yolov8添加注意力机制三部曲：添加localwindowsattention注意力

选择过程，最先使用的BIformer，but，显存炸了，搜索发现BI方法就是存在内存占用大计算代价高的特点，所以选了localwindowsattention

那么如何在yolov8中加入注意力呢，和把大象装进冰箱的步骤一样

第一步：加入注意力代码，代码位置nn-modules-conv.py

最上方添加

from torch import Tensor
from typing import Tuple
from einops import rearrange#没有下载的需要先下载
import torch.nn.functional as F
import itertools

然后，在代码最下方添加注意力代码，源码放在下方，注意：conv2d_BN也需要

class Conv2d_BN(torch.nn.Sequential):
    def __init__(self, a, b, ks=1, stride=1, pad=0, dilation=1,
                 groups=1, bn_weight_init=1, resolution=-10000):
        super().__init__()
        self.add_module('c', torch.nn.Conv2d(
            a, b, ks, stride, pad, dilation, groups, bias=False))
        self.add_module('bn', torch.nn.BatchNorm2d(b))
        torch.nn.init.constant_(self.bn.weight, bn_weight_init)
        torch.nn.init.constant_(self.bn.bias, 0)

    @torch.no_grad()
    def switch_to_deploy(self):
        c, bn = self._modules.values()
        w = bn.weight / (bn.running_var + bn.eps)**0.5
        w = c.weight * w[:, None, None, None]
        b = bn.bias - bn.running_mean * bn.weight / \
            (bn.running_var + bn.eps)**0.5
        m = torch.nn.Conv2d(w.size(1) * self.c.groups, w.size(
            0), w.shape[2:], stride=self.c.stride, padding=self.c.padding, dilation=self.c.dilation, groups=self.c.groups)
        m.weight.data.copy_(w)
        m.bias.data.copy_(b)
        return m
class CascadedGroupAttention(torch.nn.Module):
    r""" Cascaded Group Attention.

    Args:
        dim (int): Number of in