![\"O\\\\left](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc70090d0.png\")
\u964d\u4f4e\u5230![\"O\\\\left](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc701da46.png\")
，\u5176\u4e2dM \u4e3a\u7a97\u53e3\u5927\u5c0f\u3002<\/li>\n\u4e8c\u3001Swin Transformer\u6570\u5b66\u516c\u5f0f<\/h4>\n（1）Patch Partition（\u5206\u5757）<\/h5>\n
\u5c06\u539f\u59cb\u56fe\u50cf（H\u00d7W\u00d73）\u5212\u5206\u4e3a\u5927\u5c0f\u4e3a 4\u00d74 \u7684\u4e0d\u91cd\u53e0 Patch，\u6bcf\u4e2a Patch \u5c55\u5e73\u4e3a\u4e00\u7ef4\u5411\u91cf：<\/p>\n
![\"Patch(i,j)=Image\\\\left](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc702d044.png\")
<\/p>\n
\u6700\u7ec8\u5f97\u5230 (H\u200b\/4\u00d7W\/4\u200b)\u00d748 \u7684\u7279\u5f81\u56fe，\u8bb0\u4e3a![\"H_{0},W_{0},C_{0}\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc7040f4c.png\")
\u3002<\/p>\n
（2）Window Attention \u8ba1\u7b97<\/h5>\n
\u5728\u6bcf\u4e2a\u7a97\u53e3\u5185\u8ba1\u7b97\u81ea\u6ce8\u610f\u529b，\u6838\u5fc3\u516c\u5f0f\u4e0e\u6807\u51c6\u81ea\u6ce8\u610f\u529b\u4e00\u81f4，\u4f46\u4f5c\u7528\u57df\u9650\u5236\u5728\u7a97\u53e3\u5185：<\/p>\n
![\"Attention(Q,K,V)=Softmax\\\\left](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc704d32b.png\")
<\/p>\n
- \n
- Q,K,V：\u67e5\u8be2 \/ \u952e \/ \u503c\u77e9\u9635，\u7531\u8f93\u5165\u7279\u5f81\u7ebf\u6027\u53d8\u6362\u5f97\u5230，dk\u200b \u4e3a Q\/K \u7684\u7ef4\u5ea6；<\/li>\n
- M：\u63a9\u7801\u77e9\u9635\u5728\u4ec5 Shifted Window \u65f6\u751f\u6548，\u7528\u4e8e\u5c4f\u853d\u79fb\u4f4d\u540e\u8de8\u539f\u59cb\u7a97\u53e3\u7684\u65e0\u6548\u6ce8\u610f\u529b\u8ba1\u7b97\u3002<\/li>\n<\/ul>\n
（3）Shifted Window \u79fb\u4f4d\u64cd\u4f5c<\/h5>\n
\u8bbe\u7a97\u53e3\u5927\u5c0f\u4e3a M，\u7279\u5f81\u56fe\u5c3a\u5bf8\u4e3a H\u00d7W，\u79fb\u4f4d\u91cf\u4e3a \u230aM\/2\u230b，\u79fb\u4f4d\u540e\u5750\u6807\u53d8\u6362：<\/p>\n
![\"\\\\left](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc705d0cb.png\")
;<\/p>\n（4）Patch Merging（\u5206\u5c42\u4e0b\u91c7\u6837）<\/h5>\n
\u5c06 2\u00d72 \u76f8\u90bb Patch \u62fc\u63a5，\u901a\u9053\u6570\u7ffb\u500d，\u5c3a\u5bf8\u51cf\u534a：<\/p>\n
![\"Out\\\\left](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc706c71c.png\")
<\/p>\n\u6700\u7ec8\u7279\u5f81\u56fe\u5c3a\u5bf8\u53d8\u4e3a H\/2\u00d7W\/2\u00d72C\u3002<\/p>\n
\u4e09\u3001\u5b9e\u4f8b\u4ee3\u7801\u89e3\u91ca<\/h4>\n
\u6a21\u5757\u4e00：\u6838\u5fc3\u4ee3\u7801\u5bfc\u5165<\/h5>\n
import torch
\nimport torch.nn as nn
\nimport torch.nn.functional as F
\nimport numpy as np <\/p>\n\u6a21\u5757\u4e8c：\u7a97\u53e3\u6ce8\u610f\u529b<\/h5>\n
class WindowAttention(nn.Module):
\n def __init__(self, dim, window_size, num_heads):
\n """
\n \u7a97\u53e3\u6ce8\u610f\u529b\u6a21\u5757\u521d\u59cb\u5316
\n Args:
\n dim: \u8f93\u5165\u7279\u5f81\u7684\u901a\u9053\u6570（\u5982 96\u3001192）
\n window_size: \u7a97\u53e3\u5927\u5c0f（M），\u9ed8\u8ba47，\u4ee3\u88687\u00d77\u7684\u7a97\u53e3
\n num_heads: \u591a\u5934\u6ce8\u610f\u529b\u7684\u5934\u6570，\u9700\u6ee1\u8db3 dim % num_heads == 0
\n """
\n super().__init__()
\n # \u4fdd\u5b58\u57fa\u7840\u53c2\u6570
\n self.dim = dim # \u8f93\u5165\u901a\u9053\u6570
\n self.window_size = window_size # \u7a97\u53e3\u5927\u5c0f M
\n self.num_heads = num_heads # \u6ce8\u610f\u529b\u5934\u6570
\n self.head_dim = dim \/\/ num_heads # \u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u7ef4\u5ea6
