![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260117075002-696b3f2ad59d4.png\")
<\/p>\niRMB（Inverted Residual Mobile Block）\u7684\u6846\u67b6\u539f\u7406，\u662f\u4e00\u79cd\u7ed3\u5408\u8f7b\u91cf\u7ea7CNN\u548c\u6ce8\u610f\u529b\u673a\u5236\u7684\u65b9\u6cd5，\u7528\u4e8e\u6539\u8fdb\u79fb\u52a8\u8bbe\u5907\u4e0a\u7684\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u3002IRMB\u901a\u8fc7\u5012\u7f6e\u6b8b\u5dee\u5757\u548c\u5143\u79fb\u52a8\u5757\u5b9e\u73b0\u9ad8\u6548\u4fe1\u606f\u5904\u7406，\u540c\u65f6\u4fdd\u6301\u6a21\u578b\u8f7b\u91cf\u5316\u3002\u672c\u6587\u4e2d\u63d0\u51fa\u4e00\u4e2a\u65b0\u7684\u4e3b\u5e72\u7f51\u7edcEMO，\u4e3b\u8981\u601d\u60f3\u662f\u5c06\u8f7b\u91cf\u7ea7\u7684CNN\u67b6\u6784\u4e0e\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u6a21\u578b\u7ed3\u6784\u76f8\u7ed3\u5408\u3002<\/p>\n
\u76ee\u5f55<\/p>\n
1. \u7b80\u4ecb<\/p>\n
2. iRMB\u4e3b\u8981\u601d\u60f3<\/p>\n
3. iRMB\u7ed3\u6784<\/p>\n
<\/p>\n
\u539f\u59cb\u8bba\u6587：https:\/\/arxiv.org\/pdf\/2301.01146<\/p>\n
\u4ee3\u7801\u5730\u5740：https:\/\/github.com\/zhangzjn\/EMO<\/p>\n
\u5012\u7f6e\u6b8b\u5dee\u5757（IRB）\u4f5c\u4e3a\u8f7b\u91cf\u7ea7CNN\u7684\u57fa\u7840\u8bbe\u65bd，\u5728\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u7814\u7a76\u4e2d\u5c1a\u672a\u6709\u5bf9\u5e94\u7684\u90e8\u5206\u3002\u672c\u6587\u4ece\u7edf\u4e00\u7684\u89c6\u89d2\u91cd\u65b0\u8003\u8651\u4e86\u57fa\u4e8e\u9ad8\u6548IRB\u548cTransformer\u6709\u6548\u7ec4\u4ef6\u7684\u8f7b\u91cf\u7ea7\u57fa\u7840\u8bbe\u65bd，\u5c06\u57fa\u4e8eCNN\u7684IRB\u6269\u5c55\u5230\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u6a21\u578b，\u5e76\u62bd\u8c61\u51fa\u4e00\u4e2a\u6b8b\u5dee\u7684\u5143\u79fb\u52a8\u5757（MMB）\u7528\u4e8e\u8f7b\u91cf\u7ea7\u6a21\u578b\u8bbe\u8ba1\u3002\u9075\u5faa\u7b80\u5355\u4f46\u6709\u6548\u7684\u8bbe\u8ba1\u51c6\u5219，\u672c\u6587\u63a8\u5bfc\u51fa\u4e86\u73b0\u4ee3\u5316\u7684\u5012\u7f6e\u6b8b\u5dee\u79fb\u52a8\u5757（IRMB），\u5e76\u4ee5\u6b64\u6784\u5efa\u4e86\u7c7b\u4f3cResNet\u7684\u9ad8\u6548\u6a21\u578b（EMO）\u7528\u4e8e\u4e0b\u6e38\u4efb\u52a1\u3002<\/p>\n
iRMB（Inverted Residual Mobile Block）\u7684\u4e3b\u8981\u601d\u60f3\u662f\u5c06\u8f7b\u91cf\u7ea7\u7684CNN\u67b6\u6784\u4e0e\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u6a21\u578b\u7ed3\u6784\u76f8\u7ed3\u5408（\u4f18\u70b9\u7c7b\u4f3c\u4e8eACmix），\u4ee5\u521b\u5efa\u9ad8\u6548\u7684\u79fb\u52a8\u7f51\u7edc\u3002Irmb\u901a\u8fc7\u91cd\u65b0\u8003\u8651\u5bfc\u81f4\u6b8b\u5dee（IRB）\u548cTransformer\u7684\u6709\u6548\u7ec4\u4ef6，\u5b9e\u73b0\u4e86\u4e00\u79cd\u7edf\u4e00\u7684\u89c6\u89d2，\u4ece\u800c\u6269\u5c55\u4e86CNN\u7684IRB\u5230\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u6a21\u578b\u3002Irmb\u7684\u8bbe\u8ba1\u76ee\u6807\u662f\u5728\u4fdd\u6301\u6a21\u578b\u8f7b\u91cf\u7ea7\u7684\u540c\u65f6，\u5b9e\u73b0\u5bf9\u8ba1\u7b97\u8d44\u6e90\u7684\u6709\u6548\u5229\u7528\u548c\u9ad8\u51c6\u786e\u7387\u3002\u8fd9\u4e00\u65b9\u6cd5\u901a\u8fc7\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u7684\u5e7f\u6cdb\u5b9e\u9a8c\u5f97\u5230\u9a8c\u8bc1，\u5c55\u793a\u51fa\u5176\u5728\u8f7b\u91cf\u7ea7\u6a21\u578b\u9886\u57df\u7684\u4f18\u8d8a\u6027\u80fd\u3002<\/p>\n
