size", data.edge_index.size(1))
data = model.split_edges(data)
node_features, train_pos_edge_index = data.x.to(
device), data.train_pos_edge_index.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
def train_epoch():
"""
Performing training over a single epoch and optimize over loss
:return: log - loss of training loss
"""
# Todo: Add logging of results
model.train()
optimizer.zero_grad()
# Compute latent embedding Z
latent_embeddings = model.encode(node_features, train_pos_edge_index)
# Calculate loss and
loss = model.recon_loss(latent_embeddings, train_pos_edge_index)
if args.model in ['VGAE']:
loss = loss + (1 / data.num_nodes) * model.kl_loss()
# Compute gradients
loss.backward()
# Perform optimization step
optimizer.step()
# print("Train-Epoch: {} Loss: {}".format(epoch, loss))
# ToDo: Add logging via Tensorboard
log = {'loss': loss}
return log
def test(pos_edge_index, neg_edge_index):
model.eval()
with torch.no_grad():
# compute latent var
z = model.encode(node_features, train_pos_edge_index)
# model.test return - AUC, AP
return model.test(z, pos_edge_index, neg_edge_index)
def test_naive_graph(z, sample_size=1000):
if args.sample_dense_evaluation:
graph_type = "sampled"
z_sample, index_mapping = sample_graph(z, sample_size)
t = time.time()
adjacency = model.decoder.forward_all(
z_sample, sigmoid=(args.decoder == 'dot'))
else:
graph_type = "full"
t = time.time()
adjacency = model.decoder.forward_all(
z, sigmoid=(args.decoder == 'dot'))
print(f"Computing {graph_type} graph took {time.time() - t} seconds.")
print(
f"Adjacency matrix takes {adjacency.element_size() * adjacency.nelement() / 10 ** 6} MB of memory."
)
if args.min_sim_absolute_value is None:
args.min_sim_absolute_value, _ = sample_percentile(
args.min_sim,
adjacency,
dist_measure=args.decoder,
sample_size=sample_size)
if args.sample_dense_evaluation:
precision, recall = sampled_dense_precision_recall(
data, adjacency, index_mapping, args.min_sim_absolute_value)
else:
precision, recall = dense_precision_recall(
data, adjacency, args.min_sim_absolute_value)
print("Predicted {} adjacency matrix has precision {} and recall {}!".
format(graph_type, precision, recall))
return precision, recall
def sample_graph(z, sample_size):
N, D = z.shape
sample_size = min(sample_size, N)
sample_ix = np.random.choice(np.arange(N),
size=sample_size,
replace=False)
# Returns the sampled embeddings, and a mapping from their indices to the originals
return z[sample_ix], {i: sample_ix[i] for i in np.arange(sample_size)}
def test_compare_lsh_naive_graphs(z, assure_correctness=True):
"""
:param z:
:param assure_correctness:
:return:
"""
# Naive Adjacency-Matrix (Non-LSH-Version)
t = time.time()
# Don't use sigmoid in order to directly compare thresholds with LSH
naive_adjacency = model.decoder.forward_all(
z, sigmoid=(args.decoder == 'dot'))
naive_time = time.time() - t
naive_size = naive_adjacency.element_size() * naive_adjacency.nelement(
) / 10**6
if args.min_sim_absolute_value is None:
args.min_sim_absolute_value, _ = sample_percentile(
args.min_sim, z, dist_measure=args.decoder)
print(
"______________________________Naive Graph Computation KPI____________________________________________"
)
print(f"Computing naive graph took {naive_time} seconds.")
print(f"Naive adjacency matrix takes {naive_size} MB of memory.")
# LSH-Adjacency-Matrix:
t = time.time()
lsh_adjacency = LSHDecoder(bands=args.lsh_bands,
rows=args.lsh_rows,
verbose=True,
assure_correctness=assure_correctness,
sim_thresh=args.min_sim_absolute_value)(z)
lsh_time = time.time() - t
lsh_size = lsh_adjacency.element_size() * lsh_adjacency._nnz() / 10**6
print(
"__________________________________LSH Graph Computation KPI__________________________________________"
)
print(f"Computing LSH graph took {lsh_time} seconds.")
print(f"Sparse adjacency matrix takes {lsh_size} MB of memory.")
print(
"________________________________________Precision-Recall_____________________________________________"
)
# 1) Evaluation: Both Adjacency matrices against ground truth graph
naive_precision, naive_recall = dense_precision_recall(
data, naive_adjacency, args.min_sim_absolute_value)
lsh_precision, lsh_recall = sparse_precision_recall(
data, lsh_adjacency)
print(
f"Naive-Precision {naive_precision}; Naive-Recall {naive_recall}")
print(f"LSH-Precision {lsh_precision}; LSH-Recall {lsh_recall}")
print(
"_____________________________Comparison Sparse vs Dense______________________________________________"
)
# 2) Evation: Compare both adjacency matrices against each other
compare_precision, compare_recall = sparse_v_dense_precision_recall(
naive_adjacency, lsh_adjacency, args.min_sim_absolute_value)
print(
f"LSH sparse matrix has {compare_precision} precision and {compare_recall} recall w.r.t. the naively generated dense matrix!"
