…
runner:
  type: PoseTrainingRunner
  gpus:
  - 0
  - 1
  - 2
  - 3
  key_metric: test.mAP
  key_metric_asc: true
  eval_interval: 300
…

…
'model':{'backbone': {'type': 'ResNet', 'model_name': 'resnet50_gn', 'output_stride': 16, 'freeze_bn_stats': True, 'freeze_bn_weights': False}, 'backbone_output_channels': 2048, 'heads': {'bodypart': {...}}}
'net_type':'resnet_50'
'runner':{'type': 'PoseTrainingRunner', 'gpus': [0, 1, 2, 3], 'key_metric': 'test.mAP', 'key_metric_asc': True, 'eval_interval': 300, 'optimizer': {'type': 'AdamW', 'params': {...}}, 'scheduler': {'type': 'LRListScheduler', 'params': {...}}, 'snapshots': {'max_snapshots': 5, 'save_epochs': 25, 'save_optimizer_state': False}}
'train_settings':{'batch_size': 8, 'dataloader_workers': 0, 'dataloader_pin_memory': False, 'display_iters': 100, 'epochs': 200, 'seed': 42}
len():8

def build_training_runner(
    runner_config: dict,
    model_folder: Path,
    task: Task,
    model: nn.Module,
    device: str,
    gpus: list[int] | None = None,
    snapshot_path: str | None = None,
    logger: BaseLogger | None = None,
) -> TrainingRunner:
    """
    Build a runner object according to a pytorch configuration file

    Args:
        runner_config: the configuration for the runner
        model_folder: the folder where models should be saved
        task: the task the runner will perform
        model: the model to run
        device: the device to use (e.g. {'cpu', 'cuda:0', 'mps'})
        gpus: the list of GPU indices to use for multi-GPU training
        snapshot_path: the snapshot from which to load the weights
        logger: the logger to use, if any

    Returns:
        the runner that was built
    """