feat(net): 替换 token_mixer 为 SCSA 模块
- 引入新的 SCSA（空间和通道协同注意力）模块 - 用 SCSA 替换原有的 Pooling层作为 token_mixer - 删除了未使用的 SEBlock.py 文件- 移除了与当前项目无关的 TIAM(CV).py 文件
This commit is contained in:
zjut
parent
aae81d97fd
commit
23293e9138
156
componets/SCSA.py
Normal file
156
componets/SCSA.py
Normal file
@ -0,0 +1,156 @@
|
||||
import typing as t
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from einops.einops import rearrange
|
||||
from mmengine.model import BaseModule
|
||||
__all__ = ['SCSA']
|
||||
|
||||
"""SCSA：探索空间注意力和通道注意力之间的协同作用
|
||||
通道和空间注意力分别在为各种下游视觉任务提取特征依赖性和空间结构关系方面带来了显着的改进。
|
||||
虽然它们的结合更有利于发挥各自的优势，但通道和空间注意力之间的协同作用尚未得到充分探索，缺乏充分利用多语义信息的协同潜力来进行特征引导和缓解语义差异。
|
||||
我们的研究试图在多个语义层面揭示空间和通道注意力之间的协同关系，提出了一种新颖的空间和通道协同注意力模块（SCSA）。我们的SCSA由两部分组成：可共享的多语义空间注意力（SMSA）和渐进式通道自注意力（PCSA）。
|
||||
SMSA 集成多语义信息并利用渐进式压缩策略将判别性空间先验注入 PCSA 的通道自注意力中，有效地指导通道重新校准。此外，PCSA 中基于自注意力机制的稳健特征交互进一步缓解了 SMSA 中不同子特征之间多语义信息的差异。
|
||||
我们在七个基准数据集上进行了广泛的实验，包括 ImageNet-1K 上的分类、MSCOCO 2017 上的对象检测、ADE20K 上的分割以及其他四个复杂场景检测数据集。我们的结果表明，我们提出的 SCSA 不仅超越了当前最先进的注意力机制，
|
||||
而且在各种任务场景中表现出增强的泛化能力。
|
||||
"""
|
||||
|
||||
class SCSA(BaseModule):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
dim: int,
|
||||
head_num: int,
|
||||
window_size: int = 7,
|
||||
group_kernel_sizes: t.List[int] = [3, 5, 7, 9],
|
||||
qkv_bias: bool = False,
|
||||
fuse_bn: bool = False,
|
||||
norm_cfg: t.Dict = dict(type='BN'),
|
||||
act_cfg: t.Dict = dict(type='ReLU'),
|
||||
down_sample_mode: str = 'avg_pool',
|
||||
attn_drop_ratio: float = 0.,
|
||||
gate_layer: str = 'sigmoid',
|
||||
):
|
||||
super(SCSA, self).__init__()
|
||||
self.dim = dim
|
||||
self.head_num = head_num
|
||||
self.head_dim = dim // head_num
|
||||
self.scaler = self.head_dim ** -0.5
|
||||
self.group_kernel_sizes = group_kernel_sizes
|
||||
self.window_size = window_size
|
||||
self.qkv_bias = qkv_bias
|
||||
self.fuse_bn = fuse_bn
|
||||
self.down_sample_mode = down_sample_mode
|
||||
|
||||
assert self.dim // 4, 'The dimension of input feature should be divisible by 4.'