iRMB\u7684\u4e3b\u8981\u521b\u65b0\u70b9\u5728\u4e8e\u4ee5\u4e0b\u4e09\u4e2a\u70b9：<\/p>\n
\u7ed3\u5408CNN\u7684\u8f7b\u91cf\u7ea7\u7279\u5f81\u548cTransformer\u7684\u52a8\u6001\u6a21\u578b\u80fd\u529b，\u521b\u65b0\u63d0\u51fa\u4e86iRMB\u7ed3\u6784，\u9002\u7528\u4e8e\u79fb\u52a8\u8bbe\u5907\u4e0a\u7684\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1<\/p>\n<\/li>\n
\u4f7f\u7528\u5012\u7f6e\u6b8b\u5dee\u5757\u8bbe\u8ba1，\u6269\u5c55\u4e86\u4f20\u7edfCNN\u7684IRB\u5230\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u6a21\u578b，\u589e\u5f3a\u4e86\u6a21\u578b\u5904\u7406\u957f\u8ddd\u79bb\u4fe1\u606f\u7684\u80fd\u529b<\/p>\n<\/li>\n
\u63d0\u51fa\u4e86\u5143\u79fb\u52a8\u5757（Meta-Mobile Block），\u901a\u8fc7\u4e0d\u540c\u7684\u6269\u5c55\u6bd4\u7387\u548c\u9ad8\u6548\u64cd\u4f5c\u7b26，\u5b9e\u73b0\u4e86\u6a21\u578b\u7684\u6a21\u5757\u5316\u8bbe\u8ba1，\u4f7f\u5f97\u6a21\u578b\u66f4\u52a0\u7075\u6d3b\u548c\u9ad8\u6548<\/p>\n<\/li>\n
iRMB\u7ed3\u6784\u7684\u4e3b\u8981\u521b\u65b0\u70b9\u662f\u5b83\u7ed3\u5408\u4e86\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\u8f7b\u91cf\u7ea7\u7279\u6027\u548cTransformer\u6a21\u578b\u7684\u52a8\u6001\u5904\u7406\u80fd\u529b\u3002\u8fd9\u79cd\u7ed3\u6784\u7279\u522b\u9002\u7528\u4e8e\u79fb\u52a8\u8bbe\u5907\u4e0a\u7684\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1，\u56e0\u4e3a\u5b83\u65e8\u5728\u8ba1\u7b97\u80fd\u529b\u6709\u9650\u7684\u73af\u5883\u4e2d\u63d0\u4f9b\u9ad8\u6548\u7684\u6027\u80fd\u3002IRMB\u901a\u8fc7\u5176\u5012\u7f6e\u6b8b\u5dee\u8bbe\u8ba1\u6539\u8fdb\u4e86\u4fe1\u606f\u6d41\u7684\u5904\u7406，\u5141\u8bb8\u5728\u4fdd\u6301\u6a21\u578b\u8f7b\u91cf\u7684\u540c\u65f6\u6355\u6349\u548c\u5229\u7528\u957f\u8ddd\u79bb\u4f9d\u8d56，\u8fd9\u5bf9\u4e8e\u56fe\u50cf\u5206\u7c7b，\u5bf9\u8c61\u68c0\u6d4b\u548c\u8bed\u4e49\u5206\u5272\u7b49\u4efb\u52a1\u81f3\u5173\u91cd\u8981\u3002\u8fd9\u79cd\u8bbe\u8ba1\u4f7f\u5f97\u6a21\u578b\u5728\u8d44\u6e90\u53d7\u9650\u7684\u8bbe\u5907\u4e0a\u4e5f\u80fd\u591f\u9ad8\u6548\u8fd0\u884c，\u540c\u65f6\u4fdd\u6301\u6216\u63d0\u9ad8\u9884\u6d4b\u51c6\u786e\u6027\u3002<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260117075002-696b3f2ae6050.png\")
<\/p>\n
import math
\nimport torch
\nimport torch.nn as nn
\nimport torch.nn.functional as F
\nfrom functools import partial
\nfrom einops import rearrange
\nfrom timm.models._efficientnet_blocks import SqueezeExcite
\nfrom timm.models.layers import DropPath<\/p>\n
__all__ = ['iRMB', 'C2PSA_iRMB', 'C3k2_iRMB']<\/p>\n
inplace = True # \u5168\u5c40\u53d8\u91cf<\/p>\n
class LayerNorm2d(nn.Module):<\/p>\n
def __init__(self, normalized_shape, eps=1e-6, elementwise_affine=True):
\n super().__init__()
\n self.norm = nn.LayerNorm(normalized_shape, eps, elementwise_affine)<\/p>\n
