how to train model by using my data

[wang-kangkang]

Can you write a tutorial on how to organize my data? I want to add my data to the training, but I don't know how to do it

[flymin]

Our current implementation relies heavily on the nuScenes format. You can refer to the readme for the data format.

To use other data, our solution is to transfer the format and organize it like nuScenes. Otherwise, it may not be trivial to reuse the dataloader.

[wang-kangkang]

Our current implementation relies heavily on the nuScenes format. You can refer to the readme for the data format.

To use other data, our solution is to transfer the format and organize it like nuScenes. Otherwise, it may not be trivial to reuse the dataloader.

Can you explain how MagicDrive organizes labels? I think the prepare_data script has modifying the original label format of nuscenes. If I can quickly understand the label format of magicdrive with your help, I think I can train my data faster

[flymin]

There are too many details to cover. Please understand that I cannot type all of them here. Another resource you can use is the pre-processed metadata (pkl files from mmdet). However, the original format is still needed, especially for the BEV maps.

You may also run the inference code and check the model inputs from this line.