|
||||
self.group_chans = group_chans = self.dim // 4
|
||||
|
||||
self.local_dwc = nn.Conv1d(group_chans, group_chans, kernel_size=group_kernel_sizes[0],
|
||||
padding=group_kernel_sizes[0] // 2, groups=group_chans)
|
||||
self.global_dwc_s = nn.Conv1d(group_chans, group_chans, kernel_size=group_kernel_sizes[1],
|
||||
padding=group_kernel_sizes[1] // 2, groups=group_chans)
|
||||
self.global_dwc_m = nn.Conv1d(group_chans, group_chans, kernel_size=group_kernel_sizes[2],
|
||||
padding=group_kernel_sizes[2] // 2, groups=group_chans)
|
||||
self.global_dwc_l = nn.Conv1d(group_chans, group_chans, kernel_size=group_kernel_sizes[3],
|
||||
padding=group_kernel_sizes[3] // 2, groups=group_chans)
|
||||
self.sa_gate = nn.Softmax(dim=2) if gate_layer == 'softmax' else nn.Sigmoid()
|
||||
self.norm_h = nn.GroupNorm(4, dim)
|
||||
self.norm_w = nn.GroupNorm(4, dim)
|
||||
|
||||
self.conv_d = nn.Identity()
|
||||
self.norm = nn.GroupNorm(1, dim)
|
||||
self.q = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, bias=qkv_bias, groups=dim)
|
||||
self.k = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, bias=qkv_bias, groups=dim)
|
||||
self.v = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, bias=qkv_bias, groups=dim)
|
||||
self.attn_drop = nn.Dropout(attn_drop_ratio)
|
||||
self.ca_gate = nn.Softmax(dim=1) if gate_layer == 'softmax' else nn.Sigmoid()
|
||||
|
||||
if window_size == -1:
|
||||
self.down_func = nn.AdaptiveAvgPool2d((1, 1))
|
||||
else:
|
||||
if down_sample_mode == 'recombination':
|
||||
self.down_func = self.space_to_chans
|
||||
# dimensionality reduction
|
||||
self.conv_d = nn.Conv2d(in_channels=dim * window_size ** 2, out_channels=dim, kernel_size=1, bias=False)
|
||||
elif down_sample_mode == 'avg_pool':
|
||||
self.down_func = nn.AvgPool2d(kernel_size=(window_size, window_size), stride=window_size)
|
||||
elif down_sample_mode == 'max_pool':
|
||||
self.down_func = nn.MaxPool2d(kernel_size=(window_size, window_size), stride=window_size)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""
|
||||
The dim of x is (B, C, H, W)
|
||||
"""
|
||||
# Spatial attention priority calculation
|
||||
b, c, h_, w_ = x.size()
|
||||
# (B, C, H)
|
||||
x_h = x.mean(dim=3)
|
||||
l_x_h, g_x_h_s, g_x_h_m, g_x_h_l = torch.split(x_h, self.group_chans, dim=1)
|
||||
# (B, C, W)
|
||||
x_w = x.mean(dim=2)
|
||||
l_x_w, g_x_w_s, g_x_w_m, g_x_w_l = torch.split(x_w, self.group_chans, dim=1)
|
||||
|
||||
x_h_attn = self.sa_gate(self.norm_h(torch.cat((
|
||||
self.local_dwc(l_x_h),
|
||||
self.global_dwc_s(g_x_h_s),
|
||||
self.global_dwc_m(g_x_h_m),
|
||||
self.global_dwc_l(g_x_h_l),
|
||||
), dim=1)))
|
||||
x_h_attn = x_h_attn.view(b, c, h_, 1)
|
||||
|
||||
x_w_attn = self.sa_gate(self.norm_w(torch.cat((
|
||||
self.local_dwc(l_x_w),
|
||||
self.global_dwc_s(g_x_w_s),
|
||||
self.global_dwc_m(g_x_w_m),
|
||||
self.global_dwc_l(g_x_w_l)
|
||||
), dim=1)))
|
||||
x_w_attn = x_w_attn.view(b, c, 1, w_)
|
||||
|
||||
x = x * x_h_attn * x_w_attn
|
||||
|
||||
# Channel attention based on self attention
|
||||
# reduce calculations
|
||||
y = self.down_func(x)
|
||||
y = self.conv_d(y)
|
||||
_, _, h_, w_ = y.size()
|
||||
|
||||
# normalization first, then reshape -> (B, H, W, C) -> (B, C, H * W) and generate q, k and v
|
||||
y = self.norm(y)
|
||||
q = self.q(y)
|
||||
k = self.k(y)
|
||||
v = self.v(y)
|
||||
# (B, C, H, W) -> (B, head_num, head_dim, N)
|
||||
q = rearrange(q, 'b (head_num head_dim) h w -> b head_num head_dim (h w)', head_num=int(self.head_num),
|
||||
head_dim=int(self.head_dim))
|
||||
