vita-epfl / unitraj Goto Github PK

Hi,

Thank you for the amazing work! I have a question regarding training a trajectory prediction model using the waymo dataset. I followed the steps here to download the dataset and ran python -m scenarionet.convert_waymo -d UniTraj/unitraj/data_samples/waymo --raw_data_path ../dataset/waymo/training_20s/ --num_workers 64. Below is the result after the data processing code is finished.

INFO:scenarionet/scenarionet/converter/utils.py:Worker 63 finished! Files are saved at: UniTraj/unitraj/data_samples/waymo/waymo_63
Merge Data: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [00:03<00:00, 16.60it/s]
Filter Scenarios: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 70541/70541 [00:00<00:00, 2440651.33it/s]
INFO:root:
 ================ Dataset Summary and Mapping are saved at: UniTraj/unitraj/data_samples/waymo/dataset_summary.pkl ================

Now that the processed waymo dataset has a the same structure as the nuscenes folder under the data_samples folder. However, when I run python train.py (after changing the dataset path in configs/config.yaml), I get the following error and warning for every file in the waymo folder:

Error: 'NoneType' object is not iterable in sd_waymo_v1.2_470050f61839dd13.pkl
Warning: no center objects at time step 20, scene_id=c9281284c99138ef
Error: 'NoneType' object is not iterable in sd_waymo_v1.2_c9281284c99138ef.pkl

which means that we are not getting any training trajectories. I was wondering if I have missed any steps to run training on the waymo dataset. Any help is appreaciated, thank you!

Also, are there any plans to release some model checkpoints soon?

Dear author:
Sorry to disturb you. You have show me the new link of your work. In your paper, you evaluate the result on Nuscenes of the model trained on the Nuscenes dataset--"3.36" for brier-minFDE in the following picture. Is the evaluation dataset same as training dataset? To solve my confusion, I have checked the config file in your project.I saw the parameters "train_data_path" and "val_data_path" are the same, set "data_samples/nuscenes".I want to ask if it is correct?Thanks!

Thanks for your work, I am wondering how to use argoverse data to train.

What is the method to start training with multiple graphics cards? Should I modify the devices: [ 0 ] parameter in config.yaml and directly specify multiple graphics card numbers?

Or can I use the following script to start?

‘’
python -m torch.distributed.launch --nproc_per_node=${NGPUS} --rdzv_endpoint=localhost:${PORT} train.py --launcher pytorch ${PY_ARGS}
‘’