FightingCV Codebase For Attention,Backbone-CNN, MLP, Re-parameter, Convolution
Hello，大家好，我是小马🚀🚀🚀
For 小白（Like Me）： 最近在读论文的时候会发现一个问题，有时候论文核心思想非常简单，核心代码可能也就十几行。但是打开作者release的源码时，却发现提出的模块嵌入到分类、检测、分割等任务框架中，导致代码比较冗余，对于特定任务框架不熟悉的我，很难找到核心代码，导致在论文和网络思想的理解上会有一定困难。
For 进阶者（Like You）： 如果把Conv、FC、RNN这些基本单元看做小的Lego积木，把Transformer、ResNet这些结构看成已经搭好的Lego城堡。那么本项目提供的模块就是一个个具有完整语义信息的Lego组件。让科研工作者们避免反复造轮子，只需思考如何利用这些“Lego组件”，搭建出更多绚烂多彩的作品。
For 大神（May Be Like You）： 能力有限，不喜轻喷！！！
For All： 本项目就是要实现一个既能让深度学习小白也能搞懂，又能服务科研和工业社区的代码库。作为【论文复现项目】的补充，本项目的宗旨是从代码角度，实现🚀让世界上没有难读的论文🚀。
（同时也非常欢迎各位科研工作者将自己的工作的核心代码整理到本项目中，推动科研社区的发展，会在readme中注明代码的作者~）
欢迎大家关注公众号：FightingCV
公众号每天都会进行论文、算法和代码的干货分享哦~
已建立机器学习/深度学习算法/计算机视觉/多模态交流群微信交流群！
每天在群里分享一些近期的论文和解析，欢迎大家一起学习交流哈~~~
强烈推荐大家关注知乎账号和FightingCV公众号，可以快速了解到最新优质的干货资源。
If this project is helpful to you, welcome to give a star.
Don't forget to follow me to learn about project updates.
-
Pytorch implementation of "Beyond Self-attention: External Attention using Two Linear Layers for Visual Tasks---arXiv 2021.05.05"
-
Pytorch implementation of "Attention Is All You Need---NIPS2017"
-
Pytorch implementation of "Squeeze-and-Excitation Networks---CVPR2018"
-
Pytorch implementation of "Selective Kernel Networks---CVPR2019"
-
Pytorch implementation of "CBAM: Convolutional Block Attention Module---ECCV2018"
-
Pytorch implementation of "BAM: Bottleneck Attention Module---BMCV2018"
-
Pytorch implementation of "ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks---CVPR2020"
-
Pytorch implementation of "Dual Attention Network for Scene Segmentation---CVPR2019"
-
Pytorch implementation of "EPSANet: An Efficient Pyramid Split Attention Block on Convolutional Neural Network---arXiv 2021.05.30"
-
Pytorch implementation of "ResT: An Efficient Transformer for Visual Recognition---arXiv 2021.05.28"
-
Pytorch implementation of "SA-NET: SHUFFLE ATTENTION FOR DEEP CONVOLUTIONAL NEURAL NETWORKS---ICASSP 2021"
-
Pytorch implementation of "MUSE: Parallel Multi-Scale Attention for Sequence to Sequence Learning---arXiv 2019.11.17"
-
Pytorch implementation of "Spatial Group-wise Enhance: Improving Semantic Feature Learning in Convolutional Networks---arXiv 2019.05.23"
-
Pytorch implementation of "A2-Nets: Double Attention Networks---NIPS2018"
-
Pytorch implementation of "An Attention Free Transformer---ICLR2021 (Apple New Work)"
-
Pytorch implementation of VOLO: Vision Outlooker for Visual Recognition---arXiv 2021.06.24" 【论文解析】
-
Pytorch implementation of Vision Permutator: A Permutable MLP-Like Architecture for Visual Recognition---arXiv 2021.06.23 【论文解析】
-
Pytorch implementation of CoAtNet: Marrying Convolution and Attention for All Data Sizes---arXiv 2021.06.09 【论文解析】
-
Pytorch implementation of Scaling Local Self-Attention for Parameter Efficient Visual Backbones---CVPR2021 Oral 【论文解析】
-
Pytorch implementation of Polarized Self-Attention: Towards High-quality Pixel-wise Regression---arXiv 2021.07.02 【论文解析】
-
Pytorch implementation of Contextual Transformer Networks for Visual Recognition---arXiv 2021.07.26 【论文解析】
-
Pytorch implementation of Residual Attention: A Simple but Effective Method for Multi-Label Recognition---ICCV2021
-
Pytorch implementation of S²-MLPv2: Improved Spatial-Shift MLP Architecture for Vision---arXiv 2021.08.02 【论文解析】
-
Pytorch implementation of Global Filter Networks for Image Classification---arXiv 2021.07.01
-
Pytorch implementation of Rotate to Attend: Convolutional Triplet Attention Module---WACV 2021
"Beyond Self-attention: External Attention using Two Linear Layers for Visual Tasks"
from fightingcv.attention.ExternalAttention import ExternalAttention
import torch
input=torch.randn(50,49,512)
ea = ExternalAttention(d_model=512,S=8)