)
return naive_precision, naive_recall, naive_time, naive_size, lsh_precision, lsh_recall, lsh_time, lsh_size, compare_precision, compare_recall
# Training routine
early_stopping = EarlyStopping(args.use_early_stopping,
patience=args.early_stopping_patience,
verbose=True)
logs = []
if args.load_model and os.path.isfile("checkpoint.pt"):
print("Loading model from savefile...")
model.load_state_dict(torch.load("checkpoint.pt"))
if not (args.load_model and args.early_stopping_patience == 0):
for epoch in range(1, args.epochs):
log = train_epoch()
logs.append(log)
# Validation metrics
val_auc, val_ap = test(data.val_pos_edge_index,
data.val_neg_edge_index)
print('Validation-Epoch: {:03d}, AUC: {:.4f}, AP: {:.4f}'.format(
epoch, val_auc, val_ap))
# Stop training if validation scores have not improved
early_stopping(val_ap, model)
if early_stopping.early_stop:
print("Applying early-stopping")
break
else:
epoch = 0
# Load best encoder
print("Load best model for evaluation.")
model.load_state_dict(torch.load('checkpoint.pt'))
print(
"__________________________________________________________________________"
)
# Training is finished, calculate test metrics
test_auc, test_ap = test(data.test_pos_edge_index,
data.test_neg_edge_index)
print('Test Results: {:03d}, AUC: {:.4f}, AP: {:.4f}'.format(
epoch, test_auc, test_ap))
# Check if early stopping was applied or not - if not: model might not be done with training
if args.epochs == epoch + 1:
print("Model might need more epochs - Increase number of Epochs!")
# Evaluate full graph
latent_embeddings = model.encode(node_features, train_pos_edge_index)
# Save embeddings to embeddings folder if flag is set
if args.save_embeddings:
embeddings_folder = osp.join(osp.dirname(osp.abspath(__file__)),
'embeddings')
if not osp.isdir(embeddings_folder):
os.makedirs(embeddings_folder)
torch.save(
latent_embeddings,
osp.join(embeddings_folder,
args.dataset + "_" + args.decoder + ".pt"))
if not args.lsh:
# Compute precision recall w.r.t the ground truth graph
graph_precision, graph_recall = test_naive_graph(latent_embeddings)
del model
del encoder
del decoder
torch.cuda.empty_cache()
else:
# Precision w.r.t. the generated graph
naive_precision, naive_recall, naive_time, naive_size, lsh_precision, \
lsh_recall, lsh_time, lsh_size, \
compare_precision, compare_recall = test_compare_lsh_naive_graphs(
latent_embeddings)
del model
del encoder
del decoder
torch.cuda.empty_cache()
return {
'args': args,
'test_auc': test_auc,
'test_ap': test_ap,
'naive_precision': naive_precision,
'naive_recall': naive_recall,
'naive_time': naive_time,
'naive_size': naive_size,
'lsh_precision': lsh_precision,
'lsh_recall': lsh_recall,
'lsh_time': lsh_time,
'lsh_size': lsh_size,
'compare_precision': compare_precision,
'compare_recall': compare_recall
}
def run_GAE(input_data, output_dir, epochs=1000, lr=0.01, weight_decay=0.0005):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Device: '.ljust(32), device)
print('Model Name: '.ljust(32), 'GAE')
print('Model params:{:19} lr: {} weight_decay: {}'.format(
'', lr, weight_decay))
print('Total number of epochs to run: '.ljust(32), epochs)
print('*' * 70)
data = input_data.clone().to(device)
in_channels = data.num_features
out_channels = data.num_classes.item()
model = GAE(GAEncoder(in_channels, out_channels)).to(device)
data = input_data.clone().to(device)
split_data = model.split_edges(data)
x, train_pos_edge_index, edge_attr = split_data.x.to(
device), split_data.train_pos_edge_index.to(device), data.edge_attr.to(
device)
split_data.train_idx = split_data.test_idx = data.y = None
optimizer = torch.optim.Adam(model.parameters(),
lr=lr,
weight_decay=weight_decay)
train_losses, test_losses = [], []
aucs = []
aps = []
model.train()
for epoch in range(1, epochs + 1):
