def run_fix_mask(args, seed, adj_percent, wei_percent):
pruning.setup_seed(seed)
adj, features, labels, idx_train, idx_val, idx_test = load_data(
args['dataset'])
node_num = features.size()[0]
class_num = labels.numpy().max() + 1
adj = adj.cuda()
features = features.cuda()
labels = labels.cuda()
loss_func = nn.CrossEntropyLoss()
net_gcn = net.net_gcn(embedding_dim=args['embedding_dim'], adj=adj)
pruning.add_mask(net_gcn)
net_gcn = net_gcn.cuda()
pruning.random_pruning(net_gcn, adj_percent, wei_percent)
adj_spar, wei_spar = pruning.print_sparsity(net_gcn)
for name, param in net_gcn.named_parameters():
if 'mask' in name:
param.requires_grad = False
optimizer = torch.optim.Adam(net_gcn.parameters(),
lr=args['lr'],
weight_decay=args['weight_decay'])
acc_test = 0.0
best_val_acc = {'val_acc': 0, 'epoch': 0, 'test_acc': 0}
for epoch in range(args['total_epoch']):
optimizer.zero_grad()
output = net_gcn(features, adj)
loss = loss_func(output[idx_train], labels[idx_train])
loss.backward()
optimizer.step()
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
acc_val = f1_score(labels[idx_val].cpu().numpy(),
output[idx_val].cpu().numpy().argmax(axis=1),
average='micro')
acc_test = f1_score(labels[idx_test].cpu().numpy(),
output[idx_test].cpu().numpy().argmax(axis=1),
average='micro')
if acc_val > best_val_acc['val_acc']:
best_val_acc['val_acc'] = acc_val
best_val_acc['test_acc'] = acc_test
best_val_acc['epoch'] = epoch
print(
"(Fix Mask) Epoch:[{}] Val:[{:.2f}] Test:[{:.2f}] | Final Val:[{:.2f}] Test:[{:.2f}] at Epoch:[{}]"
.format(epoch, acc_val * 100, acc_test * 100,
best_val_acc['val_acc'] * 100,
best_val_acc['test_acc'] * 100, best_val_acc['epoch']))
return best_val_acc['val_acc'], best_val_acc['test_acc'], best_val_acc[
'epoch'], adj_spar, wei_spar
def run(args, seed):
setup_seed(seed)
adj, features, labels, idx_train, idx_val, idx_test = load_data(args['dataset'])
node_num = features.size()[0]
class_num = labels.numpy().max() + 1
adj = adj.cuda()
features = features.cuda()
labels = labels.cuda()
loss_func = nn.CrossEntropyLoss()
early_stopping = 10
net_gcn = net.net_gcn(embedding_dim=args['embedding_dim'])
net_gcn = net_gcn.cuda()
optimizer = torch.optim.Adam(net_gcn.parameters(), lr=args['lr'], weight_decay=args['weight_decay'])
loss_val = []
for epoch in range(1000):
optimizer.zero_grad()
output = net_gcn(features, adj)
loss = loss_func(output[idx_train], labels[idx_train])
# print('epoch', epoch, 'loss', loss_train.data)
loss.backward()
optimizer.step()
# validation
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
loss_val.append(loss_func(output[idx_val], labels[idx_val]).cpu().numpy())
# print('val acc', f1_score(labels[idx_val].cpu().numpy(), output[idx_val].cpu().numpy().argmax(axis=1), average='micro'))
# early stopping
if epoch > early_stopping and loss_val[-1] > np.mean(loss_val[-(early_stopping+1):-1]):
break
# test
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
acc_val = f1_score(labels[idx_val].cpu().numpy(), output[idx_val].cpu().numpy().argmax(axis=1), average='micro')
acc_test = f1_score(labels[idx_test].cpu().numpy(), output[idx_test].cpu().numpy().argmax(axis=1), average='micro')
return acc_val, acc_test
