Spaces:
Runtime error
Runtime error
File size: 6,264 Bytes
ac6e446 |
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 |
from __future__ import division, absolute_import
import torch
import torch.utils.model_zoo as model_zoo
from torch import nn
from torch.nn import functional as F
__all__ = ['shufflenet']
model_urls = {
# training epoch = 90, top1 = 61.8
'imagenet':
'https://mega.nz/#!RDpUlQCY!tr_5xBEkelzDjveIYBBcGcovNCOrgfiJO9kiidz9fZM',
}
class ChannelShuffle(nn.Module):
def __init__(self, num_groups):
super(ChannelShuffle, self).__init__()
self.g = num_groups
def forward(self, x):
b, c, h, w = x.size()
n = c // self.g
# reshape
x = x.view(b, self.g, n, h, w)
# transpose
x = x.permute(0, 2, 1, 3, 4).contiguous()
# flatten
x = x.view(b, c, h, w)
return x
class Bottleneck(nn.Module):
def __init__(
self,
in_channels,
out_channels,
stride,
num_groups,
group_conv1x1=True
):
super(Bottleneck, self).__init__()
assert stride in [1, 2], 'Warning: stride must be either 1 or 2'
self.stride = stride
mid_channels = out_channels // 4
if stride == 2:
out_channels -= in_channels
# group conv is not applied to first conv1x1 at stage 2
num_groups_conv1x1 = num_groups if group_conv1x1 else 1
self.conv1 = nn.Conv2d(
in_channels,
mid_channels,
1,
groups=num_groups_conv1x1,
bias=False
)
self.bn1 = nn.BatchNorm2d(mid_channels)
self.shuffle1 = ChannelShuffle(num_groups)
self.conv2 = nn.Conv2d(
mid_channels,
mid_channels,
3,
stride=stride,
padding=1,
groups=mid_channels,
bias=False
)
self.bn2 = nn.BatchNorm2d(mid_channels)
self.conv3 = nn.Conv2d(
mid_channels, out_channels, 1, groups=num_groups, bias=False
)
self.bn3 = nn.BatchNorm2d(out_channels)
if stride == 2:
self.shortcut = nn.AvgPool2d(3, stride=2, padding=1)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.shuffle1(out)
out = self.bn2(self.conv2(out))
out = self.bn3(self.conv3(out))
if self.stride == 2:
res = self.shortcut(x)
out = F.relu(torch.cat([res, out], 1))
else:
out = F.relu(x + out)
return out
# configuration of (num_groups: #out_channels) based on Table 1 in the paper
cfg = {
1: [144, 288, 576],
2: [200, 400, 800],
3: [240, 480, 960],
4: [272, 544, 1088],
8: [384, 768, 1536],
}
class ShuffleNet(nn.Module):
"""ShuffleNet.
Reference:
Zhang et al. ShuffleNet: An Extremely Efficient Convolutional Neural
Network for Mobile Devices. CVPR 2018.
Public keys:
- ``shufflenet``: ShuffleNet (groups=3).
"""
def __init__(self, num_classes, loss='softmax', num_groups=3, **kwargs):
super(ShuffleNet, self).__init__()
self.loss = loss
self.conv1 = nn.Sequential(
nn.Conv2d(3, 24, 3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(24),
nn.ReLU(),
nn.MaxPool2d(3, stride=2, padding=1),
)
self.stage2 = nn.Sequential(
Bottleneck(
24, cfg[num_groups][0], 2, num_groups, group_conv1x1=False
),
Bottleneck(cfg[num_groups][0], cfg[num_groups][0], 1, num_groups),
Bottleneck(cfg[num_groups][0], cfg[num_groups][0], 1, num_groups),
Bottleneck(cfg[num_groups][0], cfg[num_groups][0], 1, num_groups),
)
self.stage3 = nn.Sequential(
Bottleneck(cfg[num_groups][0], cfg[num_groups][1], 2, num_groups),
Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups),
Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups),
Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups),
Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups),
Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups),
Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups),
Bottleneck(cfg[num_groups][1], cfg[num_groups][1], 1, num_groups),
)
self.stage4 = nn.Sequential(
Bottleneck(cfg[num_groups][1], cfg[num_groups][2], 2, num_groups),
Bottleneck(cfg[num_groups][2], cfg[num_groups][2], 1, num_groups),
Bottleneck(cfg[num_groups][2], cfg[num_groups][2], 1, num_groups),
Bottleneck(cfg[num_groups][2], cfg[num_groups][2], 1, num_groups),
)
self.classifier = nn.Linear(cfg[num_groups][2], num_classes)
self.feat_dim = cfg[num_groups][2]
def forward(self, x):
x = self.conv1(x)
x = self.stage2(x)
x = self.stage3(x)
x = self.stage4(x)
x = F.avg_pool2d(x, x.size()[2:]).view(x.size(0), -1)
if not self.training:
return x
y = self.classifier(x)
if self.loss == 'softmax':
return y
elif self.loss == 'triplet':
return y, x
else:
raise KeyError('Unsupported loss: {}'.format(self.loss))
def init_pretrained_weights(model, model_url):
"""Initializes model with pretrained weights.
Layers that don't match with pretrained layers in name or size are kept unchanged.
"""
pretrain_dict = model_zoo.load_url(model_url)
model_dict = model.state_dict()
pretrain_dict = {
k: v
for k, v in pretrain_dict.items()
if k in model_dict and model_dict[k].size() == v.size()
}
model_dict.update(pretrain_dict)
model.load_state_dict(model_dict)
def shufflenet(num_classes, loss='softmax', pretrained=True, **kwargs):
model = ShuffleNet(num_classes, loss, **kwargs)
if pretrained:
# init_pretrained_weights(model, model_urls['imagenet'])
import warnings
warnings.warn(
'The imagenet pretrained weights need to be manually downloaded from {}'
.format(model_urls['imagenet'])
)
return model
|