def run_GCN(args,
gpu_id=None,
exp_name=None,
number=0,
return_model=False,
return_time_series=False):
random.seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
final_acc = 0
best_acc = 0
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
running_device = "cpu" if gpu_id is None \
else torch.device('cuda:{}'.format(gpu_id) if torch.cuda.is_available() else 'cpu')
dataset_kwargs = {}
train_d, adj_list, x_list = get_dataset(args, dataset_kwargs)
lable = train_d.data.y
A_I = adj_list[0]
A_I_nomal = adj_list[1]
nb_edges = train_d.data.num_edges
nb_nodes = train_d.data.num_nodes
nb_feature = train_d.data.num_features
nb_classes = int(lable.max() - lable.min()) + 1
lable_matrix = (lable.view(nb_nodes, 1).repeat(1, nb_nodes) == lable.view(
1, nb_nodes).repeat(nb_nodes, 1)) + 0
I = (torch.eye(nb_nodes).to(lable_matrix.device) == 1)
lable_matrix[I] = 0
zero_vec = 0.0 * torch.ones_like(A_I_nomal)
if args.dataset_name in [
'Photo', 'DBLP', 'Crocodile', 'CoraFull', 'WikiCS'
]:
useA = True
else:
useA = False
model = UGRL_GCN_test(nb_nodes,
nb_feature,
args.dim,
dim_x=args.dim_x,
useact=args.usingact,
liner=args.UsingLiner,
dropout=args.dropout,
useA=useA)
optimiser = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=0.0001)
model.to(running_device)
lable = lable.to(running_device)
if args.dataset_name == 'WikiCS':
train_lbls = lable[train_d.data.train_mask[:, args.NewATop]] # capture
test_lbls = lable[train_d.data.test_mask]
elif args.dataset_name in ['Cora', 'CiteSeer', 'PubMed']:
train_lbls = lable[train_d.data.train_mask]
test_lbls = lable[train_d.data.test_mask]
elif args.dataset_name in ['Photo', 'DBLP', 'Crocodile', 'CoraFull']:
train_index = []
test_index = []
for j in range(lable.max().item() + 1):
num = ((lable == j) + 0).sum().item()
index = torch.range(0, len(lable) - 1)[(lable == j)]
x_list0 = random.sample(list(index), int(len(index) * 0.1))
for x in x_list0:
train_index.append(int(x))
for c in range(len(lable)):
if int(c) not in train_index:
test_index.append(int(c))
train_lbls = lable[train_index]
test_lbls = lable[test_index]
val_lbls = lable[train_index]
A_I_nomal_dense = A_I_nomal
I_input = torch.eye(A_I_nomal.shape[1]) # .to(A_I_nomal.device)
if args.dataset_name in ['PubMed', 'CoraFull', 'DBLP']:
pass
elif args.dataset_name in ['Crocodile', 'Photo', 'WikiCS']:
A_I_nomal_dense = A_I_nomal_dense.to(running_device)
######################sparse################
if args.dataset_name in [
'PubMed', 'Crocodile', 'CoraFull', 'DBLP', 'Photo', 'WikiCS'
]:
A_I_nomal = A_I_nomal.to_sparse()
model.sparse = True
I_input = I_input.to_sparse()
######################sparse################
A_I_nomal = A_I_nomal.to(running_device)
I_input = I_input.to(A_I_nomal.device)
mask_I = I.to(running_device)
zero_vec = zero_vec.to(running_device)
my_margin = args.margin1
my_margin_2 = my_margin + args.margin2
margin_loss = torch.nn.MarginRankingLoss(margin=my_margin, reduce=False)
num_neg = args.NN
for current_iter, epoch in enumerate(
tqdm(range(args.start_epoch, args.start_epoch + args.epochs + 1))):
model.train()
optimiser.zero_grad()
