logoChat01.ai

YOLOv11没有改进前P,R,MAP50,MAP50-95的结果分别如下：0.923，0.828，...

Создано: 26 декабря 2025 г.

Отвечено с использованием GPT-5.2 Pro от Chat01

Вопрос

YOLOv11没有改进前P,R,MAP50,MAP50-95的结果分别如下：0.923，0.828，0.89，0.719，我现在的代码如下：class SPD(nn.Module):
def init(self, scale=2):
super().init()
assert scale >= 2 and int(scale) == scale
self.scale = int(scale)

text
def forward(self, x):
    b, c, h, w = x.shape
    s = self.scale

    # 若你训练时可能遇到非整除尺寸（不常见，但做个保护），这里右/下补齐到能整除
    pad_h = (s - h % s) % s
    pad_w = (s - w % s) % s
    if pad_h or pad_w:
        x = F.pad(x, (0, pad_w, 0, pad_h))  # (left,right,top,bottom)
        b, c, h, w = x.shape

    # [B,C,H,W] -> [B, C*s*s, H/s, W/s]
    x = x.view(b, c, h // s, s, w // s, s)
    x = x.permute(0, 1, 3, 5, 2, 4).contiguous()
    return x.view(b, c * s * s, h // s, w // s)

class EMA(nn.Module):
def init(self, channels, c2=None, factor=32):
super(EMA, self).init()
self.groups = factor
assert channels % factor == 0
assert channels // self.groups > 0
self.softmax = nn.Softmax(-1)
self.agp = nn.AdaptiveAvgPool2d((1, 1))
self.pool_h = nn.AdaptiveAvgPool2d((None, 1))
self.pool_w = nn.AdaptiveAvgPool2d((1, None))
self.gn = nn.GroupNorm(channels // self.groups, channels // self.groups)
self.conv1x1 = nn.Conv2d(channels // self.groups, channels // self.groups, kernel_size=1, stride=1, padding=0)
self.conv3x3 = nn.Conv2d(channels // self.groups, channels // self.groups, kernel_size=3, stride=1, padding=1)

text
def forward(self, x):
    b, c, h, w = x.size()
    group_x = x.reshape(b * self.groups, -1, h, w)  # b*g,c//g,h,w
    x_h = self.pool_h(group_x)
    x_w = self.pool_w(group_x).permute(0, 1, 3, 2)
    hw = self.conv1x1(torch.cat([x_h, x_w], dim=2))
    x_h, x_w = torch.split(hw, [h, w], dim=2)
    x1 = self.gn(group_x * x_h.sigmoid() * x_w.permute(0, 1, 3, 2).sigmoid())
    x2 = self.conv3x3(group_x)
    x11 = self.softmax(self.agp(x1).reshape(b * self.groups, -1, 1).permute(0, 2, 1))
    x12 = x2.reshape(b * self.groups, c // self.groups, -1)  # b*g, c//g, hw
    x21 = self.softmax(self.agp(x2).reshape(b * self.groups, -1, 1).permute(0, 2, 1))
    x22 = x1.reshape(b * self.groups, c // self.groups, -1)  # b*g, c//g, hw
    weights = (torch.matmul(x11, x12) + torch.matmul(x21, x22)).reshape(b * self.groups, 1, h, w)
    return (group_x * weights.sigmoid()).reshape(b, c, h, w)

class SPDConvLite(nn.Module):
"""
SPDConvLite + EMA
args 兼容你 YAML: [c2, k, s, ratio, ema_factor]
"""
def init(self, c1, c2, k=3, s=2, ratio=0.5, ema_factor=16, act=True):
super().init()
self.spd = SPD(scale=s) # 你已有的 SPD（或 space_to_depth/PixelUnshuffle）
cin = c1 * s * s
hidden = max(16, int(c2 * ratio))

text
    self.pw1 = Conv(cin, hidden, 1, 1, act=act)
    self.dw  = Conv(hidden, hidden, k, 1, g=hidden, act=act)
    self.pw2 = Conv(hidden, c2, 1, 1, act=act)

    # EMA 注意力：作用在输出通道 c2 上
    # 如果你的 EMA 定义是 EMA(channels, c2=None, factor=32)，这里传 channels=c2 即可
    self.attn = EMA(c2, factor=ema_factor)
    self.alpha=nn.Parameter(torch.zeros((1)))

def forward(self, x):
    x = self.spd(x)
    x = self.pw1(x)
    x = self.dw(x)
    y = self.pw2(x)
    y_attn = self.attn(y)
    return y+self.alpha*(y_attn-y)

class MSCA(nn.Module):
def init(self, c1, c2, reduction=8):
super().init()
assert c1 == c2, "MSCAPlus 目前假设输入输出通道相同"
dim = c2

text
    # 1. 多尺度空间分支
    # 分支0: 局部 5x5 深度卷积
    self.branch0 = nn.Conv2d(dim, dim, 5, padding=2, groups=dim)

