what is the best way to use my build func to build model?

[Fei-Wang]

This is my code.

from mmengine.registry import MODELS as MMENGINE_MODELS
MODELS = Registry('model', build_func=build_model, parent=MMENGINE_MODELS, locations=['myproject.models'])
def build_model(cfg, registry, *args, **kwargs):
    pass

but, when I use default runner in myproject, here is the error

mmengine/mmengine/runner/runner.py

Line 800 in 5753cd9

model = MODELS.build(model)

this Models registry is mmengine MODELS, not myproject MODELS, so it can't find my build_func, so what is the best way to use my build func?

[C1rN09]

Currently you can't use your own build_func. Can you share your codes or ideas with us so that we can find out why you need to change MODEL's build function? If it is necessary, we may consider changing the logic here.

[Fei-Wang]

I want to use in my nlp project, and I use the popular nlp library called transformers, here's my code in my build_func,

    pretrained = args.pop('pretrained')
    if pretrained:
        obj = obj_cls.from_pretrained(pretrained, **args)
    else:
        obj = obj_cls(**args)  # type: ignore

    return obj

so, I can init the model use from_pretrained method directly, of course, my model is inherited from transformers.PreTrainedModel and BaseModel from mmengine.
but I meet some problems, and I'm thinking about solving it.

1. use my own build_func
2. self._init_model_weights() repeated initialization in runner.train()
3. save and load model use save_pretrained and from_pretrained
4. add save and load other staff such as tokenizer...

My current idea is to rewrite the runner, but in this way, I will not enjoy the advantages and convenience brought by mmengine. Is there a better way to solve it?

[C1rN09]

For Q1, you can actually register a function to your own Registry. For example

# suppose MODELS is your Registry rather than MMEngine's root Registry
@MODELS.register_module(name='MyNLPModel')
def build_my_nlp_model(*args, **kwargs):
    # args and kwargs are your own parameters in config files, you can make your own specification
    # However, `type` filed has been popped in MODELS.build(), so you will have to use some other names
    pass

# In your config file, you should write
model = dict(type='MyNLPModel', ...)

Actually, this is a "build function in build function". Feel free to give us more feedback if you are confused, or you have any suggestions!

[C1rN09]

For Q2, you can manually remove the init_weights() method inherited from BaseModel

[C1rN09]

For Q3/Q4, we are considering making save/load checkpoints as BaseModel's method so that you can modify it on your need. This is under investigation and may not come out soon.