| 914 | return x |
| 915 | |
| 916 | class PatchEmbed(nn.Module): |
| 917 | def __init__(self, in_chans=3, embed_dim=96, norm_layer=None): |
| 918 | super().__init__() |
| 919 | |
| 920 | self.embed_dim = embed_dim |
| 921 | self.proj = nn.Sequential( |
| 922 | ConvBNLayer( |
| 923 | in_channels=in_chans, |
| 924 | out_channels=embed_dim // 2, |
| 925 | kernel_size=3, |
| 926 | stride=2, |
| 927 | padding=1, |
| 928 | act=nn.GELU, |
| 929 | bias_attr=False), |
| 930 | ConvBNLayer( |
| 931 | in_channels=embed_dim // 2, |
| 932 | out_channels=embed_dim, |
| 933 | kernel_size=3, |
| 934 | stride=2, |
| 935 | padding=1, |
| 936 | act=nn.GELU, |
| 937 | bias_attr=False) |
| 938 | ) |
| 939 | |
| 940 | def forward(self, x): |
| 941 | x = self.proj(x).permute(0, 2, 3, 1).contiguous() |
| 942 | return x |
| 943 | |
| 944 | |
| 945 | class VIPTRNet(nn.Module): |