![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260125070846-6975c17eb09e8.png\")
<\/p>\n\u672c\u6587\u4ecb\u7ecdCBSA\u6536\u7f29 – \u5e7f\u64ad\u81ea\u6ce8\u610f\u529b\u673a\u5236（Contract-and-Broadcast Self-Attention），\u5e76\u5c06\u5176\u96c6\u6210\u5230ultralytics\u6700\u65b0\u53d1\u5e03\u7684YOLO26\u76ee\u6807\u68c0\u6d4b\u7b97\u6cd5\u4e2d，\u6784\u5efaC3k2-LWGA\u521b\u65b0\u6a21\u5757\u3002\u4f20\u7edf\u6ce8\u610f\u529b\u673a\u5236\u5b58\u5728\u9ed1\u76d2\u96be\u7406\u89e3\u3001\u8ba1\u7b97\u590d\u6742\u5ea6\u9ad8\u7684\u95ee\u9898，CBSA\u901a\u8fc7\u7b97\u6cd5\u5c55\u5f00\u63a8\u5bfc\u51fa\u672c\u8d28\u4e0a\u53ef\u89e3\u91ca\u4e14\u9ad8\u6548\u7684\u6ce8\u610f\u529b\u673a\u5236\u3002\u5b83\u5148\u4ece\u8f93\u5165\u6570\u636e\u4e2d\u9009\u51fa\u5c11\u91cf\u4ee3\u8868\u6027tokens，\u63a5\u7740\u5bf9\u4ee3\u8868\u8fdb\u884c\u6536\u7f29\u8ba1\u7b97，\u518d\u5c06\u7ed3\u679c\u5e7f\u64ad\u7ed9\u6240\u6709\u539f\u59cb\u6570\u636e\u3002\u8be5\u673a\u5236\u8ba1\u7b97\u91cf\u7ebf\u6027\u589e\u957f，\u5177\u6709\u660e\u786e\u6570\u5b66\u89e3\u91ca，\u8fd8\u80fd\u7edf\u4e00\u591a\u79cd\u6ce8\u610f\u529b\u673a\u5236\u3002<\/p>\n
\u6ce8\u610f\u529b\u673a\u5236\u5728\u591a\u4e2a\u9886\u57df\u53d6\u5f97\u4e86\u663e\u8457\u7684\u5b9e\u8bc1\u6210\u529f，\u4f46\u5b83\u4eec\u7684\u57fa\u7840\u4f18\u5316\u76ee\u6807\u4ecd\u4e0d\u660e\u786e\u3002\u6b64\u5916，\u81ea\u6ce8\u610f\u529b\u7684\u4e8c\u6b21\u590d\u6742\u5ea6\u53d8\u5f97\u8d8a\u6765\u8d8a\u96be\u4ee5\u627f\u53d7\u3002\u867d\u7136\u53ef\u89e3\u91ca\u6027\u548c\u6548\u7387\u662f\u4e24\u4e2a\u76f8\u4e92\u4fc3\u8fdb\u7684\u8ffd\u6c42，\u4f46\u5148\u524d\u7684\u5de5\u4f5c\u901a\u5e38\u5206\u5f00\u7814\u7a76\u5b83\u4eec\u3002\u5728\u672c\u6587\u4e2d，\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u7edf\u4e00\u7684\u4f18\u5316\u76ee\u6807，\u901a\u8fc7\u7b97\u6cd5\u5c55\u5f00\u63a8\u5bfc\u51fa\u672c\u8d28\u4e0a\u53ef\u89e3\u91ca\u4e14\u9ad8\u6548\u7684\u6ce8\u610f\u529b\u673a\u5236\u3002\u5177\u4f53\u800c\u8a00，\u6211\u4eec\u6784\u5efa\u4e86\u6240\u63d0\u51fa\u76ee\u6807\u7684\u4e00\u4e2a\u68af\u5ea6\u6b65\u9aa4，\u5176\u4e2d\u5305\u542b\u6211\u4eec\u7684\u6536\u7f29-\u5e7f\u64ad\u81ea\u6ce8\u610f\u529b(CBSA)\u7684\u4e00\u7cfb\u5217\u524d\u5411\u4f20\u9012\u64cd\u4f5c，\u8be5\u673a\u5236\u901a\u8fc7\u6536\u7f29\u8f93\u5165\u6807\u8bb0\u7684\u5c11\u91cf\u4ee3\u8868\u6765\u5c06\u8f93\u5165\u6807\u8bb0\u538b\u7f29\u5230\u4f4e\u7ef4\u7ed3\u6784\u3002\u8fd9\u79cd\u65b0\u9896\u7684\u673a\u5236\u4e0d\u4ec5\u53ef\u4ee5\u901a\u8fc7\u56fa\u5b9a\u4ee3\u8868\u6570\u91cf\u6765\u5b9e\u73b0\u7ebf\u6027\u6269\u5c55，\u8fd8\u53ef\u4ee5\u5728\u4f7f\u7528\u4e0d\u540c\u4ee3\u8868\u96c6\u65f6\u6db5\u76d6\u5404\u79cd\u6ce8\u610f\u529b\u673a\u5236\u7684\u5b9e\u4f8b\u5316\u3002\u6211\u4eec\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u5b9e\u9a8c，\u8bc1\u660e\u4e86\u4e0e\u9ed1\u76d2\u6ce8\u610f\u529b\u673a\u5236\u76f8\u6bd4，\u5728\u89c6\u89c9\u4efb\u52a1\u4e0a\u5177\u6709\u53ef\u6bd4\u7684\u6027\u80fd\u548c\u4f18\u8d8a\u7684\u4f18\u52bf\u3002\u6211\u4eec\u7684\u5de5\u4f5c\u9610\u660e\u4e86\u53ef\u89e3\u91ca\u6027\u548c\u6548\u7387\u7684\u6574\u5408，\u4ee5\u53ca\u6ce8\u610f\u529b\u673a\u5236\u7684\u7edf\u4e00\u516c\u5f0f\u3002<\/p>\n
