Tao Train Bevfusion

BEVFusion

Standalone install? If this session was not initialized by the TAO skill bank plugin, run the tao-setup skill first (host preflight, credentials, cross-skill discovery).

BEVFusion for multi-sensor 3D object detection. Fuses LiDAR point clouds and camera images in bird's-eye-view (BEV) space. Used in autonomous driving for robust 3D perception.

Set pretrained backbone paths for Swin image backbone.

BEVFusion requires the BEVFusion-specific TAO container nvcr.io/nvidia/tao/tao-toolkit:5.5.0-pyt. <!-- unpinned: BEVFusion-only container --> Shared TAO PyTorch 7.x images do not package mmdet3d and fail before any BEVFusion action can parse its spec. The model-skill action is named dataset_convert, but the 5.5 container CLI subtask is bevfusion convert -e <spec>.

Dataclass Schemas

Generated TAO Core schemas are packaged in schemas/<action>.schema.json, with schemas/manifest.json listing available actions. Each generated schema also emits references/spec_template_<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skill_info.yaml via automl_enabled. Runnable AutoML for an action requires schemas/<action>.schema.json and references/spec_template_<action>.yaml to exist and parse. Use the packaged selected-action schema for automl_default_parameters, automl_disabled_parameters, defaults, min/max bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect ~/tao-core at runtime; maintainers regenerate schemas/templates before packaging the skill bank.

Train Action Policy

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read references/skill_info.yaml and resolve the run override from either an explicit automl_policy value or the user's workflow request. Use automl_policy: on by default and only expose on / off in new launch prompts. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as automl_policy: off for this run only. When automl_policy: on, automl_enabled: true, and both schemas/train.schema.json and references/spec_template_train.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skill_dir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automl_policy. Use direct model training only when automl_policy: off or the packaged train schema/template is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.

Non-train actions such as evaluate, inference, export, and deploy flows stay in this model skill. The per-run automl_policy override does not change model metadata.

Training Requirements

• Dataset type: bevfusion
• Formats: default
• Monitoring metric: AP11

Per-Action Dataset Requirements

Typical Spec Overrides

Data source overrides are mandatory for every action — the agent MUST construct data source paths from the Per-Action Dataset Requirements table above and include them in spec_overrides.

DATA_ROOT = "/path/to/kitti_root"
CONVERTED = DATA_ROOT  # BEVFusion 5.5 writes info pickles into root_dir.
DATA_PREFIX = {"pts": "training/velodyne_reduced", "img": "training/image_2"}

dataset_convert (mandatory data sources):

{
    "root_dir": DATA_ROOT,
    "results_dir": DATA_ROOT,
    "mode": "training",
}

train (mandatory data sources):

{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.root_dir": DATA_ROOT,
    "dataset.train_dataset": {"ann_file": f"{CONVERTED}/kitti_person_infos_train.pkl", "data_prefix": DATA_PREFIX},
    "dataset.val_dataset": {"ann_file": f"{CONVERTED}/kitti_person_infos_val.pkl", "data_prefix": DATA_PREFIX},
    "dataset.test_dataset": {"ann_file": f"{CONVERTED}/kitti_person_infos_val.pkl", "data_prefix": DATA_PREFIX},
}

evaluate (mandatory data sources):

{
    "dataset.root_dir": DATA_ROOT,
    "dataset.test_dataset": {"ann_file": f"{CONVERTED}/kitti_person_infos_val.pkl", "data_prefix": DATA_PREFIX},
}

inference (mandatory data sources):

{
    "dataset.root_dir": DATA_ROOT,
    "dataset.test_dataset": {"ann_file": f"{CONVERTED}/kitti_person_infos_val.pkl", "data_prefix": DATA_PREFIX},
}

Eval Dataset

Optional. Val dataset split is configured via ann_file in dataset config.

Important Parameters

• dataset.classes: List of detection classes. Default ["person"]. Must match the annotation categories.
• dataset.type: Dataset type. Options: KittiPersonDataset, TAO3DSyntheticDataset, TAO3DDataset.
• dataset.root_dir: Root directory of the KITTI-style dataset.
• dataset.box_type_3d: 3D box coordinate frame. Options: lidar, camera. Default lidar.
• train.optimizer.lr: Learning rate. Default 2e-4 (AdamW). Use AmpOptimWrapper for mixed precision via optimizer.wrapper_type.
• input_modality: Dict controlling sensor modalities. Keys: use_lidar (True), use_camera (True), use_radar (False), use_map (False).
• model.img_backbone: Image backbone. Default mmdet.SwinTransformer (Swin-Tiny). embed_dims=96, depths=[2,2,6,2].
• model.view_transform.type: View transform for BEV projection. Options: DepthLSSTransform, LSSTransform. Default DepthLSSTransform.
• model.point_cloud_range: Spatial extent of LiDAR. Default [0,-40,-3,70.4,40,1].
• model.voxel_size: Voxel dimensions. Default [0.05, 0.05, 0.1].
• dataset.train_dataset.batch_size: Per-GPU batch size. Default 4.

Multi-GPU / Multi-Node

Launch method: torchrun (LIGHTNING_EXCLUDED_NETWORK). The entrypoint runs torchrun --nnodes=N --nproc-per-node=M train.py, NOT plain python.

• CUDA_VISIBLE_DEVICES is explicitly set from TAO_VISIBLE_DEVICES
• BEVFusion uses mmdet3d-based distributed training, not Lightning DDP
• NODE_RANK is copied to RANK if RANK is unset

Multi-node env vars (set by orchestrator):

Hardware

One GPU is supported for minimal smoke/AutoML validation with a small dataset and batch size. Use 2+ GPUs for ordinary training and 4 GPUs when practical. 24GB+ (A100 recommended) VRAM per GPU. BEVFusion is memory-intensive due to multi-sensor fusion.