-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathAttentionGN.py
197 lines (163 loc) · 9.28 KB
/
AttentionGN.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
import torch
import torch.nn as nn
from torch_scatter import scatter_add, scatter_max, scatter_mean, scatter_min
from torch_geometric.data import Data
from GN_Model import P_GN
from blocks import EdgeBlock, NodeBlock, GlobalBlock
class TSelfAttention(nn.Module):
def __init__(self, embed_size, heads):
super(TSelfAttention, self).__init__()
self.embed_size = embed_size
self.heads = heads
self.head_dim = embed_size // heads
assert (
self.head_dim * heads == embed_size
), "Embedding size needs to be divisible by heads"
self.values = nn.Linear(self.head_dim, self.head_dim, bias=False)
self.keys = nn.Linear(self.head_dim, self.head_dim, bias=False)
self.queries = nn.Linear(self.head_dim, self.head_dim, bias=False)
self.fc_out = nn.Linear(heads * self.head_dim, embed_size)
def forward(self, values, keys, query, device):
N, T, C = query.shape
# Split the embedding into self.heads different pieces
values = values.reshape(N, T, self.heads, self.head_dim) # embed_size维拆成 heads×head_dim
keys = keys.reshape(N, T, self.heads, self.head_dim)
query = query.reshape(N, T, self.heads, self.head_dim)
values = self.values(values).to(device) # (N, T, heads, head_dim)
keys = self.keys(keys).to(device) # (N, T, heads, head_dim)
queries = self.queries(query).to(device) # (N, T, heads, heads_dim)
# Einsum does matrix mult. for query*keys for each training example
# with every other training example, don't be confused by einsum
# it's just how I like doing matrix multiplication & bmm
energy = torch.einsum("nqhd,nkhd->nqkh", [queries, keys]).to(device) # 时间self-attention
# queries shape: (N, T, heads, heads_dim),
# keys shape: (N, T, heads, heads_dim)
# energy: (N, T, T, heads)
# Normalize energy values similarly to seq2seq + attention
# so that they sum to 1. Also divide by scaling factor for
# better stability
attention = torch.softmax(energy / (self.embed_size ** (1 / 2)), dim=2).to(device) # 在K维做softmax，和为1
# attention shape: (N, query_len, key_len, heads)
out = torch.einsum("nqkh,nkhd->nqhd", [attention, values]).reshape(
N, T, self.heads * self.head_dim
).to(device)
# attention shape: (N, T, T, heads)
# values shape: (N, T, heads, heads_dim)
# out after matrix multiply: (N, T, heads, head_dim), then
# we reshape and flatten the last two dimensions.
out = self.fc_out(out).to(device)
# Linear layer doesn't modify the shape, final shape will be
# (N, T, embed_size)
return out
class attention_gn(nn.Module):
def __init__(self, in_channel, node_attr_size, edge_num_embeddings, out_size, att_layer, device, edge_hidden_size , node_hidden_size ,
global_hidden_size, heads):
super(attention_gn, self).__init__()
self.in_channel = in_channel
self.input_size = node_attr_size
self.edge_h_dim = edge_hidden_size
self.node_h_dim = node_hidden_size
self.node_half_h_dim = int(self.node_h_dim) / 2
self.global_h_dim = global_hidden_size
self.global_half_h_dim = int(self.global_h_dim) / 2
self.device = device
# Encoder
self.edge_enc = nn.Sequential(nn.Linear(1, self.edge_h_dim), nn.ReLU())
self.node_enc = nn.Sequential(nn.Linear(self.input_size, self.node_h_dim), nn.ReLU())
self.node_enc_for_att = nn.Sequential(nn.Linear(self.in_channel, self.node_h_dim), nn.ReLU())
# self-attention
self.attention = TSelfAttention(node_hidden_size, heads)
self.eb_custom_func = nn.Sequential(nn.Linear((self.edge_h_dim + self.node_h_dim * 2) * 2 + self.global_h_dim,