def run():
setup_seed(11) # 11, 0, 0
adj, features, labels, idx_train, idx_val, idx_test = load_data('cora') # 'cora', 'citeseer', 'pubmed'
adj = adj.cuda()
features = features.cuda()
labels = labels.cuda()
net_gcn = net.net_gcn(embedding_dim=[1433,16,7]) # [1433,16,7], [3703,16,6], [500,16,3]
net_gcn = net_gcn.cuda()
optimizer = torch.optim.Adam(net_gcn.parameters(), lr=0.01, weight_decay=5e-4)
loss_func = nn.CrossEntropyLoss()
loss_val = []
early_stopping = 10
for epoch in range(1000):
optimizer.zero_grad()
output = net_gcn(features, adj)
loss_train = loss_func(output[idx_train], labels[idx_train])
print('epoch', epoch, 'loss', loss_train.data)
loss_train.backward()
optimizer.step()
# validation
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
loss_val.append(loss_func(output[idx_val], labels[idx_val]).cpu().numpy())
print('val acc', f1_score(labels[idx_val].cpu(), output[idx_val].cpu().numpy().argmax(axis=1), average='micro'))
# early stopping
if epoch > early_stopping and loss_val[-1] > np.mean(loss_val[-(early_stopping+1):-1]):
break
# test
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
print('')
print('test acc', f1_score(labels[idx_test].cpu(), output[idx_test].cpu().numpy().argmax(axis=1), average='micro'))
def run(args, seed):
setup_seed(seed)
dataset = args['dataset']
adj, features, labels, idx_train, idx_val, idx_test = load_data(dataset)
adj = adj.cuda()
features = features.cuda()
labels = labels.cuda()
net_gcn = net.net_gcn(embedding_dim=args['embedding_dim'])
net_gcn = net_gcn.cuda()
optimizer = torch.optim.Adam(net_gcn.parameters(), lr=args['lr'], weight_decay=args['weight_decay'])
loss_func = nn.CrossEntropyLoss()
loss_val = []
early_stopping = 10
for epoch in range(1000):
optimizer.zero_grad()
output = net_gcn(features, adj)
loss_train = loss_func(output[idx_train], labels[idx_train])
# print('epoch', epoch, 'loss', loss_train.data)
loss_train.backward()
optimizer.step()
# validation
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
loss_val.append(loss_func(output[idx_val], labels[idx_val]).cpu().numpy())
# print('val acc', f1_score(labels[idx_val].cpu(), output[idx_val].cpu().numpy().argmax(axis=1), average='micro'))
# early stopping
if epoch > early_stopping and loss_val[-1] > np.mean(loss_val[-(early_stopping+1):-1]):
break
# test
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
# print('')
acc = f1_score(labels[idx_test].cpu(), output[idx_test].cpu().numpy().argmax(axis=1), average='micro')
# print('test acc', acc)
# attack
w0 = np.load('./weights/' + dataset + '_w0.npy').transpose()
w1 = np.load('./weights/' + dataset + '_w1.npy').transpose()
adj_raw, features_raw, labels_raw = load_data_raw(dataset)
correct_pred_link = 0
correct_pred_feat = 0
correct_pred_link_feat = 0
n_attack = args['nattack']
for idxt, n in zip(idx_test, range(1000)):
# link
pernode = [idxt]
_, _, adj_per, features_per, _ = graph_attack(adj_raw, features_raw, labels_raw, w0, w1, False, True, pernode, n=n_attack)
features_per, adj_per = preprocess_feat_adj(features_per, adj_per)
with torch.no_grad():
output = net_gcn(features_per, adj_per, val_test=True)[idxt].cpu().numpy().argmax()
if output == labels[idxt].cpu().numpy():
