-
Notifications
You must be signed in to change notification settings - Fork 69
/
models.py
157 lines (134 loc) · 6.28 KB
/
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
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
from copy import deepcopy
import logging
import torch
import torch.nn as nn
import torch.nn.functional as F
logger = logging.getLogger(__name__)
class ModelEMA(nn.Module):
def __init__(self, model, decay=0.9999, device=None):
super().__init__()
self.module = deepcopy(model)
self.module.eval()
self.decay = decay
self.device = device
if self.device is not None:
self.module.to(device=device)
def forward(self, input):
return self.module(input)
def _update(self, model, update_fn):
with torch.no_grad():
for ema_v, model_v in zip(self.module.parameters(), model.parameters()):
if self.device is not None:
model_v = model_v.to(device=self.device)
ema_v.copy_(update_fn(ema_v, model_v))
for ema_v, model_v in zip(self.module.buffers(), model.buffers()):
if self.device is not None:
model_v = model_v.to(device=self.device)
ema_v.copy_(model_v)
def update_parameters(self, model):
self._update(model, update_fn=lambda e, m: self.decay * e + (1. - self.decay) * m)
def state_dict(self):
return self.module.state_dict()
def load_state_dict(self, state_dict):
self.module.load_state_dict(state_dict)
class BasicBlock(nn.Module):
def __init__(self, in_planes, out_planes, stride, dropout=0.0, activate_before_residual=False):
super(BasicBlock, self).__init__()
self.bn1 = nn.BatchNorm2d(in_planes, momentum=0.001)
self.relu1 = nn.LeakyReLU(negative_slope=0.1, inplace=True)
self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(out_planes, momentum=0.001)
self.relu2 = nn.LeakyReLU(negative_slope=0.1, inplace=True)
self.conv2 = nn.Conv2d(out_planes, out_planes, kernel_size=3, stride=1,
padding=1, bias=False)
self.dropout = dropout
self.equalInOut = (in_planes == out_planes)
self.convShortcut = (not self.equalInOut) and nn.Conv2d(in_planes, out_planes,
kernel_size=1, stride=stride,
padding=0, bias=False) or None
self.activate_before_residual = activate_before_residual
def forward(self, x):
if not self.equalInOut and self.activate_before_residual is True:
x = self.relu1(self.bn1(x))
else:
out = self.relu1(self.bn1(x))
out = self.relu2(self.bn2(self.conv1(out if self.equalInOut else x)))
if self.dropout > 0:
out = F.dropout(out, p=self.dropout, training=self.training)
out = self.conv2(out)
return torch.add(x if self.equalInOut else self.convShortcut(x), out)
class NetworkBlock(nn.Module):
def __init__(self, nb_layers, in_planes, out_planes, block, stride, dropout=0.0,
activate_before_residual=False):
super(NetworkBlock, self).__init__()
self.layer = self._make_layer(
block, in_planes, out_planes, nb_layers, stride, dropout, activate_before_residual)
def _make_layer(self, block, in_planes, out_planes, nb_layers, stride, dropout,
activate_before_residual):
layers = []
for i in range(int(nb_layers)):
layers.append(block(i == 0 and in_planes or out_planes, out_planes,
i == 0 and stride or 1, dropout, activate_before_residual))
return nn.Sequential(*layers)
def forward(self, x):
return self.layer(x)
class WideResNet(nn.Module):
def __init__(self, num_classes, depth=28, widen_factor=2, dropout=0.0, dense_dropout=0.0):
super(WideResNet, self).__init__()
channels = [16, 16 * widen_factor, 32 * widen_factor, 64 * widen_factor]
assert((depth - 4) % 6 == 0)
n = (depth - 4) / 6
block = BasicBlock
# 1st conv before any network block
self.conv1 = nn.Conv2d(3, channels[0], kernel_size=3, stride=1,
padding=1, bias=False)
# 1st block
self.block1 = NetworkBlock(
n, channels[0], channels[1], block, 1, dropout, activate_before_residual=True)
# 2nd block
self.block2 = NetworkBlock(
n, channels[1], channels[2], block, 2, dropout)
# 3rd block
self.block3 = NetworkBlock(
n, channels[2], channels[3], block, 2, dropout)
# global average pooling and classifier
self.bn1 = nn.BatchNorm2d(channels[3], momentum=0.001)
self.relu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
self.drop = nn.Dropout(dense_dropout)
self.fc = nn.Linear(channels[3], num_classes)
self.channels = channels[3]
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight,
mode='fan_out',
nonlinearity='leaky_relu')
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1.0)
nn.init.constant_(m.bias, 0.0)
elif isinstance(m, nn.Linear):
nn.init.xavier_normal_(m.weight)
nn.init.constant_(m.bias, 0.0)
def forward(self, x):
out = self.conv1(x)
out = self.block1(out)
out = self.block2(out)
out = self.block3(out)
out = self.relu(self.bn1(out))
out = F.adaptive_avg_pool2d(out, 1)
out = out.view(-1, self.channels)
return self.fc(self.drop(out))
def build_wideresnet(args):
if args.dataset == "cifar10":
depth, widen_factor = 28, 2
elif args.dataset == 'cifar100':
depth, widen_factor = 28, 8
model = WideResNet(num_classes=args.num_classes,
depth=depth,
widen_factor=widen_factor,
dropout=0,
dense_dropout=args.dense_dropout)
if args.local_rank in [-1, 0]:
logger.info(f"Model: WideResNet {depth}x{widen_factor}")
logger.info(f"Total params: {sum(p.numel() for p in model.parameters())/1e6:.2f}M")
return model