\n self.scale = self.head_dim ** -0.5 # \u7f29\u653e\u56e0\u5b50 1\/\u221ad_k，\u9632\u6b62\u5185\u79ef\u503c\u8fc7\u5927<\/p>\n# \u7ebf\u6027\u53d8\u6362\u5c42：\u5c06\u8f93\u5165\u7279\u5f81\u4e00\u6b21\u6027\u6620\u5c04\u4e3aQ\u3001K\u3001V（\u6548\u7387\u9ad8\u4e8e3\u4e2a\u72ec\u7acb\u7ebf\u6027\u5c42）
\n # \u8f93\u5165dim \u2192 \u8f93\u51fa3*dim（Q\/K\/V\u5404\u5360dim）
\n self.qkv = nn.Linear(dim, dim * 3)
\n # \u8f93\u51fa\u6295\u5f71\u5c42：\u5c06\u6ce8\u610f\u529b\u8ba1\u7b97\u540e\u7684\u7279\u5f81\u6620\u5c04\u56de\u539f\u7ef4\u5ea6
\n self.proj = nn.Linear(dim, dim)<\/p>\n# \u76f8\u5bf9\u4f4d\u7f6e\u504f\u7f6e\u8868：\u89e3\u51b3\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801\u7684\u5c40\u9650\u6027，\u6355\u6349\u7a97\u53e3\u5185\u4f4d\u7f6e\u5173\u7cfb
\n # \u5c3a\u5bf8：(2M-1)\u00d7(2M-1) \u00d7 num_heads \u2192 \u8986\u76d6\u7a97\u53e3\u5185\u6240\u6709\u53ef\u80fd\u7684\u76f8\u5bf9\u4f4d\u7f6e
\n self.relative_position_bias_table = nn.Parameter(
\n torch.zeros((2 * window_size – 1) * (2 * window_size – 1), num_heads)
\n )<\/p>\n# ————————– \u9884\u8ba1\u7b97\u76f8\u5bf9\u4f4d\u7f6e\u7d22\u5f15 ————————–
\n # \u751f\u6210\u7a97\u53e3\u5185\u7684\u5750\u6807\u7f51\u683c：[0,1,…,M-1]
\n coords = torch.arange(self.window_size)
\n # \u751f\u62102\u00d7M\u00d7M\u7684\u5750\u6807\u77e9\u9635：coords_grid[0]\u662f\u884c\u5750\u6807，coords_grid[1]\u662f\u5217\u5750\u6807
\n coords_grid = torch.stack(torch.meshgrid([coords, coords], indexing="ij"))
\n # \u5c55\u5e73\u5750\u6807：2 \u00d7 M\u00b2（\u628aM\u00d7M\u7684\u5750\u6807\u62c9\u6210\u4e00\u7ef4）
\n coords_flatten = torch.flatten(coords_grid, 1)
\n # \u8ba1\u7b97\u6240\u6709\u4f4d\u7f6e\u5bf9\u7684\u76f8\u5bf9\u5750\u6807：(2, M\u00b2, M\u00b2) \u2192 \u6bcf\u4e2a\u4f4d\u7f6e\u76f8\u5bf9\u4e8e\u5176\u4ed6\u4f4d\u7f6e\u7684\u504f\u79fb
\n relative_coords = coords_flatten[:, :, None] – coords_flatten[:, None, :]
\n # \u7ef4\u5ea6\u91cd\u6392：(M\u00b2, M\u00b2, 2) \u2192 [\u4f4d\u7f6ei, \u4f4d\u7f6ej, \u884c\/\u5217\u504f\u79fb]
\n relative_coords = relative_coords.permute(1, 2, 0).contiguous()
\n # \u5c06\u76f8\u5bf9\u5750\u6807\u4ece[-M+1, M-1]\u6620\u5c04\u5230[0, 2M-2]（\u907f\u514d\u8d1f\u6570\u7d22\u5f15）
\n relative_coords[:, :, 0] += self.window_size – 1 # \u884c\u504f\u79fb\u4fee\u6b63
\n relative_coords[:, :, 1] += self.window_size – 1 # \u5217\u504f\u79fb\u4fee\u6b63
\n # \u884c\u7d22\u5f15\u7f16\u7801：\u884c\u504f\u79fb \u00d7 (2M-1) + \u5217\u504f\u79fb \u2192 \u552f\u4e00\u6807\u8bc6\u6bcf\u4e2a\u76f8\u5bf9\u4f4d\u7f6e
\n relative_coords[:, :, 0] *= 2 * self.window_size – 1
\n # \u6c42\u548c\u5f97\u5230\u6700\u7ec8\u7684\u76f8\u5bf9\u4f4d\u7f6e\u7d22\u5f15：(M\u00b2, M\u00b2)
\n relative_position_index = relative_coords.sum(-1)
\n # \u6ce8\u518c\u4e3a\u7f13\u51b2\u533a（\u4e0d\u53c2\u4e0e\u68af\u5ea6\u66f4\u65b0）
\n self.register_buffer("relative_position_index", relative_position_index)<\/p>\ndef forward(self, x, mask=None):
\n """
\n \u7a97\u53e3\u6ce8\u610f\u529b\u524d\u5411\u4f20\u64ad
\n Args:
\n x: \u8f93\u5165\u7279\u5f81，\u5f62\u72b6 [num_windows*B, M\u00b2, dim]
\n – num_windows: \u7279\u5f81\u56fe\u5212\u5206\u7684\u7a97\u53e3\u603b\u6570
\n – B: batch size
\n – M\u00b2: \u5355\u4e2a\u7a97\u53e3\u7684\u50cf\u7d20\u6570
\n – dim: \u901a\u9053\u6570
\n mask: \u63a9\u7801\u77e9\u9635（\u4ec5\u79fb\u4f4d\u7a97\u53e3\u65f6\u4f7f\u7528），\u5f62\u72b6 [num_windows, M\u00b2, M\u00b2]