def forward(self, x):
\n x = rearrange(x, 'b c h w -> b h w c').contiguous()
\n x = self.norm(x)
\n x = rearrange(x, 'b h w c -> b c h w').contiguous()
\n return x<\/p>\n
def get_norm(norm_layer='in_1d'):
\n eps = 1e-6
\n norm_dict = {
\n 'none': nn.Identity,
\n 'in_1d': partial(nn.InstanceNorm1d, eps=eps),
\n 'in_2d': partial(nn.InstanceNorm2d, eps=eps),
\n 'in_3d': partial(nn.InstanceNorm3d, eps=eps),
\n 'bn_1d': partial(nn.BatchNorm1d, eps=eps),
\n 'bn_2d': partial(nn.BatchNorm2d, eps=eps),
\n # 'bn_2d': partial(nn.SyncBatchNorm, eps=eps),
\n 'bn_3d': partial(nn.BatchNorm3d, eps=eps),
\n 'gn': partial(nn.GroupNorm, eps=eps),
\n 'ln_1d': partial(nn.LayerNorm, eps=eps),
\n 'ln_2d': partial(LayerNorm2d, eps=eps),
\n }
\n return norm_dict[norm_layer]<\/p>\n
def get_act(act_layer='relu'):
\n act_dict = {
\n 'none': nn.Identity,
\n 'relu': nn.ReLU,
\n 'relu6': nn.ReLU6,
\n 'silu': nn.SiLU
\n }
\n return act_dict[act_layer]<\/p>\n
class ConvNormAct(nn.Module):<\/p>\n
def __init__(self, dim_in, dim_out, kernel_size, stride=1, dilation=1, groups=1, bias=False,
\n skip=False, norm_layer='bn_2d', act_layer='relu', inplace=True, drop_path_rate=0.):
\n super(ConvNormAct, self).__init__()
\n self.has_skip = skip and dim_in == dim_out
\n padding = math.ceil((kernel_size – stride) \/ 2)
\n self.conv = nn.Conv2d(dim_in, dim_out, kernel_size, stride, padding, dilation, groups, bias)
\n self.norm = get_norm(norm_layer)(dim_out)
\n self.act = get_act(act_layer)(inplace=inplace)
\n self.drop_path = DropPath(drop_path_rate) if drop_path_rate else nn.Identity()<\/p>\n
def forward(self, x):
\n shortcut = x
\n x = self.conv(x)
\n x = self.norm(x)
\n x = self.act(x)
\n if self.has_skip:
\n x = self.drop_path(x) + shortcut
\n return x<\/p>\n
class iRMB(nn.Module):<\/p>\n
def __init__(self, dim_in, norm_in=True, has_skip=True, exp_ratio=1.0, norm_layer='bn_2d',
\n act_layer='relu', v_proj=True, dw_ks=3, stride=1, dilation=1, se_ratio=0.0, dim_head=8, window_size=7,
\n attn_s=True, qkv_bias=False, attn_drop=0., drop=0., drop_path=0., v_group=False, attn_pre=False):
\n super().__init__()
\n dim_out = dim_in
\n self.norm = get_norm(norm_layer)(dim_in) if norm_in else nn.Identity()
\n dim_mid = int(dim_in * exp_ratio)
\n self.has_skip = (dim_in == dim_out and stride == 1) and has_skip
\n self.attn_s = attn_s
\n if self.attn_s:
\n assert dim_in % dim_head == 0, 'dim should be divisible by num_heads'
\n self.dim_head = dim_head
\n self.window_size = window_size
\n self.num_head = dim_in \/\/ dim_head
\n self.scale = self.dim_head ** -0.5
\n self.attn_pre = attn_pre
\n self.qk = ConvNormAct(dim_in, int(dim_in * 2), kernel_size=1, bias=qkv_bias, norm_layer='none',
\n act_layer='none')
\n self.v = ConvNormAct(dim_in, dim_mid, kernel_size=1, groups=self.num_head if v_group else 1, bias=qkv_bias,
\n norm_layer='none', act_layer=act_layer, inplace=inplace)
\n self.attn_drop = nn.Dropout(attn_drop)
\n else:
\n if v_proj:
