Scene Generation

Training

The primary model training codebase is under lidardm/core/.

Training runs are launched with scripts/train.py.

The specific datasets, loss functions, and models that are used are specified with experiment configs lidardm/core/configs/experiment/.

The codebase supports multinode and multigpu training by passing in the argument ++trainer.devices=NUM_DEVICES ++trainer.nodes=NUM_NODES.

Conditional Generation (Waymo)

The Waymo pipeline has 3 learnable components: a Map VAE, the Waymo field VAE, and a diffusion model that generates Waymo Fields based on a Map condition.

• Waymo Map VAE: python scripts/train.py +experiment=map_vae_waymo
• Waymo VAE: python scripts/train.py +experiment=wf_s_vae
• Waymo Diffusion Model: python scripts/train.py +experiment=wf_s_unetc

Unconditional Generation (KITTI-360)

• KITTI-360 VAE: python scripts/train.py +experiment=kf_s_vae
• KITTI-360 Diffusion Model: python scripts/train.py +experiment=kf_s_unet