self.edge_h_dim),
nn.ReLU(),
)
self.nb_custom_func = nn.Sequential(nn.Linear(self.node_h_dim * 2 + self.edge_h_dim * 2 + self.global_h_dim,
self.node_h_dim),
nn.ReLU(),
)
self.gb_custom_func = nn.Sequential(nn.Linear(self.node_h_dim + self.edge_h_dim + self.global_h_dim,
self.global_h_dim),
nn.ReLU(),
)
self.eb_module = EdgeBlock((self.edge_h_dim + self.node_h_dim * 2) * 2 + self.global_h_dim,
self.edge_h_dim,
use_edges=True,
use_sender_nodes=True,
use_receiver_nodes=True,
use_globals=True,
custom_func=self.eb_custom_func)
self.nb_module = NodeBlock(self.node_h_dim * 2 + self.edge_h_dim * 2 + self.global_h_dim,
self.node_h_dim,
use_nodes=True,
use_sent_edges=True,
use_received_edges=True,
use_globals=True,
sent_edges_reducer=scatter_add,
received_edges_reducer=scatter_add,
custom_func=self.nb_custom_func)
self.gb_module = GlobalBlock(self.node_h_dim + self.edge_h_dim + self.global_h_dim,
self.global_h_dim,
edge_reducer = scatter_mean,
node_reducer = scatter_mean,
custom_func=self.gb_custom_func,
device=device)
self.gn = P_GN(self.eb_module,
self.nb_module,
self.gb_module,
use_edge_block=True,
use_node_block=True,
use_global_block=True)
##Decoder
self.node_dec = nn.Sequential(nn.Linear(self.node_h_dim, self.node_h_dim),
nn.ReLU(),
nn.Linear(self.node_h_dim, out_size)
)
self.feed_forward_att = nn.Sequential(nn.Linear(self.node_h_dim * att_layer, self.node_h_dim * att_layer),
nn.ReLU(),
nn.Linear(self.node_h_dim * att_layer, in_channel)
)
self.node_dec_for_input = nn.Sequential(nn.Linear(self.node_h_dim, self.node_h_dim),
nn.ReLU(),
nn.Linear(self.node_h_dim, self.input_size))
def forward(self, data, sp_L, t, num_processing_steps, coeff, pde='diff'):
from utils_tool import decompose_graph
input_graphs = []
node_attrs = []
edge_indexs = []
edge_attrs = []
nodes_num = sp_L.shape[0]
self_attention_nodes = []
for step_t in range(num_processing_steps):
# node_attr (nodes_num,hidden_num)
node_attr, edge_index, edge_attr, global_attr = decompose_graph(data[step_t])
node_attrs.append(node_attr)
edge_indexs.append(edge_index)
edge_attrs.append(edge_attr)
#### Encoder for attention
encoded_node_for_att = self.node_enc_for_att(node_attr.unsqueeze(2))
### Self-Attention
node_attrs_attention = self.attention(encoded_node_for_att, encoded_node_for_att, encoded_node_for_att, self.device) #(N,T,embed)
node_attrs_attention = self.feed_forward_att(node_attrs_attention) #(N,T,C)
self_attention_nodes.append(node_attrs_attention)
### residual_connection
node_attr = node_attrs_attention.squeeze(2) + node_attr
#### Input for GN
encoded_edge = self.edge_enc(edge_attr)
encoded_node = self.node_enc(node_attr)
input_graph = Data(x= encoded_node, edge_index=edge_index, edge_attr=encoded_edge)
if step_t == 0:
input_graph.global_attr = global_attr
input_graphs.append(input_graph)
init_graph = input_graphs[0]
# h_init is zero tensor
h_init = Data(x=torch.zeros(init_graph.x.size(), dtype=torch.float32, device=self.device),
edge_index=init_graph.edge_index,
edge_attr=torch.zeros(init_graph.edge_attr.size(), dtype=torch.float32, device=self.device))
h_init.global_attr = init_graph.global_attr
### GN
output_graphs, time_derivatives, spatial_derivatives = self.gn(input_graphs, sp_L, h_init, coeff, pde)
output_nodes, pred_inputs = [], []
for output_graph in output_graphs:
output_nodes.append(self.node_dec(output_graph.x))
pred_inputs.append(self.node_dec_for_input(output_graph.x))
return output_nodes, time_derivatives, spatial_derivatives