Tips for hyper-parameter tuning

[bmankirLinker]

I'm recently trying to train the model for NuScenes dataset after switching the structure to the KITTI. Are there any tips for FG threshold, bin size, etc. hyperparameters tweaking for smaller (pedestrian, traffic cone) and bigger objects (bus, truck), or just have to validate the optimal parameters by trial-error? Recently, I only update the CLS_MEAN_SIZE according to the target class but the results aren't satisfying.

[nuoloveheng]

Hi I want also train this model on my own dataset， but i don‘t know how to write a yaml file. Can you share your yaml file and Dataloader to me？ My Email: [email protected]

[bmankirLinker]

@nuoloveheng You can simply follow the KITTI format in your data and update the class name in YAML file and Imagesets sequences accordingly. That's all I can share, also forNuScenes you can follow the dev-kit instructions for KITTI transformation.

[muzi2045]

@bmankirLinker
It's difficult to directly train the RPN layer with original Nuscenes dataset structure?
If I want to do that , which part of the repo shuold be modifiled.
thanks for any advice.