train_loss = 0
test_loss = 0
optimizer.zero_grad()
z = model.encode(x, train_pos_edge_index)
train_loss = model.recon_loss(z, train_pos_edge_index)
train_losses.append(train_loss)
train_loss.backward()
optimizer.step()
model.eval()
with torch.no_grad():
z = model.encode(x, train_pos_edge_index)
auc, ap = model.test(z, split_data.test_pos_edge_index,
split_data.test_neg_edge_index)
test_loss = model.recon_loss(z, data.test_pos_edge_index)
test_losses.append(test_loss.item())
aucs.append(auc)
aps.append(ap)
figname = os.path.join(
output_dir, "_".join((GAE.__name__, str(lr), str(weight_decay))))
makepath(output_dir)
if (epoch % int(epochs / 10) == 0):
print(
'Epoch: {} Train loss: {} Test loss: {} AUC: {} AP: {}'
.format(epoch, train_loss, test_loss, auc, ap))
if (epoch == epochs):
print(
'-' * 65,
'\nFinal epoch: {} Train loss: {} Test loss: {} AUC: {} AP: {}'
.format(epoch, train_loss, test_loss, auc, ap))
log = 'Final epoch: {} Train loss: {} Test loss: {} AUC: {} AP: {}'.format(
epoch, train_loss, test_loss, auc, ap)
write_log(log, figname)
print('-' * 65)
plot_linkpred(train_losses, test_losses, aucs, aps, output_dir, epochs,
figname)
return
def perturb_edges(data,
name,
remove_pct,
add_pct,
hidden_channels=16,
epochs=400):
if remove_pct == 0 and add_pct == 0:
return
try:
cached = pickle.load(
open(f'{ROOT}/cache/edge/{name}_{remove_pct}_{add_pct}.pt', 'rb'))
print(f'Use cached edge augmentation for dataset {name}')
if data.setting == 'inductive':
data.train_edge_index = cached
else:
data.edge_index = cached
return
except FileNotFoundError:
try:
A_pred, adj_orig = pickle.load(
open(f'{ROOT}/cache/edge/{name}.pt', 'rb'))
A = sample_graph_det(adj_orig, A_pred, remove_pct, add_pct)
data.edge_index, _ = from_scipy_sparse_matrix(A)
pickle.dump(
data.edge_index,
open(f'{ROOT}/cache/edge/{name}_{remove_pct}_{add_pct}.pt',
'wb'))
return
except FileNotFoundError:
print(
f'cache/edge/{name}_{remove_pct}_{add_pct}.pt not found! Regenerating it now'
)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if data.setting == 'inductive':
train_data = Data(x=data.train_x,
ori_x=data.ori_x,
edge_index=data.train_edge_index,
y=data.train_y)
else:
train_data = deepcopy(data)
edge_index = deepcopy(train_data.edge_index)
train_data = train_test_split_edges(train_data,
val_ratio=0.1,
test_ratio=0)
num_features = train_data.ori_x.shape[1]
model = GAE(GCNEncoder(num_features, hidden_channels))
model = model.to(device)
x = train_data.ori_x.to(device)
train_pos_edge_index = train_data.train_pos_edge_index.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
best_val_auc = 0
best_z = None
for epoch in range(1, epochs + 1):
model.train()
optimizer.zero_grad()
z = model.encode(x, train_pos_edge_index)
loss = model.recon_loss(z, train_pos_edge_index)
loss.backward()
optimizer.step()
model.eval()
with torch.no_grad():
z = model.encode(x, train_pos_edge_index)
auc, ap = model.test(z, train_data.val_pos_edge_index,
train_data.val_neg_edge_index)
print('Val | Epoch: {:03d}, AUC: {:.4f}, AP: {:.4f}'.format(
epoch, auc, ap))
if auc > best_val_auc:
best_val_auc = auc
best_z = deepcopy(z)
A_pred = torch.sigmoid(torch.mm(z, z.T)).cpu().numpy()
adj_orig = to_scipy_sparse_matrix(edge_index).asformat('csr')
adj_pred = sample_graph_det(adj_orig, A_pred, remove_pct, add_pct)
if data.setting == 'inductive':
data.train_edge_index, _ = from_scipy_sparse_matrix(adj_pred)
else:
data.edge_index, _ = from_scipy_sparse_matrix(adj_pred)
pickle.dump((A_pred, adj_orig), open(f'{ROOT}/cache/edge/{name}.pt', 'wb'))
if data.setting == 'inductive':
pickle.dump(
data.train_edge_index,
open(f'{ROOT}/cache/edge/{name}_{remove_pct}_{add_pct}.pt', 'wb'))
else:
pickle.dump(
data.edge_index,
open(f'{ROOT}/cache/edge/{name}_{remove_pct}_{add_pct}.pt', 'wb'))