\n self.v = ConvNormAct(dim_in, dim_mid, kernel_size=1, bias=qkv_bias, norm_layer='none',
\n act_layer=act_layer, inplace=inplace)
\n else:
\n self.v = nn.Identity()
\n self.conv_local = ConvNormAct(dim_mid, dim_mid, kernel_size=dw_ks, stride=stride, dilation=dilation,
\n groups=dim_mid, norm_layer='bn_2d', act_layer='silu', inplace=inplace)
\n self.se = SqueezeExcite(dim_mid, rd_ratio=se_ratio, act_layer=get_act(act_layer)) if se_ratio > 0.0 else nn.Identity()<\/p>\n
self.proj_drop = nn.Dropout(drop)
\n self.proj = ConvNormAct(dim_mid, dim_out, kernel_size=1, norm_layer='none', act_layer='none', inplace=inplace)
\n self.drop_path = DropPath(drop_path) if drop_path else nn.Identity()<\/p>\n
def forward(self, x):
\n shortcut = x
\n x = self.norm(x)
\n B, C, H, W = x.shape
\n if self.attn_s:
\n # padding
\n if self.window_size <= 0:
\n window_size_W, window_size_H = W, H
\n else:
\n window_size_W, window_size_H = self.window_size, self.window_size
\n pad_l, pad_t = 0, 0
\n pad_r = (window_size_W – W % window_size_W) % window_size_W
\n pad_b = (window_size_H – H % window_size_H) % window_size_H
\n x = F.pad(x, (pad_l, pad_r, pad_t, pad_b, 0, 0,))
\n n1, n2 = (H + pad_b) \/\/ window_size_H, (W + pad_r) \/\/ window_size_W
\n x = rearrange(x, 'b c (h1 n1) (w1 n2) -> (b n1 n2) c h1 w1', n1=n1, n2=n2).contiguous()
\n # attention
\n b, c, h, w = x.shape
\n qk = self.qk(x)
\n qk = rearrange(qk, 'b (qk heads dim_head) h w -> qk b heads (h w) dim_head', qk=2, heads=self.num_head,
\n dim_head=self.dim_head).contiguous()
\n q, k = qk[0], qk[1]
\n attn_spa = (q @ k.transpose(-2, -1)) * self.scale
\n attn_spa = attn_spa.softmax(dim=-1)
\n attn_spa = self.attn_drop(attn_spa)
\n if self.attn_pre:
\n x = rearrange(x, 'b (heads dim_head) h w -> b heads (h w) dim_head', heads=self.num_head).contiguous()
\n x_spa = attn_spa @ x
\n x_spa = rearrange(x_spa, 'b heads (h w) dim_head -> b (heads dim_head) h w', heads=self.num_head, h=h,
\n w=w).contiguous()
\n x_spa = self.v(x_spa)
\n else:
\n v = self.v(x)
\n v = rearrange(v, 'b (heads dim_head) h w -> b heads (h w) dim_head', heads=self.num_head).contiguous()
\n x_spa = attn_spa @ v
\n x_spa = rearrange(x_spa, 'b heads (h w) dim_head -> b (heads dim_head) h w', heads=self.num_head, h=h,
\n w=w).contiguous()
\n # unpadding
\n x = rearrange(x_spa, '(b n1 n2) c h1 w1 -> b c (h1 n1) (w1 n2)', n1=n1, n2=n2).contiguous()
\n if pad_r > 0 or pad_b > 0:
\n x = x[:, :, :H, :W].contiguous()
\n else:
\n x = self.v(x)<\/p>\n
x = x + self.se(self.conv_local(x)) if self.has_skip else self.se(self.conv_local(x))<\/p>\n
x = self.proj_drop(x)
\n x = self.proj(x)<\/p>\n
x = (shortcut + self.drop_path(x)) if self.has_skip else x
\n return x<\/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 """Standard convolution with args(ch_in, ch_out, kernel, stride, padding, groups, dilation, activation)."""<\/p>\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 """Initialize Conv layer with given arguments including activation."""