\n Returns:
\n output: \u6ce8\u610f\u529b\u8ba1\u7b97\u540e\u7684\u7279\u5f81，\u5f62\u72b6 [num_windows*B, M\u00b2, dim]
\n """
\n # \u83b7\u53d6\u8f93\u5165\u7ef4\u5ea6：B_=num_windows*B, N=M\u00b2, C=dim
\n B_, N, C = x.shape<\/p>\n# ————————– \u751f\u6210Q\/K\/V ————————–
\n # 1. \u7ebf\u6027\u53d8\u6362：[B_, N, C] \u2192 [B_, N, 3*C]
\n # 2. \u7ef4\u5ea6\u91cd\u6392：[B_, N, 3, num_heads, head_dim] \u2192 \u62c6\u52063\u4e2a\u7ef4\u5ea6\u7ed9Q\/K\/V
\n # 3. \u7ef4\u5ea6\u7f6e\u6362：[3, B_, num_heads, N, head_dim]
\n qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
\n # \u62c6\u5206Q\/K\/V：\u6bcf\u4e2a\u7684\u5f62\u72b6\u90fd\u662f [B_, num_heads, N, head_dim]
\n q, k, v = qkv[0], qkv[1], qkv[2]<\/p>\n# ————————– \u8ba1\u7b97\u6ce8\u610f\u529b\u5206\u6570 ————————–
\n q = q * self.scale # \u7f29\u653eQ，\u5bf9\u5e94\u516c\u5f0f\u4e2d\u7684 1\/\u221ad_k
\n # Q @ K^T：[B_, num_heads, N, head_dim] \u00d7 [B_, num_heads, head_dim, N] \u2192 [B_, num_heads, N, N]
\n attn = (q @ k.transpose(-2, -1))<\/p>\n# ————————– \u6dfb\u52a0\u76f8\u5bf9\u4f4d\u7f6e\u504f\u7f6e ————————–
\n # \u4ece\u504f\u7f6e\u8868\u4e2d\u53d6\u51fa\u5bf9\u5e94\u4f4d\u7f6e\u7684\u504f\u7f6e：[M\u00b2*M\u00b2, num_heads] \u2192 [M\u00b2, M\u00b2, num_heads]
\n relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
\n self.window_size * self.window_size, self.window_size * self.window_size, -1
\n )
\n # \u7ef4\u5ea6\u91cd\u6392：[num_heads, M\u00b2, M\u00b2]
\n relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
\n # \u6dfb\u52a0\u504f\u7f6e：[B_, num_heads, N, N] + [1, num_heads, N, N] \u2192 \u5e7f\u64ad\u76f8\u52a0
\n attn = attn + relative_position_bias.unsqueeze(0)<\/p>\n# ————————– \u5e94\u7528\u63a9\u7801（\u79fb\u4f4d\u7a97\u53e3\u4e13\u7528） ————————–
\n if mask is not None:
\n nW = mask.shape[0] # \u83b7\u53d6\u7a97\u53e3\u6570\u91cf
\n # 1. \u7ef4\u5ea6\u9002\u914d：\u5c06attn\u62c6\u5206\u4e3a [B\/\/nW, nW, num_heads, N, N]
\n # 2. \u63a9\u7801\u5e7f\u64ad：mask [nW, N, N] \u2192 [1, nW, 1, N, N]
\n # 3. \u76f8\u52a0：\u5c06\u63a9\u7801\u503c（-100）\u52a0\u5230\u8de8\u7a97\u53e3\u7684\u6ce8\u610f\u529b\u5206\u6570\u4e0a
\n attn = attn.view(B_ \/\/ nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
\n # \u8fd8\u539f\u7ef4\u5ea6：[B_, num_heads, N, N]
\n attn = attn.view(-1, self.num_heads, N, N)
\n # Softmax\u5f52\u4e00\u5316：\u63a9\u7801\u4f4d\u7f6e\u7684-100\u4f1a\u88abSoftmax\u4e3a0，\u4e0d\u53c2\u4e0e\u8ba1\u7b97
\n attn = F.softmax(attn, dim=-1)
\n else:
\n # \u65e0\u63a9\u7801\u65f6\u76f4\u63a5Softmax
\n attn = F.softmax(attn, dim=-1)<\/p>\n# ————————– \u6ce8\u610f\u529b\u52a0\u6743\u6c42\u548cV ————————–
\n # attn [B_, num_heads, N, N] \u00d7 v [B_, num_heads, N, head_dim] \u2192 [B_, num_heads, N, head_dim]
\n # \u7ef4\u5ea6\u7f6e\u6362：[B_, N, num_heads, head_dim] \u2192 \u5408\u5e76\u5934\u7ef4\u5ea6 \u2192 [B_, N, dim]
\n x = (attn @ v).transpose(1, 2).reshape(B_, N, C)<\/p>\n# ————————– \u8f93\u51fa\u6295\u5f71 ————————–
\n x = self.proj(x) # \u7ebf\u6027\u53d8\u6362，\u4fdd\u6301\u7ef4\u5ea6\u4e0d\u53d8
\n return x<\/p>\n# ========================== \u8f85\u52a9\u51fd\u6570：\u7a97\u53e3\u5212\u5206\u4e0e\u8fd8\u539f ==========================
\ndef window_partition(x, window_size):
\n """
\n \u5c06\u7279\u5f81\u56fe\u5212\u5206\u4e3a\u4e0d\u91cd\u53e0\u7684\u7a97\u53e3
