在https://github.com/facebookresearch/multiface/blob/main/models.py#L580找到了临时解决方案
所以不使用常规的nn.conv2d，而是使用

class Conv2dWN(nn.Conv2d):
    def __init__(
        self,
        in_channels,
        out_channels,
        kernel_size,
        stride=1,
        padding=0,
        dilation=1,
        groups=1,
        bias=True,
    ):
        super(Conv2dWN, self).__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride,
            padding,
            dilation,
            groups,
            True,
        )
        self.g = nn.Parameter(torch.ones(out_channels))

    def forward(self, x):
        wnorm = torch.sqrt(torch.sum(self.weight**2))
        return F.conv2d(
            x,
            self.weight * self.g[:, None, None, None] / wnorm,
            bias=self.bias,
            stride=self.stride,
            padding=self.padding,
            dilation=self.dilation,
            groups=self.groups,
        )

字符串