\u4e00\u3001\u8bba\u6587\u4fe1\u606f<\/p>\n
\u4e8c\u3001\u8bba\u6587\u6458\u8981\u6982\u51b5<\/p>\n
\u4e09\u3001PATConv\u6a21\u5757\u7ed3\u6784\u56fe<\/p>\n
\u56db\u3001PATConv\u6a21\u5757\u7684\u4f5c\u7528<\/p>\n
\u4e94\u3001PATConv\u6a21\u5757\u7684\u539f\u7406<\/p>\n
\u516d\u3001PATConv\u6a21\u5757\u7684\u4f18\u52bf<\/p>\n
\u4e03\u3001\u5373\u63d2\u5373\u7528\u6a21\u5757\u4ee3\u7801<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260128154053-697a2e058fa41.png\")
<\/p>\n
\u8bba\u6587\u9898\u76ee：<\/span>Partial Channel Network: Compute Fewer, Perform Better<\/p>\n \u4e2d\u6587\u9898\u76ee：<\/span>\u90e8\u5206\u901a\u9053\u7f51\u7edc：\u8ba1\u7b97\u66f4\u5c11，\u6027\u80fd\u66f4\u4f18<\/p>\n \u8bba\u6587\u94fe\u63a5：<\/span>https:\/\/arxiv.org\/pdf\/2502.01303<\/p>\n \u6240\u5c5e\u5355\u4f4d：<\/span>\u897f\u5b89\u4ea4\u901a\u5927\u5b66\u00b7\u4eba\u5de5\u667a\u80fd\u4e0e\u673a\u5668\u4eba\u7814\u7a76\u6240<\/p>\n \u8bbe\u8ba1\u4e00\u79cd\u80fd\u8ba9\u7f51\u7edc\u5728\u4e0d\u727a\u7272\u7cbe\u5ea6\u548c\u541e\u5410\u91cf\u7684\u524d\u63d0\u4e0b，\u4fdd\u6301\u4f4e\u53c2\u6570\u6570\u91cf\u548c\u8ba1\u7b97\u91cf（FLOPs）\u7684\u6a21\u5757\u6216\u673a\u5236，\u4ecd\u662f\u4e00\u9879\u6311\u6218\u3002\u4e3a\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\u5e76\u5145\u5206\u5229\u7528\u7279\u5f81\u56fe\u901a\u9053\u95f4\u7684\u5197\u4f59\u6027，\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u89e3\u51b3\u65b9\u6848 \u2014\u2014 \u90e8\u5206\u901a\u9053\u673a\u5236（PCM）<\/span>\u3002\u5177\u4f53\u800c\u8a00，\u901a\u8fc7\u62c6\u5206\u64cd\u4f5c\u5c06\u7279\u5f81\u56fe\u901a\u9053\u5212\u5206\u4e3a\u4e0d\u540c\u90e8\u5206，\u6bcf\u4e2a\u90e8\u5206\u5bf9\u5e94\u5377\u79ef\u3001\u6ce8\u610f\u529b\u3001\u6c60\u5316\u548c\u6052\u7b49\u6620\u5c04\u7b49\u4e0d\u540c\u64cd\u4f5c\u3002\u57fa\u4e8e\u8fd9\u4e00\u601d\u8def，\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u90e8\u5206\u6ce8\u610f\u529b\u5377\u79ef（PATConv）<\/span>，\u80fd\u591f\u9ad8\u6548\u5730\u5c06\u5377\u79ef\u4e0e\u89c6\u89c9\u6ce8\u610f\u529b\u76f8\u7ed3\u5408\u3002\u7814\u7a76\u8868\u660e，PATConv \u53ef\u5b8c\u5168\u66ff\u4ee3\u5e38\u89c4\u5377\u79ef\u548c\u5e38\u89c4\u89c6\u89c9\u6ce8\u610f\u529b，\u540c\u65f6\u51cf\u5c11\u6a21\u578b\u53c2\u6570\u548c\u8ba1\u7b97\u91cf\u3002\u6b64\u5916，PATConv \u8fd8\u80fd\u884d\u751f\u51fa\u4e09\u79cd\u65b0\u578b\u6a21\u5757：<\/span>\u90e8\u5206\u901a\u9053\u6ce8\u610f\u529b\u6a21\u5757（PAT_ch）\u3001\u90e8\u5206\u7a7a\u95f4\u6ce8\u610f\u529b\u6a21\u5757（PAT_sp）\u548c\u90e8\u5206\u81ea\u6ce8\u610f\u529b\u6a21\u5757（PAT_sf）\u3002\u6211\u4eec\u8fd8\u63d0\u51fa\u4e86\u4e00\u79cd\u52a8\u6001\u90e8\u5206\u5377\u79ef（DPConv），\u5176\u80fd\u81ea\u9002\u5e94\u5b66\u4e60\u4e0d\u540c\u5c42\u7684\u901a\u9053\u62c6\u5206\u6bd4\u4f8b，\u4ee5\u5b9e\u73b0\u66f4\u4f18\u7684\u6027\u80fd\u5e73\u8861\u3002\u57fa\u4e8e PATConv \u548c DPConv，\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u65b0\u7684\u6df7\u5408\u7f51\u7edc\u5bb6\u65cf PartialNet\u3002\u8be5\u7f51\u7edc\u5728 ImageNet-1K \u5206\u7c7b\u4efb\u52a1\u4e2d\u53d6\u5f97\u4e86\u4f18\u4e8e\u73b0\u6709\u90e8\u5206 SOTA \u6a21\u578b\u7684 Top-1 \u7cbe\u5ea6\u548c\u63a8\u7406\u901f\u5ea6<\/span>，\u5e76\u4e14\u5728 COCO \u6570\u636e\u96c6\u7684\u76ee\u6807\u68c0\u6d4b\u548c\u5b9e\u4f8b\u5206\u5272\u4efb\u52a1<\/span>\u4e2d\u4e5f\u8868\u73b0\u51fa\u8272\u3002<\/p>\n PATConv（Partial Attention Convolution，\u90e8\u5206\u6ce8\u610f\u529b\u5377\u79ef）\u662f\u4e00\u79cd\u9002\u7528\u4e8e\u6df1\u5ea6\u5b66\u4e60\u89c6\u89c9\u4efb\u52a1\u7684\u5373\u63d2\u5373\u7528\u6a21\u5757，\u6838\u5fc3\u4f5c\u7528\u662f\u5728\u4e0d\u727a\u7272\u6a21\u578b\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b，\u5b9e\u73b0\u53c2\u6570\u6570\u91cf\u3001\u8ba1\u7b97\u91cf（FLOPs）\u4e0e\u63a8\u7406\u901f\u5ea6\u7684\u6700\u4f18\u5e73\u8861，\u5177\u4f53\u8868\u73b0\u4e3a：<\/p>\n PATConv \u57fa\u4e8e\u90e8\u5206\u901a\u9053\u673a\u5236（PCM） \u8bbe\u8ba1，\u6838\u5fc3\u601d\u8def\u662f \u201c\u901a\u9053\u62c6\u5206 – \u5e76\u884c\u8fd0\u7b97 – \u878d\u5408\u8f93\u51fa\u201d，\u5177\u4f53\u539f\u7406\u5982\u4e0b：<\/p>\n \u5c06\u8f93\u5165\u7279\u5f81\u56fe\u7684\u901a\u9053（c\u200b）\u6309\u6bd4\u4f8b p（\u53ef\u81ea\u9002\u5e94\u5b66\u4e60\u7684\u8d85\u53c2\u6570）\u62c6\u5206\u4e3a\u4e24\u90e8\u5206：<\/p>\n \u4e24\u90e8\u5206\u901a\u9053\u5206\u522b\u7ecf\u8fc7\u5bf9\u5e94\u8fd0\u7b97\u540e，\u901a\u8fc7\u62fc\u63a5（Concatenation）\u64cd\u4f5c\u878d\u5408\u4e3a\u5b8c\u6574\u8f93\u51fa\u7279\u5f81\u56fe，\u6570\u5b66\u5b9a\u4e49\u4e3a：<\/p>\n \u5176\u4e2d，F \u4e3a\u8f93\u5165\u7279\u5f81\u56fe（\u7ef4\u5ea6 h\u00d7w\u00d7c\u200b），O \u4e3a\u8f93\u51fa\u7279\u5f81\u56fe（\u7ef4\u5ea6 h\u00d7w\u00d7c），\u222a \u8868\u793a\u62fc\u63a5，Atten \u8868\u793a\u89c6\u89c9\u6ce8\u610f\u529b\u64cd\u4f5c\u3002<\/p>\n \u57fa\u4e8e\u4e0a\u8ff0\u539f\u7406，PATConv \u53ef\u7075\u6d3b\u7ec4\u5408\u5377\u79ef\u4e0e\u4e0d\u540c\u7c7b\u578b\u6ce8\u610f\u529b，\u884d\u751f\u51fa\u4e09\u79cd\u4e13\u7528\u5757，\u9002\u914d\u7f51\u7edc\u4e0d\u540c\u5c42\u9700\u6c42：<\/p>\n \u4e3a\u89e3\u51b3\u56fa\u5b9a\u62c6\u5206\u6bd4\u4f8b rp\u200b \u9002\u914d\u6027\u4e0d\u8db3\u7684\u95ee\u9898，PATConv \u53ef\u642d\u914d\u52a8\u6001\u90e8\u5206\u5377\u79ef（DPConv）：\u901a\u8fc7\u53ef\u5b66\u4e60\u7684\u95e8\u5411\u91cf（Gate Vector）\u548c\u514b\u7f57\u5185\u514b\u79ef（Kronecker Product）\u751f\u6210\u4e8c\u8fdb\u5236\u5173\u7cfb\u77e9\u9635，\u81ea\u9002\u5e94\u8c03\u6574\u4e0d\u540c\u7f51\u7edc\u5c42\u7684\u901a\u9053\u62c6\u5206\u6bd4\u4f8b，\u6ee1\u8db3\u53c2\u6570\u3001 latency \u7b49\u7ea6\u675f\u6761\u4ef6，\u5b9e\u73b0\u7cbe\u5ea6\u4e0e\u6548\u7387\u7684\u52a8\u6001\u5e73\u8861\u3002<\/p>\n import torch has_mmdet = True def hard_sigmoid(x, inplace: bool = False): def _make_divisible(v, divisor, min_value=None): class SqueezeExcite(nn.Module): def forward(self, x): class RPEAttention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.0, proj_drop=0.0, rpe_config=None): self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) # image relative position encoding def forward(self, x): q *= self.scale<\/p>\n attn = (q @ k.transpose(-2, -1))<\/p>\n # image relative position on keys attn = attn.softmax(dim=-1) out = attn @ v<\/p>\n # image relative position on values x = out.transpose(1, 2).reshape(B, N, C) class SRM(nn.Module): def forward(self, x): class PATConv(nn.Module): if use_attn: def forward_atten(self, x: Tensor) -> Tensor: def forward_slicing(self, x: Tensor) -> Tensor: def forward_split_cat(self, x: Tensor) -> Tensor: if __name__ == '__main__': \u4e00\u3001\u8bba\u6587\u4fe1\u606f\u4e8c\u3001\u8bba\u6587\u6458\u8981\u6982\u51b5<\/h2>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260128154053-697a2e05bbce5.png\")