<\/p>\n
\u539f\u59cb\u8bba\u6587：https:\/\/arxiv.org\/pdf\/2509.16875<\/p>\n
\u539f\u59cb\u4ee3\u7801：https:\/\/github.com\/QishuaiWen\/CBSA<\/p>\n<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260125070847-6975c17f04bd0.png\")
<\/p>\n
\u4e00\u3001\u6838\u5fc3\u5b9a\u4f4d：\u7ed9\u6ce8\u610f\u529b\u673a\u5236\u201c\u62c6\u9ed1\u76d2\u3001\u63d0\u901f\u5ea6\u3001\u505a\u7edf\u4e00\u201d<\/p>\n
\u4f20\u7edf\u6ce8\u610f\u529b\u673a\u5236（\u6bd4\u5982Transformer\u91cc\u7684softmax\u6ce8\u610f\u529b）\u6709\u4e24\u4e2a\u5927\u95ee\u9898：<\/p>\n
\u4e4b\u524d\u7684\u7814\u7a76\u8981\u4e48\u53ea\u89e3\u51b3\u201c\u7406\u89e3\u201d，\u8981\u4e48\u53ea\u89e3\u51b3\u201c\u901f\u5ea6\u201d，CBSA\u5219\u60f3\u4e00\u6b21\u6027\u641e\u5b9a\u2014\u2014\u65e2\u8ba9\u6ce8\u610f\u529b\u673a\u5236\u7684\u5de5\u4f5c\u8fc7\u7a0b\u201c\u8bf4\u5f97\u901a\u201d，\u53c8\u8ba9\u5b83\u201c\u8dd1\u5f97\u98de\u5feb\u201d，\u751a\u81f3\u8fd8\u80fd\u628a\u4e0d\u540c\u7c7b\u578b\u7684\u6ce8\u610f\u529b（\u6bd4\u5982softmax\u6ce8\u610f\u529b\u3001\u7ebf\u6027\u6ce8\u610f\u529b）\u7edf\u4e00\u6210\u4e00\u4e2a\u6846\u67b6\u3002<\/p>\n
\u4e8c\u3001\u8bbe\u8ba1\u903b\u8f91：\u201c\u6293\u5c11\u6570\u4ee3\u8868，\u641e\u5b9a\u6240\u6709\u6570\u636e\u201d<\/p>\n
CBSA\u7684\u6838\u5fc3\u601d\u8def\u7279\u522b\u50cf\u201c\u8001\u5e08\u6279\u6539\u4f5c\u4e1a\u201d：\u4e0d\u7528\u9010\u9898\u770b\u6bcf\u4e2a\u5b66\u751f\u7684\u4f5c\u4e1a，\u5148\u6311\u51e0\u4e2a\u6709\u4ee3\u8868\u6027\u7684\u540c\u5b66（\u6bd4\u5982\u4e2d\u7b49\u6c34\u5e73\u3001\u80fd\u53cd\u6620\u5168\u73ed\u5171\u6027\u95ee\u9898\u7684），\u53ea\u6279\u6539\u8fd9\u51e0\u4e2a\u201c\u4ee3\u8868\u201d\u7684\u4f5c\u4e1a，\u518d\u628a\u6279\u6539\u601d\u8def\u544a\u8bc9\u5168\u73ed，\u5927\u5bb6\u5404\u81ea\u4fee\u6b63\u3002\u5177\u4f53\u5206\u4e24\u6b65：<\/p>\n
\u4e09\u3001\u5173\u952e\u4f18\u52bf：\u53c8\u5feb\u3001\u53c8\u61c2\u3001\u8fd8\u201c\u4e07\u80fd\u201d<\/p>\n
\u4e07\u80fd：\u80fd\u7edf\u4e00\u6240\u6709\u7ecf\u5178\u6ce8\u610f\u529b\u3002\u6700\u5389\u5bb3\u7684\u662f，CBSA\u80fd\u201c\u53d8\u201d\u6210\u5176\u4ed6\u6ce8\u610f\u529b\u673a\u5236\u2014\u2014\u53ea\u8981\u6362\u4e0d\u540c\u7684\u201c\u4ee3\u8868\u201d\u9009\u62e9\u65b9\u5f0f：\u9009\u201c\u6240\u6709\u6570\u636e\u5f53\u4ee3\u8868\u201d，CBSA\u5c31\u7b49\u540c\u4e8e\u4f20\u7edf\u7684softmax\u6ce8\u610f\u529b；\u9009\u201c\u6b63\u4ea4\u7684\u4ee3\u8868\u201d（\u6bd4\u5982\u6570\u636e\u7684\u4e3b\u65b9\u5411），CBSA\u5c31\u53d8\u6210\u7ebf\u6027\u6ce8\u610f\u529b；\u9009\u201c\u56fa\u5b9a\u65b9\u5411\u7684\u4ee3\u8868\u201d，CBSA\u5c31\u53d8\u6210\u901a\u9053\u6ce8\u610f\u529b（\u6bd4\u5982SE\u3001CBAM\u91cc\u7684\u901a\u9053\u4f18\u5316）\u3002\u76f8\u5f53\u4e8e\u7ed9\u4e0d\u540c\u7684\u6ce8\u610f\u529b\u673a\u5236\u627e\u4e86\u4e2a\u201c\u7edf\u4e00\u516c\u5f0f\u201d，\u80fd\u6e05\u6670\u770b\u5230\u5b83\u4eec\u7684\u672c\u8d28\u533a\u522b（\u53ea\u662f\u201c\u4ee3\u8868\u201d\u7684\u9009\u62e9\u65b9\u5f0f\u4e0d\u540c）\u3002<\/p>\n<\/li>\n
