成功解决RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should

复现BEVDepth报错

File "BEVDepth/bevdepth/layers/backbones/base_lss_fpn.py", line 309, in forward
x = self.reduce_conv(x)
RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same

原因

由于在使用PyTorch的自动混合精度训练（AMP）时，输入张量和权重张量的数据类型不一致所致。在这种情况下，输入是半精度浮点数（HalfTensor），而权重是单精度浮点数（FloatTensor）。

解决办法

    @autocast(Flase)
    def forward(self, x):
        x = self.reduce_conv(x)
        x = self.conv(x) + x
        x = self.out_conv(x)
        return x

改为

    @autocast(True)
    def forward(self, x):
        x = self.reduce_conv(x)
        x = self.conv(x) + x
        x = self.out_conv(x)
        return x