[paper : https://arxiv.org/abs/1809.10732]

pip install -r requirement.txt

You should download Nuscenes/Fulldataset(v1.0)/Trainval/Metadata (Direct download link is here)

Unzip downloaded meta filt to [DATAROOT] then run:

python gen_dataset.py --dataroot [DATAROOT] --split [SPLIT]

[SPLIT] should be one of 'train' or 'train_val' or 'val'

• 'train' : training dataset : 32k
• 'train_val' : validation dataset : 8k
• 'val' : test dataset :

python train_mtp.py --name [NAME] 

checkpoint of experiment is saved in './exps/[NAME]'

You can add more argument such as optimizer, backbone, and so on (Refer to train_mtp.py)

visdom

run in another window:

python visualize.py --name [NAME] --ep [EP]

It loads saved weight from checkpoint of experiment named [NAME].

[EP] should be 'best' or 'last' or a specific epoch (integer)

• 'best' : w/ lowest validation loss
• 'last' : w/ last epoch
• specific epoch : 1, 10, ...

Generated trajectories and its logits are displayed on visdom server.

You can see the results at '[YOUR_IP_ADDRESS]:8097' (default port of visdom is 8097)