CLFT examples

Each example is a complete JSON file with one epoch, batch size 2, and no pretrained weight download. Prepare your dataset first, then activate .venv and run commands from the Fusion checkout.

The config

ZOD with both camera and LiDAR (configs/examples/zod/clft/fusion.json):

configs/examples/zod/clft/fusion.json

{
  "extends": "clft",
  "Summary": "ZOD: clft fusion, one epoch",
  "tags": [
    "zod",
    "clft",
    "fusion",
    "one-epoch"
  ],
  "CLI": {
    "mode": "fusion"
  },
  "Dataset": {
    "dataset_root": "$ZOD_DATA_DIR",
    "annotation_path": "annotation_fusion"
  },
  "Log": {
    "logdir": "logs/zod/clft/fusion"
  },
  "General": {
    "device": "cuda:0",
    "epochs": 1,
    "batch_size": 2,
    "seed": 0,
    "early_stop_patience": 1,
    "max_checkpoints": 1
  },
  "CLFT": {
    "warmup_epochs": 0,
    "pretrained": false
  }
}

Other datasets and inputs

Download the file for your dataset and input. Save it under any name and pass it to -c.

Dataset	Camera only	LiDAR only	Camera + LiDAR
ZOD ($ZOD_DATA_DIR)	rgb.json	lidar.json	fusion.json
Waymo ($WAYMO_DATA_DIR)	rgb.json	lidar.json	fusion.json
Iseauto ($ISEAUTO_DATA_DIR)	rgb.json	lidar.json	fusion.json

The files differ only in these keys:

Key	Values
CLI.mode	rgb, lidar, or fusion
Dataset.dataset_root	$ZOD_DATA_DIR, $WAYMO_DATA_DIR, or $ISEAUTO_DATA_DIR: export the folder you prepared
Dataset.annotation_path	ZOD: annotation_camera_only, annotation_lidar_only, annotation_fusion; Waymo and Iseauto: annotation
Log.logdir	logs/<dataset>/clft/<mode>, so runs do not overwrite each other
Summary, tags	the run's name and labels in Visin

Run

Run all four stages

visin-fusion run -c configs/examples/zod/clft/fusion.json --upload --benchmark-device cuda

On a CPU host, replace cuda with cpu. --upload enables prediction-image uploads when Visin reporting is configured. One epoch is a pipeline check, not an accuracy baseline. extends names a built-in model preset, so no parent config file is required.

Next: Run and inspect the outputs or customize a run.