Skip to content

train_test.md

Latest commit

 

History

History
57 lines (42 loc) · 2.16 KB

train_test.md

File metadata and controls

57 lines (42 loc) · 2.16 KB

Train & Test

Data Preparation

The same as VAD data format.

Trajectory Clustering

Refer to tools/traj_clustering.py, which will cluster all trajectories to 4096 classes. The following component will be saved:

  • Cluster centroids: saved in npy format
  • MiniBatchKMeans sklearn model: saved in pkl format
  • Visualization as follows:

traj cluster

Train script

Assume run on 2-GPUs machine. RTX4090 or higher is recommended.

torchrun --nproc_per_node=2 \
    --master_port=28510 \
    adzoo/vad/train.py \
    adzoo/vad/configs/VAD/VADv2_voca4096_config.py \
    --launcher=pytorch \
    --deterministic

Notes For Config File

Here we use VADv2_voca4096_config.py as example to show the most frequently-used config

  • num_cams: camera config, 8 for nuplan and 6 for nuscenes.
  • class_names: detection classes, here we combine the objects in both nusc and nuplan dataset.
  • col_class: objects that involved in collision penalty loss.
  • map_classes: layers in maps (.JSON). For example ['divider', 'ped_crossing', 'boundary'] for nuplan; ['lane_divider', 'road_edge', 'crosswalk', 'centerline'] for nuscenes.
  • plan_fut_mode: number of clusters for VADv2.
  • plan_anchors_path: path to store clusters. Note: Different numpy version will cause issue loading/saving the .npy file. The example use numpy==1.24.1 for cluster_centers_ori. -data_root: root path for dataset. -ann_file_train/ann_file_test: path to annotation files, usually in .pkl format. Essentially is a list of dict where each dict is meta infos of single frame. -ann_file_map: check Tools section in README.md for Convert Nuplan Map to Json Tools.
  • load_from and resume_from: if to load pre-trained models. See documents for mmcv.

Test script

torchrun --nproc_per_node=1 --master_port=28512 \
    ./adzoo/vad/test.py \
    ./adzoo/vad/configs/VAD/VADv2_voca4096_config.py \
    path-2-your-model \
    --launcher=pytorch \
    --eval=bbox