LiDAR-Fusion Chapter 3 - Train PointPainting Model

In this chapter, we will train the PointPainting model based on the fused Painted Kitti Dataset.

Now, we will change our working directory from painting to detector.

Generate Training Configuration

Here is a script gen_painted_kitti_config.py under the detector directory. It will scan the whole dataset and filter some useless elements. It will also generate some configurations for later training based on the values in tools/cfgs/dataset_configs/painted_kitti_dataset. Just run the following command:

python gen_painted_kitti_config.py tools/cfgs/dataset_configs/painted_kitti_dataset.yaml

It will generate the configurations to some .pkl files under detector/data/kitti/.

Train the Model

Go to detector/tools, and run the command below to train the model:

python train.py --cfg_file cfgs/kitti_models/pointpillar_painted.yaml --epochs 1

Here I set the training epoch to 1 for quickly running through the whole training procedure in order to prevent discovering some bugs when the training is about to finish which may waste a bunch of time.

There are more configurations that you can adjust. For more details, you can run this command to see:

python train.py --help

Multi-GPU Training

If you are using a multi-GPU platform, you can use the following command to accelerate training:

OMP_NUM_THREADS=8 torchrun  --nproc_per_node=4  train.py --cfg_file cfgs/kitti_models/pointpillar_painted.yaml --launcher pytorch

Get the Result

Once finishing the training, you can see the result stored in detector/output.