-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathdecoders.py
187 lines (163 loc) · 6.53 KB
/
decoders.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
from typing import Optional
import torch.nn as nn
import torch
def aggregated_sum(
data: torch.Tensor,
index: torch.LongTensor,
num_segments: int,
mean: bool = False
):
index = index.unsqueeze(1).repeat(1, data.size(1))
agg = data.new_full((num_segments, data.size(1)), 0).scatter_add_(0, index, data)
if mean:
counts = data.new_full((num_segments, data.size(1)), 0).scatter_add_(0, index, torch.ones_like(data))
agg = agg / counts.clamp(min=1)
return agg
class EuclideanDecoder(torch.nn.Module):
def __init__(
self,
d1: Optional[float] = 1.0,
d2: Optional[float] = 1.0,
threshold: Optional[float] = 0.5,
learnable: Optional[bool] = False,
sqrt: Optional[bool] = False
):
super(EuclideanDecoder, self).__init__()
self.d1 = nn.Parameter(torch.ones(1) * d1) if learnable else d1
self.d2 = nn.Parameter(torch.ones(1) * d2) if learnable else d2
self.threshold = threshold
self.learnable = learnable
self.sqrt = sqrt
self.criterion = nn.BCELoss(reduction='none')
def sigmoid(
self,
dist: torch.Tensor
):
# https://www.desmos.com/calculator/mkp3ewfmiu
exp = (self.d2 * (dist - self.d1))
return torch.sigmoid(-exp)
def decode_edge(
self,
z: torch.Tensor,
edge_index: torch.Tensor,
sigmoid: Optional[bool] = True
):
if z.dim() == 3:
return self.decode_edge_dense(z, edge_index, sigmoid)
else:
return self.decode_edge_sparse(z, edge_index, sigmoid)
def decode_adj(
self,
z: torch.Tensor,
n_nodes: Optional[torch.LongTensor] = None
):
if z.dim() == 3:
return self.decode_adj_dense(z, n_nodes)
else:
return self.decode_adj_sparse(z, n_nodes)
def decode_edge_sparse(
self,
z: torch.Tensor,
edge_index: torch.LongTensor,
sigmoid: Optional[bool] = True
):
dist = ((z[edge_index[0]] - z[edge_index[1]]) ** 2).sum(dim=1)
if self.sqrt:
dist = dist.sqrt()
return self.sigmoid(dist) if sigmoid else dist
def decode_edge_dense(
self,
z: torch.Tensor,
adj: torch.Tensor,
sigmoid: Optional[bool] = True
):
dist = ((z.unsqueeze(2) - z.unsqueeze(1)) ** 2).sum(dim=-1)
if self.sqrt:
dist = dist.sqrt()
out = (self.sigmoid(dist) if sigmoid else dist) * adj
return out
@torch.no_grad()
def decode_adj_sparse(
self,
z: torch.Tensor,
n_nodes: Optional[torch.LongTensor] = None,
sort: Optional[bool] = False
):
if n_nodes is None:
edge_index = torch.ones(z.size(0), z.size(0)).tril(-1).nonzero().T.to(z.device)
else:
offset = [0] + torch.Tensor.tolist(n_nodes[:-1].cumsum(0))
edge_index = torch.cat(
[torch.ones(n, n).tril(-1).nonzero().T + o for o, n in zip(offset, n_nodes)], dim=1).to(z.device)
edge_weight = self.decode_edge_sparse(z, edge_index, sigmoid=True)
# the higher the threshold the sparser the graph
mask = edge_weight >= self.threshold
edge_index = edge_index[:, mask]
edge_index = torch.cat([edge_index, torch.cat([edge_index[1:], edge_index[:1]], dim=0)], dim=1)
edge_weight = edge_weight[mask].repeat(2)
if sort:
perm = edge_index[0].argsort()
edge_index = edge_index[:, perm]
edge_weight = edge_weight[perm]
return edge_index, edge_weight
def decode_adj_dense(
self,
z: torch.Tensor,
n_nodes: Optional[torch.LongTensor] = None
):
assert z.dim() == 3
if n_nodes is None:
n_nodes = torch.full((z.size(0),), z.size(1)).to(z.device)
max_n_nodes = n_nodes.max()
assert z.size(1) == max_n_nodes
dist = ((z.unsqueeze(2) - z.unsqueeze(1)) ** 2).sum(-1)
if self.sqrt:
dist = dist.sqrt()
weight = self.sigmoid(dist)
adj = torch.zeros(n_nodes.size(0), max_n_nodes, max_n_nodes, dtype=torch.uint8).to(z.device)
for i in range(n_nodes.size(0)):
adj[i, :n_nodes[i], :n_nodes[i]] = 1
adj[i].fill_diagonal_(0)
adj = torch.logical_and(adj, weight >= self.threshold).to(torch.uint8)
adj_weight = weight * adj
# adj has no gradients, adj_weight does
return adj, adj_weight
def bce(
self,
z: torch.Tensor,
pos_edge_index: torch.LongTensor,
neg_edge_index: torch.LongTensor,
neg_weight: Optional[float] = 1.0
):
pos_loss = self.criterion(
input=self.decode_edge_sparse(z, pos_edge_index, sigmoid=True),
target=torch.ones(1, dtype=z.dtype, device=z.device).expand(pos_edge_index.size(1))).mean()
neg_loss = self.criterion(
input=self.decode_edge_sparse(z, neg_edge_index, sigmoid=True),
target=torch.zeros(1, dtype=z.dtype, device=z.device).expand(neg_edge_index.size(1))).mean()
loss = pos_loss + neg_weight * neg_loss
return loss
def bce_per_graph(
self,
z: torch.Tensor,
pos_edge_index: torch.LongTensor,
neg_edge_index: torch.LongTensor,
n_pos_edges: torch.LongTensor,
n_neg_edges: torch.LongTensor
):
assert n_pos_edges.ndim == n_neg_edges.ndim == 1 and len(n_pos_edges) == len(n_neg_edges)
num_graphs = len(n_pos_edges)
pos_bce_per_edge = self.criterion(
input=self.decode_edge_sparse(z, pos_edge_index, sigmoid=True),
target=torch.ones(1, dtype=z.dtype, device=z.device).expand(pos_edge_index.size(1))).unsqueeze(1)
neg_bce_per_edge = self.criterion(
input=self.decode_edge_sparse(z, neg_edge_index, sigmoid=True),
target=torch.zeros(1, dtype=z.dtype, device=z.device).expand(neg_edge_index.size(1))).unsqueeze(1)
pos_index = torch.arange(num_graphs, device=n_pos_edges.device).repeat_interleave(n_pos_edges)
pos_bce_per_graph = aggregated_sum(pos_bce_per_edge, pos_index, num_graphs, mean=True)
neg_index = torch.arange(num_graphs, device=n_neg_edges.device).repeat_interleave(n_neg_edges)
neg_bce_per_graph = aggregated_sum(neg_bce_per_edge, neg_index, num_graphs, mean=True)
bce_per_graph = pos_bce_per_graph + neg_bce_per_graph
return bce_per_graph
def __repr__(self) -> str:
return f'{self.__class__.__name__}(learnable={self.learnable})'