-
Notifications
You must be signed in to change notification settings - Fork 5
/
rc_explainer_pool.py
291 lines (214 loc) · 11.8 KB
/
rc_explainer_pool.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
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
from torch.nn import Sequential, Linear, ReLU, ModuleList, Softmax, ELU, Sigmoid
import torch.nn.functional as F
from module.utils import *
from torch_geometric.utils import softmax
from torch_scatter import scatter_max
from module.utils.reorganizer import relabel_graph, filter_correct_data
import os.path as osp
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
class RC_Explainer(torch.nn.Module):
def __init__(self, _model, _num_labels, _hidden_size, _use_edge_attr=False):
super(RC_Explainer, self).__init__()
self.model = _model
self.model = self.model.to(device)
self.model.eval()
self.num_labels = _num_labels
self.hidden_size = _hidden_size
self.use_edge_attr = _use_edge_attr
self.temperature = 0.1
self.edge_action_rep_generator = Sequential(
Linear(self.hidden_size * 2, self.hidden_size * 2),
ReLU(),
Linear(self.hidden_size * 2, self.hidden_size),
).to(device)
self.edge_action_prob_generator = self.build_edge_action_prob_generator()
def build_edge_action_prob_generator(self):
edge_action_prob_generator = Sequential(
Linear(self.hidden_size * (1 + self.use_edge_attr), self.hidden_size),
ReLU(),
Linear(self.hidden_size, self.num_labels)
).to(device)
return edge_action_prob_generator
def forward(self, graph, state, target_y):
ocp_edge_index = graph.edge_index.T[state].T
ocp_edge_attr = graph.edge_attr[state]
ava_edge_index = graph.edge_index.T[~state].T
ava_edge_attr = graph.edge_attr[~state]
ava_node_reps_0 = self.model.get_node_reps(graph.x, graph.edge_index, graph.edge_attr, graph.batch)
ava_node_reps_1 = self.model.get_node_reps(graph.x, ocp_edge_index, ocp_edge_attr, graph.batch)
ava_node_reps = ava_node_reps_0 - ava_node_reps_1
ava_action_reps = torch.cat([ava_node_reps[ava_edge_index[0]],
ava_node_reps[ava_edge_index[1]]], dim=1).to(device)
ava_action_reps = self.edge_action_rep_generator(ava_action_reps)
if self.use_edge_attr:
ava_edge_reps = self.model.edge_emb(ava_edge_attr)
ava_action_reps = torch.cat([ava_action_reps, ava_edge_reps], dim=1)
ava_action_probs = self.predict(ava_action_reps, target_y)
return ava_action_probs
def predict(self, ava_action_reps, target_y):
action_probs = self.edge_action_prob_generator(ava_action_reps)[:, target_y]
action_probs = action_probs.reshape(-1)
action_probs = F.softmax(action_probs)
return action_probs
def get_optimizer(self, lr=0.01, weight_decay=1e-5):
params = list(self.edge_action_rep_generator.parameters()) + \
list(self.edge_action_prob_generator.parameters())
optimizer = torch.optim.Adam(params, lr=lr, weight_decay=weight_decay)
return optimizer
class RC_Explainer_pro(RC_Explainer):
def __init__(self, _model, _num_labels, _hidden_size, _use_edge_attr=False):
super(RC_Explainer_pro, self).__init__(_model, _num_labels, _hidden_size, _use_edge_attr=False)
def build_edge_action_prob_generator(self):
edge_action_prob_generator = nn.ModuleList()
for i in range(self.num_labels):
i_explainer = Sequential(
Linear(self.hidden_size, self.hidden_size),
ELU(),
Linear(self.hidden_size, 1)).to(device)
edge_action_prob_generator.append(i_explainer)
return edge_action_prob_generator
def predict(self, ava_action_reps, target_y):
i_explainer = self.edge_action_prob_generator[target_y]
action_probs = i_explainer(ava_action_reps)
action_probs = action_probs.reshape(-1)
action_probs = F.softmax(action_probs)
return action_probs
def get_optimizer(self, lr=0.01, weight_decay=1e-5):
params = list(self.edge_action_rep_generator.parameters())
for i_explainer in self.edge_action_prob_generator:
params += list(i_explainer.parameters())
optimizer = torch.optim.Adam(params, lr=lr, weight_decay=weight_decay)
return optimizer
"""
Using the mini-batch training the RC-Explainer, which is much efficient than the one-by-one training.
