selected index k out of range

[tahirashehzadi]

Question

Hi,

I encountered an error while training the model on a custom dataset:

select_tgt_index = torch.topk(mc_score, self.topk_sa, dim=1)[1]
RuntimeError: selected index k out of range

Additional

No response

[xiuqhou]

Hello @tahirashehzadi
This is because the input image is too small and there are not enough queries to select top_sa. We use (800,1333) as the default input shape, and topk_sa=300. Did you change the input shape and "topk_sa"?

[xiuqhou]

You can try to increase the input shape (see transforms/presets.py), or decrease topk_sa of SalienceTransformerEncoderLayer.

[Kunghaoyu]

topk_sa

我也遇到了这个问题，但是我现在跑的是DOTA那个遥感数据集，里面目标挺多的，不应该是你说的这个情形。训练的时候，我没做裁剪，是不是图像太大了也会出现这种情况？