\u5728ultralytics\\\\nn\\\\modules\\\\\u76ee\u5f55\u4e0b，\u65b0\u5efa\u4e00\u4e2aC3k2_CBSA.py\u6587\u4ef6<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260125070847-6975c17fe2305.png\")
<\/p>\n
\u7136\u540e，\u628a\u4ee5\u4e0bC3k2_CBSA\u6a21\u5757\u6838\u5fc3\u4ee3\u7801\u62f7\u5165\u8fdb\u53bb：<\/p>\n
# https:\/\/github.com\/QishuaiWen\/CBSA
\nimport torch
\nfrom torch import nn
\nfrom einops import rearrange, repeat<\/p>\n
class TSSA(nn.Module):
\n # https:\/\/github.com\/RobinWu218\/ToST\/blob\/main\/tost_vision\/tost.py
\n def __init__(self, dim, heads, dim_head):
\n super().__init__()
\n num_heads = heads
\n self.heads = num_heads
\n self.attend = nn.Softmax(dim=1)
\n self.qkv = nn.Linear(dim, dim, bias=False)
\n self.temp = nn.Parameter(torch.ones(num_heads, 1))
\n self.to_out = nn.Linear(dim, dim)
\n self.scale = dim_head ** -0.5<\/p>\n
def forward(self, x, return_attn=False):
\n w = rearrange(self.qkv(x), 'b n (h d) -> b h n d', h=self.heads)
\n b, h, N, d = w.shape
\n if return_attn:
\n dots = w @ w.transpose(-1, -2)
\n return self.attend(dots)<\/p>\n
w_normed = torch.nn.functional.normalize(w, dim=-2)
\n w_sq = w_normed ** 2
\n # Pi from Eq. 10 in the paper
\n Pi = self.attend(torch.sum(w_sq, dim=-1) * self.temp) # b * h * n<\/p>\n
dots = torch.matmul((Pi \/ (Pi.sum(dim=-1, keepdim=True) + 1e-8)).unsqueeze(-2), w ** 2)
\n attn = 1. \/ (1 + dots)
\n out = -torch.mul(w.mul(Pi.unsqueeze(-1)), attn)
\n out = rearrange(out, 'b h n d -> b n (h d)')
\n return self.to_out(out)<\/p>\n
@torch.jit.ignore
\n def no_weight_decay(self):
\n return {'temp'}<\/p>\n
class MSSA(nn.Module):
\n # https:\/\/github.com\/Ma-Lab-Berkeley\/CRATE\/blob\/main\/model\/crate.py
\n def __init__(self, dim, heads=8, dim_head=64):
\n super().__init__()
\n inner_dim = dim_head * heads
\n project_out = not (heads == 1 and dim_head == dim)
\n self.heads = heads
\n self.scale = dim_head ** -0.5
\n self.attend = nn.Softmax(dim=-1)
\n self.qkv = nn.Linear(dim, inner_dim, bias=False)
\n self.to_out = nn.Linear(inner_dim, dim)<\/p>\n
def forward(self, x, return_attn=False):
\n w = rearrange(self.qkv(x), 'b n (h d) -> b h n d', h=self.heads)
\n dots = torch.matmul(w, w.transpose(-1, -2)) * self.scale
\n attn = self.attend(dots)<\/p>\n
if return_attn:
\n return attn
\n out = torch.matmul(attn, w)
\n out = rearrange(out, 'b h n d -> b n (h d)')
\n return self.to_out(out)<\/p>\n
class CBSA(nn.Module):
\n """
\n Cross-Block Self-Attention module.