output=ea(input)
print(output.shape)
from fightingcv.attention.SelfAttention import ScaledDotProductAttention
import torch
input=torch.randn(50,49,512)
sa = ScaledDotProductAttention(d_model=512, d_k=512, d_v=512, h=8)
output=sa(input,input,input)
print(output.shape)
from fightingcv.attention.SimplifiedSelfAttention import SimplifiedScaledDotProductAttention
import torch
input=torch.randn(50,49,512)
ssa = SimplifiedScaledDotProductAttention(d_model=512, h=8)
output=ssa(input,input,input)
print(output.shape)"Squeeze-and-Excitation Networks"
from fightingcv.attention.SEAttention import SEAttention
import torch
input=torch.randn(50,512,7,7)
se = SEAttention(channel=512,reduction=8)
output=se(input)
print(output.shape)
from fightingcv.attention.SKAttention import SKAttention
import torch
input=torch.randn(50,512,7,7)
se = SKAttention(channel=512,reduction=8)
output=se(input)
print(output.shape)"CBAM: Convolutional Block Attention Module"
from fightingcv.attention.CBAM import CBAMBlock
import torch
input=torch.randn(50,512,7,7)
kernel_size=input.shape[2]
cbam = CBAMBlock(channel=512,reduction=16,kernel_size=kernel_size)
output=cbam(input)
print(output.shape)"BAM: Bottleneck Attention Module"
from fightingcv.attention.BAM import BAMBlock
import torch
input=torch.randn(50,512,7,7)
bam = BAMBlock(channel=512,reduction=16,dia_val=2)
output=bam(input)
print(output.shape)"ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks"
from fightingcv.attention.ECAAttention import ECAAttention
import torch
input=torch.randn(50,512,7,7)
eca = ECAAttention(kernel_size=3)
output=eca(input)
print(output.shape)"Dual Attention Network for Scene Segmentation"
from fightingcv.attention.DANet import DAModule
import torch
input=torch.randn(50,512,7,7)
danet=DAModule(d_model=512,kernel_size=3,H=7,W=7)
print(danet(input).shape)"EPSANet: An Efficient Pyramid Split Attention Block on Convolutional Neural Network"
from fightingcv.attention.PSA import PSA
import torch
input=torch.randn(50,512,7,7)
psa = PSA(channel=512,reduction=8)
output=psa(input)
print(output.shape)"ResT: An Efficient Transformer for Visual Recognition"
from fightingcv.attention.EMSA import EMSA
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,64,512)
emsa = EMSA(d_model=512, d_k=512, d_v=512, h=8,H=8,W=8,ratio=2,apply_transform=True)
output=emsa(input,input,input)
print(output.shape)
"SA-NET: SHUFFLE ATTENTION FOR DEEP CONVOLUTIONAL NEURAL NETWORKS"
from fightingcv.attention.ShuffleAttention import ShuffleAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,512,7,7)
se = ShuffleAttention(channel=512,G=8)
output=se(input)
print(output.shape)
"MUSE: Parallel Multi-Scale Attention for Sequence to Sequence Learning"
from fightingcv.attention.MUSEAttention import MUSEAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,49,512)
sa = MUSEAttention(d_model=512, d_k=512, d_v=512, h=8)
output=sa(input,input,input)
print(output.shape)Spatial Group-wise Enhance: Improving Semantic Feature Learning in Convolutional Networks
from fightingcv.attention.SGE import SpatialGroupEnhance
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,512,7,7)
sge = SpatialGroupEnhance(groups=8)
output=sge(input)
print(output.shape)A2-Nets: Double Attention Networks
from fightingcv.attention.A2Atttention import DoubleAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,512,7,7)
a2 = DoubleAttention(512,128,128,True)
output=a2(input)
print(output.shape)
from fightingcv.attention.AFT import AFT_FULL
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,49,512)
aft_full = AFT_FULL(d_model=512, n=49)
output=aft_full(input)
print(output.shape)VOLO: Vision Outlooker for Visual Recognition"
