-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathChainNet.py
169 lines (145 loc) · 8.2 KB
/
ChainNet.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
import torch
import torch.nn as nn
import torch_scatter
import torch.nn.functional as F
import numpy as np
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
from torch_geometric.nn.dense.linear import Linear
from torch.nn import Parameter
from inits import glorot, zeros
class Net(nn.Module):
def __init__(self, num_iterations, size_realnode, size_hypernode, num_readoutneurons, num_heads, negative_slope, dropout):
super().__init__()
self.phiC = nn.GRU(size_realnode*2, size_hypernode, batch_first = True)
self.phiF = nn.GRUCell(size_hypernode+size_realnode, size_realnode)
self.phiD = nn.GRUCell(size_hypernode+size_realnode, size_realnode)
self.N = num_iterations
self.size_realnode = size_realnode
self.size_hypernode = size_hypernode
self.initmapping_device = nn.Linear(5, size_realnode)
self.initmapping_fragment = nn.Linear(5, size_realnode)
self.initmapping_service = nn.Linear(5, size_hypernode)
self.num_heads = num_heads
self.lin_source = Linear(size_hypernode+size_realnode, num_heads * (size_hypernode+size_realnode), bias=True, weight_initializer='glorot')
self.lin_target = Linear(size_realnode, num_heads * (size_hypernode+size_realnode), bias=True, weight_initializer='glorot')
self.att = Parameter(torch.Tensor(1, num_heads, size_hypernode+size_realnode))
self.bias = Parameter(torch.Tensor(size_hypernode+size_realnode))
self.negative_slope = negative_slope
self.dropout = dropout
self.throughput_mlp = nn.Sequential(
nn.Linear(size_hypernode, num_readoutneurons),
nn.ReLU(),
nn.Dropout(self.dropout),
nn.Linear(num_readoutneurons, num_readoutneurons),
nn.ReLU(),
nn.Dropout(self.dropout),
nn.Linear(num_readoutneurons, 1)
)
self.latency_mlp = nn.Sequential(
nn.Linear(size_realnode, num_readoutneurons),
nn.ReLU(),
nn.Dropout(self.dropout),
nn.Linear(num_readoutneurons, num_readoutneurons),
nn.ReLU(),
nn.Dropout(self.dropout),
nn.Linear(num_readoutneurons, 1)
)
self.reset_parameters()
def reset_parameters(self):
self.lin_source.reset_parameters()
self.lin_target.reset_parameters()
glorot(self.att)
zeros(self.bias)
def forward(self, x):
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
realnode = x[0]
index_devices = x[7]
index_fragments = x[8]
devices = torch.index_select(realnode, dim=0, index=index_devices-1)
fragments = torch.index_select(realnode, dim=0, index=index_fragments-1)
devices = self.initmapping_device(devices)
fragments = self.initmapping_fragment(fragments)
select = torch.cat([devices,fragments],dim=0)
select_index = torch.cat((index_devices,index_fragments))
realnode_state = torch_scatter.scatter(src=select, index=select_index, dim=0, reduce='sum')
len_exeSeq = x[11]
len_exeSeq_list = [int(element/2) for element in len_exeSeq]
max_len_exeSeq = max(len_exeSeq_list)
arr = x[5]
service_state = self.initmapping_service(arr).unsqueeze(0)
exeSeqs = x[4]
fragTodev_edges = x[9]
fragTodev_device_index = x[10]
for _ in range(self.N):
exeSeqs_matrix = realnode_state[exeSeqs]
exeSeqs_matrix_dim1 = exeSeqs_matrix.size(0)
exeSeqs_matrix_dim2 = exeSeqs_matrix.size(1)
exeSeqs_matrix_dim3 = exeSeqs_matrix.size(2)
exeSeqs_matrix = exeSeqs_matrix.view(exeSeqs_matrix_dim1,int(exeSeqs_matrix_dim2/2),exeSeqs_matrix_dim3*2)
# message-passing for service chains
packed = pack_padded_sequence(exeSeqs_matrix, len_exeSeq_list, batch_first=True, enforce_sorted=False)