\n Adapted to work with 2D feature maps (B, C, H, W) instead of sequences.
\n """
\n def __init__(self, dim, heads=8, dim_head=64):
\n super().__init__()
\n inner_dim = heads * dim_head
\n self.heads = heads
\n self.dim_head = dim_head
\n self.scale = dim_head ** -0.5
\n self.attend = nn.Softmax(dim=-1)
\n self.proj = nn.Linear(dim, inner_dim, bias=False)<\/p>\n
self.step_x = nn.Parameter(torch.randn(heads, 1, 1))
\n self.step_rep = nn.Parameter(torch.randn(heads, 1, 1))<\/p>\n
self.to_out = nn.Linear(inner_dim, dim)<\/p>\n
self.pool = nn.AdaptiveAvgPool2d(output_size=(8, 8))<\/p>\n
self.qkv = nn.Identity()<\/p>\n
def attention(self, query, key, value):
\n dots = (query @ key.transpose(-1, -2)) * self.scale
\n attn = self.attend(dots)
\n out = attn @ value
\n return out, attn<\/p>\n
def forward(self, x, return_attn=False):
\n """
\n Forward pass for CBSA.<\/p>\n
Args:
\n x: Input tensor of shape (B, C, H, W) – 2D feature map
\n return_attn: Whether to return attention weights<\/p>\n
Returns:
\n Output tensor of shape (B, C, H, W) – 2D feature map
\n """
\n b, c, h, w = x.shape
\n width = w # avoid name collision with projected tensor
\n n = h * w
\n inner_dim = self.heads * self.dim_head<\/p>\n
# Convert 2D feature map to sequence format: (B, C, H, W) -> (B, H*W, C)
\n x_seq = rearrange(x, 'b c h w -> b (h w) c')<\/p>\n
# Project to inner dimension
\n proj = self.proj(x_seq) # (B, n, inner_dim)
\n self.qkv(proj)<\/p>\n
# Create representation tokens using pooling
\n # Use full feature map to avoid shape mismatch; pool to fixed 8×8 tokens
\n if n > 1:
\n proj_2d = proj.reshape(b, h, width, inner_dim).permute(0, 3, 1, 2) # (B, inner_dim, h, w)
\n rep = self.pool(proj_2d) # (B, inner_dim, 8, 8)
\n rep = rep.reshape(b, inner_dim, -1).permute(0, 2, 1) # (B, 64, inner_dim)
\n else:
\n # Handle edge case when H*W = 1
\n rep = proj.reshape(b, 1, inner_dim).repeat(1, 64, 1) # (B, 64, inner_dim) – repeat single token<\/p>\n
# Reshape for attention
\n proj = proj.reshape(b, n, self.heads, self.dim_head).permute(0, 2, 1, 3) # (B, heads, n, dim_head)
\n rep = rep.reshape(b, 64, self.heads, self.dim_head).permute(0, 2, 1, 3) # (B, heads, 64, dim_head)<\/p>\n
# Cross attention: rep attends to w
\n rep_delta, attn = self.attention(rep, proj, proj)<\/p>\n
if return_attn:
\n return attn.transpose(-1, -2) @ attn<\/p>\n
# Update representation
\n rep = rep + self.step_rep * rep_delta<\/p>\n
# Self attention on representation
\n x_delta, _ = self.attention(rep, rep, rep)
\n x_delta = attn.transpose(-1, -2) @ x_delta
\n x_delta = self.step_x * x_delta<\/p>\n
# Reshape back to sequence: (B, heads, n, dim_head) -> (B, n, heads*dim_head)
\n x_delta = rearrange(x_delta, 'b h n k -> b n (h k)')
\n x_out = self.to_out(x_delta)<\/p>\n
# Convert back to 2D feature map: (B, H*W, C) -> (B, C, H, W)
\n x_out = rearrange(x_out, 'b (h w) c -> b c h w', h=h, w=w)<\/p>\n
return x_out<\/p>\n
def autopad(k, p=None, d=1): # kernel, padding, dilation
\n """Pad to 'same' shape outputs."""
