Samples in split doesn't match samples in predictions

[jiumozhi123]

Hi junjie: thanks for your open-source code Now I train the cmt_voxel0100_r50_800x320_cbgs.py. The datasets have been initialized. My hardware is A6000 * 4. samples_per_gpu and workers_per_gpu is 8. When the second epoll completed and evaluation start, the following error occur the nuscenes datasets is downloaded from official website. I wish to know why it was wrong? Thanks a lot!

[junjie18]

Hi, Thanks for reminding us. That's because I usually evaluate CMT with 1/10 datasets in model training. The problem can be simply solved by modifying load_interval back to 1 in the config.

[vetsasai]

Can I not do the evaluation with load_interval set to 10? Is there a config setting to do such evaluation using below command ?

bash tools/dist_test.sh /path_to_your_config /path_to_your_pth 2 --eval bbox