idx = np.random.permutation(nb_nodes)
feature_X = x_list[0].to(running_device)
lbl_z = torch.tensor([0.]).to(running_device)
feature_a = feature_X
feature_p = feature_X
feature_n = []
idx_list = []
idx_lable = []
for i in range(num_neg):
idx_0 = np.random.permutation(nb_nodes)
idx_list.append(idx_0)
idx_lable.append(lable[idx_0])
feature_temp = feature_X[idx_0]
feature_n.append(feature_temp)
h_a, h_p, h_n_lsit, h_a_0, h_p_0, h_n_0_list = model(feature_a,
feature_p,
feature_n,
A_I_nomal,
I=I_input)
s_p = F.pairwise_distance(h_a, h_p)
cos_0_list = []
for h_n_0 in h_n_0_list:
cos_0 = F.pairwise_distance(h_a_0, h_n_0)
cos_0_list.append(cos_0)
cos_0_stack = torch.stack(cos_0_list).detach()
cos_0_min = cos_0_stack.min(dim=0)[0]
cos_0_max = cos_0_stack.max(dim=0)[0]
gap = cos_0_max - cos_0_min
weight_list = []
for i in range(cos_0_stack.size()[0]):
weight_list.append((cos_0_stack[i] - cos_0_min) / gap)
s_n_list = []
s_n_cosin_list = []
for h_n in h_n_lsit:
if args.dataset_name in ['Cora', 'CiteSeer']:
s_n_cosin_list.append(cosine_dist(h_a, h_n)[mask_I].detach())
s_n = F.pairwise_distance(h_a, h_n)
s_n_list.append(s_n)
margin_label = -1 * torch.ones_like(s_p)
loss_mar = 0
mask_margin_N = 0
i = 0
for s_n in s_n_list:
loss_mar += (margin_loss(s_p, s_n, margin_label) *
weight_list[i]).mean()
mask_margin_N += torch.max((s_n - s_p.detach() - my_margin_2),
lbl_z).sum()
i += 1
mask_margin_N = mask_margin_N / num_neg
string_1 = " loss_1: {:.3f}||loss_2: {:.3f}||".format(
loss_mar.item(), mask_margin_N.item())
loss = loss_mar * args.w_loss1 + mask_margin_N * args.w_loss2 / nb_nodes
if args.dataset_name in ['Cora']:
loss = loss_mar * args.w_loss1 + mask_margin_N * args.w_loss2
loss.backward()
optimiser.step()
model.eval()
if args.dataset_name in ['Crocodile', 'WikiCS', 'Photo']:
h_p_d = h_p.detach()
S_new = cosine_dist(h_p_d, h_p_d)
model.A = normalize_graph(torch.mul(S_new,
A_I_nomal_dense)).to_sparse()
elif args.dataset_name in ['Cora', 'CiteSeer']:
h_a, h_p = model.embed(feature_a,
feature_p,
feature_n,
A_I_nomal,
I=I_input)
s_a = cosine_dist(h_a, h_a).detach()
S = (torch.stack(s_n_cosin_list).mean(dim=0).expand_as(A_I) -
s_a).detach()
# zero_vec = -9e15 * torch.ones_like(S)
one_vec = torch.ones_like(S)
s_a = torch.where(A_I_nomal > 0, one_vec, zero_vec)
attention = torch.where(S < 0, s_a, zero_vec)
attention_N = normalize_graph(attention)
attention[I] = 0
model.A = attention_N
if epoch % 50 == 0:
model.eval()
h_a, h_p = model.embed(feature_a,
feature_p,
feature_n,
A_I_nomal,
I=I_input)
if args.useNewA:
embs = h_p #torch.cat((h_a,h_p),dim=1)
else:
embs = h_a
if args.dataset_name in ['Cora', 'CiteSeer', 'PubMed']:
embs = embs / embs.norm(dim=1)[:, None]
if args.dataset_name == 'WikiCS':
train_embs = embs[train_d.data.train_mask[:, args.NewATop]]
test_embs = embs[train_d.data.test_mask]
elif args.dataset_name in ['Cora', 'CiteSeer', 'PubMed']:
train_embs = embs[train_d.data.train_mask]
test_embs = embs[train_d.data.test_mask]
# #val_embs = embs[train_d.data.val_mask]
elif args.dataset_name in [
'Photo', 'DBLP', 'Crocodile', 'CoraFull'
]:
train_embs = embs[train_index]
