def task(embeds):
train_embs = embeds[0, idx_train]
val_embs = embeds[0, idx_val]
test_embs = embeds[0, idx_test]
log = LogReg(hid_units, nb_classes)
opt = torch.optim.Adam(log.parameters(), lr=0.01, weight_decay=0.0)
if torch.cuda.is_available():
log.cuda()
for _ in range(100):
log.train()
opt.zero_grad()
logits = log(train_embs)
loss = xent(logits, train_lbls)
loss.backward()
opt.step()
logits = log(test_embs)
preds = torch.argmax(logits, dim=1)
acc = torch.sum(preds == test_lbls).float() / test_lbls.shape[0]
return acc.detach().cpu().numpy()
val_embs = embeds[0, idx_val]
test_embs = embeds[0, idx_test]
train_lbls = torch.argmax(labels[0, idx_train], dim=1)
val_lbls = torch.argmax(labels[0, idx_val], dim=1)
test_lbls = torch.argmax(labels[0, idx_test], dim=1)
tot = torch.zeros(1)
tot = tot.cuda()
accs = []
for _ in range(50):
log = LogReg(hid_units, nb_classes)
opt = torch.optim.Adam(log.parameters(), lr=0.01, weight_decay=0.0)
log.cuda()
pat_steps = 0
best_acc = torch.zeros(1)
best_acc = best_acc.cuda()
for _ in range(100):
log.train()
opt.zero_grad()
logits = log(train_embs)
loss = xent(logits, train_lbls)
loss.backward()
opt.step()
logits = log(test_embs)
def main():
saved_graph = os.path.join('assets', 'saved_graphs', 'best_dgi.pickle')
saved_logreg = os.path.join('assets', 'saved_graphs', 'best_logreg.pickle')
dataset = 'cora'
# training params
batch_size = 1
nb_epochs = 10000
patience = 25
lr = 0.001
l2_coef = 0.0
drop_prob = 0.0
hid_units = 512
sparse = True
nonlinearity = 'prelu' # special name to separate parameters
adj, features, labels, idx_train, idx_test, idx_val = process.load_data(dataset)
features, _ = process.preprocess_features(features)
nb_nodes = features.shape[0]
ft_size = features.shape[1]
nb_classes = labels.shape[1]
adj = process.normalize_adj(adj + sp.eye(adj.shape[0]))
if sparse:
adj = process.sparse_mx_to_torch_sparse_tensor(adj)
else:
adj = (adj + sp.eye(adj.shape[0])).todense()
features = torch.FloatTensor(features[np.newaxis])
if not sparse:
adj = torch.FloatTensor(adj[np.newaxis])
labels = torch.FloatTensor(labels[np.newaxis])
idx_train = torch.LongTensor(idx_train)
idx_val = torch.LongTensor(idx_val)
idx_test = torch.LongTensor(idx_test)
print("Training Nodes: {}, Testing Nodes: {}, Validation Nodes: {}".format(len(idx_train), len(idx_test), len(idx_val)))
model = DGI(ft_size, hid_units, nonlinearity)
optimiser = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=l2_coef)
if torch.cuda.is_available():
print('Using CUDA')
model.cuda()
features = features.cuda()
if sparse:
sp_adj = sp_adj.cuda()
else:
adj = adj.cuda()
labels = labels.cuda()
idx_train = idx_train.cuda()
idx_val = idx_val.cuda()
idx_test = idx_test.cuda()
b_xent = nn.BCEWithLogitsLoss()
xent = nn.CrossEntropyLoss()
cant_wait = 0
best = 1e9
best_t = 0
if not os.path.exists(saved_graph):
pbar = trange(nb_epochs)
for epoch in pbar:
model.train()
optimiser.zero_grad()
idx = np.random.permutation(nb_nodes)
shuf_fts = features[:, idx, :]
lbl_1 = torch.ones(batch_size, nb_nodes)
lbl_2 = torch.zeros(batch_size, nb_nodes)
lbl = torch.cat((lbl_1, lbl_2), 1)
if torch.cuda.is_available():
shuf_fts = shuf_fts.cuda()
lbl = lbl.cuda()
logits = model(features, shuf_fts, adj, sparse, None, None, None)
loss = b_xent(logits, lbl)
pbar.desc = 'Loss: {:.4f}'.format(loss)
if loss < best:
best = loss
best_t = epoch
cnt_wait = 0
torch.save(model.state_dict(), saved_graph)
else:
cant_wait += 1
if cant_wait == patience:
tqdm.write('Early stopping!')