correct_pred_link = correct_pred_link + 1
print(output, labels[idxt].cpu().numpy())
print(correct_pred_link, n + 1)
# feat
pernode = [idxt]
_, _, adj_per, features_per, _ = graph_attack(adj_raw, features_raw, labels_raw, w0, w1, True, False, pernode, n=n_attack)
features_per, adj_per = preprocess_feat_adj(features_per, adj_per)
with torch.no_grad():
output = net_gcn(features_per, adj_per, val_test=True)[idxt].cpu().numpy().argmax()
if output == labels[idxt].cpu().numpy():
correct_pred_feat = correct_pred_feat + 1
print(output, labels[idxt].cpu().numpy())
print(correct_pred_feat, n + 1)
# link & feat
pernode = [idxt]
_, _, adj_per, features_per, _ = graph_attack(adj_raw, features_raw, labels_raw, w0, w1, True, True, pernode, n=n_attack)
features_per, adj_per = preprocess_feat_adj(features_per, adj_per)
with torch.no_grad():
output = net_gcn(features_per, adj_per, val_test=True)[idxt].cpu().numpy().argmax()
if output == labels[idxt].cpu().numpy():
correct_pred_link_feat = correct_pred_link_feat + 1
print(output, labels[idxt].cpu().numpy())
print(correct_pred_link_feat, n + 1)
adv_acc_link = correct_pred_link / 1000
adv_acc_feat = correct_pred_feat / 1000
adv_acc_link_feat = correct_pred_link_feat / 1000
return acc, adv_acc_link, adv_acc_feat, adv_acc_link_feat
def run(args, seed):
setup_seed(seed)
dataset = args['dataset']
adj, features, labels, idx_train, idx_val, idx_test = load_data(dataset)
idx_unlabeled = list(range(len(idx_train), features.size()[0]))
# print(len(idx_train), features.size()[0])
idx_unlabeled = np.random.permutation(idx_unlabeled)
idx_clean = list(idx_unlabeled[:100])
idx_adv = list(idx_unlabeled[100:300])
adj = adj.cuda()
features = features.cuda()
labels = labels.cuda()
net_gcn = net.net_gcn(embedding_dim=args['embedding_dim'])
net_gcn = net_gcn.cuda()
optimizer = torch.optim.Adam(net_gcn.parameters(),
lr=args['lr'],
weight_decay=args['weight_decay'])
loss_func = nn.CrossEntropyLoss()
loss_val = []
early_stopping = 10
for epoch in range(1000):
optimizer.zero_grad()
output = net_gcn(features, adj)
loss_train = loss_func(output[idx_train], labels[idx_train])
# print('epoch', epoch, 'loss', loss_train.data)
loss_train.backward()
optimizer.step()
# validation
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
loss_val.append(
loss_func(output[idx_val], labels[idx_val]).cpu().numpy())
# print('val acc', f1_score(labels[idx_val].cpu(), output[idx_val].cpu().numpy().argmax(axis=1), average='micro'))
# early stopping
if epoch > early_stopping and loss_val[-1] > np.mean(
loss_val[-(early_stopping + 1):-1]):
break
# test
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
# print('')
acc = f1_score(labels[idx_test].cpu(),
output[idx_test].cpu().numpy().argmax(axis=1),
average='micro')
# print('test acc', acc)
#########
# robust training
w0 = np.load('./weights/' + dataset + '_w0.npy').transpose()
w1 = np.load('./weights/' + dataset + '_w1.npy').transpose()
adj_raw, features_raw, _ = load_data_raw(dataset)
pseudo_labels = output.argmax(dim=1).cpu().numpy()
# print(pseudo_labels)
_, _, adj_per, features_per, _ = graph_attack(adj_raw,
features_raw,
pseudo_labels,
w0,
w1,
True,
True,
idx_adv,