<\/p>\n\u4e09\u3001PATConv\u6a21\u5757\u7ed3\u6784\u56fe<\/h2>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260128154053-697a2e05d929b.png\")
<\/p>\n\u56db\u3001PATConv\u6a21\u5757\u7684\u4f5c\u7528<\/h2>\n
\u4e94\u3001PATConv\u6a21\u5757\u7684\u539f\u7406<\/h2>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260128154054-697a2e0613de3.png\")
<\/p>\n1. \u901a\u9053\u62c6\u5206（Split Operation）<\/h4>\n
\n
2. \u5e76\u884c\u8fd0\u7b97\u4e0e\u878d\u5408<\/h4>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260128154054-697a2e064ee5f.png\")
<\/p>\n3. \u884d\u751f\u4e09\u79cd\u6ce8\u610f\u529b\u5757<\/h4>\n
\n
4. \u52a8\u6001\u901a\u9053\u6bd4\u4f8b\u4f18\u5316（\u7ed3\u5408 DPConv）<\/h4>\n
\u516d\u3001PATConv\u6a21\u5757\u7684\u4f18\u52bf<\/h2>\n
1. \u6548\u7387\u4f18\u52bf：\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u5f00\u9500\u4e0e\u63a8\u7406 latency<\/h4>\n
\n
2. \u6027\u80fd\u4f18\u52bf：\u7cbe\u5ea6\u4e0d\u964d\u53cd\u5347<\/h4>\n
\n
3. \u7075\u6d3b\u6027\u4e0e\u901a\u7528\u6027\u4f18\u52bf<\/h4>\n
\n
4. \u521b\u65b0\u6027\u4f18\u52bf<\/h4>\n
\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260128154054-697a2e0664a79.png\")
<\/p>\n\u4e03\u3001\u5373\u63d2\u5373\u7528\u6a21\u5757\u4ee3\u7801<\/h2>\n
\nimport torch.nn as nn
\nimport timm
\nfrom torch import Tensor
\nimport torch.nn.functional as F
\nfrom exp.irpe import build_rpe, get_rpe_config
\ntry:
\n from mmdet.models.builder import BACKBONES as det_BACKBONES
\n from mmdet.utils import get_root_logger
\n from mmcv.runner import _load_checkpoint<\/p>\n
\nexcept ImportError:
\n print("If for detection, please install mmdetection first")
\n has_mmdet = False<\/p>\n
\n if inplace:
\n return x.add_(3.).clamp_(0., 6.).div_(6.)
\n else:
\n return F.relu6(x + 3.) \/ 6.<\/p>\n
\n """
\n This function is taken from the original tf repo.
\n It ensures that all layers have a channel number that is divisible by 4
\n It can be seen here:
\n https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/slim\/nets\/mobilenet\/mobilenet.py
\n """
\n if min_value is None:
\n min_value = divisor
\n new_v = max(min_value, int(v + divisor \/ 2) \/\/ divisor * divisor)
\n # Make sure that round down does not go down by more than 10%.
