def test(test_adj, test_fea, idx_test, labels):
unsupervised_model.eval()
classifier_model.eval()
if sampler.learning_type=='inductive':
unsupervised_model.cpu()
classifier_model.cpu()
labels = labels.cpu()
# construct g from test adj
if sampler.learning_type=='inductive':
test_edges = test_adj._indices().data.numpy()
else:
test_edges = test_adj._indices().data.cpu().numpy()
test_edges = sp.coo_matrix((np.ones(test_edges.shape[1]),
(test_edges[0], test_edges[1])),
shape=(test_adj.shape[0], test_adj.shape[0]),
dtype=np.float32)
test_g = nx.from_scipy_sparse_matrix(test_edges, create_using=nx.DiGraph())
test_g = DGLGraph(test_g)
feats = unsupervised_model(test_fea, test_g)
output = classifier_model(feats)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def test(test_adj, test_fea):
model.eval()
output = model(test_fea, test_adj)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:", "loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
return (loss_test.item(), acc_test.item())
def test_sampling(model, test_g, val_batch_size):
model.eval()
output = model.inference(test_g, test_g.ndata['features'], val_batch_size,
'cpu')
loss_test = F.nll_loss(output[idx_test.cpu()], labels[idx_test].cpu())
acc_test = accuracy(output[idx_test.cpu()], labels[idx_test].cpu())
auc_test = roc_auc_compute_fn(output[idx_test.cpu()],
labels[idx_test].cpu())
if args.debug:
print("Test set results:", "loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def test(test_adj, test_fea):
unsupervised_model.eval()
classifier_model.eval()
feats = unsupervised_model(test_fea, test_adj)
output = classifier_model(feats)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def test_sampling(test_g, val_batch_size):
unsupervised_model.eval()
classifier_model.eval()
feats = unsupervised_model.inference(test_g, test_g.ndata['features'], val_batch_size, 'cpu')
output = classifier_model(feats[idx_test.cpu()].cuda())
loss_test = F.nll_loss(output, labels[idx_test])
acc_test = accuracy(output, labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def test_full(model, test_g, idx_test, labels):
model.eval()
if sampler.dataset in ['coauthor_phy']:
model.cpu()
labels = labels.cpu()
output = model(test_g.ndata['features'], test_g)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.debug:
print("Test set results:", "loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def test(test_adj, test_fea):
model.eval()
# output = model(test_fea, test_adj)
output, embeddings = model.myforward(test_fea, test_adj)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if args.pca:
args.ssl += '~PCA'
if args.write_res:
probs = torch.exp(output)
np.save(f'preds/{args.dataset}_{args.seed}_pred.npy', probs.detach().cpu().numpy())
print("Test set results:",
"loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())
def test(test_adj, test_fea):
model.eval()
# output = model(test_fea, test_adj)
output, embeddings = model.myforward(test_fea, test_adj)
# import ipdb
# ipdb.set_trace()
# loss_ssl = ssl_agent.make_loss(embeddings)
# print(loss_ssl)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
auc_test = roc_auc_compute_fn(output[idx_test], labels[idx_test])
if False:
# if not args.write_json:
from utils import get_pairwise_sim
labelled_sim = get_pairwise_sim(embeddings[idx_train]).item()
unlabelled_sim = get_pairwise_sim(embeddings[np.union1d(
idx_test.cpu(), idx_val.cpu())]).item()
all_sim = labelled_sim
# all_sim = get_pairwise_sim(embeddings)
print(
"Labeled Node Similarity: {:.4f}".format(labelled_sim.item()),
"Unlabeled Node Similarity: {:.4f}".format(unlabelled_sim.item()),
'Total Similarity: {:.4f}'.format(all_sim.item()))
else:
labelled_sim = 0
unlabelled_sim = 0
all_sim = 0
if args.finetune:
args.load_pretrain = 3
if args.pca:
args.ssl += '~PCA'
if args.write_res:
np.save(f'preds/{args.dataset}_{args.seed}_pred.npy',
output.detach().cpu().numpy())
if args.write_json:
import json
nlayers = args.nhiddenlayer * args.nbaseblocklayer + 2
res = {
'labelled_sim': labelled_sim,
'unlabelled_sim': unlabelled_sim,
'all_sim': all_sim,
'test_loss': loss_test.item(),
'test_acc': acc_test.item(),
'loss_train': loss_train[epoch],
'loss_ssl': loss_ssl[epoch]
}
# with open('results_ssl_norm_glorot/{0}_{1}-{2}_{3}_{4}.json'.format(args.dataset, args.type, args.ssl, nlayers, args.seed), 'w') as f:
# with open('results-pretrain/{0}_{1}-{2}_{3}_{4}.json'.format(args.dataset, args.label_rate, args.ssl, args.load_pretrain, args.seed), 'w') as f:
# with open('results-nbm/{0}_{1}-{2}_{3}_{4}.json'.format(args.dataset, args.label_rate, args.ssl, args.load_pretrain, args.seed), 'w') as f:
# with open('results-baseline/{0}_{1}-{2}_{3}_{4}.json'.format(args.dataset, args.label_rate, args.ssl, args.load_pretrain, args.seed), 'w') as f:
# with open('results-pca/{0}_{1}-{2}_{3}_{4}.json'.format(args.dataset, args.label_rate, args.ssl, args.load_pretrain, args.seed), 'w') as f:
# with open('results-val/{0}_{1}-{2}_{3}_{4}.json'.format(args.dataset, args.label_rate, args.ssl, args.load_pretrain, args.seed), 'w') as f:
# with open('results-new-seeds/{0}_{1}-{2}_{3}_{4}.json'.format(args.dataset, args.label_rate, args.ssl, args.load_pretrain, args.seed), 'w') as f:
# with open('results-3layers/validation/{0}_{1}-{2}_{3}_{4}.json'.format(args.dataset, args.label_rate, args.ssl, args.load_pretrain, args.seed), 'w') as f:
with open(
'results-3layers/train_size/{0}_{1}-{2}_{3}_{4}.json'.format(
args.dataset, args.train_size, args.ssl,
args.load_pretrain, args.seed), 'w') as f:
json.dump(res, f)
print("Test set results:", "loss= {:.4f}".format(loss_test.item()),
"auc= {:.4f}".format(auc_test),
"accuracy= {:.4f}".format(acc_test.item()))
print("accuracy=%.5f" % (acc_test.item()))
return (loss_test.item(), acc_test.item())