\n if d > 1:
\n k = d * (k – 1) + 1 if isinstance(k, int) else [d * (x – 1) + 1 for x in k] # actual kernel-size
\n if p is None:
\n p = k \/\/ 2 if isinstance(k, int) else [x \/\/ 2 for x in k] # auto-pad
\n return p<\/p>\n
class Conv(nn.Module):
\n default_act = nn.SiLU() # default activation<\/p>\n
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
\n super().__init__()
\n self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p, d), groups=g, dilation=d, bias=False)
\n self.bn = nn.BatchNorm2d(c2)
\n self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()<\/p>\n
def forward(self, x):
\n return self.act(self.bn(self.conv(x)))<\/p>\n
def forward_fuse(self, x):
\n return self.act(self.conv(x))<\/p>\n
class Bottleneck(nn.Module):
\n """Standard bottleneck."""<\/p>\n
def __init__(
\n self, c1: int, c2: int, shortcut: bool = True, g: int = 1, k: tuple[int, int] = (3, 3), e: float = 0.5
\n ):
\n super().__init__()
\n c_ = int(c2 * e) # hidden channels
\n self.cv1 = Conv(c1, c_, k[0], 1)
\n self.cv2 = Conv(c_, c2, k[1], 1, g=g)
\n self.add = shortcut and c1 == c2<\/p>\n
def forward(self, x: torch.Tensor) -> torch.Tensor:
\n """Apply bottleneck with optional shortcut connection."""
\n return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))<\/p>\n
class Attention(nn.Module):<\/p>\n
def __init__(self, dim: int, num_heads: int = 8, attn_ratio: float = 0.5):
\n super().__init__()
\n self.num_heads = num_heads
\n self.head_dim = dim \/\/ num_heads
\n self.key_dim = int(self.head_dim * attn_ratio)
\n self.scale = self.key_dim**-0.5
\n nh_kd = self.key_dim * num_heads
\n h = dim + nh_kd * 2
\n self.qkv = Conv(dim, h, 1, act=False)
\n self.proj = Conv(dim, dim, 1, act=False)
\n self.pe = Conv(dim, dim, 3, 1, g=dim, act=False)<\/p>\n
def forward(self, x: torch.Tensor) -> torch.Tensor:
\n B, C, H, W = x.shape
\n N = H * W
\n qkv = self.qkv(x)
\n q, k, v = qkv.view(B, self.num_heads, self.key_dim * 2 + self.head_dim, N).split(
\n [self.key_dim, self.key_dim, self.head_dim], dim=2
\n )<\/p>\n
attn = (q.transpose(-2, -1) @ k) * self.scale
\n attn = attn.softmax(dim=-1)
\n x = (v @ attn.transpose(-2, -1)).view(B, C, H, W) + self.pe(v.reshape(B, C, H, W))
\n x = self.proj(x)
\n return x<\/p>\n
class C2f(nn.Module):<\/p>\n
def __init__(self, c1: int, c2: int, n: int = 1, shortcut: bool = False, g: int = 1, e: float = 0.5):<\/p>\n
super().__init__()
\n self.c = int(c2 * e) # hidden channels
\n self.cv1 = Conv(c1, 2 * self.c, 1, 1)
\n self.cv2 = Conv((2 + n) * self.c, c2, 1) # optional act=FReLU(c2)
\n self.m = nn.ModuleList(Bottleneck(self.c, self.c, shortcut, g, k=((3, 3), (3, 3)), e=1.0) for _ in range(n))<\/p>\n
def forward(self, x: torch.Tensor) -> torch.Tensor:
\n """Forward pass through C2f layer."""
\n y = list(self.cv1(x).chunk(2, 1))
\n y.extend(m(y[-1]) for m in self.m)
\n return self.cv2(torch.cat(y, 1))<\/p>\n
def forward_split(self, x: torch.Tensor) -> torch.Tensor:
\n """Forward pass using split() instead of chunk()."""
