-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathNNFEA_net_x_c_trans_new.py
333 lines (303 loc) · 11.1 KB
/
NNFEA_net_x_c_trans_new.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
323
324
325
326
327
328
329
330
331
332
333
import torch
import torch.nn as nn
import torch.nn.functional as nnF
from NNFEA_net_base import LinearRes, Sin, preprocess
from NNFEA_net_base import BaseNet0, BaseNet1, BaseNet1a, BaseNet2, BaseNet2a, BaseNet2b, BaseNet3, BaseNet3b
from NNFEA_net_base import BaseNet4, BaseNet4a, BaseNet4b, BaseNet5a, BaseNet5b
from NNFEA_net_base import BaseNet6a, BaseNet6b
#%%
class AttentionBlock(nn.Module):
def __init__(self, embed_dim, num_heads):
super().__init__()
self.embed_dim=embed_dim
self.num_heads=num_heads
self.MHA=nn.MultiheadAttention(embed_dim, num_heads,
dropout=0.0,
bias=True,
add_bias_kv=False,
add_zero_attn=False,
kdim=None,
vdim=None,
batch_first=True,
device=None,
dtype=None)
self.MLP=nn.Sequential(nn.Linear(embed_dim, 2*embed_dim),
nn.GELU(),
nn.Linear(2*embed_dim, embed_dim)
)
self.Norm1q=nn.LayerNorm(embed_dim)
self.Norm1k=nn.LayerNorm(embed_dim)
self.Norm2=nn.LayerNorm(embed_dim)
def forward(self, q, k, v):
#x.shape is (B, L, E), E=embed_dim
#usually, the shape of a feature map is (B,C,H,W)
#assume permute and reshape/view have been applied, then L=H*W and C=E=embed_dim
#--------------------------------------------
x=q+self.MHA(self.Norm1q(q), self.Norm1k(k), v)[0]#(B,L,E)
#add&norm
y=x+self.MLP(self.Norm2(x))
return y #(B,L,E)
#%%
class SelfAttentionBlock(nn.Module):
def __init__(self, embed_dim, num_heads):
super().__init__()
self.embed_dim=embed_dim
self.num_heads=num_heads
self.MHA=nn.MultiheadAttention(embed_dim, num_heads,
dropout=0.0,
bias=True,
add_bias_kv=False,
add_zero_attn=False,
kdim=None,
vdim=None,
batch_first=True,
device=None,
dtype=None)
self.MLP=nn.Sequential(nn.Linear(embed_dim, 2*embed_dim),
nn.GELU(),
nn.Linear(2*embed_dim, embed_dim)
)
self.Norm1=nn.LayerNorm(embed_dim)
self.Norm2=nn.LayerNorm(embed_dim)
def forward(self, x):
#x.shape is (B, L, E), E=embed_dim
#usually, the shape of a feature map is (B,C,H,W)
#assume permute and reshape/view have been applied, then L=H*W and C=E=embed_dim
#--------------------------------------------
x1=self.Norm1(x)
x2=x+self.MHA(x1,x1,x1)[0]#(B,L,E)
#add&norm
y=x2+self.MLP(self.Norm2(x2))
return y #(B,L,E)
#%%
class TransEncoder4(nn.Module):
def __init__(self, x_layers, h_dim, n_layers, num_heads, alpha=1):
super().__init__()
self.x_layers=x_layers
self.h_dim=h_dim
self.n_layers=n_layers
#not good compared to a single linear
#self.patch_layer=nn.Sequential(Sin(50*3*x_layers, h_dim, alpha, True),
# Sin(h_dim, h_dim, 1, False),
# nn.Linear(h_dim, h_dim))
#self.patch_layer=nn.Sequential(Sin(50*3*x_layers, h_dim, alpha, True),
