Train on a custom dataset in 1.x branch (and other missing functionality in the 1.x branch)

[gfreund123]

What's the feature?

Hey, I recently started using the 1.x branch for RTMDet, but there seems to be lots of missing functionality.

How can I train on a custom dataset when i cant register DATASETS or PIPELINES with mmrotate.datasets.builder? More generally there seems to be many things i now cant use (e.g. from mmcv import Config, mmdet.models.build_detector and more).

So how should one go about editing the models config and training on a custom dataset?

The documentation doesnt seem helpful. BTW in the 1.x docs here they use from mmrotate.models.builder import ROTATED_BACKBONES which also does not exist in the branch?!. I'm a little confused.

Advice would be appreciated.

I followed the intallation here Version are: mmcv '2.0.0rc4' mmrotate '1.0.0rc1' mmdet '3.0.0rc6'

Any other context?

No response

[zytx121]

Hi @gfreund123! I'm sorry for the bad experience.

• First, you need to convert the custom dataset into DOTA format. Finally, then modify the category of the dataset through metainfo like https://github.com/open-mmlab/mmrotate/blob/5d0491c826ad52b7490a7c1146fb65a7d78e2347/configs/_base_/datasets/hrsid.py#L45 https://github.com/open-mmlab/mmrotate/blob/5d0491c826ad52b7490a7c1146fb65a7d78e2347/configs/_base_/datasets/hrsid.py#L55
• The two functions you mentioned have been relocated or removed in version 1. x.
• Due to staffing problems, the 1. x document has not been updated. We will update it ASAP.