\n y = self.cv1(x).split((self.c, self.c), 1)
\n y = [y[0], y[1]]
\n y.extend(m(y[-1]) for m in self.m)
\n return self.cv2(torch.cat(y, 1))<\/p>\n
class C3(nn.Module):
\n """CSP Bottleneck with 3 convolutions."""<\/p>\n
def __init__(self, c1: int, c2: int, n: int = 1, shortcut: bool = True, g: int = 1, e: float = 0.5):<\/p>\n
super().__init__()
\n c_ = int(c2 * e) # hidden channels
\n self.cv1 = Conv(c1, c_, 1, 1)
\n self.cv2 = Conv(c1, c_, 1, 1)
\n self.cv3 = Conv(2 * c_, c2, 1) # optional act=FReLU(c2)
\n self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, k=((1, 1), (3, 3)), e=1.0) for _ in range(n)))<\/p>\n
def forward(self, x: torch.Tensor) -> torch.Tensor:
\n """Forward pass through the CSP bottleneck with 3 convolutions."""
\n return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))<\/p>\n
class PSABlock(nn.Module):<\/p>\n
def __init__(self, c: int, attn_ratio: float = 0.5, num_heads: int = 4, shortcut: bool = True) -> None:
\n super().__init__()<\/p>\n
self.attn = Attention(c, attn_ratio=attn_ratio, num_heads=num_heads)
\n self.ffn = nn.Sequential(Conv(c, c * 2, 1), Conv(c * 2, c, 1, act=False))
\n self.add = shortcut<\/p>\n
def forward(self, x: torch.Tensor) -> torch.Tensor:
\n x = x + self.attn(x) if self.add else self.attn(x)
\n x = x + self.ffn(x) if self.add else self.ffn(x)
\n return x<\/p>\n
class C3k2(C2f):
\n def __init__(
\n self,
\n c1: int,
\n c2: int,
\n n: int = 1,
\n c3k: bool = False,
\n e: float = 0.5,
\n attn: bool = False,
\n g: int = 1,
\n shortcut: bool = True,
\n ):
\n super().__init__(c1, c2, n, shortcut, g, e)
\n self.m = nn.ModuleList(
\n nn.Sequential(
\n Bottleneck(self.c, self.c, shortcut, g),
\n PSABlock(self.c, attn_ratio=0.5, num_heads=max(self.c \/\/ 64, 1)),
\n )
\n if attn
\n else C3k(self.c, self.c, 2, shortcut, g)
\n if c3k
\n else Bottleneck(self.c, self.c, shortcut, g)
\n for _ in range(n)
\n )<\/p>\n
class C3k(C3):<\/p>\n
def __init__(self, c1: int, c2: int, n: int = 1, shortcut: bool = True, g: int = 1, e: float = 0.5, k: int = 3):
\n super().__init__(c1, c2, n, shortcut, g, e)
\n c_ = int(c2 * e) # hidden channels
\n # self.m = nn.Sequential(*(RepBottleneck(c_, c_, shortcut, g, k=(k, k), e=1.0) for _ in range(n)))
\n self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, k=(k, k), e=1.0) for _ in range(n)))<\/p>\n
class C3k_CBSA(C3k):<\/p>\n
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, k=3, heads=8, dim_head=64):
\n super().__init__(c1, c2, n, shortcut, g, e, k)
\n c_ = int(c2 * e) # hidden channels
\n self.m = nn.Sequential(*(CBSA(c_, heads=heads, dim_head=dim_head) for _ in range(n)))<\/p>\n
class C3k2_CBSA(C3k2):<\/p>\n
def __init__(
\n self,
\n c1: int,
\n c2: int,
\n n: int = 1,
\n c3k: bool = False,
\n e: float = 0.5,
\n attn: bool = False,
\n g: int = 1,
\n shortcut: bool = True,
\n heads=8,
\n dim_head=64
\n ):
\n super().__init__(c1, c2, n, shortcut, g, e)<\/p>\n
if attn:<\/p>\n
print(attn , 'attn=True \u6ce8\u610f\u529b\u6a21\u5f0f – \u4f7f\u7528CBSA + Bottleneck')
\n # \u4f18\u5148\u7ea71：attn=True \u6ce8\u610f\u529b\u6a21\u5f0f – \u4f7f\u7528\u4f7f\u7528CBSA + Bottleneck，\u8fd9\u91cc\u53ef\u4ee5\u81ea\u884c\u7ec4\u5408\u8bd5\u8bd5\u770b