test_embs = embs[test_index]
#val_embs = embs[train_index]
accs = []
accs_small = []
xent = nn.CrossEntropyLoss()
for _ in range(2):
log = LogReg(args.dim, nb_classes)
opt = torch.optim.Adam(log.parameters(),
lr=1e-2,
weight_decay=args.wd)
log.to(running_device)
for _ in range(args.num1):
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)
ac = []
for i in range(nb_classes):
acc_small = torch.sum(
preds[test_lbls == i] == test_lbls[test_lbls == i]
).float() / test_lbls[test_lbls == i].shape[0]
ac.append(acc_small * 100)
accs_small = ac
accs = torch.stack(accs)
string_3 = ""
for i in range(nb_classes):
string_3 = string_3 + "|{:.1f}".format(accs_small[i].item())
string_2 = Fore.GREEN + " epoch: {},accs: {:.1f},std: {:.2f} ".format(
epoch,
accs.mean().item(),
accs.std().item())
tqdm.write(string_1 + string_2 + string_3)
final_acc = accs.mean().item()
best_acc = max(best_acc, final_acc)
return final_acc, best_acc
def evaluate(embeds, idx_train, idx_val, idx_test, labels, device, isTest=True):
hid_units = embeds.shape[2]
nb_classes = labels.shape[2]
xent = nn.CrossEntropyLoss()
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)
accs = []
micro_f1s = []
macro_f1s = []
macro_f1s_val = [] ##
for _ in range(50):
log = LogReg(hid_units, nb_classes)
opt = torch.optim.Adam(log.parameters(), lr=0.01, weight_decay=0.0)
log.to(device)
val_accs = []; test_accs = []
val_micro_f1s = []; test_micro_f1s = []
val_macro_f1s = []; test_macro_f1s = []
for iter_ in range(50):
# train
log.train()
opt.zero_grad()
logits = log(train_embs)
loss = xent(logits, train_lbls)
loss.backward()
opt.step()
# val
logits = log(val_embs)
preds = torch.argmax(logits, dim=1)
val_acc = torch.sum(preds == val_lbls).float() / val_lbls.shape[0]
val_f1_macro = f1_score(val_lbls.cpu(), preds.cpu(), average='macro')
val_f1_micro = f1_score(val_lbls.cpu(), preds.cpu(), average='micro')
val_accs.append(val_acc.item())
val_macro_f1s.append(val_f1_macro)
val_micro_f1s.append(val_f1_micro)
# test
logits = log(test_embs)
preds = torch.argmax(logits, dim=1)
test_acc = torch.sum(preds == test_lbls).float() / test_lbls.shape[0]
test_f1_macro = f1_score(test_lbls.cpu(), preds.cpu(), average='macro')
test_f1_micro = f1_score(test_lbls.cpu(), preds.cpu(), average='micro')
test_accs.append(test_acc.item())
test_macro_f1s.append(test_f1_macro)
test_micro_f1s.append(test_f1_micro)
max_iter = val_accs.index(max(val_accs))
accs.append(test_accs[max_iter])
max_iter = val_macro_f1s.index(max(val_macro_f1s))
macro_f1s.append(test_macro_f1s[max_iter])
macro_f1s_val.append(val_macro_f1s[max_iter]) ###
max_iter = val_micro_f1s.index(max(val_micro_f1s))
micro_f1s.append(test_micro_f1s[max_iter])
if isTest:
print("\t[Classification] Macro-F1: {:.4f} ({:.4f}) | Micro-F1: {:.4f} ({:.4f})".format(np.mean(macro_f1s),
np.std(macro_f1s),
np.mean(micro_f1s),
np.std(micro_f1s)))
else:
return np.mean(macro_f1s_val), np.mean(macro_f1s)
test_embs = np.array(test_embs.cpu())
test_lbls = np.array(test_lbls.cpu())
run_kmeans(test_embs, test_lbls, nb_classes)
run_similarity_search(test_embs, test_lbls)
