Skip to content

TRAIN.md

Latest commit

 

History

History
30 lines (23 loc) · 1.52 KB

TRAIN.md

File metadata and controls

30 lines (23 loc) · 1.52 KB

Training on the nuScenes dataset

Before you start, please follow the instructions to prepare the dataset as described here.

You can download some pre-trained weights by running

cd $PCP_MV_DIR
bash ./tools/download_pretrained.sh

You can then try training a baseline model with the script:

cd $PCP_MV_DIR
bash ./scripts/train.sh

This script defaults to local training and you can adjust it for multi-node distributed training.

To use the TransFusionHead, refer to the configurations in the transfusion folder. Set use_density_weights and use_relationship_targets to True to use the proposed representation learning approaches. The following shows an exemplar training configuration file, a training log, and a model checkpoint:

Configuration File Training Log Model Checkpoint
Google Drive Link Google Drive Link Google Drive Link

Training using the configurations of PCP-MV

While the raw data of PCP-MV is not released, you can train on your dataset with the experimental configurations of PCP-MV. The following is a simulation using nuScenes.

cd $PCP_MV_DIR
bash ./scripts/train_pcp_mv.sh