[MultiModal] Cloud Predictor doesn't work for object detection use cases

[tonyhoo]

If I want to for example adapt the AutoMM Detection - High Performance Finetune on COCO Format Dataset example to MultiModalCloudPredictor, what’s the required train_data format? I tried using the same format as in the local example (train_data="./pothole/pothole/Annotations/usersplit_train_cocoformat.json") but it doesn’t look like that filetype is supported:
ValueError: json file type is not supported.

And for the init and fit args, trying to do something to the effect of:

predictor_init_args = dict(
    hyperparameters={
        "model.mmdet_image.checkpoint_name": checkpoint_name,
        "env.num_gpus": num_gpus,
        "optimization.val_metric": val_metric,
    },
    problem_type="object_detection",
    sample_data_path=train_path,

)

predictor_fit_args = dict(
    train_data=train_path,
    hyperparameters={
        "optimization.learning_rate": 5e-6, # we use two stage and detection head has 100x lr
        "optimization.max_epochs": 1,
        "env.per_gpu_batch_size": 2,  # decrease it when model is large
    },
)