[paddleseg source code reading] paddleseg custom data class

Feel how to change it first ：

Config In the document item Class related __init__.py The parameters correspond one by one

stay PaddleSeg Of Config In file ：

val_dataset:
  type: Dataset
  dataset_root: /root/share/program/save/data/portraint
  val_path: /root/share/program/save/data/portraint/txtfiles/test.txt
  num_classes: 2
  transforms:
    - type: Resize
      target_size: [224, 224]
    - type: Normalize
  mode: val

The above Config Parameters are passed in Dataset Class __init__ Parameters in

    def __init__(self,
                 transforms,
                 dataset_root,
                 num_classes,
                 mode='train',
                 train_path=None,
                 val_path=None,
                 test_path=None,
                 separator=' ',
                 ignore_index=255,
                 edge=False):
        self.dataset_root = dataset_root
        self.transforms = Compose(transforms)
        self.file_list = list()
        self.mode = mode.lower()
        self.num_classes = num_classes
        self.ignore_index = ignore_index
        self.edge = edge
        
		......

If this Dataset It doesn't meet your needs , You can customize a class , Such as XXDataset

Will Config In the document Dataset type Partial change to Your custom class XXDataset

This class needs to be in paddleseg/datasets/__init__.py Import , Can be directly in __init__.py In the definition of , You can also create a new file , And then in __init__.py Import in .

Be careful , At the beginning of the custom data class , Put this decorator ,（ Register like this ?）

@manager.DATASETS.add_component

( Ah, this , I remember not using it before , Or before PaddleClas no need ?

for instance ：

paddleseg/datasets/dataset.py in ：

config file ：

__init__.py Import ：

Then pay attention here ：

If these two attributes are not defined , There will be an error ,transforms attribute , It should be the parameters that pass in , But here is the image of direct preprocessing , So I gave it to None Generally speaking, it needs to be transferred

So I have to change a place here , Anyway, there is no error , Just change where you want

Here's a little more , In general , With incoming Transform Parameter preprocessing , Don't be like me ：

class myDataset(Dataset):
    def __init__(self, data_root, 
                        img_name="images", 
                        ann_name="annotations"):
        
        self.img_path = os.path.join(data_root, img_name)
        self.ann_path = os.path.join(data_root, ann_name)
        
        #  Be sure to pass in the relative path , Absolute path  OpenCV  I can't read 
        self.imgs = [os.path.join(self.img_path, img) for img in os.listdir(self.img_path)]
        
        self.anns = [img.replace(".jpg", ".png") for img in self.imgs]
        self.anns = [img.replace(img_name, ann_name) for img in self.anns]
        

    def __len__(self):
        return len(self.anns)

    def __getitem__(self, idx):
        img = cv2.imread(self.imgs[idx])
        ann = cv2.imread(self.anns[idx], 0)
        
        img = self.img_preprocess(img)
        ann = self.ann_preprocess(ann)
        
        return img, ann
    
    @staticmethod
    def img_preprocess(img):
        img = cv2.resize(img, (224, 224)).astype(np.float32)
        img = img.transpose(2, 0, 1)
        img = img[None]
        img /= 255
        return img
    
    @staticmethod
    def ann_preprocess(ann):
        ann = cv2.resize(ann, (224, 224))
        return ann

I specifically define here self.img_preprocess and self.ann_preprocess To read pictures , And do the pretreatment .

So the code above , Will be able to self.transform to None, Then the passed parameters are given to []

That's OK. , Anyway, I'm too lazy to be transforms Register a new transform

however ,PaddleSeg stay transforms I did a little lazy work in , Will help us automatically read in pictures and Label

Take a look at , What is the transforms Which class in will read pictures ：

stay Compose Of __call__ Function :

If the parameter passed in is str Then it will read the corresponding figure

( Be careful , The pictures read here are 0-255 Of , yes uint8 Of )

( Be careful ,for Operation under cycle , It's all about Transform finish , And then Transpose)
in other words , Of all the Transform after , The output is CHW, That is, the number of channels is at the top
in addition , The final transpose is np The operation of , in other words PaddleSeg Of Transform Mostly based on numpy Realized ( Only when reading pictures , use PIL.Image class )

---

Finally, let's see what decorators there are ：

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

Probably , Add in the back .add_component Then you can add a new object