（四）mmdetection源码解读：MMCV 核心组件 Registry

一、四大注册器类DATASETS、PIPELINES、MODELS、HOOKS

Registry 用于提供全局类注册器功能，Registry 类内部其实维护的是一个全局 key-value 对。通过 Registry 类，用户可以通过字符串方式实例化任何想要的模块。

#先构建全局的 DATASETS 注册器类
##只是设置了self._name = 'dataset'，self.build_func = build_from_cfg
DATASETS = Registry('dataset')     

#通过装饰器方式作用在想要加入注册器的具体类中
@DATASETS.register_module()
class KittiTinyDataset(CustomDataset):
##只是设置了self._name = 'pipeline'，self.build_func = build_from_cfg
PIPELINES = Registry('pipeline')
#mmdet/models/builder.py
MODELS = Registry('models', parent=MMCV_MODELS)
#mmcv/runner/hooks/hook.py
HOOKS = Registry('hook')

二、四大注册器类实例化过程

通过register_module装饰器， 把key-value添加到字典中self._module_dict[name] = module_class

    def register_module(self, name=None, force=False, module=None):
       '省略。。。'
       #register_module是用作一种方法
        # use it as a normal method: x.register_module(module=SomeClass)
        if module is not None:
            self._register_module(
                module_class=module, module_name=name, force=force)
            return module
       #register_module是用作一个装饰器
        # use it as a decorator: @x.register_module()
        def _register(cls):
            self._register_module(
                module_class=cls, module_name=name, force=force)
            return cls

        return _register

    def _register_module(self, module_class, module_name=None, force=False):
            if module_name is None:
                module_name = module_class.__name__
            self._module_dict[name] = module_class

下面是数据集类的实例化过程

datasets = [build_dataset(cfg.data.train)]
def build_dataset(cfg, default_args=None):
    dataset = build_from_cfg(cfg, DATASETS, default_args)
def build_from_cfg(cfg, registry, default_args=None):
    """Build a module from config dict.

    args = cfg.copy()
#获取类和类名
    obj_type = args.pop('type')
    if isinstance(obj_type, str):
        obj_cls = registry.get(obj_type)
        if obj_cls is None:
            raise KeyError(
                f'{obj_type} is not in the {registry.name} registry')
    elif inspect.isclass(obj_type):
        obj_cls = obj_type
#在这里进行实例化，我定义的类KittiTinyDataset并没有初始化函数，他的父类CustomDataset有对应初始化函数
    try:
        return obj_cls(**args)
    except Exception as e:
        # Normal TypeError does not print class name.
        raise type(e)(f'{obj_cls.__name__}: {e}')

#先构建全局的 MODELS 注册器类
MODELS = Registry('models', parent=MMCV_MODELS)
这里的MMCV_MODELS就是下面的MODELS
MODELS = Registry('model', build_func=build_model_from_cfg)
BACKBONES = MODELS
NECKS = MODELS
ROI_EXTRACTORS = MODELS
SHARED_HEADS = MODELS
HEADS = MODELS
LOSSES = MODELS
DETECTORS = MODELS

cfg是list类型的情况下，循环调用build_from_cfg

def build_model_from_cfg(cfg, registry, default_args=None):

    if isinstance(cfg, list):
        modules = [
            build_from_cfg(cfg_, registry, default_args) for cfg_ in cfg
        ]
        return Sequential(*modules)
    else:
        return build_from_cfg(cfg, registry, default_args)

下面时配置文件中的部分代码，根据type内容，去实例化对应的模型

backbone=dict(
        type='ResNet',...),
    neck=dict(
        type='FPN',...),
    rpn_head=dict(
        type='RPNHead',...),
        bbox_coder=dict(
            type='DeltaXYWHBBoxCoder',...),
        loss_cls=dict(
            type='CrossEntropyLoss', ...),
    roi_head=dict(
        type='StandardRoIHead',...)

这里需要注意，Models通过register_module装饰器， 把key-value添加到字典中的函数都在mmdet/models/文件路径下，下面代码在mmdet/models/faster_rcnn.py中

@DETECTORS.register_module()
class FasterRCNN(TwoStageDetector):

总结：1、Registry创建多个全局注册器类，比如DATASETS = Registry(‘dataset’), PIPELINES = Registry(‘pipeline’),MODELS = Registry(‘models’, parent=MMCV_MODELS)
2、通过@DETECTORS.register_module()这种类似的装饰器，把类名和类添加到Registry._module_dict中
3、用配置文件内容去实例化对应的类