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
}
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
