Zhuofan Zong comments

Results 111 comments of


                                            Zhuofan Zong

Can‘t find pth file on Google Drive

The checkpoints trained with 2x schedule are not released.

作者您好，训练过程可视化怎么实现呢？

MMDet支持tensorboard，在你的config里加一段: ``` log_config = dict( interval=50, hooks=[ dict(type='TextLoggerHook'), dict(type='TensorboardLoggerHook') ]) ```

Co-DETR in mmdetection projects

The implementation in the original mmdet repo is developed by the mmdet team.

Co-DETR in mmdetection projects

Most of my time is occupied by other projects, and I don't have much time to integrate this project into the latest mmdetection repository.

Question about the sequence of decoder operations

We follow the decoder design of Deformable-DETR. In Deformable-DETR, the correct decoder operation order is: self_attn -> cross_attn -> ffn. Specifically, self_attn is implemented by the MultiheadSelfAttention and cross_attn is...

custom dataset

Please follow the [tutorial](https://github.com/Sense-X/Co-DETR/blob/main/docs/en/2_new_data_model.md) to prepare your data format.

The model and loaded state dict do not match exactly

These mismatched weights are the classifier branches. This is normal for the fine-tuning setting and does not affect the performance.

Test-Time augmentation

It's not supported yet.

Ensure Polygon Mask Output Instead of Bitmap in ViT-L Instance Segmentation

The `bitmap_to_polygon` [function](https://github.com/Sense-X/Co-DETR/blob/26653521e47157fdd819e764b4bb949057ea9f3b/mmdet/core/mask/structures.py#L1075) can help you convert the output bitmap mask to polygon mask.