\n if new_v < 0.9 * v:
\n new_v += divisor
\n return new_v<\/p>\n
\n def __init__(self, in_chs, se_ratio=0.25, reduced_base_chs=None, act_layer=nn.ReLU, gate_fn=hard_sigmoid, divisor=4,
\n **_):
\n super(SqueezeExcite, self).__init__()
\n self.gate_fn = gate_fn
\n reduced_chs = _make_divisible((reduced_base_chs or in_chs) * se_ratio, divisor)
\n self.avg_pool = nn.AdaptiveAvgPool2d(1)
\n self.conv_reduce = nn.Conv2d(in_chs, reduced_chs, 1, bias=True)
\n self.act1 = act_layer(inplace=True)
\n self.conv_expand = nn.Conv2d(reduced_chs, in_chs, 1, bias=True)<\/p>\n
\n x_se = self.avg_pool(x)
\n x_se = self.conv_reduce(x_se)
\n x_se = self.act1(x_se)
\n x_se = self.conv_expand(x_se)
\n x = x * self.gate_fn(x_se)
\n return x<\/p>\n
\n '''Attention with image relative position encoding '''<\/p>\n
\n super().__init__()
\n self.num_heads = num_heads
\n head_dim = dim \/\/ num_heads
\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
\n self.scale = qk_scale or head_dim ** -0.5<\/p>\n
\n self.attn_drop = nn.Dropout(attn_drop)
\n self.proj = nn.Linear(dim, dim)
\n self.proj_drop = nn.Dropout(proj_drop)<\/p>\n
\n self.rpe_q, self.rpe_k, self.rpe_v = build_rpe(rpe_config, head_dim=head_dim, num_heads=num_heads)<\/p>\n
\n B, C, h, w = x.shape
\n x = x.view(B, C, h * w).transpose(1, 2)
\n B, N, C = x.shape
\n qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C \/\/ self.num_heads).permute(2, 0, 3, 1, 4)
\n q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)<\/p>\n
\n if self.rpe_k is not None:
\n # attn += self.rpe_k(q)
\n attn += self.rpe_k(q, h, w)
\n # image relative position on queries
\n if self.rpe_q is not None:
\n attn += self.rpe_q(k * self.scale).transpose(2, 3)<\/p>\n
\n attn = self.attn_drop(attn)<\/p>\n
\n if self.rpe_v is not None:
\n out += self.rpe_v(attn)<\/p>\n
\n x = self.proj(x)
\n x = self.proj_drop(x)
\n x = x.transpose(1, 2).view(B, C, h, w)
\n return x<\/p>\n
\n def __init__(self, channel):
\n super().__init__()
\n self.cfc1 = nn.Conv2d(channel, channel, kernel_size=(1, 2), bias=False)
\n #self.cfc2 = nn.Conv2d(channel, channel, kernel_size=1, bias=False)
\n self.bn = nn.BatchNorm2d(channel)
\n self.sigmoid = nn.Hardsigmoid()<\/p>\n
\n b, c, h, w = x.shape
\n # style pooling
\n mean = x.reshape(b, c, -1).mean(-1).view(b, c, 1, 1)
\n std = x.reshape(b, c, -1).std(-1).view(b, c, 1, 1)
\n # max_value = torch.max(x.reshape(b, c, -1), -1)[0].view(b,c,1,1)
\n u = torch.cat([mean, std], dim=-1)
\n # style integration
\n z = self.cfc1(u)
\n # z = self.act(z)
\n # z = self.cfc2(z)
\n # z = self.bn(z)
\n g = self.sigmoid(z)
\n g = g.reshape(b, c, 1, 1)
\n return x * g.expand_as(x)<\/p>\n
\n def __init__(self, dim, n_div=4, forward_type='split_cat', use_attn=True, channel_type='se', patnet_t0=True): #'se' if i_stage <= 2 else 'self',
\n super().__init__()
\n self.dim_conv3 = dim \/\/ n_div
\n self.dim = dim
\n self.n_div = n_div
\n self.dim_untouched = dim – self.dim_conv3
\n self.partial_conv3 = nn.Conv2d(self.dim_conv3, self.dim_conv3, 3, 1, 1, bias=False)
\n self.use_attn = use_attn
\n self.channel_type = channel_type<\/p>\n
