Closed nhw649 closed 9 months ago
if not self.mixed_selection:
query_pos, query = torch.split(pos_trans_out, c, dim=2)
else:
# query_pos here is the content embed for deformable DETR
query = query_embed.unsqueeze(0).expand(bs, -1, -1)
query_pos, _ = torch.split(pos_trans_out, c, dim=2)
是DINO这篇工作里提到的一个涨点技巧,可以参考一下:
是DINO这篇工作里提到的一个涨点技巧,可以参考一下:
- DINO arXiv: https://arxiv.org/abs/2203.03605
- DINO project in detrex: https://github.com/IDEA-Research/detrex/tree/main/projects/dino
好的,谢谢。
作者你好,h_deformable_detr中使用的mixed_selection是哪篇文章提出的呢?是Conditional DETR for Fast Training Convergence这篇的idea还是一个提点的trick呢?