\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 """Apply convolution, batch normalization and activation to input tensor."""
\n return self.act(self.bn(self.conv(x)))<\/p>\n
def forward_fuse(self, x):
\n """Perform transposed convolution of 2D data."""
\n return self.act(self.conv(x))<\/p>\n
class PSABlock(nn.Module):
\n """
\n PSABlock class implementing a Position-Sensitive Attention block for neural networks.
\n This class encapsulates the functionality for applying multi-head attention and feed-forward neural network layers
\n with optional shortcut connections.
\n Attributes:
\n attn (Attention): Multi-head attention module.
\n ffn (nn.Sequential): Feed-forward neural network module.
\n add (bool): Flag indicating whether to add shortcut connections.
\n Methods:
\n forward: Performs a forward pass through the PSABlock, applying attention and feed-forward layers.
\n Examples:
\n Create a PSABlock and perform a forward pass
\n >>> psablock = PSABlock(c=128, attn_ratio=0.5, num_heads=4, shortcut=True)
\n >>> input_tensor = torch.randn(1, 128, 32, 32)
\n >>> output_tensor = psablock(input_tensor)
\n """<\/p>\n
def __init__(self, c, attn_ratio=0.5, num_heads=4, shortcut=True) -> None:
\n """Initializes the PSABlock with attention and feed-forward layers for enhanced feature extraction."""
\n super().__init__()<\/p>\n
self.attn = iRMB(c)
\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):
\n """Executes a forward pass through PSABlock, applying attention and feed-forward layers to the input 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 C2PSA_iRMB(nn.Module):
\n """
\n C2PSA module with attention mechanism for enhanced feature extraction and processing.
\n This module implements a convolutional block with attention mechanisms to enhance feature extraction and processing
\n capabilities. It includes a series of PSABlock modules for self-attention and feed-forward operations.
\n Attributes:
\n c (int): Number of hidden channels.
\n cv1 (Conv): 1×1 convolution layer to reduce the number of input channels to 2*c.
\n cv2 (Conv): 1×1 convolution layer to reduce the number of output channels to c.
\n m (nn.Sequential): Sequential container of PSABlock modules for attention and feed-forward operations.
\n Methods:
\n forward: Performs a forward pass through the C2PSA module, applying attention and feed-forward operations.
\n Notes:
\n This module essentially is the same as PSA module, but refactored to allow stacking more PSABlock modules.
\n Examples:
\n >>> c2psa = C2PSA(c1=256, c2=256, n=3, e=0.5)
\n >>> input_tensor = torch.randn(1, 256, 64, 64)
\n >>> output_tensor = c2psa(input_tensor)
\n """<\/p>\n
def __init__(self, c1, c2, n=1, e=0.5):
\n """Initializes the C2PSA module with specified input\/output channels, number of layers, and expansion ratio."""
\n super().__init__()
\n assert c1 == c2
\n self.c = int(c1 * e)
\n self.cv1 = Conv(c1, 2 * self.c, 1, 1)
\n self.cv2 = Conv(2 * self.c, c1, 1)<\/p>\n
self.m = nn.Sequential(*(PSABlock(self.c, attn_ratio=0.5, num_heads=self.c \/\/ 64) for _ in range(n)))<\/p>\n
def forward(self, x):
\n """Processes the input tensor 'x' through a series of PSA blocks and returns the transformed tensor."""
\n a, b = self.cv1(x).split((self.c, self.c), dim=1)
\n b = self.m(b)
\n return self.cv2(torch.cat((a, b), 1))<\/p>\n
class Bottleneck(nn.Module):
\n """Standard bottleneck."""<\/p>\n
def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):
\n """Initializes a standard bottleneck module with optional shortcut connection and configurable parameters."""
\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):
\n """Applies the YOLO FPN to input data."""
\n return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))<\/p>\n
class C3(nn.Module):
\n """CSP Bottleneck with 3 convolutions."""<\/p>\n
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
\n """Initialize the CSP Bottleneck with given channels, number, shortcut, groups, and expansion values."""
