-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
129 lines (108 loc) · 5.25 KB
/
model.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
import pdb
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
from conv_mol import (GINMolHeadEncoder, GNNMolTailEncoder, GraphMolMasker,
vGINMolHeadEncoder)
from conv_syn import GINConv, GNNSynEncoder, GraphSynMasker
from graphon import estimate_graphon
from torch_geometric.nn import MessagePassing, global_mean_pool
import time
import numpy as np
class CausalGraphon(torch.nn.Module):
def __init__(self, args,
num_class,
in_dim,
emb_dim=300,
fro_layer=2,
bac_layer=2,
cau_layer=2,
dropout_rate=0.5,
cau_gamma=0.4,
env_gamma=1.0,
use_linear=False,
graphon=True,
N=15):
super(CausalGraphon, self).__init__()
self.args = args
self.cau_gamma = cau_gamma
self.env_gamma = env_gamma
self.dropout_rate = dropout_rate
self.emb_dim = emb_dim
self.num_class = num_class
self.graphon = graphon
self.N = N
self.graph_front = GNNSynEncoder(fro_layer, in_dim, emb_dim, dropout_rate)
self.graph_backs = GNNSynEncoder(bac_layer, emb_dim, emb_dim, dropout_rate)
self.causaler = GraphSynMasker(cau_layer, in_dim, emb_dim, dropout_rate)
self.pool = global_mean_pool
if use_linear:
self.predictor = torch.nn.Linear(emb_dim, num_class)
else:
self.predictor = torch.nn.Sequential(torch.nn.Linear(emb_dim, 2*emb_dim), torch.nn.BatchNorm1d(2*emb_dim), nn.ReLU(), torch.nn.Dropout(), torch.nn.Linear(2*emb_dim, num_class))
def forward(self, data, epoch=0, eval_random=True):
x, edge_index, batch = data.x, data.edge_index, data.batch
x_encode = self.graph_front(x, edge_index)
causaler_output = self.causaler(data)
node_cau, edge_cau = causaler_output["node_key"], causaler_output["edge_key"]
node_cau_num, node_env_num = causaler_output["node_key_num"], causaler_output["node_env_num"]
edge_cau_num, edge_env_num = causaler_output["edge_key_num"], causaler_output["edge_env_num"]
cau_node_reg = self.reg_mask_loss(node_cau_num, node_env_num, self.cau_gamma, self.causaler.non_zero_node_ratio)
cau_edge_reg = self.reg_mask_loss(edge_cau_num, edge_env_num, self.cau_gamma, self.causaler.non_zero_edge_ratio)
cau_loss_reg = cau_node_reg + cau_edge_reg
node_env = (1 - node_cau)
edge_env = (1 - edge_cau)
h_node_cau = self.graph_backs(x_encode, edge_index, node_cau, edge_cau)
h_node_env = self.graph_backs(x_encode, edge_index, node_env, edge_env)
h_graph_cau = self.pool(h_node_cau, batch)
h_graph_env = self.pool(h_node_env, batch)
pred_cau = self.predictor(h_graph_cau)
pred_env = self.predictor(h_graph_env)
pred_add = self.random_layer(h_graph_cau, h_graph_env, eval_random=eval_random)
"""
edge_cau is very important for our model, we use edge_cau to calculate the causal adjacency matrix,
later using adjacency matrix to estimate the graphons.
"""
graphon_loss = 0
output = {'pred_cau': pred_cau,
'pred_env': pred_env,
'pred_add': pred_add,
'cau_loss_reg': cau_loss_reg,
'graphon_loss': graphon_loss,
'causal': causaler_output,
'h_graph_env': h_graph_env,
'args': self.args}
if self.args.graphon and epoch >= self.args.graphon_pretrain and epoch % self.args.graphon_frequency == 0:
graphons, ys, envs = estimate_graphon(data, edge_cau.squeeze(), self.N, h_graph_env, self.args.num_env)
if graphons == None:
return output
intra_y = []
for y in range(len(ys)):
for env1 in range(len(envs)):
for env2 in range(len(envs)):
if env1 < env2:
graphon1 = graphons['{}{}'.format(y, env1)]
graphon2 = graphons['{}{}'.format(y, env2)]
intra_y.append(torch.norm(graphon1 - graphon2, p=2))
output['graphon_loss'] = torch.mean(torch.tensor(intra_y))
return output
def random_layer(self, xc, xo, eval_random):
if self.args.random_add == 'shuffle':
num = xc.shape[0]
l = [i for i in range(num)]
if self.args.with_random:
if eval_random:
random.shuffle(l)
random_idx = torch.tensor(l)
x = xc[random_idx] + xo
elif self.args.random_add == 'everyadd':
x = (xo.unsqueeze(1) + xc.unsqueeze(0)).view(-1, xo.shape[1])
else:
assert False
x_logis = self.predictor(x)
return x_logis
def reg_mask_loss(self, key_mask, env_mask, gamma, non_zero_ratio):
loss_reg = torch.abs(key_mask / (key_mask + env_mask) - gamma * torch.ones_like(key_mask)).mean()
loss_reg += (non_zero_ratio - gamma * torch.ones_like(key_mask)).mean()
return loss_reg