2.Factors of Efficient Network Design
在設計輕量級網絡時,FLOPs和模型參數是主要考慮因素,但是減少模型大小和FLOPs不等同於減少推理時間和降低能耗。比如ShuffleNetv2與MobileNetv2在相同的FLOPs下,前者在GPU上速度更快。所以除了FLOPs和模型大小外,還需要考慮其他因素對能耗和模型推理速度的影響。這裏考慮兩個重要的因素:內存訪問成本(Memory Access Cost,MAC)和GPU計算效率。
2.1. Memory Access Cost
對於CNN,能耗在內存訪問而不是計算上。影響MAC的主要是是內存佔用(intermediate activation memory footprint),它主要受卷積核和feature map大小的影響。c爲輸入輸出通道。
2.2. GPUComputation
通過減少FLOP是來加速的前提是,每個flop point的計算效率是一致的。
GPU特性:
(1)擅長parallel computation,tensor越大,GPU使用效率越高。
(2)把大的卷積操作拆分成碎片的小操作將不利於GPU計算。
因此,設計layer數量少的網絡是更好的選擇。mobileNet使用1x1卷積來減少計算量,不過這不利於GPU計算。爲了衡量GPU使用效率,我們使用Flops/s指標。
3. Proposed Method
3.1. Rethinking Dense Connection
當固定卷積層參數量B=k^2hwcico的時候,MAC可以表示如下,根據均值不等式,可以知道當輸入和輸出的channel數相同時MAC才取下界,此時的設計是最高效的。
使用Dense connect模塊,輸出的channel size不變,但是輸入的channel size一直在線性增加,因此DenseNet有很高的MAC。bottleneck connection同樣不利於GPU計算。在模型很大的時候,計算量隨着深度指數二階增長。bottleneck 把一個3x3的卷積分成了兩個計算,相當於增加了一次序列運算。
3.2. One Shot Aggregation
對於DenseNet來說,其核心模塊就是Dense
Block,如下圖1a所示,這種密集連接會聚合前面所有的layer,這導致每個layer的輸入channel數線性增長。受限於FLOPs和模型參數,每層layer的輸出channel數是固定大小,這帶來的問題就是輸入和輸出channel數不一致,如前面所述,此時的MAC不是最優的。另外,由於輸入channel數較大,DenseNet採用了1x1卷積層先壓縮特徵,這個額外層的引入對GPU高效計算不利。所以,雖然DenseNet的FLOPs和模型參數都不大,但是推理卻並不高效,當輸入較大時往往需要更多的顯存和推理時間。
DenseNet的一大問題就是密集連接太重了,而且每個layer都會聚合前面層的特徵,其實造成的是特徵冗餘,而且從模型weights的L1範數會發現中間層對最後的分類層貢獻較少,這不難理解,因爲後面的特徵其實已經學習到了這些中間層的核心信息。這種信息冗餘反而是可以優化的方向,據此這裏提出了OSA(One-Shot Aggregation)模塊,如圖1b所示,簡單來說,就是隻在最後一次性聚合前面所有的layer。這一改動將會解決DenseNet前面所述的問題,因爲每個layer的輸入channel數是固定的,這裏可以讓輸出channel數和輸入一致而取得最小的MAC,而且也不再需要1x1卷積層來壓縮特徵,所以OSA模塊是GPU計算高效的。
那麼OSA模塊效果如何,論文中拿DenseNet-40來做對比,DenseBlock層數是12,OSA模塊也設計爲12層,但是保持和Dense
Block類似的參數大小和計算量,此時OSA模塊的輸出將更大。最終發現在CIFAR-10數據集上acc僅比DenseNet下降了1.2%。但是如果將OSA模塊的層數降至5,而提升layer的通道數爲43,會發現與DenseNet-40模型效果相當。這說明DenseNet中很多中間特徵可能是冗餘的。儘管OSA模塊性能沒有提升,但是MAC低且計算更高效,這對於目標檢測非常重要,因爲檢測模型一般的輸入都是較大的。
3.3. Configuration of VoVNet
VoVNet由OSA模塊構成,主要有三種不同的配置,如下表所示。VoVNet首先是一個由3個3x3卷積層構成的stem block,然後4個階段的OSA模塊,每個stage的最後會採用一個stride爲2的3x3 max pooling層進行降採樣,模型最終的output stride是32。與其他網絡類似,每次降採樣後都會提升特徵的channel數。VoVNet-27-slim是一個輕量級模型,而VoVNet-39/57在stage4和stage5包含更多的OSA模塊,所以模型更大。
4. Experiments
相比於DenseNet-67,PeleeNet減少了Flops,但是推斷速度沒有提升,與之相反,VoVNet-27-slim稍微增加了Flops,而推斷速度提升了一倍。同時,VoVNet-27-slim的精度比其他模型都高。VoVNet-27-slim的內存佔用、能耗、GPU使用效率都是最好的。相比其他模型,VoVNet做到了準確率和效率的均衡,提升了目標檢測的整體性能。
import torch
import torch.nn as nn
import torch.nn.functional as F
from collections import OrderedDict
__all__ = ['VoVNet', 'vovnet27_slim', 'vovnet39', 'vovnet57']
model_urls = {
'vovnet39': 'https://dl.dropbox.com/s/1lnzsgnixd8gjra/vovnet39_torchvision.pth?dl=1',
'vovnet57': 'https://dl.dropbox.com/s/6bfu9gstbwfw31m/vovnet57_torchvision.pth?dl=1'
}
def conv3x3(in_channels, out_channels, module_name, postfix,
stride=1, groups=1, kernel_size=3, padding=1):
"""3x3 convolution with padding"""
return [
('{}_{}/conv'.format(module_name, postfix),
nn.Conv2d(in_channels, out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=False)),
('{}_{}/norm'.format(module_name, postfix),
nn.BatchNorm2d(out_channels)),
('{}_{}/relu'.format(module_name, postfix),
nn.ReLU(inplace=True)),
]
def conv1x1(in_channels, out_channels, module_name, postfix,
stride=1, groups=1, kernel_size=1, padding=0):
"""1x1 convolution"""
return [
('{}_{}/conv'.format(module_name, postfix),
nn.Conv2d(in_channels, out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=False)),
('{}_{}/norm'.format(module_name, postfix),
nn.BatchNorm2d(out_channels)),
('{}_{}/relu'.format(module_name, postfix),
nn.ReLU(inplace=True)),
]
class _OSA_module(nn.Module):
def __init__(self,
in_ch,
stage_ch,
concat_ch,
layer_per_block,
module_name,
identity=False):
super(_OSA_module, self).__init__()
self.identity = identity
self.layers = nn.ModuleList()
in_channel = in_ch
for i in range(layer_per_block):