\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):
\n """Forward pass through the CSP bottleneck with 2 convolutions."""
\n return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))<\/p>\n
class C3k(C3):
\n """C3k is a CSP bottleneck module with customizable kernel sizes for feature extraction in neural networks."""<\/p>\n
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, k=3):
\n """Initializes the C3k module with specified channels, number of layers, and configurations."""
\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 C2f(nn.Module):
\n """Faster Implementation of CSP Bottleneck with 2 convolutions."""<\/p>\n
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):
\n """Initializes a CSP bottleneck with 2 convolutions and n Bottleneck blocks for faster processing."""
\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):
\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):
\n """Forward pass using split() instead of chunk()."""
\n y = list(self.cv1(x).split((self.c, self.c), 1))
\n y.extend(m(y[-1]) for m in self.m)
\n return self.cv2(torch.cat(y, 1))<\/p>\n
class C3k2(C2f):
\n """Faster Implementation of CSP Bottleneck with 2 convolutions."""<\/p>\n
def __init__(self, c1, c2, n=1, c3k=False, e=0.5, g=1, shortcut=True):
\n """Initializes the C3k2 module, a faster CSP Bottleneck with 2 convolutions and optional C3k blocks."""
\n super().__init__(c1, c2, n, shortcut, g, e)
\n self.m = nn.ModuleList(
\n C3k(self.c, self.c, 2, shortcut, g) if c3k else Bottleneck(self.c, self.c, shortcut, g) for _ in range(n)
\n )<\/p>\n
class C3k_iRMB(C3k):
\n def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, k=3):
\n super().__init__(c1, c2, n, shortcut, g, e, k)
\n c_ = int(c2 * e) # hidden channels
\n self.m = nn.Sequential(*(iRMB(c_, c_) for _ in range(n)))<\/p>\n
class C3k2_iRMB(C3k2):
\n def __init__(self, c1, c2, n=1, c3k=False, e=0.5, g=1, shortcut=True):
\n super().__init__(c1, c2, n, c3k, e, g, shortcut)
\n self.m = nn.ModuleList(C3k_iRMB(self.c, self.c, 2, shortcut, g) if c3k else iRMB(self.c, self.c) for _ in range(n)) <\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260117075003-696b3f2b2616a.png\")
<\/p>\n
\u5728iRMB.py\u6587\u4ef6\u91cc\u6dfb\u52a0\u7ed9\u51fa\u7684iRMB\u4ee3\u7801<\/p>\n
\u6dfb\u52a0\u5b8ciRMB\u4ee3\u7801\u540e，\u5728ultralytics\/nn\/modules\/__init__.py\u6587\u4ef6\u4e2d\u5f15\u7528<\/p>\n
from .iRMB import * <\/p>\n
\u7136\u540e，\u5728ultralytics\/nn\/tasks.py\u91cc\u5f15\u7528<\/p>\n
from .modules import * <\/p>\n
\u5728tasks.py\u627e\u5230parse_model\u51fd\u6570（\u53ef\u4ee5ctrl+f \u76f4\u63a5\u641c\u7d22parse_model\u4f4d\u7f6e）\u6dfb\u52a0：<\/p>\n
elif m in {iRMB}:
\n c2 = ch[f]
\n args = [c2, *args] <\/p>\n
\u6b64\u5904\u4ee5YOLOv11\u4ee3\u7801\u4e3a\u4f8b，\u521b\u5efaYOLO11_iRMB.yaml<\/p>\n
# Ultralytics YOLO 🚀, AGPL-3.0 license
\n# YOLO11 object detection model with P3-P5 outputs. For Usage examples see https:\/\/docs.ultralytics.com\/tasks\/detect<\/p>\n
# Parameters
\nnc: 80 # number of classes
\nscales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'
\n # [depth, width, max_channels]
\n n: [0.50, 0.25, 1024] # summary: 319 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
\n s: [0.50, 0.50, 1024] # summary: 319 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
\n m: [0.50, 1.00, 512] # summary: 409 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