\n Args:
\n x: \u8f93\u5165\u7279\u5f81，\u5f62\u72b6 [B, H, W, C]
\n window_size: \u7a97\u53e3\u5927\u5c0f M
\n Returns:
\n windows: \u7a97\u53e3\u5316\u7279\u5f81，\u5f62\u72b6 [num_windows*B, M, M, C]
\n num_windows = (H\/M) \u00d7 (W\/M)
\n """
\n B, H, W, C = x.shape
\n # \u7ef4\u5ea6\u62c6\u5206：[B, H, W, C] \u2192 [B, H\/\/M, M, W\/\/M, M, C]
\n x = x.view(B, H \/\/ window_size, window_size, W \/\/ window_size, window_size, C)
\n # \u7ef4\u5ea6\u7f6e\u6362：[B, H\/\/M, W\/\/M, M, M, C] \u2192 \u5408\u5e76\u524d\u4e09\u7ef4 \u2192 [num_windows*B, M, M, C]
\n windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
\n return windows<\/p>\ndef window_reverse(windows, window_size, H, W):
\n """
\n \u5c06\u7a97\u53e3\u5316\u7279\u5f81\u8fd8\u539f\u4e3a\u5b8c\u6574\u7279\u5f81\u56fe（window_partition\u7684\u9006\u64cd\u4f5c）
\n Args:
\n windows: \u7a97\u53e3\u5316\u7279\u5f81，\u5f62\u72b6 [num_windows*B, M, M, C]
\n window_size: \u7a97\u53e3\u5927\u5c0f M
\n H, W: \u539f\u59cb\u7279\u5f81\u56fe\u7684\u9ad8\u548c\u5bbd
\n Returns:
\n x: \u8fd8\u539f\u540e\u7684\u7279\u5f81\u56fe，\u5f62\u72b6 [B, H, W, C]
\n """
\n # \u8ba1\u7b97batch size：num_windows = (H*W)\/(M*M) \u2192 B = total_windows \/ num_windows
\n B = int(windows.shape[0] \/ (H * W \/ window_size \/ window_size))
\n # \u7ef4\u5ea6\u62c6\u5206：[num_windows*B, M, M, C] \u2192 [B, H\/\/M, W\/\/M, M, M, C]
\n x = windows.view(B, H \/\/ window_size, W \/\/ window_size, window_size, window_size, C)
\n # \u7ef4\u5ea6\u7f6e\u6362：[B, H\/\/M, M, W\/\/M, M, C] \u2192 \u5408\u5e76\u7ef4\u5ea6 \u2192 [B, H, W, C]
\n x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, C)
\n return x <\/p>\n\u6a21\u5757\u4e09：Swin Transformer Block<\/h5>\n
class SwinTransformerBlock(nn.Module):
\n def __init__(self, dim, num_heads, window_size=7, shift_size=0):
\n """
\n Swin Transformer \u57fa\u7840\u5757（\u5305\u542b\u7a97\u53e3\u6ce8\u610f\u529b\/\u79fb\u4f4d\u7a97\u53e3\u6ce8\u610f\u529b）
\n Args:
\n dim: \u8f93\u5165\u901a\u9053\u6570
\n num_heads: \u6ce8\u610f\u529b\u5934\u6570
\n window_size: \u7a97\u53e3\u5927\u5c0f M
\n shift_size: \u79fb\u4f4d\u91cf（0=\u666e\u901a\u7a97\u53e3，M\/\/2=\u79fb\u4f4d\u7a97\u53e3）
\n """
\n super().__init__()
\n self.dim = dim
\n self.num_heads = num_heads
\n self.window_size = window_size
\n self.shift_size = shift_size # \u79fb\u4f4d\u91cf，\u6838\u5fc3\u53c2\u6570<\/p>\n# \u5c42\u5f52\u4e00\u5316（Transformer\u6807\u51c6\u64cd\u4f5c，\u653e\u5728\u6ce8\u610f\u529b\u524d）
\n self.norm1 = nn.LayerNorm(dim)
\n # \u5b9e\u4f8b\u5316\u7a97\u53e3\u6ce8\u610f\u529b\u6a21\u5757
\n self.attn = WindowAttention(dim, window_size, num_heads)<\/p>\ndef forward(self, x):
\n """
\n Swin Block \u524d\u5411\u4f20\u64ad
\n Args:
\n x: \u8f93\u5165\u7279\u5f81，\u5f62\u72b6 [B, H, W, C]
\n Returns:
\n x: \u8f93\u51fa\u7279\u5f81，\u5f62\u72b6 [B, H, W, C]（\u6b8b\u5dee\u8fde\u63a5\u540e）
\n """
\n B, H, W, C = x.shape
\n shortcut = x # \u4fdd\u5b58\u6b8b\u5dee\u8fde\u63a5\u7684\u8f93\u5165<\/p>\n# 1. \u5c42\u5f52\u4e00\u5316
\n x = self.norm1(x)<\/p>\n# ————————– \u79fb\u4f4d\u64cd\u4f5c（Shifted Window） ————————–
\n if self.shift_size > 0:
\n # \u5faa\u73af\u79fb\u4f4d：\u5411\u5de6\u3001\u5411\u4e0a\u79fb\u52a8shift_size\u4e2a\u50cf\u7d20（\u8d1f\u6570\u8868\u793a\u5de6\/\u4e0a\u79fb）
\n # \u4f8b\u5982M=7，shift_size=3 \u2192 \u5de6\u79fb3，\u4e0a\u79fb3
