Junru Gu

The testing module loads the dataset [here](https://github.com/Tsinghua-MARS-Lab/DenseTNT/blob/294596b0b141fb7b18248889e5c409630c5f7ed6/src/do_eval.py#L60). You need to modify the [dataset](https://github.com/Tsinghua-MARS-Lab/DenseTNT/blob/294596b0b141fb7b18248889e5c409630c5f7ed6/src/dataset_argoverse.py#L695) to your own format.

Map information is converted into vectors during training and inference. You need to modify [map processing](https://github.com/Tsinghua-MARS-Lab/DenseTNT/blob/294596b0b141fb7b18248889e5c409630c5f7ed6/src/dataset_argoverse.py#L43) to support your own map format.

This is FDE, not minFDE

``` OUTPUT_DIR=argoverse2.densetnt.1; \ GPU_NUM=2; \ python src/run.py --argoverse --argoverse2 --future_frame_num 60 \ --do_train --data_dir data/train/ --output_dir ${OUTPUT_DIR} \ --hidden_size 128 --train_batch_size 64 --use_map \ --core_num 16 --use_centerline --distributed_training ${GPU_NUM} \...

GPU memory is less than 10 GB. 65G RAM is required for 2 GPU training, 230G RAM is required for 8 GPU training.

Thanks for pointing out. The online mode has been released now.

1. Thanks for pointing it out! 2. This produces the same results as placing in the end of the loop 3. Why "1"? "Argmin" indicates the closest predicted trajectory. 4....

We found that these two examples `02ecafff-012c-45f8-bb6e-5ace2c6e3d88` and `014c02b0-3315-4faf-aca5-b239f5b7f0ca` are both 110 frames.

What about using gloo as backend? It may be due to a NCCL error.

The 'pass' does not take any effect here. What is the value of `'complete_traj' in args.other_params`, is it `True`?