Skip to content

Latest commit

 

History

History
117 lines (84 loc) · 4.75 KB

GET_STARTED.md

File metadata and controls

117 lines (84 loc) · 4.75 KB
Raw

Getting Started on nuScenes

Modified from CenterPoint's original document.

Prepare data

Download data and organise as follows

# For nuScenes Dataset         
└── NUSCENES_DATASET_ROOT
       ├── samples       <-- key frames
       ├── sweeps        <-- frames without annotation
       ├── maps          <-- unused
       ├── v1.0-trainval <-- metadata

Create a symlink to the dataset root

mkdir data && cd data
ln -s DATA_ROOT 
mv DATA_ROOT nuScenes # rename to nuScenes

Remember to change the DATA_ROOT to the actual path in your system.

Create data

Data creation should be under the gpu environment.

# nuScenes
python tools/create_data.py nuscenes_data_prep --root_path=NUSCENES_TRAINVAL_DATASET_ROOT --version="v1.0-trainval" --nsweeps=10

In the end, the data and info files should be organized as follows

# For nuScenes Dataset 
└── LinK/detection/
       └── data    
              └── nuScenes 
                     ├── samples       <-- key frames
                     ├── sweeps        <-- frames without annotation
                     ├── maps          <-- unused
                     |── v1.0-trainval <-- metadata and annotations
                     |── infos_train_10sweeps_withvelo_filter_True.pkl <-- train annotations
                     |── infos_val_10sweeps_withvelo_filter_True.pkl <-- val annotations
                     |── dbinfos_train_10sweeps_withvelo.pkl <-- GT database info files
                     |── gt_database_10sweeps_withvelo <-- GT database

Training

Run the following command to train LinK using 4 GPUs.

python -m torch.distributed.launch --nproc_per_node=4 --master_port 23456 ./tools/train.py \
  configs/nusc/voxelnet/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_elkv3.py \
  --work_dir work_dirs/0075cbgs_cos3x7_group2

The model checkpoints and logs will be saved to the work_dir above.

Fading strategy of GT sampling

We use the fading strategy of GT sampling, i.e., 15 epochs with GT sampling and then 5 epochs without this. To do so, after completing training 15 epochs, we break the training process and resume it by the following steps.

  1. Turn off the GT sampling in config. See configs/nusc/voxelnet/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_elkv3.py, line 92.

  2. Resume the training process by

    python -m torch.distributed.launch --nproc_per_node=4 --master_port 23456 ./tools/train.py \
  configs/nusc/voxelnet/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_elkv3.py \
  --work_dir work_dirs/0075cbgs_cos3x7_group2_15+5 \
  --resume_from work_dirs/0075cbgs_cos3x7_group2/epoch_15.pth

Evaluation

Evaluate on validation set

Run the following command to evaluate LinK on validation set, using 4 GPUs.

python -m torch.distributed.launch --nproc_per_node=4 --master_port 23456 tools/dist_test.py \
  configs/nusc/voxelnet/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_elkv3.py \
  --work_dir work_dirs/0075cbgs_cos3x7_group2_15+5/atEp20 \
  --checkpoint work_dirs/0075cbgs_cos3x7_group2_15+5/epoch_20.pth

The evaluation result and metrics will be saved to work_dirs/0075cbgs_cos3x7_group2_15+5/atEp20/. Change the arguments accordingly for your own model.

Evaluate on test set

Run the following command to evaluate LinK on the test set.

python -m torch.distributed.launch --nproc_per_node=4 --master_port 23456 tools/dist_test.py \
  configs/nusc/voxelnet/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_elkv3.py \
  --work_dir work_dirs/0075cbgs_cos3x7g2_trainval2_15+5/atEp20 \
  --checkpoint work_dirs/0075cbgs_cos3x7g2_trainval2_15+5/epoch_20.pth \
  --testset

This model is trained using the training set and validation set of nuScenes. Change the arguments accordingly for your own model.

The detection result will be saved to work_dirs/0075cbgs_cos3x7g2_trainval2_15+5/atEp20/. To get the metrics of your own model on test set, submit it to the official nuScenes detection server.

Evaluate with flipping and rotation TTA

For the evaluation with flipping and rotation TTA, we provide a bash script fuse_rot_flip_results.sh. This script will evaluate the model with 7 angles (0, ±6.25, ±12.5, ±25), each with 4 kinds of flipping (no flip, x-flip, y-flip, xy-flip). The final detection result is fused by nms_better2.py (ref: https://github.com/dvlab-research/FocalsConv/blob/master/CenterPoint/test_aug_examples/nms_better2.py).

Config the bash shell accordingly for your use case. The final detection result will be saved to {work_dir}/flip_rot_fuse/.

The pretrained models and configurations are in MODEL ZOO.