n=2)
partition_labels = partition(adj_per, args['partition_num'])
# partition_labels = partition(adj_raw, args['partition_num'])
features_per, adj_per = preprocess_feat_adj(features_per, adj_per)
pseudo_labels = torch.tensor(pseudo_labels).cuda()
net_gcn = net.net_gcn_2task(embedding_dim=args['embedding_dim'],
ss_class_num=args['partition_num'])
net_gcn = net_gcn.cuda()
optimizer = torch.optim.Adam(net_gcn.parameters(),
lr=args['lr'],
weight_decay=args['weight_decay'])
for epoch in range(1000):
optimizer.zero_grad()
output, output_ss = net_gcn(features, features_per, adj, adj_per)
output_adv, _ = net_gcn(features_per, features_per, adj_per, adj_per)
loss_train = loss_func(
output[idx_train],
labels[idx_train]) * args['task_ratio'] + loss_func(
output_ss, partition_labels) * (1 - args['task_ratio'])
loss_adv_1 = loss_func(output_adv[idx_clean], pseudo_labels[idx_clean])
loss_adv_2 = loss_func(output_adv[idx_adv], pseudo_labels[idx_adv])
loss = loss_train + 1 * (loss_adv_1 + loss_adv_2)
# print('epoch', epoch, 'loss', loss_train.data)
loss.backward()
optimizer.step()
# validation
with torch.no_grad():
output, _ = net_gcn(features,
features_per,
adj,
adj_per,
val_test=True)
loss_val.append(
loss_func(output[idx_val], labels[idx_val]).cpu().numpy())
# print('val acc', f1_score(labels[idx_val].cpu(), output[idx_val].cpu().numpy().argmax(axis=1), average='micro'))
# early stopping
if epoch > early_stopping and loss_val[-1] > np.mean(
loss_val[-(early_stopping + 1):-1]):
break
# test
with torch.no_grad():
output, _ = net_gcn(features,
features_per,
adj,
adj_per,
val_test=True)
# print('')
acc = f1_score(labels[idx_test].cpu(),
output[idx_test].cpu().numpy().argmax(axis=1),
average='micro')
# print('test acc', acc)
#########
# attack
w0 = np.load('./weights/' + dataset + '_w0.npy').transpose()
w1 = np.load('./weights/' + dataset + '_w1.npy').transpose()
adj_raw, features_raw, labels_raw = load_data_raw(dataset)
correct_pred_link = 0
correct_pred_feat = 0
correct_pred_link_feat = 0
n_attack = args['nattack']
for idxt, n in zip(idx_test, range(1000)):
# link
pernode = [idxt]
_, _, adj_per, features_per, _ = graph_attack(adj_raw,
features_raw,
labels_raw,
w0,
w1,
False,
True,
pernode,
n=n_attack)
features_per, adj_per = preprocess_feat_adj(features_per, adj_per)
with torch.no_grad():
output, _ = net_gcn(features_per,
features_per,
adj_per,
adj_per,
val_test=True)
output = output[idxt].cpu().numpy().argmax()
if output == labels[idxt].cpu().numpy():
correct_pred_link = correct_pred_link + 1
print(output, labels[idxt].cpu().numpy())
print(correct_pred_link, n + 1)
# feat
pernode = [idxt]
_, _, adj_per, features_per, _ = graph_attack(adj_raw,
features_raw,
labels_raw,
w0,
w1,
True,
False,
pernode,
n=n_attack)
features_per, adj_per = preprocess_feat_adj(features_per, adj_per)
with torch.no_grad():
output, _ = net_gcn(features_per,
features_per,
adj_per,
adj_per,
val_test=True)
output = output[idxt].cpu().numpy().argmax()
if output == labels[idxt].cpu().numpy():
correct_pred_feat = correct_pred_feat + 1
print(output, labels[idxt].cpu().numpy())
print(correct_pred_feat, n + 1)
# link feat
pernode = [idxt]
_, _, adj_per, features_per, _ = graph_attack(adj_raw,