    # 分支1: 条带卷积 (1x7 -> 7x1)
    self.b1_1 = nn.Conv2d(dim, dim, (1, 7), padding=(0, 3), groups=dim)
    self.b1_2 = nn.Conv2d(dim, dim, (7, 1), padding=(3, 0), groups=dim)

    # 分支2: 3x3 空洞卷积, dilation=2 (中等感受野)
    self.branch2 = nn.Conv2d(dim, dim, 3, padding=2, dilation=2, groups=dim)

    # 分支3: 3x3 空洞卷积, dilation=3 (大感受野, 替代 21x21)
    self.branch3 = nn.Conv2d(dim, dim, 3, padding=3, dilation=3, groups=dim)

    # 2. 组卷积融合多尺度特征
    # 4 个分支 concat → 通道数 4*dim，groups=4 保证参数 ≈ 原来的 1x1 conv
    self.fuse = nn.Conv2d(4 * dim, dim, kernel_size=1, groups=4, bias=False)
    self.edge=nn.Conv2d(dim, dim, 3, padding=1,  groups=dim, bias=False)

    # 3. 通道注意力 (SE-like)
    hidden = max(dim // reduction, 4)
    self.avg_pool = nn.AdaptiveAvgPool2d(1)
    self.fc1 = nn.Conv2d(dim, hidden, kernel_size=1, bias=True)
    self.fc2 = nn.Conv2d(hidden, dim, kernel_size=1, bias=True)

    # 激活与门控
    self.act = nn.SiLU()
    self.sigmoid = nn.Sigmoid()

    self.gamma_add = nn.Parameter(torch.zeros(1))
    self.gamma_gate = nn.Parameter(torch.zeros(1))

def forward(self, x):
    identity = x

    # 多尺度空间分支
    b0 = self.branch0(x)

    b1 = self.b1_1(x)
    b1 = self.b1_2(b1)

    b2 = self.branch2(x)
    b3 = self.branch3(x)

    # concat + 融合
    ms = torch.cat([b0, b1, b2, b3], dim=1)   # [B, 4C, H, W]
    ms = self.fuse(ms)                        # [B, C, H, W]
    edge=self.edge(x)
    ms=ms+edge

    # 通道注意力
    ca = self.avg_pool(ms)                    # [B, C, 1, 1]
    ca = self.fc2(self.act(self.fc1(ca)))     # [B, C, 1, 1]
    ca = self.sigmoid(ca)

    attn = ms * ca                            # 空间 + 通道联合
    gate = self.sigmoid(attn)                 # [0,1]

    # 残差门控：避免特征被破坏性缩放
    out = identity + self.gamma_add*attn + self.gamma_gate*identity*gate
    return out

class EMA(nn.Module):
def init(self, channels, c2=None, factor=32):
super(EMA, self).init()
self.groups = factor
assert channels // self.groups > 0
self.softmax = nn.Softmax(-1)
self.agp = nn.AdaptiveAvgPool2d((1, 1))
self.pool_h = nn.AdaptiveAvgPool2d((None, 1))
self.pool_w = nn.AdaptiveAvgPool2d((1, None))
self.gn = nn.GroupNorm(channels // self.groups, channels // self.groups)
self.conv1x1 = nn.Conv2d(channels // self.groups, channels // self.groups, kernel_size=1, stride=1, padding=0)
self.conv3x3 = nn.Conv2d(channels // self.groups, channels // self.groups, kernel_size=3, stride=1, padding=1)

text
def forward(self, x):
    b, c, h, w = x.size()
    group_x = x.reshape(b * self.groups, -1, h, w)  # b*g,c//g,h,w
    x_h = self.pool_h(group_x)
    x_w = self.pool_w(group_x).permute(0, 1, 3, 2)
    hw = self.conv1x1(torch.cat([x_h, x_w], dim=2))
    x_h, x_w = torch.split(hw, [h, w], dim=2)
    x1 = self.gn(group_x * x_h.sigmoid() * x_w.permute(0, 1, 3, 2).sigmoid())
    x2 = self.conv3x3(group_x)
    x11 = self.softmax(self.agp(x1).reshape(b * self.groups, -1, 1).permute(0, 2, 1))
    x12 = x2.reshape(b * self.groups, c // self.groups, -1)  # b*g, c//g, hw
    x21 = self.softmax(self.agp(x2).reshape(b * self.groups, -1, 1).permute(0, 2, 1))
    x22 = x1.reshape(b * self.groups, c // self.groups, -1)  # b*g, c//g, hw
    weights = (torch.matmul(x11, x12) + torch.matmul(x21, x22)).reshape(b * self.groups, 1, h, w)
    return (group_x * weights.sigmoid()).reshape(b, c, h, w)我的配置文件如下：当我只增加MSCA模块时

nc: 5 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'

