Failure case:

a config which is allowed by previous mmengine:

from mmdet.datasets import CocoDataset
... = dict(
    ...=dict(
        class_names_txt=CocoDataset.METAINFO['classes']
    )
)

# delete python style config to allow printing and dumping config
del CocoDataset

This is non-lazy_import, hence i provide a hack arg in train.py:

cfg = Config.from_file(arg.cfg, lazy_import=False)

But there is still auto lazy_import detected:

I find auto detecting lazy_import maintains in _file2dict().
But If a user provide lazy_import=False, why auto detecting lazy_import again? If whether lazy_import is True or False, it must be equal to Config._is_lazy_import(filename), what is the meaning of lazy_import=False arg.

Solutions:

1. also add lazy_import arg to _file2dict() (but _file2dict() seems customized for non-lazy_import, seems not pretty)
2. delete the auto detecting (but may raise anothor error)
3. add an self.lazy_import attribute.