from fightingcv.attention.OutlookAttention import OutlookAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,28,28,512)
outlook = OutlookAttention(dim=512)
output=outlook(input)
print(output.shape)Vision Permutator: A Permutable MLP-Like Architecture for Visual Recognition"
from fightingcv.attention.ViP import WeightedPermuteMLP
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(64,8,8,512)
seg_dim=8
vip=WeightedPermuteMLP(512,seg_dim)
out=vip(input)
print(out.shape)CoAtNet: Marrying Convolution and Attention for All Data Sizes"
None
from fightingcv.attention.CoAtNet import CoAtNet
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,3,224,224)
mbconv=CoAtNet(in_ch=3,image_size=224)
out=mbconv(input)
print(out.shape)Scaling Local Self-Attention for Parameter Efficient Visual Backbones"
from fightingcv.attention.HaloAttention import HaloAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,512,8,8)
halo = HaloAttention(dim=512,
block_size=2,
halo_size=1,)
output=halo(input)
print(output.shape)Polarized Self-Attention: Towards High-quality Pixel-wise Regression"
from fightingcv.attention.PolarizedSelfAttention import ParallelPolarizedSelfAttention,SequentialPolarizedSelfAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,512,7,7)
psa = SequentialPolarizedSelfAttention(channel=512)
output=psa(input)
print(output.shape)
Contextual Transformer Networks for Visual Recognition---arXiv 2021.07.26
from fightingcv.attention.CoTAttention import CoTAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,512,7,7)
cot = CoTAttention(dim=512,kernel_size=3)
output=cot(input)
print(output.shape)
Residual Attention: A Simple but Effective Method for Multi-Label Recognition---ICCV2021
from fightingcv.attention.ResidualAttention import ResidualAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,512,7,7)
resatt = ResidualAttention(channel=512,num_class=1000,la=0.2)
output=resatt(input)
print(output.shape)
S²-MLPv2: Improved Spatial-Shift MLP Architecture for Vision---arXiv 2021.08.02
from fightingcv.attention.S2Attention import S2Attention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,512,7,7)
s2att = S2Attention(channels=512)
output=s2att(input)
print(output.shape)Global Filter Networks for Image Classification---arXiv 2021.07.01
25.3. Code - Implemented by Wenliang Zhao (Author)
from fightingcv.attention.gfnet import GFNet
import torch
from torch import nn
from torch.nn import functional as F
x = torch.randn(1, 3, 224, 224)
gfnet = GFNet(embed_dim=384, img_size=224, patch_size=16, num_classes=1000)
out = gfnet(x)
print(out.shape)Rotate to Attend: Convolutional Triplet Attention Module---WACV 2021
26.3. Code - Implemented by digantamisra98
from fightingcv.attention.TripletAttention import TripletAttention
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(50,512,7,7)
triplet = TripletAttention()
output=triplet(input)
print(output.shape)-
Pytorch implementation of "Deep Residual Learning for Image Recognition---CVPR2016 Best Paper"
-
Pytorch implementation of "Aggregated Residual Transformations for Deep Neural Networks---CVPR2017"
"Deep Residual Learning for Image Recognition---CVPR2016 Best Paper"
from fightingcv.backbone_cnn.resnet import ResNet50,ResNet101,ResNet152
import torch
if __name__ == '__main__':
input=torch.randn(50,3,224,224)
resnet50=ResNet50(1000)
# resnet101=ResNet101(1000)
# resnet152=ResNet152(1000)
out=resnet50(input)
print(out.shape)"Aggregated Residual Transformations for Deep Neural Networks---CVPR2017"
from fightingcv.backbone_cnn.resnext import ResNeXt50,ResNeXt101,ResNeXt152
import torch
if __name__ == '__main__':
input=torch.randn(50,3,224,224)
resnext50=ResNeXt50(1000)
# resnext101=ResNeXt101(1000)
# resnext152=ResNeXt152(1000)
out=resnext50(input)
