train data

[wuhen777]

          > For most simple case, hard code on mmdet3d. https://github.com/open-mmlab/mmdetection3d/blob/f1107977dfd26155fc1f83779ee6535d2468f449/mmdet3d/datasets/custom_3d.py#L64C12-L64C12 for example, sub sample to 1/5. From self.data_infos = self.load_annotations(self.ann_file) to self.data_infos = self.load_annotations(self.ann_file)[::5]

For more reasonable way, check here:

https://github.com/fundamentalvision/BEVFormer/blob/32b65ffd2a85c35b5f74b753f23ef34f9a85a576/projects/mmdet3d_plugin/datasets/nuscenes_dataset_v2.py#L18

add line: self.data_infos = self.data_infos[::5]

maybe modify more variable

Or, construct a new dataset called nuscenes-subset

Thank you for your suggestion. Now, when slicing the mini dataset, I will change the 64th line of BEVFormer/mmdeletion3d/mmdet3d/datasets/custom_1d. py from # self. data_infos=self. loaduannotations (self. annunfile) to After self. data_infos=self. loadannotations (self. annunfile) [:: 5], the training time remains the same as before it was modified. Meanwhile, if I am in line 26 of BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset. py, that is, super()__ After init__ (* args, * * kwargs), add self. data_infos=self. data_infos [:: 5], and it is found that the training time has not changed. May I ask what is the reason for this Looking forward to your answer

Originally posted by @wuhen777 in https://github.com/fundamentalvision/BEVFormer/issues/209#issuecomment-2102341794