[depth, width, max_channels]

n: [0.50, 0.25, 1024] # summary: 181 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
s: [0.50, 0.50, 1024] # summary: 181 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
m: [0.50, 1.00, 512] # summary: 231 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
l: [1.00, 1.00, 512] # summary: 357 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
x: [1.00, 1.50, 512] # summary: 357 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs

YOLO11n backbone

backbone:

[from, repeats, module, args]

  • [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  • [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
  • [-1, 2, C3k2, [256, False, 0.25]]
  • [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  • [-1, 2, C3k2, [512, False, 0.25]]
  • [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  • [-1, 2, C3k2, [512, True]]
  • [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  • [-1, 2, C3k2, [1024, True]]
  • [-1, 1, SPPF, [1024, 5]] # 9
  • [-1, 2, C2PSA, [1024]] # 10
  • [-1, 1, MSCA, [1024]]

YOLO11n head

head:

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]]

  • [[-1, 6], 1, Concat, [1]] # cat backbone P4

  • [-1, 2, C3k2, [512, False]] # 13

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]]

  • [[-1, 4], 1, Concat, [1]] # cat backbone P3

  • [-1, 2, C3k2, [256, False]] # 16 (P3/8-small)

  • [-1, 1, Conv, [256, 3, 2]]

  • [[-1, 14], 1, Concat, [1]] # cat head P4

  • [-1, 2, C3k2, [512, False]] # 19 (P4/16-medium)

  • [-1, 1, Conv, [512, 3, 2]]

  • [[-1, 11], 1, Concat, [1]] # cat head P5

  • [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large)

  • [[17, 20, 23], 1, Detect, [nc]] # Detect(P3, P4, P5)
    代码结果如下：0.935，0.866，0.914，0.725，我现在用SPD的配置文件：
    nc: 5 # number of classes
    scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'

[depth, width, max_channels]

n: [0.50, 0.25, 1024] # summary: 181 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
s: [0.50, 0.50, 1024] # summary: 181 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
m: [0.50, 1.00, 512] # summary: 231 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
l: [1.00, 1.00, 512] # summary: 357 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
x: [1.00, 1.50, 512] # summary: 357 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs

YOLO11n backbone

backbone:

[from, repeats, module, args]

  • [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  • [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
  • [-1, 2, C3k2, [256, False, 0.25]]
  • [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  • [-1, 2, C3k2, [512, False, 0.25]]
  • [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  • [-1, 2, C3k2, [512, True]]
  • [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  • [-1, 2, C3k2, [1024, True]]
  • [-1, 1, SPPF, [1024, 5]] # 9
  • [-1, 2, C2PSA, [1024]] # 10

YOLO11n head

head:

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]]

  • [[-1, 6], 1, Concat, [1]] # cat backbone P4

  • [-1, 2, C3k2, [512, False]] # 13

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]]

  • [[-1, 4], 1, Concat, [1]] # cat backbone P3

  • [-1, 2, C3k2, [256, False]] # 16 (P3/8-small)

  • [-1, 1, SPDConvLite, [256, 3, 2]]

  • [[-1, 13], 1, Concat, [1]]

  • [-1, 2, C3k2, [512, False]] # 19 (P4/16-medium)

  • [-1, 1, SPDConvLite, [512, 3, 2]]

  • [[-1, 10], 1, Concat, [1]]

  • [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large)

  • [[16, 19, 22], 1, Detect, [nc]]
    结果如下：0.959，0.865，0.91，0.725
    但是当我两者同时添加后配置文件如下：# Parameters
    nc: 5 # number of classes
    scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'

[depth, width, max_channels]

n: [0.50, 0.25, 1024] # summary: 181 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
s: [0.50, 0.50, 1024] # summary: 181 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
m: [0.50, 1.00, 512] # summary: 231 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
l: [1.00, 1.00, 512] # summary: 357 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
x: [1.00, 1.50, 512] # summary: 357 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs

YOLO11n backbone

backbone:

[from, repeats, module, args]

  • [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  • [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
  • [-1, 2, C3k2, [256, False, 0.25]]
  • [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  • [-1, 2, C3k2, [512, False, 0.25]]
  • [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  • [-1, 2, C3k2, [512, True]]
  • [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  • [-1, 2, C3k2, [1024, True]]
  • [-1, 1, SPPF, [1024, 5]] # 9
  • [-1, 2, C2PSA, [1024]] # 10
  • [-1, 1, MSCA, [1024]]

