[paddleseg source code reading] add boundary IOU calculation in paddleseg validation (1) -- val.py file details tips

The last one was tossed boundary IoU Calculation method of , This article talks about PaddleSeg How to add Boundary IoU do Val, Training and infer Almost. ,hxdm Step on the pit yourself

---

First of all, the whole thing PaddleSeg Lite version ：

.travis.yml
.style.yapf
.pre-commit-config.yaml
.gitignore
.copyright.hook
LICENSE
README.md
README_CN.md
requirements.txt

export.py  #  Delete it or not , This is used to export static graph model files 
setup.py #  This is used to install  PaddleSeg  Of , Delete it 

benchmark
deploy
docs
EISeg #  An efficient and intelligent interactive segmentation and annotation software based on propeller 
slim
test_tipc #  Integrated certification of propeller training and propulsion  (Training and Inference Pipeline Certification(TIPC))  Information and testing tools 
tests

If it's just running training and reasoning , The above can be deleted directly

contrib  Just some other functions , such as  matting、PPHumanSeg  And so on. , Can be deleted 
configs  Well , That's the pile of configuration files , Understand everything 
tools, There are some commonly used tools , Label , Annotation format conversion and so on , After that, I don't want to write wheels , Just look here ( This is a treasure chest )

That's my PaddleSeg Lite version ,

val.py file

There's nothing to see , Just two sentences of code

#  Directly here   Data class   Instantiation 
val_dataset = cfg.val_dataset

#  Directly here  model  Instantiation 
model = cfg.model

Why instantiate directly here ? Obviously, it assigns an attribute , In fact, single-step debugging , You can see , He actually added a decorator to the attribute @property

@property
def val_dataset(self) -> paddle.io.Dataset:
    _val_dataset = self.val_dataset_config
    if not _val_dataset:
        return None
    return self._load_object(_val_dataset)

@property
def model(self) -> paddle.nn.Layer:
    model_cfg = self.dic.get('model').copy()
    if not model_cfg:
        raise RuntimeError('No model specified in the configuration file.')
    if not 'num_classes' in model_cfg:
        ......

    if not self._model:
        self._model = self._load_object(model_cfg)
    return self._model

Did you see? , There is one self._load_object function , All classes are loaded through this method ( Instantiation ) Of

Take a look at this method ( I hope the old irons can read it by themselves )：

def _load_object(self, cfg: dict) -> Any:
    cfg = cfg.copy()
    if 'type' not in cfg:
        raise RuntimeError('No object information in {}.'.format(cfg))

    component = self._load_component(cfg.pop('type'))
	
	#  The upper boundary is to get the class name 
	# -----------------  This is the dividing line  -----------------
	#  Below the dividing line is the parameter dictionary 
	
    params = {
    }
    for key, val in cfg.items():
        if self._is_meta_type(val):
            params[key] = self._load_object(val)
        elif isinstance(val, list):
            params[key] = [
                self._load_object(item)
                if self._is_meta_type(item) else item for item in val
            ]
        else:
            params[key] = val
	
	#  take config After taking out the classes and parameters in , stay return This step returns the object 
    return component(**params)

There's another one inside _is_meta_type Method , Guess the meaning , Is it Can be instantiated Judge

Click this method to see ：

def _is_meta_type(self, item: Any) -> bool:
	return isinstance(item, dict) and 'type' in item

If it's a dictionary , And there are type attribute , stay PaddleSeg Of Config In this architecture , Think that this class can be instantiated

---

OK, Now keep recursing to see the source code , Now let's jump back , have a look sellf._load_object in self._load_component How to get that kind of name

def _load_component(self, com_name: str) -> Any:
    com_list = [
        manager.MODELS, manager.BACKBONES, manager.DATASETS,
        manager.TRANSFORMS, manager.LOSSES
    ]

    for com in com_list:
        if com_name in com.components_dict:
            return com[com_name]
    else:
        raise RuntimeError(
            'The specified component was not found {}.'.format(com_name))

See here , I believe if you have read my blog , Then there must be a feeling of sudden enlightenment
[PaddleSeg Source code reading ] PaddleSeg Custom data class

The end of this blog ：

manager.MODELS
manager.BACKBONES
manager.DATASETS
manager.TRANSFORMS
manager.LOSSES

These components are shown , I think so ,PaddleSeg The commonly used components are abstracted , Directly abstracted as Model class ,Backbone Backbone network , Dataset class , picture Transform class and Loss function class

Abstract as a class , You need to add the corresponding decorator , For example, add custom classes , Need this decorator ：

@manager.DATASETS.add_component

Look at that English .add_component It means adding components , So in _load_component In the source code ,

for com in com_list:
    if com_name in com.components_dict:
        return com[com_name]

Go and see com_name Whether in manager.MODELS、manager.BACKBONES、manager.DATASETS、manager.TRANSFORMS and manager.LOSSES Among these components

If none of them are in , be raise This is wrong ：

raise RuntimeError(
    'The specified component was not found {}.'.format(com_name))

“ This component was not found ”

---

Here's another line , image metrics Those indicators ( such as IoU, acc And so on. ), They're just functions , Not class , So you don't need to use it .add_component To register

( This sentence is for later use Boundary IoU Pave the way )

Almost , This blog starts with these two sentences ：

val_dataset = cfg.val_dataset
model = cfg.model

Led to so many

To sum up for yourselves , Send it to the comment area