\n shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))<\/p>\n# ————————– \u751f\u6210\u79fb\u4f4d\u7a97\u53e3\u7684\u63a9\u7801 ————————–
\n # 1. \u521d\u59cb\u5316\u63a9\u7801\u77e9\u9635：[1, H, W, 1]，\u7528\u4e8e\u6807\u8bb0\u4e0d\u540c\u539f\u59cb\u533a\u57df
\n img_mask = torch.zeros((1, H, W, 1), device=x.device)
\n # 2. \u5212\u5206\u79fb\u4f4d\u540e\u7684\u533a\u57df\u5207\u7247（\u51713\u00d73=9\u4e2a\u533a\u57df）
\n h_slices = (slice(0, -self.window_size), # \u4e0a\u534a\u90e8\u5206
\n slice(-self.window_size, -self.shift_size), # \u4e2d\u95f4\u8fc7\u6e21\u533a
\n slice(-self.shift_size, None)) # \u4e0b\u534a\u90e8\u5206
\n w_slices = (slice(0, -self.window_size), # \u5de6\u534a\u90e8\u5206
\n slice(-self.window_size, -self.shift_size), # \u4e2d\u95f4\u8fc7\u6e21\u533a
\n slice(-self.shift_size, None)) # \u53f3\u534a\u90e8\u5206
\n # 3. \u4e3a\u6bcf\u4e2a\u533a\u57df\u5206\u914d\u552f\u4e00\u6807\u7b7e（0-8）
\n cnt = 0
\n for h in h_slices:
\n for w in w_slices:
\n img_mask[:, h, w, :] = cnt
\n cnt += 1
\n # 4. \u5c06\u63a9\u7801\u5212\u5206\u4e3a\u7a97\u53e3：[num_windows, M, M, 1]
\n mask_windows = window_partition(img_mask, self.window_size)
\n # 5. \u5c55\u5e73\u63a9\u7801：[num_windows, M\u00b2]
\n mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
\n # 6. \u8ba1\u7b97\u6ce8\u610f\u529b\u63a9\u7801：\u5224\u65ad\u4e24\u4e2a\u4f4d\u7f6e\u662f\u5426\u5c5e\u4e8e\u540c\u4e00\u539f\u59cb\u533a\u57df
\n # – \u540c\u4e00\u533a\u57df：mask=0 \u2192 Softmax\u540e\u6b63\u5e38\u8ba1\u7b97
\n # – \u4e0d\u540c\u533a\u57df：mask=-100 \u2192 Softmax\u540e\u4e3a0，\u4e0d\u53c2\u4e0e\u8ba1\u7b97
\n attn_mask = mask_windows.unsqueeze(1) – mask_windows.unsqueeze(2)
\n attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
\n else:
\n # \u666e\u901a\u7a97\u53e3：\u4e0d\u79fb\u4f4d，\u65e0\u63a9\u7801
\n shifted_x = x
\n attn_mask = None<\/p>\n# ————————– \u7a97\u53e3\u6ce8\u610f\u529b\u8ba1\u7b97 ————————–
\n # 1. \u5212\u5206\u7a97\u53e3：[B, H, W, C] \u2192 [num_windows*B, M, M, C]
\n x_windows = window_partition(shifted_x, self.window_size)
\n # 2. \u5c55\u5e73\u7a97\u53e3：[num_windows*B, M\u00b2, C]（\u9002\u914dWindowAttention\u8f93\u5165）
\n x_windows = x_windows.view(-1, self.window_size * self.window_size, C)
\n # 3. \u7a97\u53e3\u6ce8\u610f\u529b\u524d\u5411\u8ba1\u7b97
\n attn_windows = self.attn(x_windows, mask=attn_mask)<\/p>\n# ————————– \u8fd8\u539f\u7a97\u53e3\u4e3a\u7279\u5f81\u56fe ————————–
\n # 1. \u8fd8\u539f\u7a97\u53e3\u5f62\u72b6：[num_windows*B, M\u00b2, C] \u2192 [num_windows*B, M, M, C]
\n attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
\n # 2. \u7a97\u53e3\u8fd8\u539f\u4e3a\u7279\u5f81\u56fe：[num_windows*B, M, M, C] \u2192 [B, H, W, C]
\n shifted_x = window_reverse(attn_windows, self.window_size, H, W)<\/p>\n# ————————– \u9006\u79fb\u4f4d（\u6062\u590d\u539f\u59cb\u4f4d\u7f6e） ————————–
\n if self.shift_size > 0:
\n # \u5411\u53f3\u3001\u5411\u4e0b\u79fb\u52a8shift_size\u4e2a\u50cf\u7d20，\u8fd8\u539f\u5230\u539f\u59cb\u4f4d\u7f6e
\n x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
\n else:
\n x = shifted_x<\/p>\n# ————————– \u6b8b\u5dee\u8fde\u63a5 ————————–
\n x = shortcut + x # \u6b8b\u5dee\u76f8\u52a0，\u63d0\u5347\u68af\u5ea6\u4f20\u64ad<\/p>\nreturn x <\/p>\n
\u6a21\u5757\u56db：Patch Merging（\u5206\u5c42\u4e0b\u91c7\u6837）<\/h5>\n
class PatchMerging(nn.Module):
