目录
Inverted Residual Block:简称iRBlock,倒残差模块
Inverted Residual Mobile Block:简称iRMBlock,倒残差移动模块,是iRBlock的改进版本
1. iRBlock简介
Inverted Residual Block,也被称为MobileNet的构建模块,通常被用在轻量化网络中,方便在移动端和嵌入式端进行部署,是一种用于提取特征的深度学习模块。它结合了深度可分离卷积(depthwise separable convolution)和shortcut connections,以减少计算成本和参数数量。
以下是一个简单的Python代码示例,展示了如何实现一个Inverted Residual Block:
import torch
import torch.nn as nn
class InvertedResidual(nn.Module):
def __init__(self, in_channels, out_channels, stride, expand_ratio):
super(InvertedResidual, self).__init__()
self.stride = stride
hidden_channels = in_channels * expand_ratio
# Squeeze and Excitation
self.conv_se = nn.Sequential(
nn.Conv2d(hidden_channels, hidden_channels // ratio, 1),
nn.ReLU(inplace=True),
nn.Conv2d(hidden_channels // ratio, hidden_channels, 1),
nn.Sigmoid()
)
# Pointwise convolution
self.conv_pw = nn.Conv2d(hidden_channels, out_channels, 1)
# Depthwise convolution
self.conv_dw = nn.Conv2d(in_channels, hidden_channels, 3, stride, 1, groups=hidden_channels)
def forward(self, x):
if self.stride == 1:
shortcut = x
else:
shortcut = self.pool(x)
x = self.conv_dw(x)
x = self.conv_se(x)
x = self.conv_pw(x)
x = x * shortcut
return x
这个模块接收输入的通道数in_channels
,输出的通道数out_channels
,步长stride
,以及扩展比率expand_ratio
作为参数。它首先使用深度可分离卷积进行特征提取,然后通过Squeeze and Excitation模块进一步改善特征表示,最后使用点卷积输出。如果步长大于1,它会使用平均池化来进行shortcut connection。
2. iRMBlock简介
iRMBlock模块是这篇论文基于iRBlock进行改进的。 iRBlock模块虽然被认为是标准的高效模块代表作之一。然而,受限于静态 CNN 的归纳偏差影响,纯 CNN 模型的准确性仍然保持较低水平,随着Transformer的出现,一时间涌现了许多性能性能超群的网络,如 Swin transformer、PVT、Eatformer、EAT等。得益于其动态建模和不受归纳偏置的影响,这些方法都取得了相对 CNN 的显着改进。然而,受多头自注意(MHSA)参数和计算量的二次方限制,基于 Transformer 的模型往往具有大量资源消耗,因此也难以在轻量化环境下部署。
受此启发,作者重新考虑了 MobileNetv2 中的 Inverted Residual Block 和 Transformer 中的 MHSA/FFN 模块,归纳抽象出一个通用的 Meta Mobile Block,它采用参数扩展比 λ 和高效算子 F 来实例化不同的模块,即 IRB、MHSA 和前馈网络 (FFN)。
官网给出的iRMB模块的python类实现python代码如下:
class iRMB(nn.Module):
def __init__(self, dim_in, dim_out, norm_in=True, has_skip=True, exp_ratio=1.0, norm_layer='bn_2d',
act_layer='relu', v_proj=True, dw_ks=3, stride=1, dilation=1, se_ratio=0.0, dim_head=64, window_size=7,
attn_s=True, qkv_bias=False, attn_drop=0., drop=0., drop_path=0., v_group=False, attn_pre=False):
super().__init__()
self.norm = get_norm(norm_layer)(dim_in) if norm_in else nn.Identity()
dim_mid = int(dim_in * exp_ratio)
self.has_skip = (dim_in == dim_out and stride == 1) and has_skip
self.attn_s = attn_s
if self.attn_s:
assert dim_in % dim_head == 0, 'dim should be divisible by num_heads'
self.dim_head = dim_head
self.window_size = window_size
self.num_head = dim_in // dim_head
self.scale = self.dim_head ** -0.5
self.attn_pre = attn_pre
self.qk = ConvNormAct(dim_in, int(dim_in * 2), kernel_size=1, bias=qkv_bias, norm_layer='none', act_layer='none')
self.v = ConvNormAct(dim_in, dim_mid, kernel_size=1, groups=self.num_head if v_group else 1, bias=qkv_bias, norm_layer='none', act_layer=act_layer, inplace=inplace)
self.attn_drop = nn.Dropout(attn_drop)
else:
if v_proj:
