-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathppr.py
227 lines (184 loc) · 8.1 KB
/
ppr.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
import argparse
import numba
import numpy as np
import torch
import logging
import os
import scipy.sparse as sp
import dgl
from dgl.data import load_data, CoraGraphDataset, CiteseerGraphDataset, PubmedGraphDataset
from ogb.nodeproppred import DglNodePropPredDataset
def collect_topk_ppr(graph, nodes, topk, alpha, epsilon):
if torch.is_tensor(nodes):
nodes = nodes.numpy()
row, col = graph.edges()
row = row.numpy()
col = col.numpy()
num_nodes = graph.num_nodes()
neighbors = build_topk_ppr((row, col), alpha, epsilon, nodes, topk, num_nodes=num_nodes)
return neighbors
# modified
@numba.njit(cache=True, locals={"_val": numba.float32, "res": numba.float32, "res_vnode": numba.float32})
def _calc_ppr_node(inode, indptr, indices, deg, alpha, epsilon):
alpha_eps = alpha * epsilon
f32_0 = numba.float32(0)
p = {inode: f32_0}
r = {inode: numba.float32(1.0)}
q = [inode]
while len(q) > 0:
unode = q.pop()
res = r[unode] if unode in r else f32_0
if unode in p:
p[unode] += alpha * res
else:
p[unode] = alpha * res
r[unode] = f32_0
for vnode in indices[indptr[unode]: indptr[unode + 1]]:
_val = (1 - alpha) * res / deg[unode]
if vnode in r:
r[vnode] += _val
else:
r[vnode] = _val
res_vnode = r[vnode] if vnode in r else f32_0
if res_vnode >= alpha_eps * deg[vnode]:
if vnode not in q:
q.append(vnode)
return list(p.keys()), list(p.values())
@numba.njit(cache=True, parallel=True)
def calc_ppr_topk_parallel(indptr, indices, deg, alpha, epsilon, nodes, topk, mode="transformer"):
js = [np.zeros(0, dtype=np.int64)] * len(nodes)
vals = [np.zeros(0, dtype=np.float32)] * len(nodes)
self_val = np.array([1.0]).astype(vals[0].dtype)
mask_val = np.array([-1], dtype=js[0].dtype)
for i in numba.prange(len(nodes)):
j, val = _calc_ppr_node(nodes[i], indptr, indices, deg, alpha, epsilon)
j_np, val_np = np.array(j), np.array(val)
idx_topk = np.argsort(val_np)[-topk:][::-1]
js[i] = j_np[idx_topk]
vals[i] = val_np[idx_topk]
mask = (js[i] != nodes[i])
js[i] = js[i][mask]
vals[i] = vals[i][mask]
# if mode == "transformer" and len(js[i]) < topk:
# diff = topk - len(js[i])
# js[i] = np.concatenate((np.array([nodes[i]], dtype=js[0].dtype), js[i], mask_val.repeat(diff)))
# vals[i] = np.concatenate((self_val, vals[i], self_val.repeat(diff)))
# else:
js[i] = np.concatenate((np.array([nodes[i]]), js[i]))
vals[i] = np.concatenate((self_val, vals[i]))
return js, vals
@numba.njit(cache=True, parallel=True)
def add_padding(js, vals, topk, nodes):
v_type = vals[0].dtype
self_val = np.array([1.0], dtype=v_type)
mask_val = np.array([-1], dtype=js[0].dtype)
for i in numba.prange(len(nodes)):
if len(js[i]) < topk:
diff = topk - len(js[i])
js[i] = np.concatenate((np.array([i], dtype=js[0].dtype), js[i], mask_val.repeat(diff)))
vals[i] = np.concatenate((self_val, vals[i], self_val.reshape(diff)))
return js, vals
# def pos_embedding(neighbors, adj, hidden_size):
# nodes = neighbors
# mx = adj[nodes, :][:, nodes]
# emb = get_undirected_graph_positional_embedding(mx, hidden_size)
# return emb
def ppr_ego_graphs(adj_matrix, alpha, epsilon, nodes, topk, mode="transformer"):
out_degree = np.sum(adj_matrix > 0, axis=1).A1
neighbors, weights = calc_ppr_topk_parallel(
adj_matrix.indptr, adj_matrix.indices, out_degree, numba.float32(alpha), numba.float32(epsilon), nodes, topk, mode
)
for i in range(len(weights[0]) - 1):
assert weights[0][i] >= weights[0][i+1]
return neighbors, weights
def build_topk_ppr(edge_index, alpha, epsilon, nodes, topk, mode="transformer", num_nodes=-1):
assert num_nodes > 0
