def subgraph_gen(hg, label_idx, neighbours=[10, 10]):
"""
Subgraph sampling by graphsage_sampling
"""
nt = hg.node_types
drugs_idx = np.where(nt == 'drug')[0]
layer1_neighs, layer1_eids = graphsage_sampling(hg,
drugs_idx,
num_neighbours = neighbours[0],
etype='dti')
layer2_neighs, layer2_eids = graphsage_sampling(hg,
layer1_neighs,
num_neighbours = neighbours[1],
etype='ppi')
sub_nodes = drugs_idx.tolist() + layer1_neighs + layer2_neighs
sub_nodes_reidx = dict(zip(sub_nodes, range(len(sub_nodes))))
label_mat = np.zeros((len(sub_nodes), len(sub_nodes))).astype('float32')
for p in hg['dds'].edges.tolist():
label_mat[sub_nodes_reidx[p[0]], sub_nodes_reidx[p[1]]] = label_idx[tuple(p)]
sub_eids = {}
sub_eids['dds'] = [(sub_nodes_reidx[src], sub_nodes_reidx[dst]) for (src, dst) in hg['dds'].edges.tolist()]
sub_eids['dti'] = [(sub_nodes_reidx[src], sub_nodes_reidx[dst]) for (src, dst) in layer1_eids]
sub_eids['dti'] += [(sub_nodes_reidx[dst], sub_nodes_reidx[src]) for (src, dst) in layer1_eids]
sub_eids['ppi'] = [(sub_nodes_reidx[src], sub_nodes_reidx[dst]) for (src, dst) in layer2_eids]
sub_eids['ppi'] += [(sub_nodes_reidx[dst], sub_nodes_reidx[src]) for (src, dst) in layer2_eids]
sub_nodes_feat = hg['dds'].node_feat['features'][sub_nodes, :]
sub_graph = pgl.HeterGraph(edges=sub_eids, num_nodes=len(sub_nodes), node_feat={'features': sub_nodes_feat})
return {'sub_graph': (sub_graph, len(sub_nodes), sub_eids, sub_nodes_feat, label_mat)}
def test_tensor(self):
np.random.seed(1)
dim = 4
num_nodes = 15
edges = {}
# for test no successor
edges['c2p'] = [(1, 4), (0, 5), (1, 9), (1, 8), (2, 8), (2, 5), (3, 6),
(3, 7), (3, 4), (3, 8)]
edges['p2c'] = [(v, u) for u, v in edges['c2p']]
edges['p2a'] = [(4, 10), (4, 11), (4, 12), (4, 14), (4, 13), (6, 12),
(6, 11), (6, 14), (7, 12), (7, 11), (8, 14), (9, 10)]
edges['a2p'] = [(v, u) for u, v in edges['p2a']]
node_types = ['c' for _ in range(4)] + \
['p' for _ in range(6)] + \
['a' for _ in range(5)]
node_types = [(i, t) for i, t in enumerate(node_types)]
nfeat = {'nfeat': np.random.randn(num_nodes, dim)}
efeat = {}
for etype, _edges in edges.items():
efeat[etype] = {'efeat': np.random.randn(len(_edges), dim)}
hg = pgl.HeterGraph(edges=edges,
node_types=node_types,
node_feat=nfeat,
edge_feat=efeat)
# inplace
new_hg = hg.tensor(inplace=False)
self.assertNotIsInstance(hg.node_feat['nfeat'], paddle.Tensor)
self.assertNotIsInstance(hg.edge_feat['a2p']['efeat'], paddle.Tensor)
self.assertIsInstance(new_hg.node_feat['nfeat'], paddle.Tensor)
self.assertIsInstance(new_hg.edge_feat['a2p']['efeat'], paddle.Tensor)
self.assertIsInstance(new_hg.num_nodes, paddle.Tensor)
hg.tensor(inplace=True)
self.assertIsInstance(hg.node_feat['nfeat'], paddle.Tensor)
self.assertIsInstance(hg.edge_feat['a2p']['efeat'], paddle.Tensor)
new_hg = hg.numpy(inplace=False)
self.assertIsInstance(new_hg.node_feat['nfeat'], np.ndarray)
self.assertIsInstance(new_hg.edge_feat['a2p']['efeat'], np.ndarray)
self.assertIsInstance(hg.node_feat['nfeat'], paddle.Tensor)
self.assertIsInstance(hg.edge_feat['a2p']['efeat'], paddle.Tensor)
hg.numpy(inplace=True)
self.assertIsInstance(hg.node_feat['nfeat'], np.ndarray)
