NeurIPS-2021 workshop

---

---

1 Background and Motivation

注意力机制是近些年视觉领域研究的热门方向之一

We aim to utilize the contributing factors of weights for the improvement of attention mechanisms.

2 Related Work

However, these works neglect information from the tuned weights from training.

3 Advantages / Contributions

提出 Normalization-based Attention Modul，在 resnet 和 mobilenet 上 验证了其有效性

4 Method

a NAM module is embedded at the end of each network block

$W_{\gamma }$ 和 $W_{\lambda }$ 的计算方法如图 1

作者还对 $\gamma$ 和 $\lambda$ 进行了归一化约束

$p$ is the penalty that balances $g(\gamma )$ and $g(\lambda )$

看看作者开源的代码，https://github.com/Christian-lyc/NAM

import torch.nn as nn
import torch
from torch.nn import functional as Fclass Channel_Att(nn.Module):def __init__(self, channels, t=16):super(Channel_Att, self).__init__()self.channels = channelsself.bn2 = nn.BatchNorm2d(self.channels, affine=True)def forward(self, x):residual = xx = self.bn2(x)weight_bn = self.bn2.weight.data.abs() / torch.sum(self.bn2.weight.data.abs())x = x.permute(0, 2, 3, 1).contiguous()x = torch.mul(weight_bn, x)x = x.permute(0, 3, 1, 2).contiguous()x = torch.sigmoid(x) * residual #return xclass Att(nn.Module):def __init__(self, channels,shape, out_channels=None, no_spatial=True):super(Att, self).__init__()self.Channel_Att = Channel_Att(channels)def forward(self, x):x_out1=self.Channel_Att(x)return x_out1  

仅有 channel normalization-based attention 的部分

5 Experiments

5.1 Datasets and Metrics

• CIFAR-100

• ImageNet

top1 and top5

5.2 Experiments

单加 channel NAM 比单加 spatial 的要好

提升不是特别的明显，优势在于基本没有引入额外的参数量，下面具体看看参数量

乘以 4，仅看作者开源的代码的话，应该是乘以 2，也就是 BN 的参数量

6 Conclusion（own）

文章篇幅较短，细节未可知，eg： pixel normalization 的具体实现