Swin Transformer

最新推荐文章于 2024-11-27 15:30:58 发布

不想悲伤到天明

最新推荐文章于 2024-11-27 15:30:58 发布

阅读量1.2k

点赞数 28

CC 4.0 BY-SA版权

文章标签：深度学习人工智能

本文链接：https://blog.csdn.net/qq_41661809/article/details/137236766

1. Patch Partition & Patch Embedding

1. Patch Partition & Patch Embedding

1.1 Patch Partition

首先将输入的图像 [H,W,3] ，切割成patches ，每个Patch大小是[4,4,3],比如一张[3,224,224]大小的图像，会被分成 224/4 * 224/4 = 3136个patch 。这样图像的维度变成 [224/4,224/4,4*4*3] 即[56,56,48]。

1.2 Linear Embedding

Linear Embedding 将特征维度从48映射到C(=96) 。

class PatchEmbed(nn.Module) :
    """
    将image split into no-overlapping patch
    """
    def __init__(self,patch_size=4,in_c=3,embed_dim=96,norm_layer=None) :
        super(PatchEmbed,self).__init__()
        patch_size = (patch_size ,patch_size)
        self.patch_size = patch_size
        self.in_chans = in_c
        self.embed_dim = embed_dim
        self.proj = nn.Conv2d(in_c,embed_dim,kernel_size=patch_size,stride=patch_size)
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()


    def forward(self,x) :
        _,_,H,W = x.shape
        pad_input = (H % self.patch_size[0] !=0 ) or (W % self.patch_size[1] !=0)

        if pad_input :
            # 需要 padding
            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1],  # 表示宽度方向右侧填充数
                          0, self.patch_size[0] - H % self.patch_size[0],  # 表示高度方向底部填充数
                          0, 0))