break
loss.backward()
optimiser.step()
print('Loading {}th Epoch'.format(best_t) if best_t else 'Loading Existing Graph')
model.load_state_dict(torch.load(saved_graph))
embeds, _ = model.embed(features, adj, sparse, None)
train_embs = embeds[0, idx_train]
val_embs = embeds[0, idx_val]
test_embs = embeds[0, idx_test]
train_lbls = torch.argmax(labels[0, idx_train], dim=1)
val_lbls = torch.argmax(labels[0, idx_val], dim=1)
test_lbls = torch.argmax(labels[0, idx_test], dim=1)
tot = torch.zeros(1)
if torch.cuda.is_available():
tot = tot.cuda()
accs = []
print("\nValidation:")
pbar = trange(50)
for _ in pbar:
log = LogReg(hid_units, nb_classes)
opt = torch.optim.Adam(log.parameters(), lr=0.01, weight_decay=0.0)
pat_steps = 0
best_acc = torch.zeros(1)
if torch.cuda.is_available():
log.cuda()
best_acc = best_acc.cuda()
for _ in range(100):
log.train()
opt.zero_grad()
logits = log(train_embs)
loss = xent(logits, train_lbls)
loss.backward()
opt.step()
logits = log(test_embs)
preds = torch.argmax(logits, dim=1)
acc = torch.sum(preds == test_lbls).float() / test_lbls.shape[0]
accs.append(acc * 100)
pbar.desc = "Accuracy: {:.2f}%".format(100 * acc)
tot += acc
torch.save(log.state_dict(), saved_logreg)
accs = torch.stack(accs)
print('Average Accuracy: {:.2f}%'.format(accs.mean()))
print('Standard Deviation: {:.3f}'.format(accs.std()))
print("\nTesting")
logits = log(val_embs)
preds = torch.argmax(logits, dim=1)
acc = torch.sum(preds == val_lbls).float() / val_lbls.shape[0]
print("Accuracy: {:.2f}%".format(100 * acc))
def process_inductive(dataset, gnn_type="GCNConv", K=None, random_init=False, runs=10):
hyperparameters = get_hyperparameters()
nb_epochs = hyperparameters["nb_epochs"]
patience = hyperparameters["patience"]
lr = hyperparameters["lr"]
l2_coef = hyperparameters["l2_coef"]
drop_prob = hyperparameters["drop_prob"]
hid_units = hyperparameters["hid_units"]
nonlinearity = hyperparameters["nonlinearity"]
batch_size = hyperparameters["batch_size"]
norm_features = torch_geometric.transforms.NormalizeFeatures()
dataset_train = PPI(
"./geometric_datasets/"+dataset,
split="train",
transform=norm_features,
)
print(dataset_train)
dataset_val = PPI(
"./geometric_datasets/"+dataset,
split="val",
transform=norm_features,
)
print(dataset_val)
dataset_test = PPI(
"./geometric_datasets/"+dataset,
split="test",
transform=norm_features,
)
data = []
for d in dataset_train:
data.append(d)
for d in dataset_val:
data.append(d)
ft_size = dataset_train[0].x.shape[1]
nb_classes = dataset_train[0].y.shape[1] # multilabel
b_xent = nn.BCEWithLogitsLoss()
loader_train = DataLoader(
data,
batch_size=hyperparameters["batch_size"],
shuffle=True,
)
loader_test = DataLoader(
dataset_test,
batch_size=hyperparameters["batch_size"],
shuffle=False
)
all_accs = []
for _ in range(runs):
model = DGI(ft_size, hid_units, nonlinearity, update_rule=gnn_type, batch_size=1, K=K)
model_name = get_model_name(dataset, gnn_type, K, random_init=random_init)
print(model)
optimiser = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=l2_coef)
if torch.cuda.is_available():
print('Using CUDA')
model = model.cuda()
model.train()
torch.cuda.empty_cache()
for epoch in range(20):
if random_init:
break
total_loss = 0
batch_id = 0
model.train()
loaded = list(loader_train)
for batch in loaded:
optimiser.zero_grad()
if torch.cuda.is_available:
batch = batch.to('cuda')
nb_nodes = batch.x.shape[0]
features = batch.x
labels = batch.y
edge_index = batch.edge_index
idx = np.random.randint(0, len(data))
while idx == batch_id:
idx = np.random.randint(0, len(data))
shuf_fts = torch.nn.functional.dropout(loaded[idx].x, drop_prob)
edge_index2 = loaded[idx].edge_index
lbl_1 = torch.ones(nb_nodes)
lbl_2 = torch.zeros(shuf_fts.shape[0])
lbl = torch.cat((lbl_1, lbl_2), 0)
if torch.cuda.is_available():
shuf_fts = shuf_fts.cuda()
if edge_index2 is not None:
edge_index2 = edge_index2.cuda()
lbl = lbl.cuda()
logits = model(features, shuf_fts, edge_index, batch=batch.batch, edge_index_alt=edge_index2)
loss = b_xent(logits, lbl)
loss.backward()
optimiser.step()
batch_id += 1
total_loss += loss.item()
print(epoch, 'Train Loss:', total_loss/(len(dataset_train)))
torch.save(model.state_dict(), './trained_models/'+model_name)
torch.cuda.empty_cache()
print('Loading last epoch')
if not random_init:
model.load_state_dict(torch.load('./trained_models/'+model_name))
model.eval()