k = rearrange(k, 'b (head_num head_dim) h w -> b head_num head_dim (h w)', head_num=int(self.head_num),
|
||||
head_dim=int(self.head_dim))
|
||||
v = rearrange(v, 'b (head_num head_dim) h w -> b head_num head_dim (h w)', head_num=int(self.head_num),
|
||||
head_dim=int(self.head_dim))
|
||||
|
||||
# (B, head_num, head_dim, head_dim)
|
||||
attn = q @ k.transpose(-2, -1) * self.scaler
|
||||
attn = self.attn_drop(attn.softmax(dim=-1))
|
||||
# (B, head_num, head_dim, N)
|
||||
attn = attn @ v
|
||||
# (B, C, H_, W_)
|
||||
attn = rearrange(attn, 'b head_num head_dim (h w) -> b (head_num head_dim) h w', h=int(h_), w=int(w_))
|
||||
# (B, C, 1, 1)
|
||||
attn = attn.mean((2, 3), keepdim=True)
|
||||
attn = self.ca_gate(attn)
|
||||
return attn * x
|
||||
|
||||
if __name__ == '__main__':
|
||||
|
||||
block = SCSA(
|
||||
dim=256,
|
||||
head_num=8,
|
||||
)
|
||||
|
||||
input_tensor = torch.rand(1, 256, 32, 32)
|
||||
|
||||
# 调用模块进行前向传播
|
||||
output_tensor = block(input_tensor)
|
||||
|
||||
# 打印输入和输出张量的大小
|
||||
print("Input size:", input_tensor.size())
|
||||
print("Output size:", output_tensor.size())
|
||||
@ -1,110 +0,0 @@
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
"""Elsevier2024
|
||||
变化检测 (CD) 是地球观测中一种重要的监测方法，尤其适用于土地利用分析、城市管理和灾害损失评估。然而，在星座互联和空天协作时代，感兴趣区域 (ROI) 的变化由于几何透视旋转和时间风格差异而导致许多错误检测。
|
||||
为了应对这些挑战，我们引入了 CDNeXt，该框架阐明了一种稳健而有效的方法，用于将基于预训练主干的 Siamese 网络与用于遥感图像的创新时空交互注意模块 (TIAM) 相结合。
|
||||
CDNeXt 可分为四个主要组件：编码器、交互器、解码器和检测器。值得注意的是，由 TIAM 提供支持的交互器从编码器提取的二进制时间特征中查询和重建空间透视依赖关系和时间风格相关性，以扩大 ROI 变化的差异。
|
||||
最后，检测器集成解码器生成的分层特征，随后生成二进制变化掩码。
|
||||
"""
|
||||
|
||||
class SpatiotemporalAttentionFullNotWeightShared(nn.Module):
|
||||
def __init__(self, in_channels, inter_channels=None, dimension=2, sub_sample=False):
|
||||
super(SpatiotemporalAttentionFullNotWeightShared, self).__init__()
|
||||
assert dimension in [2, ]
|
||||
self.dimension = dimension
|
||||
self.sub_sample = sub_sample
|
||||
self.in_channels = in_channels
|
||||
self.inter_channels = inter_channels
|
||||
|
||||
if self.inter_channels is None:
|
||||
self.inter_channels = in_channels // 2
|
||||
if self.inter_channels == 0:
|
||||
self.inter_channels = 1
|
||||
|
||||
self.g1 = nn.Sequential(
|
||||
nn.BatchNorm2d(self.in_channels),
|
||||
nn.Conv2d(in_channels=self.in_channels, out_channels=self.inter_channels,
|
||||
kernel_size=1, stride=1, padding=0)
|
||||
)
|
||||
self.g2 = nn.Sequential(
|
||||
nn.BatchNorm2d(self.in_channels),
|
||||
nn.Conv2d(in_channels=self.in_channels, out_channels=self.inter_channels,
|
||||
kernel_size=1, stride=1, padding=0),
|
||||
)
|
||||
|
||||
self.W1 = nn.Sequential(
|
||||
nn.Conv2d(in_channels=self.inter_channels, out_channels=self.in_channels,
|
||||
kernel_size=1, stride=1, padding=0),
|
||||
nn.BatchNorm2d(self.in_channels)
|
||||
)
|
||||
self.W2 = nn.Sequential(
|
||||
nn.Conv2d(in_channels=self.inter_channels, out_channels=self.in_channels,
|
||||
kernel_size=1, stride=1, padding=0),
|
||||
nn.BatchNorm2d(self.in_channels)
|
||||
)
|
||||
self.theta = nn.Sequential(
|
||||
nn.BatchNorm2d(self.in_channels),
|
||||
nn.Conv2d(in_channels=self.in_channels, out_channels=self.inter_channels,
|
||||
kernel_size=1, stride=1, padding=0),
|
||||
)
|
||||
self.phi = nn.Sequential(
|
||||
nn.BatchNorm2d(self.in_channels),
|
||||
nn.Conv2d(in_channels=self.in_channels, out_channels=self.inter_channels,
|
||||
kernel_size=1, stride=1, padding=0),
|
||||
)
|
||||
|
||||
def forward(self, x1, x2):
|
||||
"""
|
||||
:param x: (b, c, h, w)
|
||||
:param return_nl_map: if True return z, nl_map, else only return z.
