-
Notifications
You must be signed in to change notification settings - Fork 26
/
utils_models.py
132 lines (107 loc) · 5.48 KB
/
utils_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
from utils_libs import *
import torchvision.models as models
class client_model(nn.Module):
def __init__(self, name, args=True):
super(client_model, self).__init__()
self.name = name
if self.name == 'Linear':
[self.n_dim, self.n_out] = args
self.fc = nn.Linear(self.n_dim, self.n_out)
if self.name == 'mnist':
self.n_cls = 10
self.fc1 = nn.Linear(1 * 28 * 28, 200)
self.fc2 = nn.Linear(200, 200)
self.fc3 = nn.Linear(200, self.n_cls)
if self.name == 'emnist':
self.n_cls = 10
self.fc1 = nn.Linear(1 * 28 * 28, 100)
self.fc2 = nn.Linear(100, 100)
self.fc3 = nn.Linear(100, self.n_cls)
if self.name == 'cifar10':
self.n_cls = 10
self.conv1 = nn.Conv2d(in_channels=3, out_channels=64 , kernel_size=5)
self.conv2 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=5)
self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
self.fc1 = nn.Linear(64*5*5, 384)
self.fc2 = nn.Linear(384, 192)
self.fc3 = nn.Linear(192, self.n_cls)
if self.name == 'cifar100':
self.n_cls = 100
self.conv1 = nn.Conv2d(in_channels=3, out_channels=64 , kernel_size=5)
self.conv2 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=5)
self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
self.fc1 = nn.Linear(64*5*5, 384)
self.fc2 = nn.Linear(384, 192)
self.fc3 = nn.Linear(192, self.n_cls)
if self.name == 'Resnet18':
resnet18 = models.resnet18()
resnet18.fc = nn.Linear(512, 10)
# Change BN to GN
resnet18.bn1 = nn.GroupNorm(num_groups = 2, num_channels = 64)
resnet18.layer1[0].bn1 = nn.GroupNorm(num_groups = 2, num_channels = 64)
resnet18.layer1[0].bn2 = nn.GroupNorm(num_groups = 2, num_channels = 64)
resnet18.layer1[1].bn1 = nn.GroupNorm(num_groups = 2, num_channels = 64)
resnet18.layer1[1].bn2 = nn.GroupNorm(num_groups = 2, num_channels = 64)
resnet18.layer2[0].bn1 = nn.GroupNorm(num_groups = 2, num_channels = 128)
resnet18.layer2[0].bn2 = nn.GroupNorm(num_groups = 2, num_channels = 128)
resnet18.layer2[0].downsample[1] = nn.GroupNorm(num_groups = 2, num_channels = 128)
resnet18.layer2[1].bn1 = nn.GroupNorm(num_groups = 2, num_channels = 128)
resnet18.layer2[1].bn2 = nn.GroupNorm(num_groups = 2, num_channels = 128)
resnet18.layer3[0].bn1 = nn.GroupNorm(num_groups = 2, num_channels = 256)
resnet18.layer3[0].bn2 = nn.GroupNorm(num_groups = 2, num_channels = 256)
resnet18.layer3[0].downsample[1] = nn.GroupNorm(num_groups = 2, num_channels = 256)
resnet18.layer3[1].bn1 = nn.GroupNorm(num_groups = 2, num_channels = 256)
resnet18.layer3[1].bn2 = nn.GroupNorm(num_groups = 2, num_channels = 256)
resnet18.layer4[0].bn1 = nn.GroupNorm(num_groups = 2, num_channels = 512)
resnet18.layer4[0].bn2 = nn.GroupNorm(num_groups = 2, num_channels = 512)
resnet18.layer4[0].downsample[1] = nn.GroupNorm(num_groups = 2, num_channels = 512)
resnet18.layer4[1].bn1 = nn.GroupNorm(num_groups = 2, num_channels = 512)
resnet18.layer4[1].bn2 = nn.GroupNorm(num_groups = 2, num_channels = 512)
assert len(dict(resnet18.named_parameters()).keys()) == len(resnet18.state_dict().keys()), 'More BN layers are there...'
self.model = resnet18
if self.name == 'shakespeare':
embedding_dim = 8
hidden_size = 100
num_LSTM = 2
input_length = 80
self.n_cls = 80
self.embedding = nn.Embedding(input_length, embedding_dim)
self.stacked_LSTM = nn.LSTM(input_size=embedding_dim, hidden_size=hidden_size, num_layers=num_LSTM)
self.fc = nn.Linear(hidden_size, self.n_cls)
def forward(self, x):
if self.name == 'Linear':
x = self.fc(x)
if self.name == 'mnist':
x = x.view(-1, 1 * 28 * 28)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
if self.name == 'emnist':
x = x.view(-1, 1 * 28 * 28)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
if self.name == 'cifar10':
x = self.pool(F.relu(self.conv1(x)))
x = self.pool(F.relu(self.conv2(x)))
x = x.view(-1, 64*5*5)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
if self.name == 'cifar100':
x = self.pool(F.relu(self.conv1(x)))
x = self.pool(F.relu(self.conv2(x)))
x = x.view(-1, 64*5*5)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
if self.name == 'Resnet18':
x = self.model(x)
if self.name == 'shakespeare':
x = self.embedding(x)
x = x.permute(1, 0, 2) # lstm accepts in this style
output, (h_, c_) = self.stacked_LSTM(x)
# Choose last hidden layer
last_hidden = output[-1,:,:]
x = self.fc(last_hidden)
return x