trj.md

HiVT

Setup

Please follow HiVT setup guide to set up environment.

Note that different maps have different encoder configurations for optimal perforamance

Training

cd HiVT_modified

# if method is 'bev', choose [MapTR, MapTRv2, MapTRv2_CL, StreamMapNet]
python train.py \
  --root ../trj_data/{maptr,maptrv2,maptrv2_cent,stream} \
  --method {base, unc, bev} \
  --map_model {MapTR, MapTRv2, MapTRv2_CL, StreamMapNet} \
  --embed_dim 128

For training MapTRv2 Centerline, add an --centerline argument.

Testing

cd HiVT_modified

# if method is 'bev', choose [MapTR, MapTRv2, MapTRv2_CL, StreamMapNet]
python eval.py \
  --root ../trj_data/{maptr, maptrv2, maptrv2_cent, stream} \
  --split {mini_val, val} \
  --method {base, unc, bev} \
  --map_model {MapTR, MapTRv2, MapTRv2_CL, StreamMapNet} \
  --batch_size 32 \
  --ckpt_path /path/to/your_checkpoint.ckpt

For evaluating MapTRv2 Centerline, add an --centerline argument.

Visualization

Please uncomment this line to save the pkl files necessary for later visualization.

DenseTNT

Setup

Please follow DenseTNT setup guide to set up environment.

Training

You can use the src/train.sh or by running the following:

cd DenseTNT_modified

# Please adjust the hyperparemters based on the paper's Appendix
epochs=12
batch=1 
lr=0.00015
wd=0.05
dropout=0.2 
output_dir=/MapBEVPrediction/DenseTNT_modified/models/maptr_al
train_dir=/MapBEVPrediction/trj_data/maptr/train/data/
val_dir=/MapBEVPrediction/trj_data/maptr/val/data/

python src/run.py \
  --method {base_unc, maptr_bev, stream_bev} \
  --nuscenes \
  --argoverse \
  --argoverse2 \
  --future_frame_num 30 \
  --do_train \
  --data_dir $train_dir \
  --data_dir_for_val $val_dir \
  --output_dir $output_dir \
  --hidden_size 128 \
  --train_batch_size $batch \
  --use_map \
  --core_num 16 \
  --use_centerline \
  --distributed_training 0 \
  --other_params semantic_lane direction l1_loss goals_2D enhance_global_graph subdivide goal_scoring laneGCN point_sub_graph lane_scoring complete_traj complete_traj-3 \
  --eval_params optimization MRminFDE=0.0 cnt_sample=9 opti_time=0.1 \
  --learning_rate $lr \
  --weight_decay $wd \
  --hidden_dropout_prob $dropout \
  --num_train_epochs $epochs

Testing

You can use src/eval.sh or by running the following:

cd DenseTNT_modified

# Please adjust the hyperparemters based on the paper's Appendix
epochs=12
batch=1 
lr=0.00015
wd=0.05
dropout=0.2 
output_dir=/MapBEVPrediction/DenseTNT_modified/models/maptr_al
train_dir=/MapBEVPrediction/trj_data/maptr/train/data/
val_dir=/MapBEVPrediction/trj_data/maptr/val/data/

CUDA_LAUNCH_BLOCKING=1
i=12    # Or any checkpoint number you want to evaluate

python src/run.py \
  --method base_unc \
  --nuscenes \
  --argoverse \
  --argoverse2 \
  --future_frame_num 30 \
  --do_eval \
  --data_dir $train_dir \
  --data_dir_for_val $val_dir \
  --output_dir $output_dir \
  --hidden_size 128 \
  --train_batch_size $batch \
  --eval_batch_size 16 \
  --use_map \
  --core_num 16 \
  --use_centerline \
  --distributed_training 0 \
  --other_params semantic_lane direction l1_loss goals_2D enhance_global_graph subdivide goal_scoring laneGCN point_sub_graph lane_scoring complete_traj complete_traj-3 \
  --eval_params optimization MRminFDE=0.0 cnt_sample=9 opti_time=0.1 \
  --learning_rate $lr \
  --weight_decay $wd \
  --hidden_dropout_prob $dropout \
  --model_recover_path $i