"""
class RC_Explainer_Batch(torch.nn.Module):
def __init__(self, _model, _num_labels, _hidden_size, _use_edge_attr=False):
super(RC_Explainer_Batch, self).__init__()
self.model = _model
self.model = self.model.to(device)
self.model.eval()
self.num_labels = _num_labels
self.hidden_size = _hidden_size
self.use_edge_attr = _use_edge_attr
self.temperature = 0.1
self.edge_action_rep_generator = Sequential(
Linear(self.hidden_size * (2 + self.use_edge_attr), self.hidden_size * 4),
ELU(),
Linear(self.hidden_size * 4, self.hidden_size * 2),
ELU(),
Linear(self.hidden_size * 2, self.hidden_size)
).to(device)
self.edge_action_prob_generator = self.build_edge_action_prob_generator()
def build_edge_action_prob_generator(self):
edge_action_prob_generator = Sequential(
Linear(self.hidden_size, self.hidden_size),
ELU(),
Linear(self.hidden_size, self.num_labels)
).to(device)
return edge_action_prob_generator
def forward(self, graph, state, train_flag=False):
ocp_edge_index = graph.edge_index.T[state].T
ocp_edge_attr = graph.edge_attr[state]
ava_edge_index = graph.edge_index.T[~state].T
ava_edge_attr = graph.edge_attr[~state]
ava_node_reps_0 = self.model.get_node_reps(graph.x, graph.edge_index, graph.edge_attr, graph.batch)
ava_node_reps_1 = self.model.get_node_reps(graph.x, ocp_edge_index, ocp_edge_attr, graph.batch)
ava_node_reps = ava_node_reps_0 - ava_node_reps_1
ava_action_reps = torch.cat([ava_node_reps[ava_edge_index[0]],
ava_node_reps[ava_edge_index[1]]], dim=1).to(device)
ava_action_reps = self.edge_action_rep_generator(ava_action_reps)
ava_action_batch = graph.batch[ava_edge_index[0]]
ava_y_batch = graph.y[ava_action_batch]
if self.use_edge_attr:
ava_edge_reps = self.model.edge_emb(ava_edge_attr)
ava_action_reps = torch.cat([ava_action_reps, ava_edge_reps], dim=1)
ava_action_probs = self.predict(ava_action_reps, ava_y_batch, ava_action_batch)
added_action_probs, added_actions = scatter_max(ava_action_probs, ava_action_batch)
if train_flag:
rand_action_probs = torch.rand(ava_action_probs.size()).to(device)
_, rand_actions = scatter_max(rand_action_probs, ava_action_batch)
return ava_action_probs, ava_action_probs[rand_actions], rand_actions
return ava_action_probs, added_action_probs, added_actions
def predict(self, ava_action_reps, target_y, ava_action_batch):
action_probs = self.edge_action_prob_generator(ava_action_reps)
action_probs = action_probs.gather(1, target_y.view(-1,1))
action_probs = action_probs.reshape(-1)
action_probs = softmax(action_probs, ava_action_batch)
return action_probs
def get_optimizer(self, lr=0.01, weight_decay=1e-5, scope='all'):
if scope in ['all']:
params = self.parameters()
else:
params = list(self.edge_action_rep_generator.parameters()) + \
list(self.edge_action_prob_generator.parameters())
optimizer = torch.optim.Adam(params, lr=lr, weight_decay=weight_decay)
return optimizer
def load_policy_net(self, name='policy.pt', path=None):
if not path:
path = osp.join(osp.dirname(__file__), '..', '..', 'params', name)
self.load_state_dict(torch.load(path))
def save_policy_net(self, name='policy.pt', path=None):
if not path:
path = osp.join(osp.dirname(__file__), '..', '..', 'params', name)
