Zhuofan Zong
Zhuofan Zong
Yes, the 5-scale model consumes much more memory than 4-scale
> I use projects/configs/co_dino/co_dino_5scale_swin_large_16e_o365tococo.py, and it seems if I freeze the backbone and set the checkpoint to False, it will OOM in a 24G A30 Co-DETR with frozen SwinL and...
It seems you install the CPU version of mmcv-full. But the code `model = init_detector(config_file, checkpoint_file, device='cuda:0')` requires GPU for inference.
Please try `pip install -e .` in the co-detr repo.
把encoder输出的不同尺度的feature抽出来，插值到最大特征的维度再concat
https://github.com/Sense-X/Co-DETR/blob/d01ef8b657cc3711c59759cf863fe5ab9aa50e76/projects/models/co_deformable_detr_head.py#L160 这里是从encoder输出中抽出不同scale特征的代码，后面把每个scale的特征尺寸都插值到`mlvl_feats[0]`的尺寸就可以
Some operations (e.g., deformable attention) in the model do not support model summary.
The total FLOPs of Co-DINO-5scale is the same as DINO-5scale since our training technique does not introduce additional inference costs. The FLOPs of the DINO model is presented in Table...
The current repository does not support model export. Please refer to this issue https://github.com/Sense-X/Co-DETR/issues/26 for more details about export.
pretrain: `eva02_L_pt_m38m_medft_in21k_ft_in1k_p14` The model is [Co-DINO](https://github.com/Sense-X/Co-DETR/blob/main/projects/configs/co_dino/co_dino_5scale_swin_large_16e_o365tococo.py) and more details about this large model are presented in the paper appendix.