\n l: [1.00, 1.00, 512] # summary: 631 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
\n x: [1.00, 1.50, 512] # summary: 631 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs<\/p>\n
# YOLO11n 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, [256, False, 0.25]]
\n – [-1, 1, Conv, [256, 3, 2]] # 3-P3\/8
\n – [-1, 2, C3k2, [512, False, 0.25]]
\n – [-1, 1, Conv, [512, 3, 2]] # 5-P4\/16
\n – [-1, 2, C3k2, [512, True]]
\n – [-1, 1, Conv, [1024, 3, 2]] # 7-P5\/32
\n – [-1, 2, C3k2, [1024, True]]
\n – [-1, 1, SPPF, [1024, 5]] # 9
\n – [-1, 2, C2PSA, [1024]] # 10<\/p>\n
# YOLO11n head
\nhead:
\n – [-1, 1, nn.Upsample, [None, 2, "nearest"]]
\n – [[-1, 6], 1, Concat, [1]] # cat backbone P4
\n – [-1, 2, C3k2, [512, False]] # 13<\/p>\n
– [-1, 1, nn.Upsample, [None, 2, "nearest"]]
\n – [[-1, 4], 1, Concat, [1]] # cat backbone P3
\n – [-1, 2, C3k2, [256, False]] # 16
\n – [-1, 1, iRMB, []] # 17 \u5c0f\u76ee\u6807\u68c0\u6d4b\u5c42\u589e\u52a0\u6ce8\u610f\u529b\u673a\u5236<\/p>\n
– [-1, 1, Conv, [256, 3, 2]]
\n – [[-1, 13], 1, Concat, [1]] # cat head P4
\n – [-1, 2, C3k2, [512, False]] # 20 (P4\/16-medium)
\n – [-1, 1, iRMB, []] # 21 \u4e2d\u76ee\u6807\u68c0\u6d4b\u5c42\u589e\u52a0\u6ce8\u610f\u529b\u673a\u5236<\/p>\n
– [-1, 1, Conv, [512, 3, 2]]
\n – [[-1, 10], 1, Concat, [1]] # cat head P5
\n – [-1, 2, C3k2, [1024, True]] # 24
\n – [-1, 1, iRMB, []] # 25 \u5927\u76ee\u6807\u68c0\u6d4b\u5c42\u589e\u52a0\u6ce8\u610f\u529b\u673a\u5236<\/p>\n
– [[17, 21, 25], 1, Detect, [nc]] # Detect(P3, P4, P5) <\/p>\n
\u4ee5\u4e0a\u662f\u4e00\u79cd\u6539\u8fdb\u65b9\u6cd5，\u4e5f\u53ef\u4ee5\u5728C3k2\u6a21\u5757\u3001C2PSA\u6a21\u5757\u7b49\u5904\u8fdb\u884c\u6539\u8fdb，\u6784\u5efa\u65b0\u7684\u7684C3k2_iRMB\u3001C2PSA_iRMB\u6a21\u5757\u3002<\/p>\n
YOLO11_C2PSA_iRMB.yaml<\/p>\n
# Ultralytics YOLO 🚀, AGPL-3.0 license
\n# YOLO11 object detection model with P3-P5 outputs. For Usage examples see https:\/\/docs.ultralytics.com\/tasks\/detect<\/p>\n
# Parameters
\nnc: 80 # number of classes
\nscales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'
\n # [depth, width, max_channels]
\n n: [0.50, 0.25, 1024] # summary: 319 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
\n s: [0.50, 0.50, 1024] # summary: 319 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
\n m: [0.50, 1.00, 512] # summary: 409 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
\n l: [1.00, 1.00, 512] # summary: 631 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
\n x: [1.00, 1.50, 512] # summary: 631 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs<\/p>\n
# YOLO11n 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, [256, False, 0.25]]
\n – [-1, 1, Conv, [256, 3, 2]] # 3-P3\/8
\n – [-1, 2, C3k2, [512, False, 0.25]]
\n – [-1, 1, Conv, [512, 3, 2]] # 5-P4\/16
\n – [-1, 2, C3k2, [512, True]]
\n – [-1, 1, Conv, [1024, 3, 2]] # 7-P5\/32
\n – [-1, 2, C3k2, [1024, True]]
\n – [-1, 1, SPPF, [1024, 5]] # 9
\n – [-1, 2, C2PSA_iRMB, [1024]] # 10<\/p>\n
# YOLO11n head
\nhead:
\n – [-1, 1, nn.Upsample, [None, 2, "nearest"]]
\n – [[-1, 6], 1, Concat, [1]] # cat backbone P4
\n – [-1, 2, C3k2, [512, False]] # 13<\/p>\n
– [-1, 1, nn.Upsample, [None, 2, "nearest"]]
\n – [[-1, 4], 1, Concat, [1]] # cat backbone P3
\n – [-1, 2, C3k2, [256, False]] # 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, [512, False]] # 19 (P4\/16-medium)<\/p>\n
