| 104 | |
| 105 | |
| 106 | class PatchEmbed(nn.Module): |
| 107 | def __init__(self, |
| 108 | in_channels, |
| 109 | out_channels, |
| 110 | stride=1): |
| 111 | super(PatchEmbed, self).__init__() |
| 112 | norm_layer = partial(nn.BatchNorm2d, eps=NORM_EPS) |
| 113 | if stride == 2: |
| 114 | self.avgpool = nn.AvgPool2d((2, 2), stride=2, ceil_mode=True, count_include_pad=False) |
| 115 | self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, bias=False) |
| 116 | self.norm = norm_layer(out_channels) |
| 117 | elif in_channels != out_channels: |
| 118 | self.avgpool = nn.Identity() |
| 119 | self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, bias=False) |
| 120 | self.norm = norm_layer(out_channels) |
| 121 | else: |
| 122 | self.avgpool = nn.Identity() |
| 123 | self.conv = nn.Identity() |
| 124 | self.norm = nn.Identity() |
| 125 | |
| 126 | def forward(self, x): |
| 127 | return self.norm(self.conv(self.avgpool(x))) |
| 128 | |
| 129 | |
| 130 | class MHCA(nn.Module): |