Zhuofan Zong comments

Results 111 comments of


                                            Zhuofan Zong

Training on custom dataset

You can set `use_checkpoint=True` instead of `with_cp=True`. For Swin Transformer, we do not use the mmdet version and follow the original implementation (https://github.com/SwinTransformer/Swin-Transformer-Object-Detection).

Could you provide config file for O365 pretrained model for inference?

I will update the O365 model configs for testing soon. However, I'm currently handling other higher-priority tasks and will push the update as quickly as possible.

How to train on custom dataset and what is the structure of annotation will be prepred?

The data structure is consistent with the COCO-style annotation. Please refer to the [document](https://github.com/Sense-X/Co-DETR/blob/main/docs/en/2_new_data_model.md) for more details.

使用co_dino_5scale_swin_large_3x_coco配置训练不收敛

如果用mmdetection里的faster-rcnn来训会收敛吗。如果还是不收敛，那要么是你的十万多数据噪音比较大，要么是自定义数据集那部分不对

HOW TO IMPLEMENT ALBUMENTATION ?

Sorry for the late reply, but I have never tried the albumentation in this codebase.

Model on Huggingface

Thanks for your suggestion. However, my focus is currently on other projects, and I don't have much time to develop new features for this project right now. Community PR is...

模型并行

暂时不支持

请问tta如何实现

我们使用的是multi-scale testing+box voting

How much VRAM is needed to finetune co_dino_5scale_vit_large_coco?

We use 56 A100 80G GPUs to pretrain the model. FSDP and deepspeed can help you reduce training memory consumption, please refer to this [issue](https://github.com/Sense-X/Co-DETR/issues/121).

5 RTX4090s are still not enough for this project?

1. Please show me your training config. 2. This repo does not support fp16 training.