# nn.Linear(h_dim, h_dim))
self.patch_layer=nn.Linear(50*3*x_layers, h_dim)
self.netT=nn.Sequential()
for n in range(0, n_layers):
self.netT.append(SelfAttentionBlock(h_dim, num_heads))
self.out=nn.Sequential(nn.LayerNorm(h_dim), #better to use norm
nn.Linear(h_dim, 50*3*x_layers))
#not good, compared to sin-sin-linear
#self.pos=nn.Parameter(0.02*torch.randn(100, h_dim))
#not good compared to sin-sin-linear
#self.pos=nn.Linear(3, h_dim)
self.pos=nn.Sequential(Sin(3, h_dim, alpha, True),
Sin(h_dim, h_dim, 1, False),
nn.Linear(h_dim, h_dim))
def get_patch(self, x):
#x.shape (N, x_dim), N=5000*2
y=x.view(-1,5000,3)
x_layers=y.shape[0]
y=y.reshape(x_layers,100,50,3)
y=y.permute(1,0,2,3) #(100,x_layers,50,3)
y=y.reshape(1,100,-1)
return y
def un_patch(sefl, y):
#y.shape (1,100,-1)
y=y.view(100,-1).view(100,-1,50,3)
y=y.permute(1,0,2,3)
y=y.reshape(-1,3)
return y
def forward(self, x1, x2):
#x1.shape (N, x_dim) x-meanshape
#x2.shape (N, x_dim) meanshape
#---------------
#"Patch"
#print('0', x1.shape)
x1=self.get_patch(x1)
#print('1', x.shape)
x1=self.patch_layer(x1) #(100, h_dim)
#---------------
if torch.is_tensor(self.pos):
x2=self.pos
else:
x2=self.get_patch(x2).reshape(100,-1,3).mean(dim=1)
x2=self.pos(x2)
#---------------
#pos embedding
x=x1+x2
#print(x.shape)
#-------------
x=self.netT(x)
y=self.out(x) #(100, 50*x_layers)
#print('2', y.shape)
y=self.un_patch(y)
return y
#%%
class TransEncoder4A(nn.Module):
def __init__(self, x_layers, h_dim, n_layers, num_heads, alpha=1):
super().__init__()
self.x_layers=x_layers
self.h_dim=h_dim
self.n_layers=n_layers
self.netA=Sin(6, h_dim, alpha=1, is_first=True)
self.patch_layer=nn.Linear(50*x_layers*6, h_dim)
self.netT=nn.Sequential()
for n in range(0, n_layers):
self.netT.append(SelfAttentionBlock(h_dim, num_heads))
self.out=nn.Sequential(nn.LayerNorm(h_dim), #better to use norm
nn.Linear(h_dim, 50*3*x_layers))
#not good, compared to sin-sin-linear
#self.pos=nn.Parameter(0.02*torch.randn(100, h_dim))
#not good compared to sin-sin-linear
#self.pos=nn.Linear(3, h_dim)
def get_patch(self, x):
#x.shape (N, x_dim), N=5000*2
y=x.view(-1,5000,3)
x_layers=y.shape[0]
y=y.reshape(x_layers,100,50,3)
y=y.permute(1,0,2,3) #(100,x_layers,50,3)
y=y.reshape(1,100,-1)
return y
def un_patch(sefl, y):
#y.shape (1,100,-1)
y=y.view(100,-1).view(100,-1,50,3)
y=y.permute(1,0,2,3)
y=y.reshape(-1,3)
return y
def forward(self, x1, x2):
#x1.shape (N, x_dim) x-meanshape
#x2.shape (N, x_dim) meanshape
#---------------
#"Patch"
#print('0', x1.shape)
x=torch.cat([x1,x2], dim=1)
#x=self.netA(x)
x=self.get_patch(x, self.x_layers)
#print('1', x.shape)
x=self.patch_layer(x) #(100, h_dim)
#-------------
x=x.view(1,x.shape[0],x.shape[1])
x=self.netT(x)
y=self.out(x) #(100, 50*x_layers)
#print('2', y.shape)
