训练报错

[ycdhqzhiai]

1.torch.arange应该是在CPU上的，下面一行to(cuda)直接报错了 https://github.com/dragonfly606/MonoCD/blob/d3ec3455838fe2035b0babf90f779e0b44922fa7/model/head/detector_loss.py#L161 2.batch为4，训练过程中N=0 https://github.com/dragonfly606/MonoCD/blob/d3ec3455838fe2035b0babf90f779e0b44922fa7/model/anno_encoder.py#L120

不知道是什么原因，难道是我版本不一致？torch1.13

[ycdhqzhiai]

(https://github.com/dragonfly606/MonoCD/blob/d3ec3455838fe2035b0babf90f779e0b44922fa7/data/datasets/kitti.py#L687) 这个地方有个问题，只有在box有效的时候，reg_mask才为true，如果某个batch内，所有样本都没有box，会导致reg_mask全部为0，这样在计算loss时候，就会出现上面N=0的情况， https://github.com/dragonfly606/MonoCD/issues/1 这个里面问题也应该是reg_mask出现全为0 batchsize越小，出现概率越大

[dragonfly606]

@ycdhqzhiai , I'm late. Thank you very much for your attention to our work and your suggestions. You are right. I have added the following code to avoid training abort. https://github.com/dragonfly606/MonoCD/blob/d44ea9c664e932be6bc4c1e1de25daf5c583c0ed/model/head/detector_loss.py#L408-L410