SauravMaheshkar commited on
Commit
b80ed00
1 Parent(s): c84c63c

feat: add link generation script

Browse files
Files changed (1) hide show
  1. link_gen.py +108 -0
link_gen.py ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import dgl
2
+ from dgl.data import AmazonCoBuyPhotoDataset
3
+ import torch
4
+ import pickle
5
+ from copy import deepcopy
6
+ import scipy.sparse as sp
7
+ import numpy as np
8
+ import os
9
+
10
+
11
+ def mask_test_edges(adj_orig, val_frac, test_frac):
12
+
13
+ # Remove diagonal elements
14
+ adj = deepcopy(adj_orig)
15
+ # set diag as all zero
16
+ adj.setdiag(0)
17
+ adj.eliminate_zeros()
18
+ # Check that diag is zero:
19
+ # assert np.diag(adj.todense()).sum() == 0
20
+
21
+ adj_triu = sp.triu(adj, 1)
22
+ edges = sparse_to_tuple(adj_triu)[0]
23
+ num_test = int(np.floor(edges.shape[0] * test_frac))
24
+ num_val = int(np.floor(edges.shape[0] * val_frac))
25
+
26
+ all_edge_idx = list(range(edges.shape[0]))
27
+ np.random.shuffle(all_edge_idx)
28
+ val_edge_idx = all_edge_idx[:num_val]
29
+ test_edge_idx = all_edge_idx[num_val : (num_val + num_test)]
30
+ test_edges = edges[test_edge_idx]
31
+ val_edges = edges[val_edge_idx]
32
+ train_edges = edges[all_edge_idx[num_val + num_test :]]
33
+
34
+ noedge_mask = np.ones(adj.shape) - adj
35
+ noedges = np.asarray(sp.triu(noedge_mask, 1).nonzero()).T
36
+ all_edge_idx = list(range(noedges.shape[0]))
37
+ np.random.shuffle(all_edge_idx)
38
+ val_edge_idx = all_edge_idx[:num_val]
39
+ test_edge_idx = all_edge_idx[num_val : (num_val + num_test)]
40
+ test_edges_false = noedges[test_edge_idx]
41
+ val_edges_false = noedges[val_edge_idx]
42
+
43
+ data = np.ones(train_edges.shape[0])
44
+ adj_train = sp.csr_matrix(
45
+ (data, (train_edges[:, 0], train_edges[:, 1])), shape=adj.shape
46
+ )
47
+ adj_train = adj_train + adj_train.T
48
+
49
+ train_mask = np.ones(adj_train.shape)
50
+ for edges_tmp in [val_edges, val_edges_false, test_edges, test_edges_false]:
51
+ for e in edges_tmp:
52
+ assert e[0] < e[1]
53
+ train_mask[edges_tmp.T[0], edges_tmp.T[1]] = 0
54
+ train_mask[edges_tmp.T[1], edges_tmp.T[0]] = 0
55
+
56
+ train_edges = np.asarray(sp.triu(adj_train, 1).nonzero()).T
57
+ train_edges_false = np.asarray(
58
+ (sp.triu(train_mask, 1) - sp.triu(adj_train, 1)).nonzero()
59
+ ).T
60
+
61
+ # NOTE: all these edge lists only contain single direction of edge!
62
+ return (
63
+ train_edges,
64
+ train_edges_false,
65
+ val_edges,
66
+ val_edges_false,
67
+ test_edges,
68
+ test_edges_false,
69
+ )
70
+
71
+
72
+ def sparse_to_tuple(sparse_mx):
73
+ if not sp.isspmatrix_coo(sparse_mx):
74
+ sparse_mx = sparse_mx.tocoo()
75
+ coords = np.vstack((sparse_mx.row, sparse_mx.col)).transpose()
76
+ values = sparse_mx.data
77
+ shape = sparse_mx.shape
78
+ return coords, values, shape
79
+
80
+
81
+ if __name__ == "__main__":
82
+ os.mkdir("links")
83
+ os.mkdir("pretrain_labels")
84
+ g = AmazonCoBuyPhotoDataset()[0]
85
+ total_pos_edges = torch.randperm(g.num_edges())
86
+ adj_train = g.adjacency_matrix(scipy_fmt="csr")
87
+ (
88
+ train_edges,
89
+ train_edges_false,
90
+ val_edges,
91
+ val_edges_false,
92
+ test_edges,
93
+ test_edges_false,
94
+ ) = mask_test_edges(adj_train, 0.1, 0.2)
95
+ tvt_edges_file = "links/co_photo_tvtEdges.pkl"
96
+ pickle.dump(
97
+ (
98
+ train_edges,
99
+ train_edges_false,
100
+ val_edges,
101
+ val_edges_false,
102
+ test_edges,
103
+ test_edges_false,
104
+ ),
105
+ open(tvt_edges_file, "wb"),
106
+ )
107
+ node_assignment = dgl.metis_partition_assignment(g, 10)
108
+ torch.save(node_assignment, "pretrain_labels/metis_label_co_photo.pt")