How can we train using both the train and validation datasets to prepare for test dataset submission?

[qifeng22]

Thank you for your excellent work. Could you share the method for training on the entire dataset (train and validation) for test dataset submission in Waymo and nuScenes?
@happinesslz

[AlmoonYsl]

Thank you for your excellent work. Could you share the method for training on the entire dataset (train and validation) for test dataset submission in Waymo and nuScenes? @happinesslz

Hi, you should combine training and validation set for generating gt database and info.

[qifeng22]

The meaning is that I should move the validation folder to train folder， and then run this command:

python -m pcdet.datasets.waymo.waymo_dataset --func create_waymo_infos \
    --cfg_file tools/cfgs/dataset_configs/waymo_dataset.yaml

or

python -m pcdet.datasets.nuscenes.nuscenes_dataset --func create_nuscenes_infos \
    --cfg_file tools/cfgs/dataset_configs/nuscenes_dataset.yaml \
    --version v1.0-trainval

[qifeng22]

If I train all the dataset, how to generate the test result for the evaluation website?

[AlmoonYsl]

If I train all the dataset, how to generate the test result for the evaluation website?

For waymo, you need to inference the test set and convert the prediction to binary file like gt.bin.
For nuScenes, you need to inference the test set and generate the json file.

[qifeng22]

Can you provide the command? Maybe I dont konw how to do?

[AlmoonYsl]

Can you provide the command? Maybe I dont konw how to do?

For Waymo, we develop the generation tool for converting the prediction to binary file. But we do not clean the code. You can refer to the MMDetection3D codebase.
For nuScenes, you can directly eval the model by the test script for generating the json file in the output directory.

[qifeng22]

okok, thanks