\n if channel_type == 'self':
\n self.partial_conv3 = nn.Conv2d(self.dim_conv3, self.dim_conv3, 3, 1, 1, bias=False)
\n rpe_config = get_rpe_config(
\n ratio=20,
\n method="euc",
\n mode='bias',
\n shared_head=False,
\n skip=0,
\n rpe_on='k',
\n )
\n if patnet_t0:
\n num_heads = 4
\n else:
\n num_heads = 6
\n self.attn = RPEAttention(self.dim_untouched, num_heads=num_heads, attn_drop=0.1, proj_drop=0.1,
\n rpe_config=rpe_config)
\n self.norm = timm.layers.LayerNorm2d(self.dim_untouched)
\n # self.norm = timm.layers.LayerNorm2d(self.dim)
\n self.forward = self.forward_atten
\n elif channel_type == 'se':
\n self.partial_conv3 = nn.Conv2d(self.dim_conv3, self.dim_conv3, 3, 1, 1, bias=False)
\n self.attn = SRM(self.dim_untouched)
\n self.norm = nn.BatchNorm2d(self.dim_untouched)
\n self.forward = self.forward_atten
\n else:
\n if forward_type == 'slicing':
\n self.forward = self.forward_slicing
\n elif forward_type == 'split_cat':
\n self.forward = self.forward_split_cat
\n else:
\n raise NotImplementedError<\/p>\n
\n if self.channel_type:
\n #print(self.channel_type)
\n if self.channel_type == 'se':
\n x1, x2 = torch.split(x, [self.dim_conv3, self.dim_untouched], dim=1)
\n x1 = self.partial_conv3(x1)
\n # x = self.partial_conv3(x)
\n x2 = self.attn(x2)
\n x2 = self.norm(x2)
\n x = torch.cat((x1, x2), 1)
\n # x = self.attn(x)
\n else:
\n x1, x2 = torch.split(x, [self.dim_conv3, self.dim_untouched], dim=1)
\n x1 = self.partial_conv3(x1)
\n x2 = self.norm(x2)
\n x2 = self.attn(x2)
\n x = torch.cat((x1, x2), 1)
\n return x<\/p>\n
\n x1 = x.clone() # !!! Keep the original input intact for the residual connection later
\n x1[:, :self.dim_conv3, :, :] = self.partial_conv3(x1[:, :self.dim_conv3, :, :])
\n return x1<\/p>\n
\n x1, x2 = torch.split(x, [self.dim_conv3, self.dim_untouched], dim=1)
\n x1 = self.partial_conv3(x1)
\n x = torch.cat((x1, x2), 1)
\n return x<\/p>\n
\n input = torch.rand(1, 64, 32, 32)
\n PATConv= PATConv(64,channel_type='se')
\n output = PATConv(input)
\n print('PATConv input_size:', input.size())
\n print('PATConv output_size:', output.size())<\/p>\n","protected":false},"excerpt":{"rendered":"
\n\u672c\u6587\u76ee\u5f55
\n\u4e00\u3001\u8bba\u6587\u4fe1\u606f
\n\u4e8c\u3001\u8bba\u6587\u6458\u8981\u6982\u51b5
\n\u4e09\u3001PATConv\u6a21\u5757\u7ed3\u6784\u56fe
\n\u56db\u3001PATConv\u6a21\u5757\u7684\u4f5c\u7528
\n\u4e94\u3001PATConv\u6a21\u5757\u7684\u539f\u7406
\n\u516d\u3001PATConv\u6a21\u5757\u7684\u4f18\u52bf
\n\u4e03\u3001\u5373\u63d2\u5373\u7528\u6a21\u5757\u4ee3\u7801 \u8bba\u6587\u9898\u76ee：Partial Channel Network: Compute Fewer, Perform Better \u4e2d\u6587\u9898\u76ee：\u90e8\u5206\u901a\u9053\u7f51\u7edc：\u8ba1\u7b97\u66f4\u5c11，\u6027\u80fd\u66f4\u4f18 \u8bba\u6587\u94fe\u63a5：https:\/\/arx<\/p>\n","protected":false},"author":2,"featured_media":67570,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[7153,7155,7152,6399,7154,5337,5804,1665],"topic":[],"class_list":["post-67576","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-aaai-2026","tag-cv","tag-patconv","tag-6399","tag-7154","tag-5337","tag-5804","tag-1665"],"yoast_head":"\n