[CodeCamp2023-327]Control whether methods of the Visualizer are executed only in the main process through a parameter

[DeliMm]

Motivation

Control whether methods of the Visualizer are executed only in the main process through a parameter.. See more details in #1081

Modification

To implement this function, we add a parameter 'img_only_master' to visualizer by all kinds of vis_backend. During runtime, it is determined whether to retain only the images in the GPU by checking if img_only_master is passed and examining its corresponding value. The modified master_only decorator is as following:

def master_only(func: Callable) -> Callable:
    """Decorate those methods which should be executed in master process.

    Args:
        func (callable): Function to be decorated.

    Returns:
        callable: Return decorated function.
    """

    @functools.wraps(func)
    def wrapper(*args, **kwargs):
        if ('img_only_master' in kwargs
                and kwargs['img_only_master'] is False) or is_main_process():
            return func(*args, **kwargs)

    return wrapper

When we build the runner, we can follow this way to pass the img_only_master:

visualizer=dict(¬
             type='Visualizer',
             vis_backends=[
                 dict(
                     type='LocalVisBackend',
                     save_dir='./vis',
                     img_only_master=False)
             ]),

When we use the add_img method, we need pass the img_only_master through img_only_master=runner.visualizer.img_only_master

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
3. If the modification has potential influence on downstream projects, this PR should be tested with downstream projects, like MMDet or MMCls.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[HAOCHENYE]

Coupling the implementation of master_only with the Visualizer is not a good practice. I believe we can avoid using the master_only decorator inside the Visualizer and instead opt for adding a master_only parameter to the Visualizer. Then, in various methods, we can check whether master_only is set to True and control whether to invoke the corresponding method only on rank 0.

class Visualizer
    def __init__(self, ..., master_only=True):
        ...
        self.master_only = master_only

    def xxx(self, ...):
        if _should_do():
            ...
    
    def _should_do(self):
        if self.master_only and get_rank() != 0:
            return False
        else:
            return True

[DeliMm]

Thank you for your response. I initially thought that I could only add 'master_only' without modifying the decorator. And I'll make the necessary adjustments as soon as possible.

[HAOCHENYE]

We need to reconsider that if it is necessary to apply self._should_do here. I think we could skip this check here.

[DeliMm]

For this, I added the check based on the original master_only decorator for this function. For dataset_meta, it might be better to keep the original master_only decorator.

[HAOCHENYE]

Hi, please check the modification will not break the CI