Train lidar_apollo_instance_segmentation CNN with Nuscenes.
Only supports python3.
- Clone original_model branch and install some packages.
git clone --recursive https://github.com/kosuke55/train_baiducnn.git
pip install -r requirements.txt
mkdir build
cd build
cmake ..
make -j
- create_dataset_from_nusc.py is for creating a dataset to train apollo cnn. Set dataroot and save_dir.
cd ~/train_baiducnn/scripts/create_dataset
python create_dataset_from_nusc.py --dataroot <downloaded nuscenes path> --save_dir <dir to save created dataset> --nusc_version <v1.0-mini or v1.0-trainval>
- Execute start_server.sh and access from a web browser. Then you can train with train_bcnn.py.
cd ~/train_baiducnn/scripts/pytorch
./start_server.sh
python train_bcnn.py --data_path <dir to save created dataset>
- Trained model can be converted to onnx by pytorch2onnx.py and converted to engine by onnx-tensorrt.
cd ~/train_baiducnn/scripts/pytorch
python pytorch2onnx --trained_model <your_trained_model.pt>
# after installing onnx-tensorrt
onnx2trt <your_trained_model.onnx> -o <your_trained_model.engine>
- Run lidar_apollo_instance_segmentation with <your_trained_model.engine>
apollo 3D Obstacle Percption description