-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy pathMultiTrans.py
196 lines (152 loc) · 7.93 KB
/
MultiTrans.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
import torch
from torch import nn
import torch.nn.functional as F
from einops import rearrange, repeat
from einops.layers.torch import Rearrange
from modules.transformer import TransformerEncoder
class MULTModel(nn.Module):
def __init__(self, img_H=256, img_W=192, patch_size = 32, dim = 1024):
"""
"""
super(MULTModel, self).__init__()
assert img_H % patch_size == 0, 'Image dimensions must be divisible by the patch size H.'
assert img_W % patch_size == 0, 'Image dimensions must be divisible by the patch size W.'
num_patches = (img_H // patch_size) * (img_W // patch_size) # (256 / 32) * (192 / 32) = 48
patch_dim_22 = 22 * patch_size * patch_size # 22 * 32 * 32 = 22528
patch_dim_3 = 3 * patch_size * patch_size # 3 * 32 * 32 = 3072
patch_dim_1 = 1 * patch_size * patch_size # 1 * 32 * 32 = 1024
self.to_patch_embedding_22 = nn.Sequential(
# [B, 22, 256, 192] -> [B, 22, 8 * 32, 6 * 32] -> [B, 8 * 6, 32 * 32 * 22]
Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_size, p2 = patch_size),
# [B, 48, 32*32*22] -> [B, 48, 2048]
nn.Linear(patch_dim_22, 11264),
)
self.to_patch_embedding_3 = nn.Sequential(
# [B, 3, 256, 192] -> [B, 3, 8 * 32, 6 * 32] -> [B, 8 * 6, 32 * 32 * 3]
Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_size, p2 = patch_size),
# [B, 48, 3072] -> [B, 48, 1024]
nn.Linear(patch_dim_3, dim),
)
self.to_patch_embedding_1 = nn.Sequential(
# [B, 3, 256, 192] -> [B, 3, 8 * 32, 6 * 32] -> [B, 8 * 6, 32 * 32 * 3]
Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_size, p2 = patch_size),
# [B, 48, 3072] -> [B, 48, 1024]
nn.Linear(patch_dim_1, dim),
)
# [B, 48, 32 * 32 * 26]
self.backRearrange = nn.Sequential(Rearrange('b (h w) (p1 p2 c) -> b c (h p1) (w p2)', p1=patch_size, p2=patch_size, h=8, w=6))
self.d_l, self.d_a, self.d_v = 1024, 1024, 1024
combined_dim = self.d_l + self.d_a + self.d_v
output_dim = 32 * 32 * 26
self.num_heads = 8
self.layers = 3
self.attn_dropout = nn.Dropout(0.1)
self.attn_dropout_a = nn.Dropout(0.0)
self.attn_dropout_v = nn.Dropout(0.0)
self.relu_dropout = nn.Dropout(0.1)
self.embed_dropout = nn.Dropout(0.25)
self.res_dropout = nn.Dropout(0.1)
self.attn_mask = True
# 2. Crossmodal Attentions
# if self.lonly:
self.trans_l_with_a = self.get_network(self_type='la')
self.trans_l_with_v = self.get_network(self_type='lv')
# if self.aonly:
self.trans_a_with_l = self.get_network(self_type='al')
self.trans_a_with_v = self.get_network(self_type='av')
# if self.vonly:
self.trans_v_with_l = self.get_network(self_type='vl')
self.trans_v_with_a = self.get_network(self_type='va')
