-
Notifications
You must be signed in to change notification settings - Fork 7
/
refine.py
252 lines (202 loc) · 8.76 KB
/
refine.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
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
import copy
import math
import numpy as np
import torch
import torch.nn as nn
from torch.nn import functional as F
from torch_geometric.nn import MessagePassing
from explainers.base import Explainer
from .common import EdgeMaskNet
EPS = 1e-6
class ReFine(Explainer):
coeffs = {
'edge_size': 1e-4,
'edge_ent': 1e-2,
}
def __init__(self, device, gnn_model,
n_in_channels=14,
e_in_channels=3,
hid=50, n_layers=2,
n_label=2, gamma=1
):
super(ReFine, self).__init__(device, gnn_model)
self.edge_mask = nn.ModuleList([
EdgeMaskNet(
n_in_channels,
e_in_channels,
hid=hid,
n_layers=n_layers) for _ in range(n_label)
]).to(device)
self.gamma = gamma
def _set_masks(self, mask, model):
for module in model.modules():
if isinstance(module, MessagePassing):
module.__explain__ = True
module.__edge_mask__ = mask
def _clear_masks(self, model):
for module in model.modules():
if isinstance(module, MessagePassing):
module.__explain__ = False
module.__edge_mask__ = None
def _reparameterize(self, log_alpha, beta=1, training=True):
if training:
random_noise = torch.rand(log_alpha.size()).to(self.device)
gate_inputs = torch.log2(random_noise) - torch.log2(1.0 - random_noise)
gate_inputs = (gate_inputs + log_alpha) / beta + EPS
gate_inputs = gate_inputs.sigmoid()
else:
gate_inputs = log_alpha.sigmoid()
return gate_inputs
def _setup_target_label(self, graph):
if not hasattr(graph, 'hat_y'):
graph.hat_y = self.model(graph).argmax(-1).to(graph.x.device)
return graph
def fidelity_loss(self, log_logits, mask, pred_label):
idx = [i for i in range(len(pred_label))]
loss = -log_logits.softmax(dim=1)[idx, pred_label.view(-1)].sum()
loss = loss + self.coeffs['edge_size'] * mask.mean()
ent = -mask * torch.log(mask + EPS) - (1 - mask) * torch.log(1 - mask + EPS)
loss = loss + self.coeffs['edge_ent'] * ent.mean()
return loss
# batch version
def pack_subgraph(self, graph, imp, top_ratio=0.2):
if abs(top_ratio - 1.0) < EPS:
return graph, imp
exp_subgraph = copy.deepcopy(graph)
top_idx = torch.LongTensor([])
graph_map = graph.batch[graph.edge_index[0, :]]
# extract ego graph
for i in range(graph.num_graphs):
edge_indicator = torch.where(graph_map == i)[0].detach().cpu()
Gi_n_edge = len(edge_indicator)
topk = max(math.ceil(top_ratio * Gi_n_edge), 1)
Gi_pos_edge_idx = np.argsort(-imp[edge_indicator])[:topk]
top_idx = torch.cat([top_idx, edge_indicator[Gi_pos_edge_idx]])
exp_subgraph.edge_attr = graph.edge_attr[top_idx]
exp_subgraph.edge_index = graph.edge_index[:, top_idx]
exp_subgraph.x, exp_subgraph.edge_index, exp_subgraph.batch, _ = \
self._relabel(exp_subgraph, exp_subgraph.edge_index)
return exp_subgraph, imp[top_idx]
def get_contrastive_loss(self, c, y, batch, tau=0.1):
c = c / c.norm(dim=1, keepdim=True)
mat = F.relu(torch.mm(c, c.T))
unique_graphs = torch.unique(batch)
ttl_scores = torch.sum(mat, dim=1)
pos_scores = torch.tensor([mat[i, y == y[i]].sum() for i in unique_graphs]).to(c.device)
# contrastive_loss = - torch.log(torch.sum(pos_scores / ttl_scores, dim=0))
contrastive_loss = - torch.logsumexp(pos_scores / (tau * ttl_scores), dim=0)
return contrastive_loss
def get_mask(self, graph):
# batch version
graph_map = graph.batch[graph.edge_index[0, :]]
mask = torch.FloatTensor([]).to(graph.x.device)
for i in range(len(graph.hat_y)):
edge_indicator = (graph_map == i).bool()
G_i_mask = self.edge_mask[graph.hat_y[i]](