|
||||
:return:
|
||||
"""
|
||||
batch_size = x1.size(0)
|
||||
g_x11 = self.g1(x1).reshape(batch_size, self.inter_channels, -1)
|
||||
g_x12 = g_x11.permute(0, 2, 1)
|
||||
g_x21 = self.g2(x2).reshape(batch_size, self.inter_channels, -1)
|
||||
g_x22 = g_x21.permute(0, 2, 1)
|
||||
|
||||
theta_x1 = self.theta(x1).reshape(batch_size, self.inter_channels, -1)
|
||||
theta_x2 = theta_x1.permute(0, 2, 1)
|
||||
|
||||
phi_x1 = self.phi(x2).reshape(batch_size, self.inter_channels, -1)
|
||||
phi_x2 = phi_x1.permute(0, 2, 1)
|
||||
|
||||
energy_time_1 = torch.matmul(theta_x1, phi_x2)
|
||||
energy_time_2 = energy_time_1.permute(0, 2, 1)
|
||||
energy_space_1 = torch.matmul(theta_x2, phi_x1)
|
||||
energy_space_2 = energy_space_1.permute(0, 2, 1)
|
||||
|
||||
energy_time_1s = F.softmax(energy_time_1, dim=-1)
|
||||
energy_time_2s = F.softmax(energy_time_2, dim=-1)
|
||||
energy_space_2s = F.softmax(energy_space_1, dim=-2)
|
||||
energy_space_1s = F.softmax(energy_space_2, dim=-2)
|
||||
# C1*S(C2) energy_time_1s * C1*H1W1 g_x12 * energy_space_1s S(H2W2)*H1W1 -> C1*H1W1
|
||||
y1 = torch.matmul(torch.matmul(energy_time_2s, g_x11), energy_space_2s).contiguous() # C2*H2W2
|
||||
# C2*S(C1) energy_time_2s * C2*H2W2 g_x21 * energy_space_2s S(H1W1)*H2W2 -> C2*H2W2
|
||||
y2 = torch.matmul(torch.matmul(energy_time_1s, g_x21), energy_space_1s).contiguous() # C1*H1W1
|
||||
y1 = y1.reshape(batch_size, self.inter_channels, *x2.size()[2:])
|
||||
y2 = y2.reshape(batch_size, self.inter_channels, *x1.size()[2:])
|
||||
return x1 + self.W1(y1), x2 + self.W2(y2)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
in_channels = 64
|
||||
batch_size = 8
|
||||
height = 32
|
||||
width = 32
|
||||
|
||||
block = SpatiotemporalAttentionFullNotWeightShared(in_channels=in_channels)
|
||||
|
||||
input1 = torch.rand(batch_size, in_channels, height, width)
|
||||
input2 = torch.rand(batch_size, in_channels, height, width)
|
||||
|
||||
output1, output2 = block(input1, input2)
|
||||
|
||||
print(f"Input1 size: {input1.size()}")
|
||||
print(f"Input2 size: {input2.size()}")
|
||||
print(f"Output1 size: {output1.size()}")
|
||||
print(f"Output2 size: {output2.size()}")
|
||||
5
net.py
5
net.py
@ -6,6 +6,8 @@ import torch.utils.checkpoint as checkpoint
|
||||
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
|
||||
from einops import rearrange
|
||||
|
||||
from componets.SCSA import SCSA
|
||||
|
||||
|
||||
def drop_path(x, drop_prob: float = 0., training: bool = False):
|
||||
if drop_prob == 0. or not training:
|
||||
@ -164,7 +166,8 @@ class BaseFeatureExtractionSAR(nn.Module):
|
||||
super().__init__()
|
||||
|
||||
self.norm1 = LayerNorm(dim, 'WithBias')
|
||||
self.token_mixer = Pooling(kernel_size=pool_size) # vits是msa，MLPs是mlp，这个用pool来替代
|
||||
self.token_mixer = SCSA(dim=dim,head_num=8)
|
||||
# self.token_mixer = Pooling(kernel_size=pool_size) # vits是msa，MLPs是mlp，这个用pool来替代
|
||||
self.norm2 = LayerNorm(dim, 'WithBias')
|
||||
mlp_hidden_dim = int(dim * mlp_ratio)
|
||||
self.poolmlp = PoolMlp(in_features=dim, hidden_features=mlp_hidden_dim,
|
||||
|
||||
Loading…
Reference in New Issue
Block a user