Open-Loop Inference

Bench2Drive-VL supports open-loop VQA evaluation via VQA dataset.

Write a config file

Write a config file under ./B2DVL_Adapter for inferencing framework.

your_infer_config.json

{
    "TASK_CONFIGS": {
        "INFER_SUBSET": false, // inference a subset of given dataset
        "USE_CHECKPOINT": true, // record the process of inference
        "SUBSET_FILE": "./infer_configs/subset.txt", // subset file, leave blank if not used
        "CHECKPOINT_FILE": "./infer_configs/finished_scenarios.txt", // checkpoint file, leave blank if not used
        "ENTRY_EXIT_FILE": "./infer_configs/entry_exits.json", // the file which specifies entry and exit point of certain scenario, 
        // you can create a file with "{}" as content if do not specify
        "FRAME_PER_SEC": 10 // sensor frame
    },
    "INFERENCE_BASICS": {
        "INPUT_WINDOW": 1,
        "CONVERSATION_WINDOW": 2,
        "USE_ALL_CAMERAS": true,
        "NO_HISTORY_MODE": false,
        "APPEND_QUESTION": true,
        "APPENDIX_FILE": "./infer_configs/append_questions.json" // not used now, to be removed
    },
    "CHAIN": {
        "NODE": [43, 50],
        "EDGE": {
            "43": [50],
            "50": []
        },
        "INHERIT": {
            "19": [43, 7],
            "15": [7]
        },
        "USE_GT": []
    }
}

Start to inference

Run the start up script you just wrote.

cd ./B2DVL_Adapter
python inference.py --model Qwen2.5VL --model_path /path/to/Qwen2.5VL-3B-Instruct --config_dir /path/to/your_infer_config.json --image_dir /path/to/Bench2Drive/dataset --vqa_dir /path/to/vqa/dataset --num_workers 4 --out_dir ./infer_outputs