-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathnetworks.py
224 lines (196 loc) · 6.94 KB
/
networks.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
import os
from os.path import join as ospj
from os.path import expanduser
import numpy as np
import torch
import torch.nn as nn
import clip
from tqdm import tqdm
from utils_lavis import *
class BayesCap_MLP(nn.Module):
'''
Baseclass to create a simple MLP
Inputs
inp_dim: int, Input dimension
out_dim: int, Output dimension
hid_dim: int, hidden dimension
num_layers: Number of hidden layers
p_drop: dropout probability
'''
def __init__(
self,
inp_dim,
out_dim,
hid_dim=512,
num_layers=1,
p_drop=0,
):
super(BayesCap_MLP, self).__init__()
mod = []
for layer in range(num_layers):
if layer==0:
incoming = inp_dim
outgoing = hid_dim
mod.append(nn.Linear(incoming, outgoing))
mod.append(nn.ReLU())
elif layer==num_layers//2:
incoming = hid_dim
outgoing = hid_dim
mod.append(nn.Linear(incoming, outgoing))
mod.append(nn.ReLU())
mod.append(nn.Dropout(p=p_drop))
elif layer==num_layers-1:
incoming = hid_dim
outgoing = out_dim
mod.append(nn.Linear(incoming, outgoing))
self.mod = nn.Sequential(*mod)
self.block_mu = nn.Sequential(
nn.Linear(out_dim, out_dim),
nn.ReLU(),
nn.Linear(out_dim, out_dim),
)
self.block_alpha = nn.Sequential(
nn.Linear(out_dim, out_dim),
nn.ReLU(),
# nn.Linear(out_dim, out_dim),
# nn.ReLU(),
nn.Linear(out_dim, out_dim),
nn.ReLU(),
)
self.block_beta = nn.Sequential(
nn.Linear(out_dim, out_dim),
nn.ReLU(),
# nn.Linear(out_dim, out_dim),
# nn.ReLU(),
nn.Linear(out_dim, out_dim),
nn.ReLU(),
)
def forward(self, x):
x_intr = self.mod(x)
# print('dbg', x_intr.shape, x.shape)
x_intr = x_intr + x
x_mu = self.block_mu(x_intr)
x_1alpha = self.block_alpha(x_intr)
x_beta = self.block_beta(x_intr)
return x_mu, x_1alpha, x_beta
class BayesCap_HF_MLP(nn.Module):
'''
Baseclass to create a simple MLP
Inputs
inp_dim: int, Input dimension
out_dim: int, Output dimension
hid_dim: int, hidden dimension
num_layers: Number of hidden layers
p_drop: dropout probability
'''
def __init__(
self,
inp_dim,
out_dim,
hid_dim=512,
num_layers=1,
p_drop=0,
):
super(BayesCap_MLP, self).__init__()
mod = []
for layer in range(num_layers):
if layer==0:
incoming = inp_dim
outgoing = hid_dim
mod.append(nn.Linear(incoming, outgoing))
mod.append(nn.ReLU())
elif layer==num_layers//2:
incoming = hid_dim
outgoing = hid_dim
mod.append(nn.Linear(incoming, outgoing))
mod.append(nn.ReLU())
mod.append(nn.Dropout(p=p_drop))
elif layer==num_layers-1:
incoming = hid_dim
outgoing = out_dim
mod.append(nn.Linear(incoming, outgoing))
self.mod = nn.Sequential(*mod)
self.block_mu = nn.Sequential(
nn.Linear(out_dim, 128),
nn.ReLU(),
nn.Linear(128, out_dim),
)
self.block_alpha = nn.Sequential(
nn.Linear(out_dim, 128),
nn.ReLU(),
# nn.Linear(out_dim, out_dim),
# nn.ReLU(),
nn.Linear(128, out_dim),
nn.ReLU(),
)
self.block_beta = nn.Sequential(
nn.Linear(out_dim, 128),
nn.ReLU(),
# nn.Linear(out_dim, out_dim),
# nn.ReLU(),
nn.Linear(128, out_dim),
nn.ReLU(),
)
def forward(self, x):
x_intr = self.mod(x)
# print('dbg', x_intr.shape, x.shape)
x_intr = x_intr + x
x_mu = self.block_mu(x_intr)
x_1alpha = self.block_alpha(x_intr)
x_beta = self.block_beta(x_intr)
return x_mu, x_1alpha, x_beta
class BayesCLIP(nn.Module):
def __init__(
self,
model_path=None,
device='cuda',
):
super(BayesCLIP, self).__init__()
self.clip_model = load_model(device, model_path)
self.clip_model.eval()
for param in self.clip_model.parameters():
param.requires_grad = False
self.img_BayesCap = BayesCap_MLP(inp_dim=512, out_dim=512, hid_dim=512, num_layers=3, p_drop=0.3).to(device)
self.txt_BayesCap = BayesCap_MLP(inp_dim=512, out_dim=512, hid_dim=512, num_layers=3, p_drop=0.3).to(device)
def forward(self, i_inputs, t_inputs):
i_features, t_features = self.clip_model(i_inputs, t_inputs)
img_mu, img_1alpha, img_beta = self.img_BayesCap(i_features)
txt_mu, txt_1alpha, txt_beta = self.txt_BayesCap(t_features)
return (img_mu, img_1alpha, img_beta), (txt_mu, txt_1alpha, txt_beta), (i_features, t_features)
class BayesCap_for_CLIP(nn.Module):
def __init__(
self,
inp_dim=512,
out_dim=512,
hid_dim=256,
num_layers=3,
p_drop=0.1,
):
super(BayesCap_for_CLIP, self).__init__()
self.img_BayesCap = BayesCap_MLP(inp_dim=inp_dim, out_dim=out_dim, hid_dim=hid_dim, num_layers=num_layers, p_drop=p_drop)
self.txt_BayesCap = BayesCap_MLP(inp_dim=inp_dim, out_dim=out_dim, hid_dim=hid_dim, num_layers=num_layers, p_drop=p_drop)
def forward(self, i_features, t_features):
# print('dbg', i_features.shape, t_features.shape)
img_mu, img_1alpha, img_beta = self.img_BayesCap(i_features)
txt_mu, txt_1alpha, txt_beta = self.txt_BayesCap(t_features)
return (img_mu, img_1alpha, img_beta), (txt_mu, txt_1alpha, txt_beta)
class BayesCap_for_HF_CLIP(nn.Module):
def __init__(
self,
# inp_i_dim=512,
# out_i_dim=512,
# hid_i_dim=256,
inp_t_dim=512,
out_t_dim=512,
hid_t_dim=256,
num_layers=3,
p_drop=0.1,
):
super(BayesCap_for_HF_CLIP, self).__init__()
# self.img_BayesCap = BayesCap_MLP(inp_dim=inp_i_dim, out_dim=out_i_dim, hid_dim=hid_i_dim, num_layers=num_layers, p_drop=p_drop)
self.txt_BayesCap = BayesCap_MLP(inp_dim=inp_t_dim, out_dim=out_t_dim, hid_dim=hid_t_dim, num_layers=num_layers, p_drop=p_drop)
def forward(self, i_features, t_features):
# img_mu, img_1alpha, img_beta = self.img_BayesCap(i_features)
txt_mu, txt_1alpha, txt_beta = self.txt_BayesCap(t_features)
# return (img_mu, img_1alpha, img_beta), (txt_mu, txt_1alpha, txt_beta)
return (None, None, None), (txt_mu, txt_1alpha, txt_beta)