Co-DETR

Co-DETR copied to clipboard

使用自己的训练数据集训练实例分割任务不收敛

3

使用自己的训练数据集，转换成coco2017格式，20个类别，checkpoint使用  ./tools/dist_train.sh /workspace/Co-DETR/projects/configs/co_dino_vit/co_dino_5scale_vit_large_coco_instance.py 4 /workspace/train_workdir 训练100 epoch，loss_mask: 20.5192，推理结果完全不对，除了数据集相关类别标签、数量之外，还有哪里需要修改吗？

gujiacheng

custom_hooks = [ dict( type='ExpMomentumEMAHook', momentum=0.0001, priority=49),] Why did you set EMA only for ViT-L, rather than SwinL?

kasteric

2025-08-07 22:38:53,659] torch.distributed.elastic.multiprocessing.api: [WARNING] Sending process 2235248 closing signal SIGTERM [2025-08-07 22:38:53,659] torch.distributed.elastic.multiprocessing.api: [WARNING] Sending process 2235250 closing signal SIGTERM [2025-08-07 22:38:53,661] torch.distributed.elastic.multiprocessing.api: [WARNING] Sending process 2235251 closing signal SIGTERM...

kasteric

用的Co-detr，转ONNX然后再推理，转化ONNX成功了，但是有些值变成了常量，导致后面推理错误，比如mlvl_feats。这种有什么方法解决吗

SoSoSaMa

Performance gap between this version and Co-Detr in mmdetection v3

3

I wonder why there is a performance gap between Co-detr and its version in mmdetection v3, I have used the same training data to train on both repo, and I...

kasteric

The user-unfriendly implementation of CoDINO head

3

This code makes the inference format of output become a class, rather than a list of tensors, which may cause errors when trying to convert the model to onnx, even...

kasteric

ImportError: cannot import name 'generate_block_target' from 'mmdet.models.losses.cross_entropy_loss' (C:\Users\wt\.conda\envs\CO-DETR\lib\site-packages\mmdet\models\losses\cross_entropy_loss.py)

2

Traceback (most recent call last): File "D:\Pycharm\Co-DETR-main\Co-DETR-main\tools\train.py", line 28, in from projects import * File "D:\Pycharm\Co-DETR-main/Co-DETR-main\projects\__init__.py", line 1, in from .models import * File "D:\Pycharm\Co-DETR-main/Co-DETR-main\projects\models\__init__.py", line 1, in from .co_detr...

tobeyourstar

Hello, when I was loading the co-detr-vit-large-coco-instance.pth model for inference, there was a warning about some unexpected weights, as shown below: It seems that the EMA pretrained weights were not...

bartbuaa

这个版本mmdet2.15的模型能转onnx吗，必须得重新在mmdetection3.0重新训练一遍才能转onnx吗

1

kasteric

inference得到了莫名奇妙的背景框。。

3

我用cascadercnn训练都至少可以区分前景和背景，用dino训练也不会出现莫名奇妙的invalid的框，但是用co-dino大尺度训练，出现了把背景预测为前景的例子，好神奇。。

kasteric