ablation_models.md

Configurations of Ablation Models

Use the following configurations to train the ablation models. You can add/update these configurations to the experiment configs. The configuration of the traffic simulation policy and the GMM-based ego policy are similar. The experiment config just overrides the configs in the model config

Data Augmentation

• Trajectory perturbation of SMART. (Open-loop)

  token_processor:
    agent_token_sampling:
      num_k: 5
      temp: 1.0
  training_loss:
    rollout_as_gt: true
  

• Noisy tokenization of Trajeglish.
  • Default. (Open-loop)

    token_processor:
      agent_token_sampling:
        num_k: 5
        temp: 1.0
    

  • Sampled from uniform distribution. (Open-loop)

    token_processor:
      agent_token_sampling:
        num_k: 5
        temp: 1e5
    

  • Sampled from policy predicted probability. (Closed-loop)

    training_rollout_sampling:
      criterium: topk_dist_sampled_with_prob
      num_k: 5
      temp: 1.0
    

Closed-Loop Supervised Fine-tuning with Top-K Sampling

• Top-5

  training_rollout_sampling:
    criterium: topk_prob
    num_k: 5
    temp: 1.0
  

• Top-5 + distance based filtering

  training_rollout_sampling:
    criterium: topk_prob
    num_k: 5
    temp: 1.0
  training_loss:
    gt_thresh_scale_length: 1.0
  

• Top-5 + distance based sampling with super low temperature is equivalent to CAT-5, and that's exactly how we implemented CAT-K rollout.

  training_rollout_sampling:
    criterium: topk_prob_sampled_with_dist
    num_k: 5
    temp: 1e-5