-
Notifications
You must be signed in to change notification settings - Fork 1
/
gnn_models.py
322 lines (289 loc) · 16.8 KB
/
gnn_models.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
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
import math
import torch
from tqdm import tqdm
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.parameter import Parameter
from torch.nn.modules.module import Module
torch.manual_seed(3)
import scipy
import dgl.nn.pytorch.conv as dglnn
from torch import nn
from scipy.special import comb
import math
np.random.seed(1)
import dgl.function as fn
import dgl
import dgl.nn.pytorch.conv as dglnn
from torch import nn
from dgl.nn import AvgPooling,MaxPooling,WeightAndSum,GlobalAttentionPooling,SortPooling
from utils import *
from torch.autograd import Variable
device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")
def log(*pargs):
with open('/apdcephfs/share_1364275/kaithgao/RLNDOCK_CODE/my_log/inference.txt', 'a+') as w:
w.write(" ".join(["{}".format(t) for t in pargs]))
w.write("\n")
class GNN_oracle(nn.Module):
def __init__(self, in_feats=13, h_feats=32,cls_h = 256, num_layers=4, dropout_rate=0.,
activation='ELU',GNN='gin',pooling = 'mean'):
super().__init__()
self.cls1 = nn.Linear(13, cls_h)
self.cls2 = nn.Linear(cls_h, 1)
self.h_feats = h_feats
self.layers = nn.ModuleList()
self.act = getattr(nn, activation)()
self.lm = nn.LayerNorm(h_feats)
self.lm_last = nn.LayerNorm(13)
if GNN=='gcn':
self.layers.append(dglnn.GraphConv(h_feats, h_feats, activation=self.act, allow_zero_in_degree=True))
for i in range(num_layers-1):
self.layers.append(dglnn.GraphConv(h_feats, h_feats, activation=self.act, allow_zero_in_degree=True))
self.layers.append(dglnn.GraphConv(h_feats, 13, activation=self.act, allow_zero_in_degree=True))
elif GNN=='gin':
self.layers.append(dglnn.GINConv(torch.nn.Linear(h_feats,h_feats),'max',learn_eps=True))
for i in range(num_layers-1):
self.layers.append(dglnn.GINConv(torch.nn.Linear(h_feats,h_feats),'max',learn_eps=True))
self.layers.append(dglnn.GINConv(torch.nn.Linear(h_feats,13),'max',learn_eps=True))
self.mlp_before = MLP_before(in_feats, h_feats, dropout_rate)
self.mlp = MLP_oracle(h_feats, h_feats, dropout_rate)
self.dropout = nn.Dropout(dropout_rate) if dropout_rate > 0 else nn.Identity()
if pooling == 'mean':
self.pool = AvgPooling()
elif pooling == 'max':
self.pool = MaxPooling()
elif pooling == 'weight':
self.pool = WeightAndSum(64)
elif pooling == 'sort':
self.pool = SortPooling(k=2)
elif pooling == 'attention':
self.pool = GlobalAttentionPooling(torch.nn.Linear(64, 1))
def reset_parameters(self):
if self.layer_1.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_1.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
if self.layer_2.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_2.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
def forward(self, graph, edge_weight = None):
# h = torch.cat((graph.ndata['mu_r_norm'],graph.ndata['res_feat']),dim=1)
h = self.mlp_before(graph.ndata['features'])
for i, layer in enumerate(self.layers):
if i != 0:
h = self.dropout(h)
if i != len(self.layers)-1:
h = self.lm(layer(graph, h, edge_weight = edge_weight))
else:
h = self.lm_last(layer(graph, h, edge_weight = edge_weight))
nodes_emb = h.clone()
h = self.pool(graph, h)
# h = self.pool(graph, self.mlp(h))
return self.cls2(self.act(self.cls1(h))), nodes_emb
def forward_action(self, edge_index_all, edge_weight_all, feature_matrix_all):
g = dgl.DGLGraph((np.array(edge_index_all).tolist()[0],np.array(edge_index_all).tolist()[1])).to(device)
g.ndata['features'] = feature_matrix_all
# pred_rmsd = self.forward(dgl.add_self_loop(g), edge_weight = edge_weight_all)
pred_rmsd,_ = self.forward(g, edge_weight = edge_weight_all)
return pred_rmsd
def update_complex(self, complex_rep_clone, edge_index_all, edge_weight_init, feature_matrix_init):
device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")
N_now, feat_dim = feature_matrix_init.shape[0], feature_matrix_init.shape[1]