seq_packed, service_state = self.phiC(packed, service_state)
# reshape output
output, lens_unpacked = pad_packed_sequence(seq_packed, batch_first=True)
output = output.reshape(-1,self.size_hypernode)
# pick meaningful output
extract_indices = []
for extract in range(len(len_exeSeq_list)):
extract_indices.extend(extract_id for extract_id in range(extract*max_len_exeSeq,extract*max_len_exeSeq+len_exeSeq_list[extract]))
output = output[extract_indices]
# message-passing for fragments
## messageF_fromservice
messageF_fromservice = output
## messageF_fromdevice
source = torch.index_select(realnode_state, dim=0, index=fragTodev_edges[1])
cap_toedge_index = torch.tensor(np.arange(0,fragTodev_edges.size(1)),dtype=torch.int64).to(device)
messageF_fromdevice = torch_scatter.scatter(src=source, index=cap_toedge_index, dim=0, reduce='sum')
## concatenation
messageF = torch.cat((messageF_fromservice, messageF_fromdevice),dim=1)
## update
fragment_state = torch.index_select(realnode_state, dim=0, index=index_fragments)
fragment_state_copy = fragment_state
fragment_state = self.phiF(messageF, fragment_state)
# message-passing for devices
## concatenation
messageD_exeSteps = torch.cat((output,fragment_state_copy),dim=1)
## linear transformation
messageD_exeSteps_trans = self.lin_source(messageD_exeSteps).view(-1, self.num_heads, self.size_hypernode+self.size_realnode) # x_j (j->i)
device_state = torch.index_select(realnode_state, dim=0, index=index_devices)
device_state_trans = self.lin_target(device_state).view(-1, self.num_heads, self.size_hypernode+self.size_realnode)
## attention coefficients of (src,dst) pairs
device_state_trans_select = torch.index_select(device_state_trans, dim=0, index=fragTodev_device_index)
x = messageD_exeSteps_trans+device_state_trans_select
x = F.leaky_relu(x, self.negative_slope)
alpha = (x * self.att).sum(dim=-1)
## transform attention coefficients to weights (softmax)
alpha_max = torch_scatter.scatter(alpha, index=fragTodev_device_index, dim=0, reduce='max')
alpha_max = torch.index_select(alpha_max, dim=0, index=fragTodev_device_index)
out = (alpha - alpha_max).exp()
out_sum = torch_scatter.scatter(out, index=fragTodev_device_index, dim=0, reduce='sum')
out_sum = torch.index_select(out_sum, dim=0, index=fragTodev_device_index)
alpha = out / (out_sum + 1e-16)
## weighted
weighted_messageD_exeSteps_trans = messageD_exeSteps_trans * alpha.unsqueeze(-1)
## weighted sum
messageD = torch_scatter.scatter(src=weighted_messageD_exeSteps_trans, index=fragTodev_device_index, dim=0, reduce='sum') # aggregate sernode, ->capnode
messageD = messageD.mean(dim=1) # for multi-head cases
messageD = messageD+self.bias
## update
device_state = self.phiD(messageD, device_state)
# renew the embeddings of real nodes
select = torch.cat([device_state,fragment_state],dim=0)
select_index = torch.cat((index_devices,index_fragments))
realnode_state = torch_scatter.scatter(src=select, index=select_index, dim=0, reduce='sum')
start_index = 0
pooled_result = []
for group in len_exeSeq_list:
pooled_tensor = fragment_state[start_index:start_index+group].mean(dim=0)
pooled_result.append(pooled_tensor)
start_index = start_index + group
latency_state = torch.stack(pooled_result)
latency = self.latency_mlp(latency_state).squeeze(-1)
latency = latency.unsqueeze(0)
throughput = self.throughput_mlp(service_state).squeeze(-1)
solution = torch.cat((latency, throughput), dim=1)
return solution