# 3. Self Attentions (Could be replaced by LSTMs, GRUs, etc.)
# [e.g., self.trans_x_mem = nn.LSTM(self.d_x, self.d_x, 1)
self.trans_l_mem = self.get_network(self_type='l_mem', layers=3)
self.trans_a_mem = self.get_network(self_type='a_mem', layers=3)
self.trans_v_mem = self.get_network(self_type='v_mem', layers=3)
# Projection layers
self.proj1 = nn.Linear(6144, 6144)
self.proj2 = nn.Linear(6144, 6144)
self.out_layer = nn.Linear(6144, output_dim)
self.projConv1 = nn.Conv1d(11264, 1024, kernel_size=1, padding=0, bias=False)
self.projConv2 = nn.Conv1d(1024, 1024, kernel_size=1, padding=0, bias=False)
self.projConv3 = nn.Conv1d(1024, 1024, kernel_size=1, padding=0, bias=False)
def get_network(self, self_type='l', layers=-1):
if self_type in ['l', 'al', 'vl']:
embed_dim, attn_dropout = self.d_l, self.attn_dropout
elif self_type in ['a', 'la', 'va']:
embed_dim, attn_dropout = self.d_a, self.attn_dropout_a
elif self_type in ['v', 'lv', 'av']:
embed_dim, attn_dropout = self.d_v, self.attn_dropout_v
elif self_type == 'l_mem':
embed_dim, attn_dropout = self.d_l, self.attn_dropout
elif self_type == 'a_mem':
embed_dim, attn_dropout = self.d_a, self.attn_dropout
elif self_type == 'v_mem':
embed_dim, attn_dropout = self.d_v, self.attn_dropout
else:
raise ValueError("Unknown network type")
return TransformerEncoder(embed_dim=embed_dim,
num_heads=self.num_heads,
layers=max(self.layers, layers),
attn_dropout=attn_dropout,
relu_dropout=self.relu_dropout,
res_dropout=self.res_dropout,
embed_dropout=self.embed_dropout,
attn_mask=self.attn_mask)
def forward(self, x1, x2, x3):
# Input:
# x1: [B, 22, 256, 192]
# x2: [B, 3, 256, 192]
# x3: [B, 1, 256, 192]
# Step1: patch_embedding
x1 = self.to_patch_embedding_22(x1) # [B, 22, 256, 192] -> [B, 48, 11264]
x2 = self.to_patch_embedding_3(x2) # [B, 3, 256, 192] -> [B, 48, 1024]
x3 = self.to_patch_embedding_1(x3) # [B, 1, 256, 192] -> [B, 48, 1024]
# Step2: Project & 1D Conv & Permute
# [B, 48, 1024] -> [B, 1024, 48]
x1 = x1.transpose(1, 2) # [B, 11264, 48]
x2 = x2.transpose(1, 2)
x3 = x3.transpose(1, 2)
# [1024]
proj_x1 = self.projConv1(x1)
proj_x2 = self.projConv2(x2)
proj_x3 = self.projConv3(x3)
# [48, B, 1024]
proj_x1 = proj_x1.permute(2, 0, 1)
proj_x2 = proj_x2.permute(2, 0, 1)
proj_x3 = proj_x3.permute(2, 0, 1)
# Self_att first [48, B, 1024]
proj_x1_trans = self.trans_l_mem(proj_x1)
proj_x2_trans = self.trans_a_mem(proj_x2)
proj_x3_trans = self.trans_v_mem(proj_x3)
# Step3: Cross Attention
# (x3,x2) --> x1
h_l_with_as = self.trans_l_with_a(proj_x1, proj_x2_trans, proj_x2_trans) # Dimension (L, N, d_l) [48, B, 1024]
h_l_with_vs = self.trans_l_with_v(proj_x1, proj_x3_trans, proj_x3_trans) # Dimension (L, N, d_l) [48, B, 1024]
cross1 = torch.cat([h_l_with_as, h_l_with_vs], 2) # [2048]
# (x1, x3) --> x2
h_a_with_ls = self.trans_a_with_l(proj_x2, proj_x1_trans, proj_x1_trans)
h_a_with_vs = self.trans_a_with_v(proj_x2, proj_x3_trans, proj_x3_trans)
cross2 = torch.cat([h_a_with_ls, h_a_with_vs], 2)
# (x1,x2) --> x3
h_v_with_ls = self.trans_v_with_l(proj_x3, proj_x1_trans, proj_x1_trans)
h_v_with_as = self.trans_v_with_a(proj_x3, proj_x2_trans, proj_x2_trans)
cross3 = torch.cat([h_v_with_ls, h_v_with_as], 2)
# Combine by cat
# 三个[48, B, 2048] -> [48, B, 6144]
# last_hs = torch.cat([last_h_l, last_h_a, last_h_v], dim=1#[N,6144]
last_hs = torch.cat([cross1, cross2, cross3], dim=2) #[48, B, 6144]
# A residual block
decompo_1 = self.proj1(last_hs)
decompo_1_relu = F.relu(decompo_1)
last_hs_proj = self.proj2(decompo_1_relu)
last_hs_proj += last_hs
# last_hs_proj = self.proj2(F.dropout(F.relu(self.proj1(last_hs)), p=self.out_dropout, training=self.training))
# last_hs_proj += last_hs
output = self.out_layer(last_hs_proj) # [48, B, 26624(26 * 32 * 32)]
output = output.permute(1, 0, 2) # [B, 8 * 6, 32 * 32 * 26]
output = self.backRearrange(output)
return output
if __name__ == '__main__':
encoder = TransformerEncoder(300, 4, 2)
x = torch.tensor(torch.rand(20, 2, 300))
print(encoder(x).shape)