Description
Is your feature request related to a problem? Please describe.
Currently, to support code execution in order, we defined expressions in a list of bundle config, like:
https://github.com/Project-MONAI/model-zoo/blob/dev/models/spleen_ct_segmentation/configs/multi_gpu_train.json#L27
But when loading the bundle in other applications, like MONAI Label, MONAI FL, we want to get the trainer or other components and hard to execute all the expressions in the list immediately, so we re-define much logic in the python parser, like:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/fl/client/monai_algo.py#L460
To easily connect with other apps, I think we should clearly define the expressions in initialize, run, finalize sections, then others can call the operations separately. For example:
{
"initialize": [
"$import torch.distributed as dist",
"$dist.is_initialized() or dist.init_process_group(backend='nccl')",
"$torch.cuda.set_device(@device)",
"$monai.utils.set_determinism(seed=123)",
"$setattr(torch.backends.cudnn, 'benchmark', True)",
"$import logging",
"$@train#trainer.logger.setLevel(logging.WARNING if dist.get_rank() > 0 else logging.INFO)",
"$@validate#evaluator.logger.setLevel(logging.WARNING if dist.get_rank() > 0 else logging.INFO)"
],
"run": [
"$@train#trainer.run()"
],
"finalize": [
"$dist.destroy_process_group()"
]
}Will also create another ticket to define a high level API Bundle to simplify the bundle loading in other apps.