\n def __init__(self, dim):
\n """
\n Patch Merging \u6a21\u5757：\u5c062\u00d72\u76f8\u90bbPatch\u5408\u5e76，\u5b9e\u73b0\u4e0b\u91c7\u6837
\n Args:
\n dim: \u8f93\u5165\u901a\u9053\u6570
\n """
\n super().__init__()
\n self.dim = dim
\n # \u7ebf\u6027\u53d8\u6362：\u5c064*dim\u901a\u9053\u538b\u7f29\u4e3a2*dim（\u4e0b\u91c7\u6837\u540e\u901a\u9053\u6570\u7ffb\u500d）
\n self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
\n # \u5c42\u5f52\u4e00\u5316：\u653e\u5728\u7ebf\u6027\u53d8\u6362\u524d，\u63d0\u5347\u7a33\u5b9a\u6027
\n self.norm = nn.LayerNorm(4 * dim)<\/p>\ndef forward(self, x):
\n """
\n Patch Merging \u524d\u5411\u4f20\u64ad
\n Args:
\n x: \u8f93\u5165\u7279\u5f81，\u5f62\u72b6 [B, H, W, C]
\n Returns:
\n x: \u4e0b\u91c7\u6837\u540e\u7684\u7279\u5f81，\u5f62\u72b6 [B, H\/2, W\/2, 2C]
\n """
\n B, H, W, C = x.shape<\/p>\n# ————————– 2\u00d72 Patch\u5408\u5e76 ————————–
\n # 1. \u7ef4\u5ea6\u62c6\u5206：[B, H, W, C] \u2192 [B, H\/\/2, 2, W\/\/2, 2, C]
\n x = x.view(B, H \/\/ 2, 2, W \/\/ 2, 2, C)
\n # 2. \u7ef4\u5ea6\u7f6e\u6362：[B, H\/\/2, W\/\/2, 2, 2, C] \u2192 \u5408\u5e76\u6700\u540e\u4e09\u7ef4 \u2192 [B, H\/\/2, W\/\/2, 4C]
\n x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H \/\/ 2, W \/\/ 2, -1)<\/p>\n# ————————– \u5f52\u4e00\u5316+\u901a\u9053\u538b\u7f29 ————————–
\n x = self.norm(x) # \u5c42\u5f52\u4e00\u5316
\n x = self.reduction(x) # 4C \u2192 2C，\u901a\u9053\u6570\u7ffb\u500d<\/p>\nreturn x<\/p>\n
\u6a21\u5757\u4e94：\u6d4b\u8bd5\u4ee3\u7801<\/h5>\n
if __name__ == "__main__":
\n # \u6a21\u62df\u8f93\u5165：batch_size=2，\u7279\u5f81\u56fe56\u00d756，\u901a\u9053\u657096（Swin-T\u7684\u7b2c\u4e00\u5c42\u7279\u5f81）
\n x = torch.randn(2, 56, 56, 96)<\/p>\n# 1. \u6d4b\u8bd5\u666e\u901a\u7a97\u53e3\u6ce8\u610f\u529b\u5757（\u65e0\u79fb\u4f4d）
\n block1 = SwinTransformerBlock(dim=96, num_heads=8, window_size=7, shift_size=0)
\n out1 = block1(x)
\n print("\u666e\u901a\u7a97\u53e3\u6ce8\u610f\u529b\u8f93\u51fa\u5f62\u72b6:", out1.shape) # \u9884\u671f：torch.Size([2, 56, 56, 96])<\/p>\n# 2. \u6d4b\u8bd5\u79fb\u4f4d\u7a97\u53e3\u6ce8\u610f\u529b\u5757（\u79fb\u4f4d\u91cf=3）
\n block2 = SwinTransformerBlock(dim=96, num_heads=8, window_size=7, shift_size=3)
\n out2 = block2(x)
\n print("\u79fb\u4f4d\u7a97\u53e3\u6ce8\u610f\u529b\u8f93\u51fa\u5f62\u72b6:", out2.shape) # \u9884\u671f：torch.Size([2, 56, 56, 96])<\/p>\n# 3. \u6d4b\u8bd5Patch Merging\u4e0b\u91c7\u6837
\n patch_merge = PatchMerging(dim=96)
\n out3 = patch_merge(x)
\n print("Patch Merging\u8f93\u51fa\u5f62\u72b6:", out3.shape) # \u9884\u671f：torch.Size([2, 28, 28, 192]) <\/p>\n\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u7528 \u201c\u5206\u7a97\u53e3\u7b97\u6ce8\u610f\u529b\u201d \u964d\u4f4e\u8ba1\u7b97\u91cf，\u7528 \u201c\u632a\u7a97\u53e3 + \u63a9\u7801\u201d \u6253\u901a\u7a97\u53e3\u95f4\u4fe1\u606f，\u7528 \u201c\u5408\u5e76\u50cf\u7d20\u5757\u201d \u6784\u5efa\u5206\u5c42\u7279\u5f81，\u6700\u7ec8\u8ba9 Transformer \u80fd\u9ad8\u6548\u5904\u7406\u56fe\u7247，\u65e2\u5feb\u53c8\u80fd\u5b66\u5230\u6709\u7528\u7684\u7279\u5f81\u3002\u3001<\/p>\n
\u8fd0\u884c\u7ed3\u679c：<\/h5>\n
\u666e\u901a\u7a97\u53e3\u6ce8\u610f\u529b\u8f93\u51fa\u5f62\u72b6: torch.Size([2, 56, 56, 96])<\/p>\n
\u79fb\u4f4d\u7a97\u53e3\u6ce8\u610f\u529b\u8f93\u51fa\u5f62\u72b6: torch.Size([2, 56, 56, 96])<\/p>\n
Patch Merging\u8f93\u51fa\u5f62\u72b6: torch.Size([2, 28, 28, 192])<\/p>\n