val = np.ones(edge_index[0].shape[0])
adj_matrix = sp.csr_matrix((val, (edge_index[0], edge_index[1])), shape=(num_nodes, num_nodes))
neighbors, weights = ppr_ego_graphs(adj_matrix, alpha, epsilon, nodes, topk, mode)
return neighbors
def lc_prepare_ego_graphs(split_idx,dataset_name, graph, topk, alpha, epsilon, nodes=None, save_dir="./lc_preprocessed"):
save_dir = "./data/ego-graphs"
os.makedirs(save_dir, exist_ok=True)
os.makedirs(os.path.join(save_dir, dataset_name), exist_ok=True)
if nodes is not None:
suffix = "ppr_eval"
else:
suffix = "ppr"
save_path = os.path.join(save_dir, dataset_name, f"k-{topk}_alpha-{alpha}_eps_{epsilon}_{suffix}.pt")
if nodes is None:
#split_idx = dataset.get_idx_split()
if dataset_name not in ["cora", "citeseer", "pubmed","film"]:
train_idx, val_idx, test_idx = split_idx["train"], split_idx["valid"], split_idx["test"]
num_used_nodes = train_idx.shape[0] + val_idx.shape[0] + test_idx.shape[0]
else:
num_used_nodes = graph.num_nodes()
if num_used_nodes < graph.num_nodes():
logging.info("use 10 times unlabelled nodes ..")
labelled_nodes = torch.cat([train_idx, val_idx, test_idx], dim=0)
sample_nodes = np.random.randint(0, int(graph.num_nodes()), 10*num_used_nodes)
sample_nodes = np.setdiff1d(sample_nodes, labelled_nodes)
sample_nodes = torch.from_numpy(sample_nodes)
nodes = torch.cat([labelled_nodes, sample_nodes], dim=0)
else:
logging.info("use full nodes")
nodes = torch.arange(graph.num_nodes()).numpy()
else:
if not torch.is_tensor(nodes):
nodes = torch.tensor(nodes)
if os.path.exists(save_path):
logging.info("--- start loading graphs ---")
ego_graph_nodes = torch.load(save_path)
else:
logging.info("--- computing ppr ---")
ego_graph_nodes = collect_topk_ppr(graph, nodes, topk, alpha, epsilon)
logging.info("--- finish computing ppr ---")
torch.save(ego_graph_nodes, save_path)
logging.info("--- ego-graph loaded ---")
avg_size = np.mean([x.shape[0] for x in ego_graph_nodes])
logging.info(f"Average ego-graph size: {avg_size:.2f}")
logging.info("--- preprocessing done ---")
return ego_graph_nodes
def preprocess(graph):
# make bidirected
if "feat" in graph.ndata:
feat = graph.ndata["feat"]
else:
feat = None
graph = dgl.to_bidirected(graph)
if feat is not None:
graph.ndata["feat"] = feat
# add self-loop
graph = graph.remove_self_loop().add_self_loop()
# graph.create_formats_()
return graph
def load_dataset(dataset_name):
start_path = ""
if dataset_name.startswith("ogbn"):
dataset = DglNodePropPredDataset(dataset_name,root=os.path.join(start_path, "dataset"))
graph, label = dataset[0]
if "year" in graph.ndata:
del graph.ndata["year"]
if not graph.is_multigraph:
graph = preprocess(graph)
# graph = graph.remove_self_loop().add_self_loop()
split_idx = dataset.get_idx_split()
label = label.view(-1)
feats = graph.ndata.pop("feat")
return feats, graph, label, split_idx
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='')
parser.add_argument('--task', type=str, default='saint')
parser.add_argument('--dataset', type=str, default='flickr')
parser.add_argument('--ego_size', type=int, default=1024)
parser.add_argument('--hidden_size', type=int, default=128)
parser.add_argument('--num_iter', type=int, default=1000)
parser.add_argument('--log_steps', type=int, default=10000)
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--method', type=str, default='acl')
parser.add_argument('--num_workers', type=int, default=50)
args = parser.parse_args()
print(args)
np.random.seed(args.seed)
feat, graph, label, split_idx = load_dataset(args.dataset)
ego_graph_nodes = lc_prepare_ego_graphs(split_idx, args.dataset, graph, args.ego_size, alpha=0.75, epsilon=0.000001)