features_raw,
labels_raw,
w0,
w1,
True,
True,
pernode,
n=n_attack)
features_per, adj_per = preprocess_feat_adj(features_per, adj_per)
with torch.no_grad():
output, _ = net_gcn(features_per,
features_per,
adj_per,
adj_per,
val_test=True)
output = output[idxt].cpu().numpy().argmax()
if output == labels[idxt].cpu().numpy():
correct_pred_link_feat = correct_pred_link_feat + 1
print(output, labels[idxt].cpu().numpy())
print(correct_pred_link_feat, n + 1)
adv_acc_link = correct_pred_link / 1000
adv_acc_feat = correct_pred_feat / 1000
adv_acc_link_feat = correct_pred_link_feat / 1000
return acc, adv_acc_link, adv_acc_feat, adv_acc_link_feat
def run_get_mask(args, seed, imp_num, rewind_weight_mask=None):
pruning.setup_seed(seed)
adj, features, labels, idx_train, idx_val, idx_test = load_data(
args['dataset'])
# adj = coo_matrix(adj)
# adj_dict = {}
# adj_dict['adj'] = adj
# torch.save(adj_dict, "./adjs/pubmed/original.pt")
# pdb.set_trace()
node_num = features.size()[0]
class_num = labels.numpy().max() + 1
adj = adj.cuda()
features = features.cuda()
labels = labels.cuda()
loss_func = nn.CrossEntropyLoss()
net_gcn = net.net_gcn(embedding_dim=args['embedding_dim'], adj=adj)
pruning.add_mask(net_gcn)
net_gcn = net_gcn.cuda()
if args['weight_dir']:
print("load : {}".format(args['weight_dir']))
encoder_weight = {}
cl_ckpt = torch.load(args['weight_dir'], map_location='cuda')
encoder_weight['weight_orig_weight'] = cl_ckpt['gcn.fc.weight']
ori_state_dict = net_gcn.net_layer[0].state_dict()
ori_state_dict.update(encoder_weight)
net_gcn.net_layer[0].load_state_dict(ori_state_dict)
if rewind_weight_mask:
net_gcn.load_state_dict(rewind_weight_mask)
pruning.soft_mask_init(net_gcn, args['init_soft_mask_type'], seed)
adj_spar, wei_spar = pruning.print_sparsity(net_gcn)
else:
pruning.soft_mask_init(net_gcn, args['init_soft_mask_type'], seed)
optimizer = torch.optim.Adam(net_gcn.parameters(),
lr=args['lr'],
weight_decay=args['weight_decay'])
acc_test = 0.0
best_val_acc = {'val_acc': 0, 'epoch': 0, 'test_acc': 0}
rewind_weight = copy.deepcopy(net_gcn.state_dict())
for epoch in range(args['mask_epoch']):
optimizer.zero_grad()
output = net_gcn(features, adj)
loss = loss_func(output[idx_train], labels[idx_train])
loss.backward()
pruning.subgradient_update_mask(net_gcn, args) # l1 norm
optimizer.step()
with torch.no_grad():
output = net_gcn(features, adj, val_test=True)
acc_val = f1_score(labels[idx_val].cpu().numpy(),
output[idx_val].cpu().numpy().argmax(axis=1),
average='micro')
acc_test = f1_score(labels[idx_test].cpu().numpy(),
output[idx_test].cpu().numpy().argmax(axis=1),
average='micro')
if acc_val > best_val_acc['val_acc']:
best_val_acc['test_acc'] = acc_test
best_val_acc['val_acc'] = acc_val
best_val_acc['epoch'] = epoch
best_epoch_mask = pruning.get_final_mask_epoch(
net_gcn,
adj_percent=args['pruning_percent_adj'],
wei_percent=args['pruning_percent_wei'])
print(
"(Get Mask) Epoch:[{}] Val:[{:.2f}] Test:[{:.2f}] | Best Val:[{:.2f}] Test:[{:.2f}] at Epoch:[{}]"
.format(epoch, acc_val * 100, acc_test * 100,
best_val_acc['val_acc'] * 100,
best_val_acc['test_acc'] * 100, best_val_acc['epoch']))
return best_epoch_mask, rewind_weight