complex_rep_clone = complex_rep_clone.to(device)
edge_weight_init = Variable(edge_weight_init, requires_grad = False)
edge_weight_expend = Variable(torch.rand(N_now).to(device), requires_grad = True)
feature_matrix_expend = Variable(torch.softmax(torch.rand((1,complex_rep_clone.shape[0])),0).to(device), requires_grad = True)
complex_rep_clone = Variable(complex_rep_clone, requires_grad = False)
# opt_action = torch.optim.Adam([edge_weight_expend, feature_matrix_expend], lr=0.01, betas=(0.9, 0.99))
opt_action = torch.optim.Adam(
[
{'params': edge_weight_expend,'lr': 1e-2},
{'params': feature_matrix_expend, 'lr': 1e-1}
]
)
for i in range(200): ### 5k 10k
edge_weight_all = torch.cat((edge_weight_init, torch.softmax(edge_weight_expend.to(device),0)),dim=0)
# edge_weight_all = torch.cat((edge_weight_init, edge_weight_expend.to(device)),dim=0)
# feature_matrix_all = torch.cat((feature_matrix_init, (torch.softmax(feature_matrix_expend,0) @ complex_rep_clone).to(device)),dim=0)
feature_matrix_all = torch.cat((feature_matrix_init, (torch.softmax(feature_matrix_expend,1) @ complex_rep_clone).to(device)),dim=0)
pred_rmsd = self.forward_action(edge_index_all, edge_weight_all, feature_matrix_all).abs()
if i % 20 ==0:
print(np.array(pred_rmsd.detach().cpu()))
opt_action.zero_grad()
pred_rmsd.backward()
opt_action.step()
# print(torch.softmax(feature_matrix_expend,1))
# print(edge_weight_expend)
# return edge_weight_expend, f_exp
return edge_weight_expend, feature_matrix_expend @ complex_rep_clone
class MLP_before(nn.Module):
def __init__(self, in_feats, h_feats, dropout_rate=0.2, activation='ELU'):
super(MLP_before, self).__init__()
self.layers = nn.ModuleList()
self.act = getattr(nn, activation)()
ra = 4
self.layers_1 = nn.Linear(in_feats, h_feats*ra)
self.layers_2 = nn.Linear(h_feats*ra, h_feats)
self.lm_1 = nn.LayerNorm(h_feats*ra)
self.lm_2 = nn.LayerNorm(h_feats)
self.dropout = nn.Dropout(dropout_rate) if dropout_rate > 0 else nn.Identity()
def reset_parameters(self):
if self.layer_1.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_1.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
if self.layer_2.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_2.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
def forward(self, h):
h = self.dropout(self.lm_1(self.act(self.layers_1(h))))
h = self.dropout(self.lm_2(self.act(self.layers_2(h))))
return h
class MLP_oracle(nn.Module):
def __init__(self, in_feats, h_feats, dropout_rate=0.2, activation='ELU'):
super(MLP_oracle, self).__init__()
self.layers = nn.ModuleList()
self.act = getattr(nn, activation)()
self.layers_1 = nn.Linear(in_feats, h_feats)
self.layers_2 = nn.Linear(h_feats, h_feats)
self.lm = nn.LayerNorm(h_feats)
self.dropout = nn.Dropout(dropout_rate) if dropout_rate > 0 else nn.Identity()
def reset_parameters(self):
if self.layer_1.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_1.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
if self.layer_2.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_2.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
def forward(self, h):
h = self.lm(self.act(self.layers_1(h)))
h = self.dropout(h)
h = self.lm(self.act(self.layers_2(h)))
h = self.dropout(h)
return h
class MLP_oracle_BWGNN(nn.Module):
def __init__(self, in_feats, h_feats, dropout_rate=0.2, activation='ELU'):
super(MLP_oracle_BWGNN, self).__init__()
self.layers = nn.ModuleList()
self.act = getattr(nn, activation)()
self.layers_1 = nn.Linear(in_feats*4, h_feats)
self.layers_2 = nn.Linear(h_feats, h_feats)
self.lm = nn.LayerNorm(h_feats)
self.dropout = nn.Dropout(dropout_rate) if dropout_rate > 0 else nn.Identity()
def reset_parameters(self):
if self.layer_1.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_1.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
if self.layer_2.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_2.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
def forward(self, h):
h = self.dropout(h)
h = self.lm(self.act(self.layers_1(h)))