graph.x,
graph.edge_index[:, edge_indicator],
graph.edge_attr[edge_indicator, :]
).view(-1)
mask = torch.cat([mask, G_i_mask])
return mask
def get_pos_edge(self, graph, mask, ratio):
num_edge = [0]
num_node = [0]
sep_edge_idx = []
graph_map = graph.batch[graph.edge_index[0, :]]
pos_idx = torch.LongTensor([])
mask = mask.detach().cpu()
for i in range(graph.num_graphs):
edge_indicator = torch.where(graph_map == i)[0].detach().cpu()
Gi_n_edge = len(edge_indicator)
topk = max(math.ceil(ratio * Gi_n_edge), 1)
Gi_pos_edge_idx = np.argsort(-mask[edge_indicator])[:topk]
pos_idx = torch.cat([pos_idx, edge_indicator[Gi_pos_edge_idx]])
num_edge.append(num_edge[i] + Gi_n_edge)
num_node.append(
num_node[i] + (graph.batch == i).sum().long()
)
sep_edge_idx.append(Gi_pos_edge_idx)
return pos_idx, num_edge, num_node, sep_edge_idx
def explain_graph(
self, graph, ratio=1.0, fine_tune=False,
lr=1e-4, epoch=50, draw_graph=0, vis_ratio=0.2):
graph = self._setup_target_label(graph)
if not fine_tune:
edge_mask = self.get_mask(graph)
edge_mask = self._reparameterize(edge_mask, training=False)
imp = edge_mask.detach().cpu().numpy()
self.last_result = (graph, imp)
if draw_graph:
self.visualize(graph, imp, vis_ratio=vis_ratio)
return imp
mask_net = copy.deepcopy(self.edge_mask[graph.hat_y.item()])
optimizer = torch.optim.Adam(mask_net.parameters(), lr=lr)
for _ in range(epoch):
optimizer.zero_grad()
edge_mask = mask_net(
graph.x,
graph.edge_index,
graph.edge_attr
).view(-1)
edge_mask = self._reparameterize(edge_mask, training=False)
pos_idx, _, _, _ = self.get_pos_edge(graph, edge_mask, ratio)
pos_edge_mask = edge_mask[pos_idx]
pos_edge_index = graph.edge_index[:, pos_idx]
pos_edge_attr = graph.edge_attr[pos_idx, :]
self._set_masks(pos_edge_mask, self.model)
G1_x, G1_pos_edge_index, G1_batch, G1_pos = self._relabel(graph, pos_edge_index)
log_logits = self.model(
x=G1_x,
edge_index=G1_pos_edge_index,
edge_attr=pos_edge_attr,
batch=G1_batch,
pos=G1_pos
)
self._clear_masks(self.model)
fid_loss = self.fidelity_loss(log_logits, edge_mask, graph.hat_y)
fid_loss.backward()
optimizer.step()
imp = edge_mask.detach().cpu().numpy()
self.last_result = (graph, imp)
if draw_graph:
self.visualize(graph, imp, vis_ratio=vis_ratio)
return imp
def pretrain(
self, graph, model=None,
ratio=1.0, reperameter=False, **kwargs
):
if model == None:
model = self.model
graph = self._setup_target_label(graph)
ori_mask = self.get_mask(graph)
edge_mask = self._reparameterize(ori_mask, training=reperameter)
# ----------------------------------------------------
# (1) compute fidelity loss
self._set_masks(edge_mask, self.model)
log_logits = self.model(graph)
fid_loss = self.fidelity_loss(log_logits, edge_mask, graph.hat_y)
self._clear_masks(self.model)
# ----------------------------------------------------
# (2) compute contrastive loss
pos_idx, _, _, _ = self.get_pos_edge(graph, edge_mask, ratio)
pos_edge_mask = edge_mask[pos_idx]
pos_edge_index = graph.edge_index[:, pos_idx]
pos_edge_attr = graph.edge_attr[pos_idx, :]
self._set_masks(pos_edge_mask, self.model)
G1_x, G1_pos_edge_index, G1_batch, G1_pos = self._relabel(graph, pos_edge_index)
graph_rep = self.model.get_graph_rep(
x=G1_x,
edge_index=G1_pos_edge_index,
edge_attr=pos_edge_attr,
batch=G1_batch,
pos=G1_pos
)
cts_loss = self.get_contrastive_loss(graph_rep, graph.hat_y, graph.batch)
self._clear_masks(self.model)
loss = fid_loss + self.gamma * cts_loss
return loss
def remap_device(self, device):
self.device = device
self.model = self.model.to(device)
self.edge_mask = self.edge_mask.to(device)