\n self.m = nn.ModuleList(
\n nn.Sequential(
\n Bottleneck(self.c, self.c, shortcut, g),
\n CBSA(self.c, heads=heads, dim_head=dim_head),
\n # C3k_CBSA(self.c, self.c, shortcut, g, e=1.0),
\n # PSABlock(self.c, attn_ratio=0.5, num_heads=max(self.c \/\/ 64, 1)),
\n )
\n for _ in range(n)
\n )
\n elif c3k:
\n print(c3k,'c3k=True C3k\u6a21\u5f0f – \u4f7f\u7528C3k_CBSA\u4ee3\u66ffC3k')
\n # \u4f18\u5148\u7ea72：c3k=True C3k\u6a21\u5f0f – \u4f7f\u7528C3k_CBSA\u4ee3\u66ffC3k
\n self.m = nn.ModuleList(
\n C3k_CBSA(self.c, self.c, 2, shortcut, g)
\n for _ in range(n)
\n )
\n else:
\n print(c3k,'c3k=False \u57fa\u7840\u6a21\u5f0f – \u4f7f\u7528C3k_CBSA\u4ee3\u66ffBottleneck')
\n # \u4f18\u5148\u7ea73：c3k=False \u57fa\u7840\u6a21\u5f0f – \u4f7f\u7528C3k_CBSA\u4ee3\u66ffBottleneck
\n self.m = nn.ModuleList(
\n C3k_CBSA(self.c, self.c, shortcut, g, e=1.0)
\n for _ in range(n)
\n )<\/p>\n
\u9996\u5148，\u5728ultralytics\/nn\/tasks.py\u4ee3\u7801\u6700\u524d\u7aef\u5bfc\u5165C3k2_CBSA\u6a21\u5757，\u4ee3\u7801\u5982\u4e0b：<\/p>\n
from ultralytics.nn.modules.C3k2_CBSA import C3k2_CBSA <\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260125070848-6975c1800506d.png\")
<\/p>\n
\u7136\u540e，tasks.py\u6587\u4ef6\u4e2d\u627e\u5230parse_model\u51fd\u6570（ctrl+f \u53ef\u4ee5\u76f4\u63a5\u641c\u7d22parse_model\u4f4d\u7f6e）\u5bfc\u5165C3k2_CBSA\u6a21\u5757：1）base_modules\u90e8\u5206；2）repeat_modules\u90e8\u5206\u3002<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260125070848-6975c1801fa34.png\")
<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260125070848-6975c1804342c.png\")
<\/p>\n
\u4ee5Ultralytics \u516c\u53f8\u4e8e 2025 \u5e74 9 \u6708\u53d1\u5e03\u7684\u6700\u65b0\u4e00\u4ee3\u76ee\u6807\u68c0\u6d4b\u6a21\u578bYOLO26\u4e3a\u4f8b（\u4ee3\u7801\u5b98\u65b9\u53ef\u4e0b\u8f7d，\u4e0d\u7136\u4f1a\u62a5\u9519），\u521b\u5efayolo26-C3k2_CBSA.yaml\u914d\u7f6e\u6587\u4ef6，\u4ee3\u7801\u5982\u4e0b：<\/p>\n
# Ultralytics 🚀 AGPL-3.0 License – https:\/\/ultralytics.com\/license<\/p>\n
# Ultralytics YOLO26 object detection model with P3\/8 – P5\/32 outputs
\n# Model docs: https:\/\/docs.ultralytics.com\/models\/yolo26
\n# Task docs: https:\/\/docs.ultralytics.com\/tasks\/detect<\/p>\n
# Parameters
\nnc: 80 # number of classes
\nend2end: True # whether to use end-to-end mode
\nreg_max: 1 # DFL bins
\nscales: # model compound scaling constants, i.e. 'model=yolo26n.yaml' will call yolo26.yaml with scale 'n'
\n # [depth, width, max_channels]
\n n: [0.50, 0.25, 1024] # summary: 260 layers, 2,572,280 parameters, 2,572,280 gradients, 6.1 GFLOPs
\n s: [0.50, 0.50, 1024] # summary: 260 layers, 10,009,784 parameters, 10,009,784 gradients, 22.8 GFLOPs
\n m: [0.50, 1.00, 512] # summary: 280 layers, 21,896,248 parameters, 21,896,248 gradients, 75.4 GFLOPs
\n l: [1.00, 1.00, 512] # summary: 392 layers, 26,299,704 parameters, 26,299,704 gradients, 93.8 GFLOPs
\n x: [1.00, 1.50, 512] # summary: 392 layers, 58,993,368 parameters, 58,993,368 gradients, 209.5 GFLOPs<\/p>\n
# YOLO26n backbone
\nbackbone:
\n # [from, repeats, module, args]
\n – [-1, 1, Conv, [64, 3, 2]] # 0-P1\/2
