-
Notifications
You must be signed in to change notification settings - Fork 1
/
layer.py
250 lines (207 loc) · 10.7 KB
/
layer.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
import torch
from torch.nn import Linear, ReLU, Dropout, LayerNorm, Module, ModuleList
from torch_geometric.utils import subgraph, k_hop_subgraph, remove_isolated_nodes, add_self_loops, subgraph, index_to_mask, sort_edge_index, remove_self_loops, dropout_adj, segregate_self_loops
from torch_sparse import SparseTensor, transpose, coalesce
from pool import global_mean_pool, global_min_pool
import numpy as np
from sweep import GetRunTime
from torch_scatter import scatter
import numpy as np
class GNN(Module):
def __init__(self, hidden_channel, num_layers, dropout, gnn_model, act_final):
super().__init__()
m = ModuleList()
for i in range(num_layers):
if gnn_model=='GCN':
from layers.gcn_conv import GCNConv
m.append(GCNConv(in_channels=hidden_channel, out_channels=hidden_channel))
elif gnn_model=='GAT':
from layers.gat_conv import GATConv
m.append(GATConv(in_channels=hidden_channel, out_channels=hidden_channel))
elif gnn_model=='GIN':
from layers.gin_conv import GINConv
m.append(GINConv(
torch.nn.Sequential(
Linear(hidden_channel, hidden_channel),
ReLU(),
Linear(hidden_channel, hidden_channel),
ReLU(),
), train_eps=False))
if i < num_layers-1 or act_final==True:
m.append(ReLU())
m.append(Dropout(dropout))
self.gnn = m
self.norm = None
if num_layers>0:
self.norm = LayerNorm(hidden_channel)
def forward(self, x, adj_t):
init_x = x
for i in range(len(self.gnn)):
if i%3==0:
x = self.gnn[i](x, adj_t)
else:
x = self.gnn[i](x)
if self.norm!=None:
x = self.norm(init_x+x)
return x
class MLP(Module):
def __init__(self, in_channel, hidden_channel, out_channel, num_layers, dropout, norm_mode, act_final):
super().__init__()
m = ModuleList()
for i in range(num_layers):
if num_layers==1:
m.append(Linear(in_channel, out_channel))
else:
if i==0:
m.append(Linear(in_channel, hidden_channel))
elif i==num_layers-1:
m.append(Linear(hidden_channel, out_channel))
else:
m.append(Linear(hidden_channel, hidden_channel))
if i < num_layers-1 or act_final==True:
if norm_mode=='insert':
m.append(LayerNorm(hidden_channel))
m.append(ReLU())
m.append(Dropout(dropout))
self.mlp = m
self.norm = None
if num_layers>0 and norm_mode=='post':
self.norm = LayerNorm(hidden_channel)
def forward(self, x):
init_x = x
for i in range(len(self.mlp)):
x = self.mlp[i](x)
if self.norm!=None:
x = self.norm(init_x+x)
return x
class ParsingNet_GPU(Module):
def __init__(self, channel, dropout_network, dropout_parsing, layer_parsingnet, link_ignore_self_loop):
super().__init__()
self.dropout_parsing = dropout_parsing
self.link_ignore_self_loop = link_ignore_self_loop
self.edge_net = MLP(in_channel=channel, hidden_channel=channel, out_channel=1, num_layers=layer_parsingnet, dropout=dropout_network, norm_mode='insert', act_final=False)
# @profile
def forward(self, x, edge_index, batch):
# remove self-loops and then (optionally) add self-loops
device = x.device
row, col, edge_attr = edge_index.coo()
edge_index = torch.stack([row, col], dim=0)
edge_index, edge_attr = remove_self_loops(edge_index, edge_attr)
if self.link_ignore_self_loop==False:
edge_index, edge_attr = add_self_loops(edge_index, edge_attr, fill_value='mean', num_nodes=x.size(0))
# drop edge
edge_index, edge_attr = dropout_adj(
edge_index = edge_index,
edge_attr = edge_attr,
p = self.dropout_parsing,
force_undirected = True,
num_nodes = batch.shape[0],
training = self.training
)
# compute edge score (on the dropped graph)
edge_score = x[edge_index[0]] * x[edge_index[1]]
edge_score = self.edge_net(edge_score)
edge_score = torch.sigmoid(edge_score).squeeze(dim=-1)
num_nodes_full = batch.shape[0]
device_original = device
device = 'cpu'
edge_index = edge_index.to(device)
edge_score = edge_score.to(device)
batch = batch.to(device)
torch.set_num_threads(1)
# remove isolated nodes and save their masks
edge_index_temp, edge_score_temp, mask = remove_isolated_nodes(edge_index, edge_score, num_nodes=num_nodes_full)
mask_nodes = (~mask).nonzero(as_tuple=False).view(-1)
unmask_nodes = (mask).nonzero(as_tuple=False).view(-1)
num_nodes_connected = unmask_nodes.shape[0]
# get edges and scores for self-loop on isolated nodes
edge_index_, edge_score_, edge_mask_ = subgraph(mask_nodes, edge_index, edge_score, num_nodes=num_nodes_full, return_edge_mask=True)
edge_index_conn, edge_score_conn, edge_index_loop, edge_score_loop = segregate_self_loops(edge_index_, edge_score_)