h = self.dropout(h)
h = self.lm(self.act(self.layers_2(h)))
h = self.dropout(h)
h = self.lm(self.act(self.layers_2(h)))
return h
class target_model(nn.Module):
def __init__(self, in_feats, h_feats, f_feats, dropout_rate=0.2, activation='ELU'):
super(target_model, self).__init__()
self.layers = nn.ModuleList()
self.act = getattr(nn, activation)()
self.layers_1 = nn.Linear(in_feats, h_feats)
self.layers_2 = nn.Linear(h_feats, h_feats)
self.layers_3 = nn.Linear(h_feats, f_feats)
self.layers_4 = nn.Linear(h_feats, h_feats)
self.layers_5 = nn.Linear(h_feats, f_feats)
self.lm1 = nn.LayerNorm(h_feats)
self.lm2 = nn.LayerNorm(h_feats)
self.lm3 = nn.LayerNorm(f_feats)
self.lm4 = nn.LayerNorm(h_feats)
self.lm5 = nn.LayerNorm(f_feats)
self.dropout = nn.Dropout(dropout_rate) if dropout_rate > 0 else nn.Identity()
self.c_a = CrossAttentionLayer(13,13)
def reset_parameters(self):
if self.layer_1.weight is not None:
nn.init.xavier_normal_(self.layers_1.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_1.bias is not None:
nn.init.constant_(self.layer_1.bias.data, 0.0)
if self.layer_2.weight is not None:
nn.init.xavier_normal_(self.layers_2.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_2.bias is not None:
nn.init.constant_(self.layer_2.bias.data, 0.0)
if self.layer_3.weight is not None:
nn.init.xavier_normal_(self.layers_3.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_3.bias is not None:
nn.init.constant_(self.layer_3.bias.data, 0.0)
if self.layer_4.weight is not None:
nn.init.xavier_normal_(self.layers_4.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_4.bias is not None:
nn.init.constant_(self.layer_4.bias.data, 0.0)
if self.layer_5.weight is not None:
nn.init.xavier_normal_(self.layers_5.weight, gain=0.02) # Implement Xavier Uniform
if self.layer_5.bias is not None:
nn.init.constant_(self.layer_5.bias.data, 0.0)
def forward(self, h):
# h1 = self.c_a(h[0,:].unsqueeze(0), h[1,:].unsqueeze(0), [1], [1])
# h0 = self.c_a(h[1,:].unsqueeze(0), h[0,:].unsqueeze(0), [1], [1])
# h = torch.cat((h0,h1),0)
h_clone = h.clone()
for b in range(int(h.size(0)/2)):
h_clone[[b*2,b*2+1]] = h_clone[[b*2+1,b*2]]
h = torch.cat((h,h_clone),1)
h = self.dropout(h)
h = self.lm1(self.act(self.layers_1(h)))
h = self.dropout(h)
h = self.lm2(self.act(self.layers_2(h)))
h = self.dropout(h)
h = self.lm3(self.act(self.layers_3(h)))
# h = self.lm4(self.act(self.layers_4(h)))
# h = self.lm5(self.act(self.layers_5(h)))
return self.act(h)
class CrossAttentionLayer(nn.Module):
def __init__(self,h_dim,num_heads):
super(CrossAttentionLayer,self).__init__()
self.h_dim = h_dim
self.h_dim_div = self.h_dim // 1
self.num_heads = num_heads
assert self.h_dim_div % self.num_heads == 0
self.head_dim = self.h_dim_div // self.num_heads
self.merge = nn.Conv1d(self.h_dim_div, self.h_dim_div, kernel_size=1)
self.proj = nn.ModuleList([nn.Conv1d(self.h_dim, self.h_dim_div, kernel_size=1) for _ in range(3)])
dropout = 0.
self.mlp = nn.Sequential(
nn.Linear(self.h_dim+self.h_dim_div, self.h_dim),
nn.Dropout(dropout),
nn.LayerNorm(self.h_dim),
nn.ReLU(),
nn.Linear(self.h_dim, self.h_dim),
nn.Dropout(dropout),
nn.LayerNorm(self.h_dim)
)
def forward(self, src_h, dst_h, src_num_verts, dst_num_verts):
h = dst_h
src_h_list = torch.split(src_h, src_num_verts)
dst_h_list = torch.split(dst_h, dst_num_verts)
h_msg = []
for idx in range(len(src_num_verts)):
src_hh = src_h_list[idx].unsqueeze(0).transpose(1, 2)
dst_hh = dst_h_list[idx].unsqueeze(0).transpose(1, 2)
query, key, value = [hh.view(1, self.head_dim, self.num_heads, -1) \
for ll, hh in zip(self.proj, (dst_hh, src_hh, src_hh))]
dim = query.shape[1]
scores = torch.einsum('bdhn,bdhm->bhnm', query, key) / (dim ** 0.5)
attn = F.softmax(scores, dim=-1)
h_dst = torch.einsum('bhnm,bdhm->bdhn', attn, value)
h_dst = h_dst.contiguous().view(1, self.h_dim_div, -1)
h_msg.append(h_dst.squeeze(0).transpose(0, 1))
h_msg = torch.cat(h_msg, dim=0)
# skip connection
h_out = h + self.mlp(torch.cat((h, h_msg), dim=-1))
return h_out