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))
class UnsGAE(object):
def __init__(self, data, embed_dim, **kwargs):
super(UnsGAE, self).__init__()
self.data = data
self.input_dim = self.data.dim
self.embed_dim = embed_dim
# for now, we only work with 2-layer encoders
self.hidden_dim = kwargs.get('hidden_dim', 2*embed_dim)
self.encoder = kwargs.get('encoder', batched_SAGEEncoder)
self.encoder = self.encoder(self.input_dim,
self.hidden_dim,
self.embed_dim)
self.model = GAE(self.encoder)
# preparing the device
device = kwargs.get('device', 'cuda')
if device=='gpu' and not(torch.cuda.is_available()):
print('CUDA is not available in PyTorch. the model ' +\
'will be initiated on CPU.')
device = 'cpu'
self.device = torch.device(device)
def init_model(self, sizes, weights_path=None):
self.model = self.model.to(self.device)
# sizes are directly used for initializing the model
# but it will be used for every feed-forward as the
# sampling size of the neighbors
assert len(sizes)==self.model.encoder.num_layers, \
'Number of sizes should be equal to the number of layers in the encoder.'
self.sizes = sizes
if not(hasattr(self.data, 'loader')):
self.data.get_neighbor_sampler(self.sizes)
if weights_path is not None:
self.model.load_state_dict(torch.load(weights_path, map_location=self.device))
def init_training(self, neg_num, optim='Adam', lr=1e-5, smooth_par=0.75):
if optim=='Adam':
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=lr)
elif optim=='SGD':
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=lr)
self.train_one_epoch = self._train_edge_batching
self.neg_num = neg_num
if not(hasattr(self.data, 'pos_pairs')):
assert 'pos_samples_path' in kwargs, 'The provided data does ' +\
'not come with positive pairs, and we need a path to the ' +\
'already selected positive samples. You can provide it through ' +\
'input pos_samples_path .'
include_nodes = kwargs.get('include_nodes', None)
self.data.load_positive_pairs(kwargs['pos_samples_path'], include_nodes)
if not(hasattr(self.data, 'neg_sampler')):
#smooth_par = kwargs.get('smooth_par', 0.75)
self.data.get_negative_sampler(smooth_par)
if not(hasattr(self.data, 'x_all')):
self.data._fetch_node_features()
def init_validation(self):
if not(hasattr(self.data, 'x_all')):
self.data._fetch_node_features()
def embed_some(self, sample_inds, b=100):
"""This will be used in the training, when the
embedding of a batch of samples are needed
"""
quot, rem = np.divmod(len(sample_inds), b)
Z = []
for i in range(quot+1):
if i<quot:
b_ids = sample_inds[i*b:(i+1)*b]
elif rem>0:
b_ids = sample_inds[i*b:]
# neighbor-sampling for each sample
_, n_id, adjs = self.data.train_loader.sample(b_ids)
adjs = [adj.to(self.device) for adj in adjs]
# get feature vectors through the neighbors sampled above
batch_X = torch.from_numpy(self.data.get_node_features(n_id))
batch_X = batch_X.to(torch.float).to(self.device)
# the encoder's output as the embedding
try:
batch_Z = self.model.encoder(batch_X, adjs)
except:
pdb.set_trace()
Z += [batch_Z]
Z = torch.cat(Z, dim=0)
return Z
def embed_all(self):
L = self.model.encoder.num_layers
pbar = tqdm(total=self.data.n_x * L, position=0, leave=True)
pbar.set_description('Evaluating')
self.model.encoder.eval()
# inference is used in the evaluation stage (not in training) when
# the embeddings for "all" nodes will be computed. It's written in a way
# that is faster than the foward-passing function which is mostly used
# for single batches in the training
with torch.no_grad():
for i in range(L):
xs = []
for batch_size, n_id, adj in self.data.test_loader:
edge_index, _, size = adj.to(self.device)
if i==0:
x = torch.from_numpy(self.data.get_node_features(n_id))
x = x.to(torch.float).to(self.device)
else:
x = x_all[n_id,:].to(self.device)
x_target = x[:size[1]]
x = self.model.encoder.convs[i]((x,x_target), edge_index)
if i != L-1:
x = F.relu(x)
xs.append(x[:batch_size,:].cpu())
pbar.update(batch_size)
x_all = torch.cat(xs, dim=0)
pbar.close()
return x_all
def _train_edge_batching(self, ep, batch_size=5000):
assert hasattr(self.data, 'pos_pairs'), 'Positive and negative ' + \
'samples must be generated before starting the training'
self.model.train()
neg_num = self.neg_num
torch.multiprocessing.set_sharing_strategy('file_system')
pbar = tqdm(total=self.data.pos_pairs.shape[1], position=0, leave=True)
pbar.set_description(f'Epoch {ep:02d}')
total_loss = 0
np.random.shuffle(self.data.pos_pairs.T)
quot, rem = np.divmod(self.data.pos_pairs.shape[1], batch_size)
for i in range(quot+1):
# positive mini-batch
# (#: batch size)
if i<quot:
batch_pos_pairs = self.data.pos_pairs[:,i*batch_size:(i+1)*batch_size]
else:
batch_pos_pairs = self.data.pos_pairs[:,i*batch_size:]
batch_pos_samples, pos_edge_index = np.unique(batch_pos_pairs,
return_inverse=True)
pos_edge_index = pos_edge_index.reshape(batch_pos_pairs.shape)
# negative mini-batch
# (#: batch_size * neg_num)
batch_neg_samples = self.data.neg_sampler.sample(
torch.Size([neg_num*batch_pos_pairs.shape[1]]))
neg_edge_index = np.array([np.repeat(pos_edge_index[0,:],neg_num),
np.arange(pos_edge_index.max()+1,
pos_edge_index.max()+len(batch_neg_samples)+1)])
# embeddings of the nodes involved in + and - edges
self.optimizer.zero_grad()
unodes = batch_pos_samples.tolist() + batch_neg_samples.tolist()
Z = self.embed_some(unodes)
# reconstruction loss
pos_edge_index = torch.from_numpy(pos_edge_index).to(self.device)
neg_edge_index = torch.from_numpy(neg_edge_index).to(self.device)
loss = self.model.recon_loss(Z, pos_edge_index, neg_edge_index)
loss.backward()
self.optimizer.step()
total_loss += float(loss)
pbar.update(batch_size)
pbar.close()
loss = total_loss / (quot+1)
return loss
def validate(self):
self.model.eval()
Z = self.embed_all()
ents_Z = Z[:-1,:][self.data.selected_inds[:-1]>=self.data.nA,:].detach().numpy()
prop_Z = Z[self.data.tags=='prop',:].detach().numpy()
scores = np.dot(ents_Z, prop_Z.T).squeeze()
sorted_ents = self.data.selected_ents[np.argsort(-scores)]
unstudied_sorted_ents = np.array([x for x in sorted_ents
if x not in self.data.studied_ents])
preds = unstudied_sorted_ents[:50]
prec = np.isin(preds,self.data.GT).sum() / len(preds)
return prec
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'))