self.v = ConvNormAct(dim_in, dim_mid, kernel_size=1, bias=qkv_bias, norm_layer='none', act_layer=act_layer, inplace=inplace)
else:
self.v = nn.Identity()
self.conv_local = ConvNormAct(dim_mid, dim_mid, kernel_size=dw_ks, stride=stride, dilation=dilation, groups=dim_mid, norm_layer='bn_2d', act_layer='silu', inplace=inplace)
self.se = SE(dim_mid, rd_ratio=se_ratio, act_layer=get_act(act_layer)) if se_ratio > 0.0 else nn.Identity()
self.proj_drop = nn.Dropout(drop)
self.proj = ConvNormAct(dim_mid, dim_out, kernel_size=1, norm_layer='none', act_layer='none', inplace=inplace)
self.drop_path = DropPath(drop_path) if drop_path else nn.Identity()
def forward(self, x):
shortcut = x
x = self.norm(x)
B, C, H, W = x.shape
if self.attn_s:
# padding
if self.window_size <= 0:
window_size_W, window_size_H = W, H
else:
window_size_W, window_size_H = self.window_size, self.window_size
pad_l, pad_t = 0, 0
pad_r = (window_size_W - W % window_size_W) % window_size_W
pad_b = (window_size_H - H % window_size_H) % window_size_H
x = F.pad(x, (pad_l, pad_r, pad_t, pad_b, 0, 0,))
n1, n2 = (H + pad_b) // window_size_H, (W + pad_r) // window_size_W
x = rearrange(x, 'b c (h1 n1) (w1 n2) -> (b n1 n2) c h1 w1', n1=n1, n2=n2).contiguous()
# attention
b, c, h, w = x.shape
qk = self.qk(x)
qk = rearrange(qk, 'b (qk heads dim_head) h w -> qk b heads (h w) dim_head', qk=2, heads=self.num_head, dim_head=self.dim_head).contiguous()
q, k = qk[0], qk[1]
attn_spa = (q @ k.transpose(-2, -1)) * self.scale
attn_spa = attn_spa.softmax(dim=-1)
attn_spa = self.attn_drop(attn_spa)
if self.attn_pre:
x = rearrange(x, 'b (heads dim_head) h w -> b heads (h w) dim_head', heads=self.num_head).contiguous()
x_spa = attn_spa @ x
x_spa = rearrange(x_spa, 'b heads (h w) dim_head -> b (heads dim_head) h w', heads=self.num_head, h=h, w=w).contiguous()
x_spa = self.v(x_spa)
else:
v = self.v(x)
v = rearrange(v, 'b (heads dim_head) h w -> b heads (h w) dim_head', heads=self.num_head).contiguous()
x_spa = attn_spa @ v
x_spa = rearrange(x_spa, 'b heads (h w) dim_head -> b (heads dim_head) h w', heads=self.num_head, h=h, w=w).contiguous()
# unpadding
x = rearrange(x_spa, '(b n1 n2) c h1 w1 -> b c (h1 n1) (w1 n2)', n1=n1, n2=n2).contiguous()
if pad_r > 0 or pad_b > 0:
x = x[:, :, :H, :W].contiguous()
else:
x = self.v(x)
x = x + self.se(self.conv_local(x)) if self.has_skip else self.se(self.conv_local(x))
x = self.proj_drop(x)
x = self.proj(x)
x = (shortcut + self.drop_path(x)) if self.has_skip else x
return x
3.如何使用iRMB模块
论文中将多个iRMB堆叠起来,形成一个stage,EMO网络将四个stage串联起来,便形成了基本的骨干特征提取网络,只需按需连接下游任务(分类、分割、目标检测)即可实现完整网络结构构建。
官网实现EMO网络代码:github
反向残差移动块 (iRMB),它吸收了 CNN 架构的效率来建模局部特征和 Transformer 架构动态建模的能力来学习长距离交互。
具体实现中,iRMB 中的 F 被建模为级联的 MHSA 和卷积运算,公式可以抽象为F(⋅)=Conv(MHSA(⋅))。这里需要考虑的问题主要有两个:
λ通常大于中间维度将是输入维度的倍数,导致参数和计算的二次增加。因此作者很自然的考虑结合 W-MHSA 和 DW-Conv 并结合残差机制设计了一种新的模块。此外,通过这种级联方式可以提高感受野的扩展率,同时可以有效降低模型计算复杂度。
总结:后续在轻量化网络设计,即可考虑在网络中设计iRBlock 或iRMBlock模块的实现。
整理不易,欢迎一键三连!!!
送你们一条美丽的--分割线--
🌷🌷🍀🍀🌾🌾🍓🍓🍂🍂🙋🙋🐸🐸🙋🙋💖💖🍌🍌🔔🔔🍉🍉🍭🍭🍋🍋🍇🍇🏆🏆📸📸⛵⛵⭐⭐🍎🍎👍👍🌷🌷