self.layers.append(nn.Sequential(
OrderedDict(conv3x3(in_channel, stage_ch, module_name, i))))
in_channel = stage_ch
# feature aggregation
in_channel = in_ch + layer_per_block * stage_ch
print("inch:",in_ch,"layer_per_block:",layer_per_block,"stage:",stage_ch)
self.concat = nn.Sequential(
OrderedDict(conv1x1(in_channel, concat_ch, module_name, 'concat')))
def forward(self, x):
identity_feat = x
output = []
output.append(x)
for layer in self.layers:
x = layer(x)
output.append(x)
print("output:",output)
x = torch.cat(output, dim=1) #按列拼
xt = self.concat(x)
if self.identity:
xt = xt + identity_feat
return xt
class _OSA_stage(nn.Sequential):
def __init__(self,
in_ch,
stage_ch,
concat_ch,
block_per_stage,
layer_per_block,
stage_num):
super(_OSA_stage, self).__init__()
if not stage_num == 2:
self.add_module('Pooling',
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)) #向上取整,默認padding爲0
module_name = f'OSA{stage_num}_1'
self.add_module(module_name,
_OSA_module(in_ch,
stage_ch,
concat_ch,
layer_per_block,
module_name))
for i in range(block_per_stage-1):
module_name = f'OSA{stage_num}_{i+2}'
self.add_module(module_name,
_OSA_module(concat_ch,
stage_ch,
concat_ch,
layer_per_block,
module_name,
identity=True))
class VoVNet(nn.Module):
def __init__(self,
config_stage_ch,
config_concat_ch,
block_per_stage,
layer_per_block,
num_classes=1000):
super(VoVNet, self).__init__()
# Stem module
stem = conv3x3(3, 64, 'stem', '1', 2)
stem += conv3x3(64, 64, 'stem', '2', 1)
stem += conv3x3(64, 128, 'stem', '3', 2)
self.add_module('stem', nn.Sequential(OrderedDict(stem)))
stem_out_ch = [128]
in_ch_list = stem_out_ch + config_concat_ch[:-1]
self.stage_names = []
for i in range(4): #num_stages
name = 'stage%d' % (i+2)
self.stage_names.append(name)
self.add_module(name,
_OSA_stage(in_ch_list[i],
config_stage_ch[i],
config_concat_ch[i],
block_per_stage[i],
layer_per_block,
i+2))
self.classifier = nn.Linear(config_concat_ch[-1], num_classes)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.constant_(m.bias, 0)
def forward(self, x):
x = self.stem(x)
for name in self.stage_names:
x = getattr(self, name)(x)
x = F.adaptive_avg_pool2d(x, (1, 1)).view(x.size(0), -1)
x = self.classifier(x)
return x
def _vovnet(arch,
config_stage_ch,
config_concat_ch,
block_per_stage,
layer_per_block,
pretrained,
progress,
**kwargs):
model = VoVNet(config_stage_ch, config_concat_ch,
block_per_stage, layer_per_block,
**kwargs)
if pretrained:
state_dict = torch.hub.load_state_dict_from_url(model_urls[arch],
progress=progress)
model.load_state_dict(state_dict)
return model
def vovnet57(pretrained=False, progress=True, **kwargs):
r"""Constructs a VoVNet-57 model as described in
`"An Energy and GPU-Computation Efficient Backbone Networks"
<https://arxiv.org/abs/1904.09730>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
return _vovnet('vovnet57', [128, 160, 192, 224], [256, 512, 768, 1024],
[1,1,4,3], 5, pretrained, progress, **kwargs)
def vovnet39(pretrained=False, progress=True, **kwargs):
r"""Constructs a VoVNet-39 model as described in
`"An Energy and GPU-Computation Efficient Backbone Networks"
<https://arxiv.org/abs/1904.09730>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
return _vovnet('vovnet39', [128, 160, 192, 224], [256, 512, 768, 1024],
[1,1,2,2], 5, pretrained, progress, **kwargs)
def vovnet27_slim(pretrained=False, progress=True, **kwargs):
r"""Constructs a VoVNet-39 model as described in
`"An Energy and GPU-Computation Efficient Backbone Networks"
<https://arxiv.org/abs/1904.09730>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
return _vovnet('vovnet27_slim', [64, 80, 96, 112], [128, 256, 384, 512],
[1,1,1,1], 5, pretrained, progress, **kwargs)
if __name__ == '__main__':
model = vovnet39()
print(model)