tomevault-io Project Q&A
Answer the user's question about this project. Use the general lookup order first, then the topic-specific maps below for deeper dives.
General lookup order
Read these in order:
.claude/docs/project_structure.md— directory layout and component overview.README.md— high-level project description, setup, and usage.DEVELOPING.md— development workflow, testing, and contribution guide.docs/source/— detailed Sphinx documentation pages.
Topic: Data generation / collection
- slurm/data_collection/collect_data.py — SLURM orchestrator launching parallel CARLA data collection jobs.
- slurm/data_collection/delete_failed_routes.py — cleans up failed route directories.
- 3rd_party/leaderboard_autopilot/leaderboard/leaderboard_evaluator_local.py — runs CARLA scenarios per route.
- lead/main.py — adapts LEAD to the CARLA leaderboard agent protocol.
- lead/expert/expert.py — main expert agent driving logic.
- lead/expert/expert_base.py — shared base class for expert variants.
- lead/expert/expert_data.py — saves frames and labels during collection.
- lead/expert/config_expert.py — expert configuration dataclass.
- lead/expert/privileged_route_planner.py — privileged route/lane planning.
Topic: Data preparation
- lead/data_loader/carla_dataset.py — CARLA Leaderboard 2.0 PyTorch dataset.
- lead/data_loader/carla_dataset_utils.py — CARLA loading and preprocessing helpers.
- lead/data_loader/training_cache.py — compressed on-disk training cache.
- lead/data_loader/navsim_dataset.py — NavSim PyTorch dataset.
- lead/data_loader/waymo_e2e_dataset.py — Waymo Open Dataset end-to-end dataset.
- lead/data_buckets/abstract_bucket_collection.py — ABC for bucket collections (serialization, loading).
- lead/data_buckets/bucket.py — single bucket holding paths and metadata.
- lead/data_buckets/route_filtering.py — route validity and deduplication.
- scripts/build_cache.py — entry point to build the training cache.
- scripts/build_buckets_pretrain.py — generate pretraining bucket collections.
Topic: Training
- lead/training/train.py — main DDP-aware training entry point (Trainer class).
- lead/training/config_training.py — training hyperparameters dataclass.
- lead/training/training_utils.py — seeding, checkpointing, data loader construction.
- lead/training/mixed_training_utils.py — multi-dataset sample scheduling.
- lead/training/logger.py — WandB metric and media logging.
- scripts/pretrain_ddp.sh — DDP pretraining launcher.
- scripts/posttrain_ddp.sh — DDP posttraining launcher.
- slurm/experiments/ — per-experiment SLURM scripts (see
/prepare-slurm-scripts).
Topic: Model architecture (TransFuser v6)
- lead/tfv6/tfv6.py — top-level model orchestrating backbone and decoder heads, loss aggregation.
- lead/tfv6/transfuser_backbone.py — image + LiDAR fusion backbone.
- lead/tfv6/transfuser_utils.py — backbone utility layers and functions.
- lead/tfv6/planning_decoder.py — waypoint, path and target speed planning head.
- lead/tfv6/bev_decoder.py — BEV semantic segmentation head.
- lead/tfv6/perspective_decoder.py — perspective view output head.
- lead/tfv6/center_net_decoder.py — CenterNet 2D bounding box detector.
- lead/tfv6/radar_detector.py — 2D radar object detector.
Topic: Evaluation (bench2drive, longest6, town13)
- slurm/evaluation/evaluate.py — core orchestrator: submit, monitor, retry SLURM eval jobs.
- slurm/evaluation/evaluate_scripts_generator.py — generates per-route shell scripts.
- slurm/evaluation/evaluate_utils.py — JSON parsing, status checks.
- slurm/evaluation/merge_route_json.py — merges per-route results into one summary.
- slurm/evaluation/evaluate_wandb_logger.py — logs aggregated eval results to WandB.
- lead/inference/closed_loop_inference.py — CARLA closed-loop evaluation with PID control.
- lead/inference/open_loop_inference.py — open-loop evaluation (NavSim / Bench2Drive).
- lead/inference/sensor_agent.py — sensor-based agent wrapper for the CARLA leaderboard protocol.
- lead/inference/infraction_recorder.py — records and serializes driving infractions.
- lead/inference/config_closed_loop.py — closed-loop eval config.
Rules
- Read in order. Start with general docs, then dive into topic-specific files.
- Be specific. Point the user to the exact file, section, or line.
- Don't guess. If none of the sources answer the question, say so and suggest where else to look.
Source: kesai-labs/lead — distributed by TomeVault.