b_xent_reg = nn.BCEWithLogitsLoss(pos_weight=torch.tensor(2.25))
train_embs, whole_train_data = preprocess_embeddings(model, dataset_train)
val_embs, whole_val_data = preprocess_embeddings(model, dataset_val)
test_embs, whole_test_data = preprocess_embeddings(model, dataset_test)
for _ in range(50):
log = LogReg(hid_units, nb_classes)
opt = torch.optim.Adam(log.parameters(), lr=0.01, weight_decay=0.0)
log.cuda()
pat_steps = 0
best = 1e9
log.train()
for _ in range(250):
opt.zero_grad()
logits = log(train_embs)
loss = b_xent_reg(logits, whole_train_data.y)
loss.backward()
opt.step()
log.eval()
val_logits = log(val_embs)
loss = b_xent_reg(val_logits, whole_val_data.y)
if loss.item() < best:
best = loss.item()
pat_steps = 0
if pat_steps >= 5:
break
pat_steps += 1
log.eval()
logits = log(test_embs)
preds = torch.sigmoid(logits) > 0.5
f1 = sklearn.metrics.f1_score(whole_test_data.y.cpu(), preds.long().cpu(), average='micro')
all_accs.append(float(f1))
print()
print('Micro-averaged f1:', f1)
all_accs = torch.tensor(all_accs)
with open("./results/"+model_name[:-4]+"_results.txt", "w") as f:
f.writelines([str(all_accs.mean().item())+'\n', str(all_accs.std().item())])
print(all_accs.mean())
print(all_accs.std())
def process_transductive(dataset, gnn_type='GCNConv', K=None, random_init=False, runs=10, drop_sigma=False, just_plot=False):
dataset_str = dataset
norm_features = torch_geometric.transforms.NormalizeFeatures()
dataset = Planetoid("./geometric_datasets"+'/'+dataset,
dataset,
transform=norm_features)[0]
# training params
batch_size = 1 # Transductive setting
hyperparameters = get_hyperparameters()
nb_epochs = hyperparameters["nb_epochs"]
patience = hyperparameters["patience"]
lr = hyperparameters["lr"]
xent = nn.CrossEntropyLoss()
l2_coef = hyperparameters["l2_coef"]
drop_prob = hyperparameters["drop_prob"]
hid_units = hyperparameters["hid_units"]
nonlinearity = hyperparameters["nonlinearity"]
nb_nodes = dataset.x.shape[0]
ft_size = dataset.x.shape[1]
nb_classes = torch.max(dataset.y).item()+1 # 0 based cnt
features = dataset.x
labels = dataset.y
edge_index = dataset.edge_index
edge_index, _ = torch_geometric.utils.add_remaining_self_loops(edge_index)
mask_train = dataset.train_mask
mask_val = dataset.val_mask
mask_test = dataset.test_mask
model_name = get_model_name(dataset_str, gnn_type, K, random_init=random_init, drop_sigma=drop_sigma)
with open("./results/"+model_name[:-4]+"_results.txt", "w") as f:
pass
accs = []
for i in range(runs):
model = DGI(ft_size, hid_units, nonlinearity, update_rule=gnn_type, K=K, drop_sigma=drop_sigma)
print(model, model_name, drop_sigma)
optimiser = torch.optim.Adam(model.parameters(), lr=lr)
if torch.cuda.is_available():
print('Using CUDA')
features = features.cuda()
labels = labels.cuda()
edge_index = edge_index.cuda()
mask_train = mask_train.cuda()
mask_val = mask_val.cuda()
mask_test = mask_test.cuda()
model = model.cuda()
best_t = train_transductive(dataset, dataset_str, edge_index, gnn_type, model_name, K=K, random_init=random_init, drop_sigma=drop_sigma)
xent = nn.CrossEntropyLoss()
print('Loading {}th epoch'.format(best_t))
print(model, model_name)
if not random_init:
model.load_state_dict(torch.load('./trained_models/'+model_name))
model.eval()
embeds, _ = model.embed(features, edge_index, None, standardise=False)
if just_plot:
plot_tsne(embeds, labels, model_name)
exit(0)
train_embs = embeds[mask_train, :]
val_embs = embeds[mask_val, :]
test_embs = embeds[mask_test, :]
train_lbls = labels[mask_train]
val_lbls = labels[mask_val]
test_lbls = labels[mask_test]
tot = torch.zeros(1)
tot = tot.cuda()
for _ in range(50):
log = LogReg(hid_units, nb_classes)
opt = torch.optim.Adam(log.parameters(), lr=0.01, weight_decay=0.0)
log.cuda()
pat_steps = 0
best_acc = torch.zeros(1)
best_acc = best_acc.cuda()
for _ in range(150):
log.train()
opt.zero_grad()
logits = log(train_embs)
loss = xent(logits, train_lbls)
loss.backward()
opt.step()
logits = log(test_embs)
preds = torch.argmax(logits, dim=1)
acc = torch.sum(preds == test_lbls).float() / test_lbls.shape[0]
accs.append(acc * 100)
print(acc)
tot += acc
print('Average accuracy:', tot / 50)
all_accs = torch.stack(accs, dim=0)
with open("./results/"+model_name[:-4]+"_results.txt", "a+") as f:
f.writelines([str(all_accs.mean().item())+'\n', str(all_accs.std().item())+'\n'])
print(all_accs.mean())
print(all_accs.std())