print(out.shape)
-
Pytorch implementation of "RepMLP: Re-parameterizing Convolutions into Fully-connected Layers for Image Recognition---arXiv 2021.05.05"
-
Pytorch implementation of "MLP-Mixer: An all-MLP Architecture for Vision---arXiv 2021.05.17"
-
Pytorch implementation of "ResMLP: Feedforward networks for image classification with data-efficient training---arXiv 2021.05.07"
-
Pytorch implementation of "Pay Attention to MLPs---arXiv 2021.05.17"
-
Pytorch implementation of "Sparse MLP for Image Recognition: Is Self-Attention Really Necessary?---arXiv 2021.09.12"
"RepMLP: Re-parameterizing Convolutions into Fully-connected Layers for Image Recognition"
from fightingcv.mlp.repmlp import RepMLP
import torch
from torch import nn
N=4 #batch size
C=512 #input dim
O=1024 #output dim
H=14 #image height
W=14 #image width
h=7 #patch height
w=7 #patch width
fc1_fc2_reduction=1 #reduction ratio
fc3_groups=8 # groups
repconv_kernels=[1,3,5,7] #kernel list
repmlp=RepMLP(C,O,H,W,h,w,fc1_fc2_reduction,fc3_groups,repconv_kernels=repconv_kernels)
x=torch.randn(N,C,H,W)
repmlp.eval()
for module in repmlp.modules():
if isinstance(module, nn.BatchNorm2d) or isinstance(module, nn.BatchNorm1d):
nn.init.uniform_(module.running_mean, 0, 0.1)
nn.init.uniform_(module.running_var, 0, 0.1)
nn.init.uniform_(module.weight, 0, 0.1)
nn.init.uniform_(module.bias, 0, 0.1)
#training result
out=repmlp(x)
#inference result
repmlp.switch_to_deploy()
deployout = repmlp(x)
print(((deployout-out)**2).sum())"MLP-Mixer: An all-MLP Architecture for Vision"
from fightingcv.mlp.mlp_mixer import MlpMixer
import torch
mlp_mixer=MlpMixer(num_classes=1000,num_blocks=10,patch_size=10,tokens_hidden_dim=32,channels_hidden_dim=1024,tokens_mlp_dim=16,channels_mlp_dim=1024)
input=torch.randn(50,3,40,40)
output=mlp_mixer(input)
print(output.shape)"ResMLP: Feedforward networks for image classification with data-efficient training"
from fightingcv.mlp.resmlp import ResMLP
import torch
input=torch.randn(50,3,14,14)
resmlp=ResMLP(dim=128,image_size=14,patch_size=7,class_num=1000)
out=resmlp(input)
print(out.shape) #the last dimention is class_num
from fightingcv.mlp.g_mlp import gMLP
import torch
num_tokens=10000
bs=50
len_sen=49
num_layers=6
input=torch.randint(num_tokens,(bs,len_sen)) #bs,len_sen
gmlp = gMLP(num_tokens=num_tokens,len_sen=len_sen,dim=512,d_ff=1024)
output=gmlp(input)
print(output.shape)"Sparse MLP for Image Recognition: Is Self-Attention Really Necessary?"
from fightingcv.mlp.sMLP_block import sMLPBlock
import torch
from torch import nn
from torch.nn import functional as F
if __name__ == '__main__':
input=torch.randn(50,3,224,224)
smlp=sMLPBlock(h=224,w=224)
out=smlp(input)
print(out.shape)-
Pytorch implementation of "RepVGG: Making VGG-style ConvNets Great Again---CVPR2021"
-
Pytorch implementation of "ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks---ICCV2019"
-
Pytorch implementation of "Diverse Branch Block: Building a Convolution as an Inception-like Unit---CVPR2021"
"RepVGG: Making VGG-style ConvNets Great Again"
from fightingcv.rep.repvgg import RepBlock
import torch
input=torch.randn(50,512,49,49)
repblock=RepBlock(512,512)
repblock.eval()
out=repblock(input)
repblock._switch_to_deploy()
out2=repblock(input)
print('difference between vgg and repvgg')
print(((out2-out)**2).sum())"ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks"
from fightingcv.rep.acnet import ACNet
import torch
from torch import nn
input=torch.randn(50,512,49,49)
acnet=ACNet(512,512)
acnet.eval()
out=acnet(input)
acnet._switch_to_deploy()
out2=acnet(input)
print('difference:')
print(((out2-out)**2).sum())"Diverse Branch Block: Building a Convolution as an Inception-like Unit"