\n – [-1, 1, Conv, [128, 3, 2]] # 1-P2\/4
\n – [-1, 2, C3k2_CBSA, [256, False, 0.25]]
\n – [-1, 1, Conv, [256, 3, 2]] # 3-P3\/8
\n – [-1, 2, C3k2_CBSA, [512, False, 0.25]]
\n – [-1, 1, Conv, [512, 3, 2]] # 5-P4\/16
\n – [-1, 2, C3k2_CBSA, [512, True]]
\n – [-1, 1, Conv, [1024, 3, 2]] # 7-P5\/32
\n – [-1, 2, C3k2_CBSA, [1024, True]]
\n – [-1, 1, SPPF, [1024, 5, 3, True]] # 9
\n – [-1, 2, C2PSA, [1024]] # 10<\/p>\n
# YOLO26n head
\nhead:
\n – [-1, 1, nn.Upsample, [None, 2, "nearest"]]
\n – [[-1, 6], 1, Concat, [1]] # cat backbone P4
\n – [-1, 2, C3k2_CBSA, [512, True]] # 13<\/p>\n
– [-1, 1, nn.Upsample, [None, 2, "nearest"]]
\n – [[-1, 4], 1, Concat, [1]] # cat backbone P3
\n – [-1, 2, C3k2_CBSA, [256, True]] # 16 (P3\/8-small)<\/p>\n
– [-1, 1, Conv, [256, 3, 2]]
\n – [[-1, 13], 1, Concat, [1]] # cat head P4
\n – [-1, 2, C3k2_CBSA, [512, True]] # 19 (P4\/16-medium)<\/p>\n
– [-1, 1, Conv, [512, 3, 2]]
\n – [[-1, 10], 1, Concat, [1]] # cat head P5
\n – [-1, 1, C3k2_CBSA, [1024, True, 0.5, True]] # 22 (P5\/32-large)<\/p>\n
– [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5)<\/p>\n
import warnings
\nwarnings.filterwarnings('ignore')
\nfrom ultralytics import YOLO<\/p>\n
if __name__ == '__main__':
\n model = YOLO('ultralytics\/cfg\/models\/26\/yolo26-C3k2_CBSA.yaml') # \u5bfc\u5165yaml\u914d\u7f6e\u6587\u4ef6
\n # model.load('yolo11n.pt') # loading pretrain weights
\n model.train(data='dataset\/data.yaml',
\n cache=False,
\n imgsz=640,
\n epochs=300,
\n batch=32,
\n close_mosaic=0,
\n workers=4, # Windows\u4e0b\u51fa\u73b0\u83ab\u540d\u5176\u5999\u5361\u4e3b\u7684\u60c5\u51b5\u53ef\u4ee5\u5c1d\u8bd5\u628aworkers\u8bbe\u7f6e\u4e3a0
\n # device='0',
\n optimizer='SGD', # using SGD
\n # patience=0, # set 0 to close earlystop.
\n # resume=True, # \u65ad\u70b9\u7eed\u8bad,YOLO\u521d\u59cb\u5316\u65f6\u9009\u62e9last.pt
\n # amp=False, # close amp
\n # fraction=0.2,
\n project='runs\/train',
\n name='yolo26-C3k2_CBSA',
\n ) <\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260125070848-6975c18075c04.png\")
<\/p><\/p>\n","protected":false},"excerpt":{"rendered":"
0. \u524d\u8a00
\n\u672c\u6587\u4ecb\u7ecdCBSA\u6536\u7f29 – \u5e7f\u64ad\u81ea\u6ce8\u610f\u529b\u673a\u5236（Contract-and-Broadcast Self-Attention），\u5e76\u5c06\u5176\u96c6\u6210\u5230ultralytics\u6700\u65b0\u53d1\u5e03\u7684YOLO26\u76ee\u6807\u68c0\u6d4b\u7b97\u6cd5\u4e2d，\u6784\u5efaC3k2-LWGA\u521b\u65b0\u6a21\u5757\u3002\u4f20\u7edf\u6ce8\u610f\u529b\u673a\u5236\u5b58\u5728\u9ed1\u76d2\u96be\u7406\u89e3\u3001\u8ba1\u7b97\u590d\u6742\u5ea6\u9ad8\u7684\u95ee\u9898，CBSA\u901a\u8fc7\u7b97\u6cd5\u5c55\u5f00\u63a8\u5bfc\u51fa\u672c\u8d28\u4e0a\u53ef\u89e3\u91ca\u4e14\u9ad8\u6548\u7684\u6ce8\u610f\u529b\u673a\u5236\u3002\u5b83\u5148\u4ece\u8f93\u5165\u6570\u636e\u4e2d\u9009\u51fa\u5c11\u91cf\u4ee3\u8868\u6027tokens，\u63a5\u7740<\/p>\n","protected":false},"author":2,"featured_media":65658,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[6942,6396,152,156,50,1665,523],"topic":[],"class_list":["post-65665","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-c3k2","tag-yolo26","tag-pytorch","tag-yolo","tag-50","tag-1665","tag-523"],"yoast_head":"\n