- \n
- \u8f93\u5165\u4e00\u5f20\u771f\u5b9e\u56fe\u7247（\u6bd4\u5982\u732b\u3001\u72d7\u7684\u7167\u7247）；<\/li>\n
- \u6a21\u578b\u8f93\u51fa\u8fd9\u5f20\u56fe\u7247\u7684 \u201c\u9ad8\u7ea7\u7279\u5f81\u201d（\u4e0d\u662f\u50cf\u7d20，\u662f\u80fd\u63cf\u8ff0 \u201c\u8fd9\u662f\u732b\u3001\u90a3\u662f\u72d7\u201d \u7684\u7279\u5f81）；<\/li>\n
- \u518d\u642d\u914d\u7b80\u5355\u7684\u5206\u7c7b \/ \u68c0\u6d4b\u5934，\u5c31\u80fd\u5b9e\u73b0\u56fe\u7247\u5206\u7c7b\u3001\u76ee\u6807\u68c0\u6d4b\u3001\u8bed\u4e49\u5206\u5272\u7b49\u89c6\u89c9\u4efb\u52a1（\u6bd4\u5982\u8bc6\u522b\u56fe\u7247\u91cc\u6709\u4ec0\u4e48\u3001\u627e\u5230\u7269\u4f53\u7684\u4f4d\u7f6e）\u3002<\/li>\n<\/ul>\n
\u56db\u3001\u603b\u7ed3<\/h4>\n
- \n
- \u6838\u5fc3\u521b\u65b0：Swin Transformer \u7528\u300c\u7a97\u53e3\u6ce8\u610f\u529b\u300d\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6，\u7528\u300c\u79fb\u4f4d\u7a97\u53e3\u6ce8\u610f\u529b\u300d\u89e3\u51b3\u7a97\u53e3\u95f4\u4fe1\u606f\u5b64\u7acb\u95ee\u9898，\u9002\u914d\u9ad8\u5206\u8fa8\u7387\u89c6\u89c9\u4efb\u52a1；<\/li>\n
- \u6570\u5b66\u6838\u5fc3：\u7a97\u53e3\u5185\u81ea\u6ce8\u610f\u529b\u516c\u5f0f，\u5176\u4e2d\u63a9\u7801 M \u662f\u79fb\u4f4d\u7a97\u53e3\u7684\u5173\u952e；<\/li>\n
- \u4ee3\u7801\u6838\u5fc3：\n
- \n
- \u7a97\u53e3\u5212\u5206 \/ \u8fd8\u539f\u662f\u7a97\u53e3\u6ce8\u610f\u529b\u7684\u57fa\u7840；<\/li>\n
- \u79fb\u4f4d + \u63a9\u7801\u662f\u5b9e\u73b0 Shifted Window \u7684\u6838\u5fc3；<\/li>\n
- Patch Merging \u5b9e\u73b0\u5206\u5c42\u4e0b\u91c7\u6837，\u6a21\u4eff CNN \u7684\u5c42\u7ea7\u7279\u5f81\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"
\u4e00\u3001Swin Transformer\u6838\u5fc3\u6982\u5ff5 Swin Transformer（Shifted Window Transformer）\u662f\u4e13\u4e3a\u89c6\u89c9\u4efb\u52a1\u8bbe\u8ba1\u7684 Transformer \u53d8\u4f53，\u89e3\u51b3\u4e86\u539f\u59cb Transformer \u5728\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u8ba1\u7b97\u91cf\u7206\u70b8\u7684\u95ee\u9898，\u6838\u5fc3\u521b\u65b0\u662f\u5206\u5c42\u7ed3\u6784\u548c\u79fb\u4f4d\u7a97\u53e3\u6ce8\u610f\u529b\u673a\u5236\u3002
\n\u6838\u5fc3\u6982\u5ff5：
\n\u5206\u5c42\u7279\u5f81\u63d0\u53d6：\u6a21\u4eff CNN \u7684\u5c42\u7ea7\u7ed3\u6784，\u901a\u8fc7 Patch Mer<\/p>\n","protected":false},"author":2,"featured_media":77551,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[841,50,523],"topic":[],"class_list":["post-77558","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-transformer","tag-50","tag-523"],"yoast_head":"\n\u8ba1\u7b97\u673a\u89c6\u89c9CV\u9886\u57df\u2014\u2014\u2014\u2014Swin Transformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/77558.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u8ba1\u7b97\u673a\u89c6\u89c9CV\u9886\u57df\u2014\u2014\u2014\u2014Swin Transformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4e00\u3001Swin Transformer\u6838\u5fc3\u6982\u5ff5 Swin Transformer（Shifted Window Transformer）\u662f\u4e13\u4e3a\u89c6\u89c9\u4efb\u52a1\u8bbe\u8ba1\u7684 Transformer \u53d8\u4f53，\u89e3\u51b3\u4e86\u539f\u59cb Transformer \u5728\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u8ba1\u7b97\u91cf\u7206\u70b8\u7684\u95ee\u9898，\u6838\u5fc3\u521b\u65b0\u662f\u5206\u5c42\u7ed3\u6784\u548c\u79fb\u4f4d\u7a97\u53e3\u6ce8\u610f\u529b\u673a\u5236\u3002 \u6838\u5fc3\u6982\u5ff5： \u5206\u5c42\u7279\u5f81\u63d0\u53d6：\u6a21\u4eff CNN \u7684\u5c42\u7ea7\u7ed3\u6784，\u901a\u8fc7 Patch Mer\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/77558.