-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
187 lines (181 loc) · 7.11 KB
/
model.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
#!git clone https://github.com/ndrplz/ConvLSTM_pytorch
#!git clone https://github.com/ndrplz/ConvLSTM_pytorch
from ConvLSTM_pytorch.convlstm import ConvLSTM
size = 3
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
model1 = EfficientNet.from_pretrained('efficientnet-b3').cuda()
for param in model1.parameters():
param.requires_grad = False
self.model1 = model1
c1 = ConvLSTM( input_size=(size,size),
input_dim=1536,
hidden_dim=[128,512],
kernel_size=(5,5),
num_layers=2,
batch_first = True,
bias=True,
return_all_layers = False)
for param in c1.parameters():
param.requires_grad = False
self.c1 = c1
self.f1 = nn.Linear(512*size*size*5,2048)
self.relu = Swish()
#fast = resnet50().cuda()
#for param in fast.parameters():
# param.requires_grad = False
#self.fast = fast
self.fc2 = nn.Linear(2048,512)
self.fc3 = nn.Linear(512,4)
#self.f3 = nn.Linear(256 + 256,4)
self.dp = nn.Dropout(0.2)
def forward(self, input):
x = self.model1.extract_features(input)
#y = self.fast(input)
x = x.view(-1,150,1536,size,size).detach()
x = self.c1(x)
x1 = []
with torch.no_grad():
for i in range(5):
x1.append(torch.mean(x[0][0][:,i*30:i*30+30,:,:,:],dim = 1))
x1 = torch.stack(x1).cuda()
#x = torch.mean(x[0][0],dim = 1)
x = x1.view(-1,512*size*size*5).detach()
x = self.dp(self.relu(self.f1(x)))
x = self.dp(self.relu(self.fc2(x)))
x = self.dp(self.fc3(x))
return x
model = Model().to('cuda')
model = nn.DataParallel(model)
pretrained_dict = torch.load('./results/effnetb4-fast/save.pth')['state_dict']
# 1. filter out unnecessary keys
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model.state_dict()}
# 2. overwrite entries in the existing state dict
model_dict.update(pretrained_dict)
#model.load_state_dict(torch.load('./results/effnetb4-fast/save.pth')['state_dict'])
from ConvLSTM_pytorch.convlstm import ConvLSTM
size = 3
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
model = models.densenet121(pretrained = True).to('cuda')
modules= list(model.children())[0]
model = nn.Sequential(*modules)
for param in model.parameters():
param.requires_grad = False
self.model = model
self.c1 = ConvLSTM( input_size=(size,size),
input_dim= 1024,
hidden_dim=[128, 512],
kernel_size=(5,5),
num_layers=2,
batch_first = True,
bias=True,
return_all_layers=False)
self.fc1 = nn.Linear(512*size*size,1000)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(1000,256)
self.fc3 = nn.Linear(256,4)
self.dp = nn.Dropout(0.2)
def forward(self, x):
x = self.model(x)
x = x.view(-1,150,1024,size,size).detach()
x = self.c1(x)
x = x[0][0][:,-1,:,:,:]
x = x.view(-1,512*size*size)
x = self.dp(self.relu(self.fc1(x)))
x = self.dp(self.relu(self.fc2(x)))
x = self.dp(self.fc3(x))
return x
size = 4
from torchvision import models
from efficientnet_pytorch import EfficientNet
class SwishImplementation(torch.autograd.Function):
@staticmethod
def forward(ctx, i):
result = i * torch.sigmoid(i)
ctx.save_for_backward(i)
return result
@staticmethod
def backward(ctx, grad_output):
i = ctx.saved_variables[0]
sigmoid_i = torch.sigmoid(i)
return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i)))
class Identity(nn.Module):
def __init__(self):
super(Identity, self).__init__()
def forward(self, x):
return x
class Swish(nn.Module):
@staticmethod
def forward(x):
return SwishImplementation.apply(x)
from ConvLSTM_pytorch.convlstm import ConvLSTM
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
model = models.resnext50_32x4d(pretrained=True)
model.fc = nn.Linear(2048,3)
model.load_state_dict(torch.load('./WACV/best.pth'))
model1 = nn.Sequential(*list(model.children())[:-2])
for param in model1.parameters():
param.requires_grad = False
self.model1 = model1
c1 = ConvLSTM( input_size=(size,size),
input_dim= 2048,
hidden_dim=[128,256],
kernel_size=(3,3),
num_layers=2,
batch_first = True,
bias= True,
return_all_layers = False)
self.c1 = c1
self.attention = nn.Linear(256*size*size,1)
self.f1 = nn.Linear(256*size*size+2048,2048)
self.relu = nn.ELU()
self.fc2 = nn.Linear(2048,512)
# self.fc3 = nn.Linear(2048,512)
self.fc4 = nn.Linear(512,4)
self.dp = nn.Dropout(0.4)
self.sig = nn.Sigmoid()
self.pool = nn.AdaptiveAvgPool2d(output_size = 1)
self.attention1 = nn.Linear(2048,1)
self.beta_attention = nn.Linear(4096,1)
def forward(self, input):
input = input.view(-1,3,112,112)
x = self.model1(input)
x_pool = self.pool(x).reshape(-1,100,2048)
x = x.view(-1,100,2048,size,size)
x = self.c1(x)
x = x[0][0].view(-1,256*size*size)
x = x.view(-1,100,256*size*size)
#x1 = []
x2 = []
v1 = []
for i in range(x_pool.shape[0]):
'''alpha = self.sig(self.attention1(self.dp(x_pool[i])))
x1 = torch.sum(x_pool[i]*alpha,dim = 0)
a1 = torch.sum(alpha , dim = 0)
x1 = torch.div(x1,a1)
x11 = x1.unsqueeze(dim = 0).repeat(100,1)
x3 = torch.cat([x_pool[i],x11],dim = 1)
betas = self.sig(self.beta_attention(self.dp(x3)))
v1.append(x3*alpha*betas)'''
alpha = self.sig(self.attention1(self.dp(x_pool[i])))
#x1.append(torch.div(torch.sum(x[i]*att[i].reshape(-1,1),dim = 0),torch.sum(att[i])))
x2.append(torch.div(torch.sum(x_pool[i]*alpha.reshape(-1,1),dim = 0),torch.sum(alpha,dim = 0)))
x2 = torch.stack(x2)
# v1 = torch.stack(v1)
# v1 = torch.sum(v1,dim = 1)
# v1 = torch.div(v1,torch.sum(alpha*betas,dim = 0))
x1 = torch.mean(x,dim = 1)
x = torch.cat([x1,x2],dim = 1)
# x = self.dp(self.relu(self.fc3(x)))
#x = torch.div(x,torch.sum(att,dim = 1))
x = self.dp(self.relu(self.f1(x)))
x = self.dp(self.relu(self.fc2(x)))
x = self.dp(self.fc4(x))
return x
model = Model().to('cuda')
model = nn.DataParallel(model)