论文笔记-Squeeze-and-Excitation Networks

作者提出为了增强网络的表达能力，现有的工作显示了加强空间编码的作用。在这篇论文里面，作者重点关注channel上的信息，提出了“Squeeze-and-Excitation"（SE）block，实际上就是显式的让网络关注channel之间的信息 （adaptively recalibrates channel-wise feature responsesby explicitly modelling interdependencies between channels.）。SEnets取得了ILSVRC2017的第一名, top-5 error 2.251%

之前的一些架构设计关注空间依赖

Inception architectures: embedding multi-scale processes in its modules

Resnet, stack hourglass

spatial attention: Spatial transformer networks

作者的设计思路：

we investigate a different

aspect of architectural design - the channel relationship

Our goal is to improve the representational power of a network by explicitly

modelling the interdependencies between the channels of its

convolutional features. To achieve this, we propose a mechanism that allows the network to perform feature recalibration, through which it can learn to use global information

to selectively emphasise informative features and suppress

less useful ones. 作者希望能够对卷积特征进行recalibration，根据后文我的理解就是对channel进行加权了。

相关工作

网络结构：

VGGNets, Inception models, BN, Resnet, Densenet, Dual path network

其他方式：Grouped convolution, Multi-branch convolution, Cross-channel correlations

This approach reflects an assumption that channel relationships can

be formulated as a composition of instance-agnostic functions with local receptive fields.

Attention, gating mechanisms

SE block

\({F_{tr}}:X \in R{^{W' \times H' \times C'}},{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} U \in {\kern 1pt} {\kern 1pt} {R^{W \times H \times C}}\)

设\(V = [v_1, v_2, ..., v_C]\)表示学习到的filter kernel, \(v_c\)表示第c个filter的参数，那么\(F_{tr}\)的输出\(U = [u_1,u_2,...,u_C]\):

\[{u_c} = {\rm{ }}{{\rm{v}}_c} * X = \sum\limits_{s = 1}^{C'} {v_c^s} * {x^s}
\]

\(v_c^s\)是一个channel的kernel，一个新产生的channel是原有所有channel与相应的filter kernel卷积的和。channel间的关系隐式的包含在\(v_c\)中，但是这些信息和空间相关性纠缠在一起了，作者的目标就是让网络更加关注有用的信息。分成了Squeeze和Excitation两步来完成目的。

Squeeze

现有网络的问题：由于卷积实在local receptive field做的，因此每个卷积单元只能关注这个field内的空间信息。

为了减轻这个问题，提出了Squeeze操作将全局的空间信息编码到channel descriptor中，具体而言是通过global average pooling操作完成的。

\[{z_c} = {F_{sq}}({u_c}) = {1 \over {W \times H}}\sum\limits_{i = 1}^W {\sum\limits_{j = 1}^H {{u_c}(i,j)} }
\]

就是求每个channel的均值，作为全局的描述。

Excitation: Adaptive Recalibration

为了利用Squeeze得到的信息，提出了第二个op,这个op需要满足2个要求：一个是足够灵活，需要能够学习channel间的非线性关系，另一个就是能够学习non-mutually-exclusive关系，这个词我的理解是非独占性，可能是说多个channnel之间会有各种各样的关系吧。

\[s = {F_{ex}}(z,W) = \sigma (g(z,W)) = \sigma ({W_2}\delta ({W_1}z))
\]

$\delta \(是ReLu,\){W_1} \in {R^{{C \over r} \times C}}\(，\){W_2} \in {R^{C \times {C \over r}}}\(，\)W_1\(是bottleneck，降低channel数，\)W_2\(是增加channel数，\)\gamma\(设置为16。最终再将\)U\(用\)s$来scale，其实也就是加权了。这样就得到了一个block的输出。

\[{x_c} = {F_{scale}}({u_c},{s_c}) = {s_c} \cdot {u_c}
\]

\(F_{scale}\)表示feature map \(u_c \in R^{W \times H}\)和\(s_c\)的channel-wise乘法

The activations act as channel weights

adapted to the input-specific descriptor z. In this regard,

SE blocks intrinsically introduce dynamics conditioned on

the input, helping to boost feature discriminability

1. Example
   
   
   
   SE block可以很方便的加到其他网络结构上。
2. Mxnet code

squeeze = mx.sym.Pooling(data=bn3, global_pool=True, kernel=(7, 7), pool_type='avg', name=name + '_squeeze')
squeeze = mx.symbol.Flatten(data=squeeze, name=name + '_flatten')
excitation = mx.symbol.FullyConnected(data=squeeze, num_hidden=int(num_filter*ratio), name=name + '_excitation1')#bottleneck
excitation = mx.sym.Activation(data=excitation, act_type='relu', name=name + '_excitation1_relu')
excitation = mx.symbol.FullyConnected(data=excitation, num_hidden=num_filter, name=name + '_excitation2')
excitation = mx.sym.Activation(data=excitation, act_type='sigmoid', name=name + '_excitation2_sigmoid')
bn3 = mx.symbol.broadcast_mul(bn3, mx.symbol.reshape(data=excitation, shape=(-1, num_filter, 1, 1)))

1. 网络结构
   
   

2. Experiments
   
   

参考文献：

[1] Hu, Jie, Li Shen, and Gang Sun. "Squeeze-and-excitation networks." arXiv preprint arXiv:1709.01507 (2017).

欢迎关注公众号：vision_home 共同学习，不定期分享论文和资源