y=y.view(-1,3)
return y
#%%
class TransDecoder4(nn.Module):
def __init__(self,):
super().__init__()
self.p=nn.Parameter(torch.zeros(1))
def forward(self, x, c):
#x: meanshape
#c: code
return c
#%%
class TransEncoder5(nn.Module):
def __init__(self, x_layers, h_dim, n_layers, num_heads, alpha=1):
super().__init__()
self.x_layers=x_layers
self.h_dim=h_dim
self.n_layers=n_layers
self.patch_layer=nn.Linear(50*3*x_layers, h_dim)
self.netT=nn.Sequential()
for n in range(0, n_layers):
self.netT.append(SelfAttentionBlock(h_dim, num_heads))
self.out=nn.Linear(h_dim, h_dim)
self.pos=nn.Sequential(Sin(3, h_dim, alpha, True),
Sin(h_dim, h_dim, 1, False),
nn.Linear(h_dim, h_dim))
def get_patch(self, x):
#x.shape (N, x_dim), N=5000*2
y=x.view(-1,5000,3)
x_layers=y.shape[0]
y=y.reshape(x_layers,100,50,3)
y=y.permute(1,0,2,3) #(100,x_layers,50,3)
y=y.reshape(1,100,-1)
return y
def forward(self, x1, x2):
#x1.shape (N, x_dim) x-meanshape
#x2.shape (N, x_dim) meanshape
#---------------
#"Patch"
#print('0', x1.shape)
x=self.get_patch(x1)
#print('1', x.shape)
x=self.patch_layer(x) #(100, h_dim)
#---------------
#pos embedding
if torch.is_tensor(self.pos):
x2=self.pos
else:
x2=self.get_patch(x2).reshape(100,-1,3).mean(dim=1)
x2=self.pos(x2)
#-------------
x=self.netT(x)
y=self.out(x) #(100, h_dim)
return y
#%%
class TransDecoder5(nn.Module):
def __init__(self, x_layers, h_dim, n_layers, num_heads):
super().__init__()
self.x_layers=x_layers
self.h_dim=h_dim
self.n_layers=n_layers
self.netQ=eval("BaseNet5b("
+'x_dim=3'
+',h_dim=128'
+',n_layers=2'
+',x_scale=1'
+',alpha=1'
+',gamma=1'
+',y_dim='+str(h_dim)
+")"
)
self.attnA=nn.ModuleList()
self.attnB=nn.ModuleList()
for n in range(0, n_layers):
self.attnA.append(AttentionBlock(h_dim, num_heads))
self.attnB.append(SelfAttentionBlock(h_dim, num_heads))
self.out=nn.Linear(h_dim, 3)
def forward(self, x, c):
#x: meanshape
#c: code
#print(c.shape)
Q=self.netQ(x) #Q.shape (N, h_dim)
Q=Q.view(1,Q.shape[0],Q.shape[1])
V=K=c
for n in range(self. n_layers):
A=self.attnA[n](Q,K,V)
B=self.attnB[n](A)
Q=B
y=self.out(B)
y=y.view(-1,3)
return y
#%%
class TransDecoder5A(nn.Module):
def __init__(self, x_layers, h_dim, n_layers, num_heads):
super().__init__()
self.x_layers=x_layers
self.h_dim=h_dim
self.n_layers=n_layers
self.netQ=nn.Linear(3, h_dim)
self.attnA=nn.ModuleList()
self.attnB=nn.ModuleList()
for n in range(0, n_layers):
self.attnA.append(AttentionBlock(h_dim, num_heads))
self.attnB.append(SelfAttentionBlock(h_dim, num_heads))
self.out=nn.Linear(h_dim, 3)
def forward(self, x, c):
#x: meanshape
#c: code
#print(c.shape)
Q=self.netQ(x) #Q.shape (N, h_dim)
Q=Q.view(1,Q.shape[0],Q.shape[1])
V=K=c
for n in range(self. n_layers):
A=self.attnA[n](Q,K,V)
B=self.attnB[n](A)
Q=B
y=self.out(B)
y=y.view(-1,3)
return y