evaluation/eval.py

# from evalscope.backend.vlm_eval_kit import VLMEvalKitBackendManager
# print(f'** All models from VLMEvalKit backend: {VLMEvalKitBackendManager.list_supported_datasets()}')

import argparse
import yaml
import importlib.util
import sys
import os
from evalscope.run import run_task
from evalscope.summarizer import Summarizer

from data_configs import DATASET_CONFIGS


def import_custom_dataset(dataset_name):
    """Dynamically import a custom dataset module for the given dataset."""
    if dataset_name in DATASET_CONFIGS:
        config = DATASET_CONFIGS[dataset_name]
        if config["custom_script"] and config["custom_eval_class"]:
            print(f"Importing custom dataset module for {dataset_name}")

            # Get the script path
            script_path = config["custom_script"]

            # Extract module name from script path
            module_name = f"custom_dataset_{dataset_name}"

            try:
                # Check if module is already loaded
                if module_name in sys.modules:
                    # Remove it to force reload
                    del sys.modules[module_name]

                # Load the module dynamically
                spec = importlib.util.spec_from_file_location(module_name, script_path)
                if spec and spec.loader:
                    module = importlib.util.module_from_spec(spec)
                    sys.modules[module_name] = module
                    spec.loader.exec_module(module)

                    # The module should automatically patch CustomVQADataset when loaded
                    print(f"Successfully imported custom dataset module for {dataset_name}")
                    return True
                else:
                    print(f"Failed to create module spec for {dataset_name}")
                    return False
            except Exception as e:
                print(f"Failed to import custom dataset module for {dataset_name}: {str(e)}")
                return False
    return False


def run_eval(config, analysis_report=False):
    # Since evalscope runs evaluations dataset by dataset, We'll hook into the evaluation process to import custom datasets as needed

    # First, let's check if this is running a specific dataset
    # We can inspect the config to see which datasets are being evaluated
    with open(config, "r") as f:
        cfg = yaml.safe_load(f)

    datasets = cfg.get("eval_config", {}).get("data", [])

    # Import VLMEvalKit's registry to hook into dataset loading
    try:
        from vlmeval.config import supported_VLM
        from vlmeval.dataset import build_dataset
        from vlmeval.dataset import DATASET_TYPE

        # Override the build_dataset function to import custom datasets on demand
        original_build_dataset = build_dataset

        def custom_build_dataset(dataset_name, **kwargs):
            # Import custom dataset if needed
            import_custom_dataset(dataset_name)
            # Call the original build_dataset function
            return original_build_dataset(dataset_name, **kwargs)

        # Monkey patch the build_dataset function
        import vlmeval.dataset

        vlmeval.dataset.build_dataset = custom_build_dataset

    except Exception as e:
        print(f"Warning: Could not hook into VLMEvalKit dataset loading: {e}")
        # Fall back to importing all custom datasets (with the overwriting issue)
        for dataset_name in datasets:
            import_custom_dataset(dataset_name)
    # Compatibility patch: ensure VLMEval receives a `reuse_aux` attribute
    # Some versions of VLMEvalKit's runner access args.reuse_aux but
    # evalscope constructs an Arguments object without that field.
    try:
        import vlmeval.run as _vlm_run

        _orig_run_task = _vlm_run.run_task

        def _run_task_with_reuse_aux(args):
            if not hasattr(args, "reuse_aux"):
                setattr(args, "reuse_aux", True)
            return _orig_run_task(args)

        _vlm_run.run_task = _run_task_with_reuse_aux
    except Exception as _e:
        print(f"Warning: Failed to patch vlmeval.run.run_task for reuse_aux: {_e}")

    run_task(task_cfg=config)

    if analysis_report:
        report_list = Summarizer.get_report_from_cfg(config)
        print(f"\n>> The report list: {report_list}")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, required=True, help="Path to evaluation config file")
    parser.add_argument(
        "--analysis_report", type=str, choices=["True", "False"], default="True", help="Generate analysis report"
    )
    args = parser.parse_args()

    analysis_report = args.analysis_report == "True"
    run_eval(args.config, analysis_report)