from fightingcv.rep.ddb import transI_conv_bn
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,64,7,7)
#conv+bn
conv1=nn.Conv2d(64,64,3,padding=1)
bn1=nn.BatchNorm2d(64)
bn1.eval()
out1=bn1(conv1(input))
#conv_fuse
conv_fuse=nn.Conv2d(64,64,3,padding=1)
conv_fuse.weight.data,conv_fuse.bias.data=transI_conv_bn(conv1,bn1)
out2=conv_fuse(input)
print("difference:",((out2-out1)**2).sum().item())from fightingcv.rep.ddb import transII_conv_branch
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,64,7,7)
#conv+conv
conv1=nn.Conv2d(64,64,3,padding=1)
conv2=nn.Conv2d(64,64,3,padding=1)
out1=conv1(input)+conv2(input)
#conv_fuse
conv_fuse=nn.Conv2d(64,64,3,padding=1)
conv_fuse.weight.data,conv_fuse.bias.data=transII_conv_branch(conv1,conv2)
out2=conv_fuse(input)
print("difference:",((out2-out1)**2).sum().item())from fightingcv.rep.ddb import transIII_conv_sequential
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,64,7,7)
#conv+conv
conv1=nn.Conv2d(64,64,1,padding=0,bias=False)
conv2=nn.Conv2d(64,64,3,padding=1,bias=False)
out1=conv2(conv1(input))
#conv_fuse
conv_fuse=nn.Conv2d(64,64,3,padding=1,bias=False)
conv_fuse.weight.data=transIII_conv_sequential(conv1,conv2)
out2=conv_fuse(input)
print("difference:",((out2-out1)**2).sum().item())from fightingcv.rep.ddb import transIV_conv_concat
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,64,7,7)
#conv+conv
conv1=nn.Conv2d(64,32,3,padding=1)
conv2=nn.Conv2d(64,32,3,padding=1)
out1=torch.cat([conv1(input),conv2(input)],dim=1)
#conv_fuse
conv_fuse=nn.Conv2d(64,64,3,padding=1)
conv_fuse.weight.data,conv_fuse.bias.data=transIV_conv_concat(conv1,conv2)
out2=conv_fuse(input)
print("difference:",((out2-out1)**2).sum().item())from fightingcv.rep.ddb import transV_avg
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,64,7,7)
avg=nn.AvgPool2d(kernel_size=3,stride=1)
out1=avg(input)
conv=transV_avg(64,3)
out2=conv(input)
print("difference:",((out2-out1)**2).sum().item())from fightingcv.rep.ddb import transVI_conv_scale
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,64,7,7)
#conv+conv
conv1x1=nn.Conv2d(64,64,1)
conv1x3=nn.Conv2d(64,64,(1,3),padding=(0,1))
conv3x1=nn.Conv2d(64,64,(3,1),padding=(1,0))
out1=conv1x1(input)+conv1x3(input)+conv3x1(input)
#conv_fuse
conv_fuse=nn.Conv2d(64,64,3,padding=1)
conv_fuse.weight.data,conv_fuse.bias.data=transVI_conv_scale(conv1x1,conv1x3,conv3x1)
out2=conv_fuse(input)
print("difference:",((out2-out1)**2).sum().item())-
Pytorch implementation of "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications---CVPR2017"
-
Pytorch implementation of "Efficientnet: Rethinking model scaling for convolutional neural networks---PMLR2019"
-
Pytorch implementation of "Involution: Inverting the Inherence of Convolution for Visual Recognition---CVPR2021"
"MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications"
from fightingcv.conv.DepthwiseSeparableConvolution import DepthwiseSeparableConvolution
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,3,224,224)
dsconv=DepthwiseSeparableConvolution(3,64)
out=dsconv(input)
print(out.shape)"Efficientnet: Rethinking model scaling for convolutional neural networks"
from fightingcv.conv.MBConv import MBConvBlock
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,3,224,224)
mbconv=MBConvBlock(ksize=3,input_filters=3,output_filters=512,image_size=224)
out=mbconv(input)
print(out.shape)
"Involution: Inverting the Inherence of Convolution for Visual Recognition"
from fightingcv.conv.Involution import Involution
import torch
from torch import nn
from torch.nn import functional as F
input=torch.randn(1,4,64,64)
involution=Involution(kernel_size=3,in_channel=4,stride=2)
out=involution(input)
print(out.shape)