self.assertIsInstance(hg.edge_feat['a2p']['efeat'], np.ndarray)
def build_heter_graph(data_path, num_nodes):
edges = {}
idx = 0
for filename in glob.glob(os.path.join(data_path, '*')):
try:
e = pd.read_csv(filename, header=None, sep="\t").values
edges['etype%s' % idx] = e
idx += 1
except Exception as e:
log.info(e)
continue
node_types = [(i, "n") for i in range(num_nodes)]
hg = pgl.HeterGraph(edges=edges, node_types=node_types)
return hg
def test_build_hetergraph(self):
np.random.seed(1)
dim = 4
num_nodes = 15
edges = {}
# for test no successor
edges['c2p'] = [(1, 4), (0, 5), (1, 9), (1, 8), (2, 8), (2, 5), (3, 6),
(3, 7), (3, 4), (3, 8)]
edges['p2c'] = [(v, u) for u, v in edges['c2p']]
edges['p2a'] = [(4, 10), (4, 11), (4, 12), (4, 14), (4, 13), (6, 12),
(6, 11), (6, 14), (7, 12), (7, 11), (8, 14), (9, 10)]
edges['a2p'] = [(v, u) for u, v in edges['p2a']]
node_types = ['c' for _ in range(4)] + \
['p' for _ in range(6)] + \
['a' for _ in range(5)]
node_types = [(i, t) for i, t in enumerate(node_types)]
nfeat = {'nfeat': np.random.randn(num_nodes, dim)}
efeat = {}
for etype, _edges in edges.items():
efeat[etype] = {'efeat': np.random.randn(len(_edges), dim)}
hg = pgl.HeterGraph(edges=edges,
node_types=node_types,
node_feat=nfeat,
edge_feat=efeat)
self.assertFalse(hg.is_tensor())
self.assertEqual(hg.indegree(5), [2])
self.assertEqual(hg.outdegree(4), [7])
self.assertEqual(hg.outdegree(4, 'c2p'), [0])
self.assertEqual(hg.successor('c2p', [4]).tolist(), [[]])
self.assertEqual(
hg.predecessor('a2p', [4]).tolist(), [[10, 11, 12, 14, 13]])
print()
# print(hg.predecessor('a2p', [4]))
for batch in hg.node_batch_iter(3):
break
def test_build_tensor_hetergraph(self):
np.random.seed(1)
dim = 4
num_nodes = paddle.to_tensor(15)
edges = {}
# for test no successor
c2p = [(1, 4), (0, 5), (1, 9), (1, 8), (2, 8), (2, 5), (3, 6), (3, 7),
(3, 4), (3, 8)]
edges['c2p'] = paddle.to_tensor(np.array(c2p))
p2c = [(v, u) for u, v in c2p]
edges['p2c'] = paddle.to_tensor(np.array(p2c))
p2a = [(4, 10), (4, 11), (4, 12), (4, 14), (4, 13), (6, 12), (6, 11),
(6, 14), (7, 12), (7, 11), (8, 14), (9, 10)]
edges['p2a'] = paddle.to_tensor(np.array(p2a))
a2p = [(v, u) for u, v in p2a]
edges['a2p'] = paddle.to_tensor(np.array(a2p))
node_types = ['c' for _ in range(4)] + \
['p' for _ in range(6)] + \
['a' for _ in range(5)]
node_types = [(i, t) for i, t in enumerate(node_types)]
hg = pgl.HeterGraph(edges=edges, node_types=node_types)
self.assertTrue(hg.is_tensor())
print()
self.assertEqual(
hg.indegree(paddle.to_tensor(5)).numpy(), np.array([2]))
self.assertEqual(
hg.indegree(paddle.to_tensor(5), 'c2p').numpy(), np.array([2]))
self.assertEqual(
hg.outdegree(paddle.to_tensor(4)).numpy(), np.array([7]))
self.assertEqual(
hg.outdegree(paddle.to_tensor(4), 'p2a').numpy(), np.array([5]))
# print(hg.outdegree(paddle.to_tensor(4)).numpy())
for batch in hg.node_batch_iter(3, n_type='c'):
break
def test_dump_and_load(self):
np.random.seed(1)
dim = 4
num_nodes = 15
edges = {}
# for test no successor
edges['c2p'] = [(1, 4), (0, 5), (1, 9), (1, 8), (2, 8), (2, 5), (3, 6),
(3, 7), (3, 4), (3, 8)]
