def __init__(self, adj, features, labels, idx_train, idx_test, args):
"""
idx_train: labeled data
idx_test: unlabeled data
"""
self.args = args
self.adj = adj.tocsr()
if args.dataset != 'reddit':
self.adj_two_hop = adj.dot(adj)
self.adj_two_hop.setdiag(0)
self.adj_two_hop.eliminate_zeros()
# self.graph = nx.from_scipy_sparse_matrix(adj)
self.pseudo_labels = np.zeros((adj.shape[0], labels.shape[1]))
# load_data = os.path.exists(f'preds/ICA_probs_{args.dataset}_{args.seed}.npy')
load_data = os.path.exists(
f'preds/ICA_probs_{args.train_size}_{args.dataset}_{args.seed}.npy'
)
print('if loading: ', load_data)
if not load_data:
st = time.time()
if args.dataset != 'cora':
features[features != 0] = 1
classifier = 'sklearn.linear_model.LogisticRegression'
aggregate = 'count' # choices=['count', 'prop'], help='Aggregation operator'
graph, domain_labels = build_graph(adj, features, labels)
y_true = [graph.node_list[t].label for t in idx_test]
local_clf = LocalClassifier(classifier)
agg = pick_aggregator(aggregate, domain_labels)
relational_clf = RelationalClassifier(classifier, agg)
ica = ICA(local_clf,
relational_clf,
bootstrap=True,
max_iteration=10)
ica.fit(graph, idx_train)
conditional_node_to_label_map = create_map(graph, idx_train)
eval_idx = np.setdiff1d(range(adj.shape[0]), idx_train)
ica_predict, probs = ica.predict(graph, eval_idx, idx_test,
conditional_node_to_label_map)
ica_accuracy = accuracy_score(y_true, ica_predict)
print('Acc: ' + str(ica_accuracy))
print('optimization consumes %s s' % (time.time() - st))
# self.ica_predict = np.array([int(x[1:]) for x in ica_predict])
dict_pred = {x: int(y[1:]) for x, y in zip(idx_test, ica_predict)}
dict_train = {x: labels.argmax(1)[x] for x in idx_train}
dict_pred.update(dict_train)
concated = sorted(dict_pred.items(), key=lambda x: x[0])
self.probs = np.vstack((labels[idx_train], probs))
self.concated = np.array([y for x, y in concated])
np.save(
f'preds/ICA_probs_{args.train_size}_{args.dataset}_{args.seed}.npy',
self.probs)
np.save(
f'preds/ICA_preds_{args.train_size}_{args.dataset}_{args.seed}.npy',
self.concated)
else:
print('loading probs/preds...')
# self.probs = np.load(f'ICA_probs_{args.dataset}_{args.seed}.npy')
# self.concated = np.load(f'ICA_preds_{args.dataset}_{args.seed}.npy')
# self.probs = np.load(f'ICA_probs_{args.dataset}_10.npy')
# self.concated = np.load(f'ICA_preds_{args.dataset}_10.npy')
self.probs = np.load(
f'preds/ICA_probs_{args.train_size}_{args.dataset}_{args.seed}.npy'
)
self.concated = np.load(
f'preds/ICA_preds_{args.train_size}_{args.dataset}_{args.seed}.npy'
)
# self.probs = np.load(f'preds/{args.dataset}_{args.seed}_pred.npy')
# self.concated = self.probs.argmax(1)
# self.concated[idx_train] = labels.argmax(1)[idx_train]
print('Acc: ',
(self.concated == labels.argmax(1))[idx_test].sum() /
len(idx_test))
import ipdb
ipdb.set_trace()
# labels = encode_onehot(labels)
#
features = normalize_feature(features)
# idx_train = np.arange(120)
# idx_val = idx_train
# # idx_test = np.arange(120, adj.shape[0])
# idx_test = np.arange(120+500, 120+500+1000)
#
# import ipdb
# ipdb.set_trace()
graph, domain_labels = build_graph(adj, features, labels)
# train / test splits
train = idx_train
if args.validation:
test = idx_val
else:
test = idx_test
eval_idx = np.setdiff1d(range(adj.shape[0]), idx_train)
# run training
ica_accuracies = list()
for run in range(args.num_trials):
t_begin = time.time()