edge_index, edge_score = edge_index_temp, edge_score_temp
# sort edges (in the connected graph) based on scores
edge_score_sorted, indices = edge_score.sort(stable=True, descending=True)
edge_index_sorted = edge_index[:,indices]
# get each node's dominant node (in the connected graph)
index, _ = transpose(torch.stack([torch.arange(indices.shape[0], device=device), indices], dim=0), None, indices.shape[0], indices.shape[0])
node2dom_edge_sorted_rank = scatter(index[1], edge_index[0], reduce='min')
node2dom_edge = edge_index_sorted[:,node2dom_edge_sorted_rank]
node2dom_score = edge_score_sorted[node2dom_edge_sorted_rank]
_, indices_ = node2dom_score.sort(stable=True, descending=True)
edge_index_dom = node2dom_edge[:,indices_]
# gen vars for parsing algo below
node2comm_temp = -1*torch.ones(num_nodes_connected, device=device).long()
edge_mask_inside = ~(torch.ones(edge_index.shape[1], device=device)).bool()
batch_temp = batch[unmask_nodes]
num_batch_temp = batch_temp.max().item()+1 if batch_temp.shape[0]!=0 else 0
index_comm = 0
iter = 0
# core parsing algo, loop until there is no more dom edges
while edge_index_dom.shape[1]!=0:
node_batch = batch_temp[edge_index_dom[0]]
node_set_idx = scatter(torch.arange(edge_index_dom[0].shape[0], device=device), node_batch, reduce='min') # select top rank edge for each batch
node_set = edge_index_dom[:,node_set_idx].unique()
while True:
iter += 1
edge_mask = np.isin(edge_index_dom[1], node_set)
subset = torch.cat([edge_index_dom[0][edge_mask], node_set]).unique()
batch_set = batch_temp[subset]
index_values_temp = index_comm+torch.arange(batch_set.unique().shape[0], device=device)
index_values = -1*torch.ones(num_batch_temp, device=device).long()
index_values[batch_set.unique()] = index_values_temp
edge_index_dom = edge_index_dom[:,~edge_mask]
node2comm_temp[subset] = index_values[batch_set]
if subset.shape[0]<=node_set.shape[0]:
edge_mask_ = np.isin(edge_index, node_set)
edge_mask_ = edge_mask_[0] & edge_mask_[1]
edge_mask_inside[edge_mask_] = True
index_comm = node2comm_temp.max()+1 if node2comm_temp.shape[0]!=0 else 0
break
else:
node_set = subset
device = device_original
edge_index = edge_index.to(device)
edge_score = edge_score.to(device)
batch = batch.to(device)
mask_nodes = mask_nodes.to(device)
unmask_nodes = unmask_nodes.to(device)
node2comm_temp = node2comm_temp.to(device)
edge_mask_inside = edge_mask_inside.to(device)
edge_index_loop = edge_index_loop.to(device)
edge_score_loop = edge_score_loop.to(device)
node2comm = -1*torch.ones(num_nodes_full, device=device).long()
node2comm[unmask_nodes] = node2comm_temp
node2comm[mask_nodes] = index_comm+torch.arange(mask_nodes.shape[0], device=device)
num_comm = node2comm.max().item()+1 if node2comm.shape[0]!=0 else 0
num_batch = batch.max().item()+1 if batch.shape[0]!=0 else 0
s = SparseTensor(
row = torch.arange(node2comm.shape[0], device=device),
col = node2comm,
sparse_sizes = (num_nodes_full, num_comm)
).to_device(batch.device)
adj_comm_temp = node2comm_temp[edge_index[:,~edge_mask_inside]]
adj_comm, _ = coalesce(adj_comm_temp, None, m=num_comm, n=num_comm)
adj_community = SparseTensor.from_edge_index(
edge_index = adj_comm,
sparse_sizes = (num_comm, num_comm)
).to_device(batch.device)
index_, _ = coalesce(torch.stack([batch, node2comm]), None, m=num_batch, n=num_comm)
new_batch = sort_edge_index(index_, sort_by_row=False)[0]
if new_batch.shape[0]==batch.shape[0]:
flag = False
else:
flag = True
uniques, indexs, counts = torch.unique(batch, return_inverse=True, return_counts=True)
mask1 = (counts==1)[indexs].nonzero().squeeze(dim=-1).tolist()
mask2 = node2comm[mask1].tolist()
edge2comm = node2comm_temp[edge_index[:,edge_mask_inside][0]]
mask_connected_comm = index_to_mask(node2comm[unmask_nodes].unique(), size=new_batch.shape[0])
node_score_temp = global_mean_pool(edge_score[edge_mask_inside], edge2comm)
node_score = torch.ones(new_batch.shape[0], device=device)
node_score[mask_connected_comm] = node_score_temp
node_score[node2comm[edge_index_loop[0]]] = edge_score_loop
uniques, indexs, link_counts_temp = torch.unique(edge2comm, return_inverse=True, return_counts=True)
link_counts = torch.ones(new_batch.shape[0], device=device).float()
link_counts[mask_connected_comm] = link_counts_temp.float()
assert adj_community.is_symmetric()
assert not torch.any(link_counts==0)
return s, adj_community, new_batch, mask1, mask2, flag, node_score, link_counts