– [-1, 1, Conv, [512, 3, 2]]
\n – [[-1, 10], 1, Concat, [1]] # cat head P5
\n – [-1, 2, C3k2, [1024, True]] # 22 (P5\/32-large)<\/p>\n
– [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5) <\/p>\n
YOLO11_C3k2_iRMB.yaml<\/p>\n
# Ultralytics YOLO 🚀, AGPL-3.0 license
\n# YOLO11 object detection model with P3-P5 outputs. For Usage examples see https:\/\/docs.ultralytics.com\/tasks\/detect<\/p>\n
# Parameters
\nnc: 80 # number of classes
\nscales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'
\n # [depth, width, max_channels]
\n n: [0.50, 0.25, 1024] # summary: 319 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
\n s: [0.50, 0.50, 1024] # summary: 319 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
\n m: [0.50, 1.00, 512] # summary: 409 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
\n l: [1.00, 1.00, 512] # summary: 631 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
\n x: [1.00, 1.50, 512] # summary: 631 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs<\/p>\n
# YOLO11n 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_iRMB, [256, False, 0.25]]
\n – [-1, 1, Conv, [256, 3, 2]] # 3-P3\/8
\n – [-1, 2, C3k2_iRMB, [512, False, 0.25]]
\n – [-1, 1, Conv, [512, 3, 2]] # 5-P4\/16
\n – [-1, 2, C3k2_iRMB, [512, True]]
\n – [-1, 1, Conv, [1024, 3, 2]] # 7-P5\/32
\n – [-1, 2, C3k2_iRMB, [1024, True]]
\n – [-1, 1, SPPF, [1024, 5]] # 9
\n – [-1, 2, C2PSA, [1024]] # 10<\/p>\n
# YOLO11n head
\nhead:
\n – [-1, 1, nn.Upsample, [None, 2, "nearest"]]
\n – [[-1, 6], 1, Concat, [1]] # cat backbone P4
\n – [-1, 2, C3k2, [512, False]] # 13<\/p>\n
– [-1, 1, nn.Upsample, [None, 2, "nearest"]]
\n – [[-1, 4], 1, Concat, [1]] # cat backbone P3
\n – [-1, 2, C3k2, [256, False]] # 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, [512, False]] # 19 (P4\/16-medium)<\/p>\n
– [-1, 1, Conv, [512, 3, 2]]
\n – [[-1, 10], 1, Concat, [1]] # cat head P5
\n – [-1, 2, C3k2, [1024, 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('YOLO11_iRMB.yaml')
\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='exp',
\n ) <\/p>\n","protected":false},"excerpt":{"rendered":"
iRMB\u5012\u7f6e\u6b8b\u5dee\u5757\u6ce8\u610f\u529b\u673a\u5236\u7b80\u4ecb
\niRMB（Inverted Residual Mobile Block）\u7684\u6846\u67b6\u539f\u7406，\u662f\u4e00\u79cd\u7ed3\u5408\u8f7b\u91cf\u7ea7CNN\u548c\u6ce8\u610f\u529b\u673a\u5236\u7684\u65b9\u6cd5，\u7528\u4e8e\u6539\u8fdb\u79fb\u52a8\u8bbe\u5907\u4e0a\u7684\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u3002IRMB\u901a\u8fc7\u5012\u7f6e\u6b8b\u5dee\u5757\u548c\u5143\u79fb\u52a8\u5757\u5b9e\u73b0\u9ad8\u6548\u4fe1\u606f\u5904\u7406，\u540c\u65f6\u4fdd\u6301\u6a21\u578b\u8f7b\u91cf\u5316\u3002\u672c\u6587\u4e2d\u63d0\u51fa\u4e00\u4e2a\u65b0\u7684\u4e3b\u5e72\u7f51\u7edcEMO，\u4e3b\u8981\u601d\u60f3\u662f\u5c06\u8f7b\u91cf\u7ea7\u7684CNN\u67b6\u6784\u4e0e\u57fa\u4e8e\u6ce8\u610f\u529b\u7684\u6a21\u578b\u7ed3\u6784\u76f8\u7ed3\u5408\u3002
\n\u76ee\u5f55
\n1. \u7b80\u4ecb
\n2. iR<\/p>\n","protected":false},"author":2,"featured_media":61420,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[6492,152,156,50,1665,1326,523],"topic":[],"class_list":["post-61423","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-irmb","tag-pytorch","tag-yolo","tag-50","tag-1665","tag-1326","tag-523"],"yoast_head":"\n