edges['p2c'] = [(v, u) for u, v in edges['c2p']]
edges['p2a'] = [(4, 10), (4, 11), (4, 12), (4, 14), (4, 13), (6, 12),
(6, 11), (6, 14), (7, 12), (7, 11), (8, 14), (9, 10)]
edges['a2p'] = [(v, u) for u, v in edges['p2a']]
node_types = ['c' for _ in range(4)] + \
['p' for _ in range(6)] + \
['a' for _ in range(5)]
node_types = [(i, t) for i, t in enumerate(node_types)]
nfeat = {'nfeat': np.random.randn(num_nodes, dim)}
efeat = {}
for etype, _edges in edges.items():
efeat[etype] = {'efeat': np.random.randn(len(_edges), dim)}
hg = pgl.HeterGraph(edges=edges,
node_types=node_types,
node_feat=nfeat,
edge_feat=efeat)
path = "./tmp"
hg.dump(path, indegree=True)
hg2 = pgl.HeterGraph.load(path)
self.assertEqual(hg.num_nodes, hg2.num_nodes)
del hg
del hg2
shutil.rmtree(path)
def collate_fn(self, ddi_data, dti_data, ppi_data, features):
"""Aggregate all needed nodes into a Hetrogenous graph"""
drug_feat = pd.read_csv(features, index_col=0)
drug_feat = drug_feat[~drug_feat.index.duplicated()]
drug_feat = drug_feat.fillna(0)
drug_feat.replace([np.inf, -np.inf], 0, inplace=True)
nm = StandardScaler()
scaled_feat = pd.DataFrame(nm.fit_transform(drug_feat))
scaled_feat = scaled_feat.fillna(0)
scaled_feat.index = drug_feat.index
edges = {'dds': [], 'dti': [], 'ppi': []}
ddi_nn, ddi_nodes = num_nodes_stat(ddi_data)
selected_drugs_feat = scaled_feat[scaled_feat.index.isin(ddi_nodes)]
ddi_nodes = set(selected_drugs_feat.index)
total_nodes = set()
label = {}
for d in ddi_data:
if d['pair'][0] in ddi_nodes and d['pair'][1] in ddi_nodes:
edges['dds'].append((d['pair'][0], d['pair'][1]))
edges['dds'].append((d['pair'][1], d['pair'][0]))
total_nodes.add(d['pair'][0])
total_nodes.add(d['pair'][1])
label[d['pair'][0], d['pair'][1]] = d['label']
label[d['pair'][1], d['pair'][0]] = d['label']
for d in dti_data:
if d['pair'][0] in ddi_nodes:
edges['dti'].append((d['pair'][0], d['pair'][1]))
edges['dti'].append((d['pair'][1], d['pair'][0]))
total_nodes.add(d['pair'][0])
total_nodes.add(d['pair'][1])
for d in ppi_data:
edges['ppi'].append((d['pair'][0], d['pair'][1]))
edges['ppi'].append((d['pair'][1], d['pair'][0]))
total_nodes.add(d['pair'][0])
total_nodes.add(d['pair'][1])
num_nodes = len(total_nodes)
nodes_dict = dict(zip(total_nodes, range(num_nodes)))
node_feat = np.zeros((num_nodes, 2325)).astype('float32')
selected_drugs_feat.index = [
nodes_dict[x] for x in selected_drugs_feat.index
]
for d in selected_drugs_feat.index:
node_feat[d, :] = selected_drugs_feat.loc[d, :].values.astype(
'float32')
node_feats = {'features': node_feat}
ek = {'dds': [], 'dti': [], 'ppi': []}
for edge_type in edges.keys():
for p in edges[edge_type]:
p1, p2 = nodes_dict[p[0]], nodes_dict[p[1]]
ek[edge_type].append((p1, p2))
node_types = []
for m in nodes_dict.keys():
if m.startswith('CID'):
node_types.append((nodes_dict[m], 'drug'))
else:
node_types.append((nodes_dict[m], 'protein'))
hg = pgl.HeterGraph(num_nodes=num_nodes,
edges=ek,
node_types=node_types,
node_feat=node_feats)
label_idx = {}
for key in label.keys():
label_idx[(nodes_dict[key[0]], nodes_dict[key[1]])] = label[key]
return {'rt': (hg, nodes_dict, label, label_idx)}