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-02-24T14:57:53+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc70090d0.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"9 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/77558.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/77558.html\",\"name\":\"\u8ba1\u7b97\u673a\u89c6\u89c9CV\u9886\u57df\u2014\u2014\u2014\u2014Swin Transformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-02-24T14:57:53+00:00\",\"dateModified\":\"2026-02-24T14:57:53+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/77558.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/77558.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/77558.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8ba1\u7b97\u673a\u89c6\u89c9CV\u9886\u57df\u2014\u2014\u2014\u2014Swin Transformer\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u8ba1\u7b97\u673a\u89c6\u89c9CV\u9886\u57df\u2014\u2014\u2014\u2014Swin Transformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/77558.html","og_locale":"zh_CN","og_type":"article","og_title":"\u8ba1\u7b97\u673a\u89c6\u89c9CV\u9886\u57df\u2014\u2014\u2014\u2014Swin Transformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4e00\u3001Swin Transformer\u6838\u5fc3\u6982\u5ff5 Swin Transformer（Shifted Window Transformer）\u662f\u4e13\u4e3a\u89c6\u89c9\u4efb\u52a1\u8bbe\u8ba1\u7684 Transformer \u53d8\u4f53，\u89e3\u51b3\u4e86\u539f\u59cb Transformer \u5728\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u8ba1\u7b97\u91cf\u7206\u70b8\u7684\u95ee\u9898，\u6838\u5fc3\u521b\u65b0\u662f\u5206\u5c42\u7ed3\u6784\u548c\u79fb\u4f4d\u7a97\u53e3\u6ce8\u610f\u529b\u673a\u5236\u3002 \u6838\u5fc3\u6982\u5ff5： \u5206\u5c42\u7279\u5f81\u63d0\u53d6：\u6a21\u4eff CNN \u7684\u5c42\u7ea7\u7ed3\u6784，\u901a\u8fc7 Patch Mer","og_url":"https:\/\/www.wsisp.com\/helps\/77558.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-02-24T14:57:53+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260224145752-699dbc70090d0.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"9 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/77558.html","url":"https:\/\/www.wsisp.com\/helps\/77558.html","name":"\u8ba1\u7b97\u673a\u89c6\u89c9CV\u9886\u57df\u2014\u2014\u2014\u2014Swin Transformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-02-24T14:57:53+00:00","dateModified":"2026-02-24T14:57:53+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/77558.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/77558.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/77558.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u8ba1\u7b97\u673a\u89c6\u89c9CV\u9886\u57df\u2014\u2014\u2014\u2014Swin Transformer"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/77558","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=77558"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/77558\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/77551"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=77558"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=77558"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=77558"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=77558"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}