YOLO11n head

head:

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]]

  • [[-1, 6], 1, Concat, [1]] # cat backbone P4

  • [-1, 2, C3k2, [512, False]] # 13

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]]

  • [[-1, 4], 1, Concat, [1]] # cat backbone P3

  • [-1, 2, C3k2, [256, False]] # 16 (P3/8-small)

  • [-1, 1, SPDConvLite, [256, 3, 2]]

  • [[-1, 14], 1, Concat, [1]] # cat head P4

  • [-1, 2, C3k2, [512, False]] # 19 (P4/16-medium)

  • [-1, 1, SPDConvLite, [512, 3, 2]]

  • [[-1, 11], 1, Concat, [1]] # cat head P5

  • [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large)

  • [[17, 20, 23], 1, Detect, [nc]] # Detect(P3, P4, P5)
    代码结果如下：0.909，0.864，0.896，0.719，我现在还想把我的检测头的改进点融入进去，我现在的检测头的改进配置文件如下：# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license

Ultralytics YOLO11 object detection model with P3/8 - P5/32 outputs

Model docs: https://docs.ultralytics.com/models/yolo11

Task docs: https://docs.ultralytics.com/tasks/detect

Parameters

nc: 5 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'

[depth, width, max_channels]

n: [0.50, 0.25, 1024] # summary: 181 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
s: [0.50, 0.50, 1024] # summary: 181 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
m: [0.50, 1.00, 512] # summary: 231 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
l: [1.00, 1.00, 512] # summary: 357 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
x: [1.00, 1.50, 512] # summary: 357 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs

YOLO11n backbone

backbone:

[from, repeats, module, args]

  • [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  • [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
  • [-1, 2, C3k2, [256, False, 0.25]]
  • [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  • [-1, 2, C3k2, [512, False, 0.25]]
  • [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  • [-1, 2, C3k2, [512, True]]
  • [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  • [-1, 2, C3k2, [1024, True]]
  • [-1, 1, SPPF, [1024, 5]] # 9
  • [-1, 2, C2PSA, [1024]] # 10

#head:

# -------- top-down FPN --------

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 11: P5 -> P4
  • [[-1, 6], 1, Concat, [1]] # 12: cat backbone P4 (layer 6)
  • [-1, 2, C3k2, [512, False]] # 13: P4_td

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 14: P4 -> P3
  • [[-1, 4], 1, Concat, [1]] # 15: cat backbone P3 (layer 4)
  • [-1, 2, C3k2, [256, False]] # 16: P3_td

# -------- add P2 detection layer (paper) --------

  • [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 17: P3 -> P2
  • [[-1, 2], 1, Concat, [1]] # 18: cat backbone P2 (layer 2, stride=4)
  • [-1, 2, C3k2, [128, False]] # 19: P2_out (P2/4)

# -------- bottom-up PAN --------

  • [-1, 1, Conv, [256, 3, 2]] # 20: P2 -> P3
  • [[-1, 16], 1, Concat, [1]] # 21: cat P3_td
  • [-1, 2, C3k2, [256, False]] # 22: P3_out (P3/8)

  • [-1, 1, Conv, [512, 3, 2]] # 23: P3 -> P4
  • [[-1, 13], 1, Concat, [1]] # 24: cat P4_td
  • [-1, 2, C3k2, [512, False]] # 25: P4_out (P4/16)

  • [-1, 1, Conv, [1024, 3, 2]] # 26: P4 -> P5
  • [[-1, 10], 1, Concat, [1]] # 27: cat backbone P5 (layer 10)
  • [-1, 2, C3k2, [1024, True]] # 28: P5_out (P5/32)

# -------- Detect (4 heads) --------

  • [[19, 22, 25, 28], 1, DetectLSDECD, [nc, 16, 16]] # Detect(P2,P3,P4,P5)
    只增加检测头结果是0.951，0.888，0.927，0.781，现在单独使用每个模块都没问题，我同时使用MSCA、SPDConvLite更改检测头效果就变差，假如我必须选一个作为不改的，你认为我应该保留那个呢，去修正哪个呢，一定要分析这其中的原因，还是说应该更改其他模块呢，去改C3K2还是其他呢，为什么1+1<2了呢，模块相加还冲突导致检测效果差呢，我现在使用检测头+MSCA的结果如下：0.949，0.874，0.918，0.763，你现在告诉我根据目前的结果我如何改算法，具体给出每次的操作，尝试了SPDConvLite+检测头但是效果也不好，结果如下：0.916，0.838，0.873，0.726，下一步怎么办呢，怎么解决呢

Поделиться этим Q&A