Skip to content

Latest commit

 

History

History
127 lines (97 loc) · 3.37 KB

transition.md

File metadata and controls

127 lines (97 loc) · 3.37 KB
Raw

How to transition

Parser

The major modification was in the Parser API, the concept of a ParserMixin was completely removed. Now, instead of starting by defining the parser, we start by defining the record:

Before:

class MyParser(
    parsers.Parser,
    parsers.FilepathMixin,
    parsers.LabelsMixin,
    parsers.BBoxesMixins,
):

After:

template_record = BaseRecord(
    (
        FilepathRecordComponent(),
        InstancesLabelsRecordComponent(),
        BBoxesRecordComponent(),
    )
)

We then continue to generate the parser template:

Before:

MyParser.generate_template()

After:

Parser.generate_template(template_record)

And here's how the parser class looks:

Before:

class ChessParser(
    parsers.Parser,
    parsers.FilepathMixin,
    parsers.LabelsMixin,
    parsers.BBoxesMixins,
):
    def __init__(self, data_dir):
        self.data_dir = data_dir
        self.df = pd.read_csv(data_dir / "annotations.csv")

        class_map = ClassMap(list(self.df['label'].unique()))
        super().__init__(class_map=class_map)
        
    def __iter__(self) -> Any:
        yield from self.df.itertuples()
        
    def __len__(self) -> int:
        return len(self.df)
        
    def imageid(self, o) -> Hashable:
        return o.filename

    def filepath(self, o) -> Union[str, Path]:
        return self.data_dir / 'images' / o.filename

    def image_width_height(self, o) -> Tuple[int, int]:
        return o.width, o.height

    def labels(self, o) -> List[int]:
        return [o.label]

    def bboxes(self, o) -> List[BBox]:
        return [BBox.from_xyxy(o.xmin, o.ymin, o.xmax, o.ymax)]

After:

class ChessParser(Parser):
    def __init__(self, template_record, data_dir):
        super().__init__(template_record=template_record)
        
        self.data_dir = data_dir
        self.df = pd.read_csv(data_dir / "annotations.csv")
        self.class_map = ClassMap(list(self.df['label'].unique()))
        
    def __iter__(self) -> Any:
        yield from self.df.itertuples()
        
    def __len__(self) -> int:
        return len(self.df)
        
    def imageid(self, o) -> Hashable:
        return o.filename
        
    def parse_fields(self, o, record):
        record.set_filepath(self.data_dir / 'images' / o.filename)
        record.set_img_size(ImgSize(width=o.width, height=o.height))
        
        record.detect.set_class_map(self.class_map)
        record.detect.add_bboxes([BBox.from_xyxy(o.xmin, o.ymin, o.xmax, o.ymax)])
        record.detect.add_labels([o.label])

Record

The attributes on the record are now separated by task, to access bboxes you now have to do record.detection.bboxes instead of record.bboxes. Common attributes (not specific to one task) can still be accessed directly: record.filepath.

You can use print to check the attributes on the record and how to access them:

BaseRecord

common: 
	- Image size ImgSize(width=416, height=416)
	- Image ID: 3
	- Filepath: /home/lgvaz/.icevision/data/chess_sample/chess_sample-master/images/e79deba8fe520409790b601ad61da4ee_jpg.rf.016bc04dee292f80d1f975931f32bc21.jpg
	- Image: None
detection: 
	- BBoxes: [<BBox (xmin:208, ymin:88, xmax:226, ymax:128)>]
	- Labels: [5]

All fields under common can be accessed directly, the others have to be specified by it's corresponding task.