torch.save(self.state_dict(), path)
class RC_Explainer_Batch_star(RC_Explainer_Batch):
def __init__(self, _model, _num_labels, _hidden_size, _use_edge_attr=False):
super(RC_Explainer_Batch_star, self).__init__(_model, _num_labels, _hidden_size, _use_edge_attr=False)
def build_edge_action_prob_generator(self):
edge_action_prob_generator = nn.ModuleList()
for i in range(self.num_labels):
i_explainer = Sequential(
Linear(self.hidden_size * (2 + self.use_edge_attr), self.hidden_size * 2),
ELU(),
Linear(self.hidden_size * 2, self.hidden_size),
ELU(),
Linear(self.hidden_size, 1)
).to(device)
edge_action_prob_generator.append(i_explainer)
return edge_action_prob_generator
def forward(self, graph, state, train_flag=False):
graph_rep = self.model.get_graph_rep(graph.x, graph.edge_index, graph.edge_attr, graph.batch)
if len(torch.where(state==True)[0]) == 0:
subgraph_rep = torch.zeros(graph_rep.size()).to(device)
else:
subgraph = relabel_graph(graph, state)
subgraph_rep = self.model.get_graph_rep(subgraph.x, subgraph.edge_index, subgraph.edge_attr, subgraph.batch)
ava_edge_index = graph.edge_index.T[~state].T
ava_edge_attr = graph.edge_attr[~state]
ava_node_reps = self.model.get_node_reps(graph.x, ava_edge_index, ava_edge_attr, graph.batch)
if self.use_edge_attr:
ava_edge_reps = self.model.edge_emb(ava_edge_attr)
ava_action_reps = torch.cat([ava_node_reps[ava_edge_index[0]],
ava_node_reps[ava_edge_index[1]],
ava_edge_reps], dim=1).to(device)
else:
ava_action_reps = torch.cat([ava_node_reps[ava_edge_index[0]],
ava_node_reps[ava_edge_index[1]]], dim=1).to(device)
ava_action_reps = self.edge_action_rep_generator(ava_action_reps)
ava_action_batch = graph.batch[ava_edge_index[0]]
ava_y_batch = graph.y[ava_action_batch]
# get the unique elements in batch, in cases where some batches are out of actions.
unique_batch, ava_action_batch = torch.unique(ava_action_batch, return_inverse=True)
ava_action_probs = self.predict_star(graph_rep, subgraph_rep, ava_action_reps, ava_y_batch, ava_action_batch)
# assert len(ava_action_probs) == sum(~state)
added_action_probs, added_actions = scatter_max(ava_action_probs, ava_action_batch)
if train_flag:
rand_action_probs = torch.rand(ava_action_probs.size()).to(device)
_, rand_actions = scatter_max(rand_action_probs, ava_action_batch)
return ava_action_probs, ava_action_probs[rand_actions], rand_actions, unique_batch
return ava_action_probs, added_action_probs, added_actions, unique_batch
def predict_star(self, graph_rep, subgraph_rep, ava_action_reps, target_y, ava_action_batch):
action_graph_reps = graph_rep - subgraph_rep
action_graph_reps = action_graph_reps[ava_action_batch]
action_graph_reps = torch.cat([ava_action_reps, action_graph_reps], dim=1)
action_probs = []
for i_explainer in self.edge_action_prob_generator:
i_action_probs = i_explainer(action_graph_reps)
action_probs.append(i_action_probs)
action_probs = torch.cat(action_probs, dim=1)
action_probs = action_probs.gather(1, target_y.view(-1,1))
action_probs = action_probs.reshape(-1)
# action_probs = softmax(action_probs, ava_action_batch)
# action_probs = F.sigmoid(action_probs)
return action_probs