ParticleDetection.modelling.export

Functions to export a Detectron2 model as a pure pytorch model and functions to

export its output to other formats.

Author: Adrian Niemann (adrian.niemann@ovgu.de)

Date: 31.10.2022

annotation_to_json(prediction: DetectionResult, image: Path | str, classes: dict | None = None, output_dir: Path | str = Path(), *, filename: str = 'extracted_meta_data.json', **_)

Saves detected object masks in the metadata format used for model training.

Hint

This function is intended to be used as a saving function with detect().

Parameters:

• prediction (DetectionResult) – Prediction output of a Detectron2 network. It can also be given as prediction["instances"] as detectron2.structures.Instances or dict, as long as the resulting dict contains at least the same keys as DetectionResult.

• image (Union[Path, str]) – Path to image that prediction was created from.

• classes (dict, optional) –

  Dictionary of classes detectable by the model with

  {key} -> Index of class in the model

  {value} -> Name of the class

  By default None.

• output_dir (Path | str, optional) –

  Path to a folder the output file will be written to.

  By default Path().

• filename (str, optional) –

  Name of the *.json file the annotation data should be saved in. Already existing data for an image in this file will be overwritten.

  By default "extracted_meta_data.json".

export_model(config_path: Path, weights_path: Path, sample_img: Path, option: Literal['cpu', 'cuda'] = 'cuda') → None

Exports a Detectron2 model to be usable with just pytorch.

Parameters:

• config_path (Path) – File that holds the model’s configuration in yaml format.

• weights_path (Path) – File that holds the trained model’s weights.

• sample_img (Path) – Image to be used to trace the model.

• option (EXPORT_OPTIONS, optional) –

  Option whether to restrict the exported model to be used on the CPU or to also allow the use of a GPU.

  By default "cuda".

Note

The GPU version then requires the pytorch GPU version to be installed. The CPU version can be run with both, pytorch’s CPU and GPU version.

get_sample_img(sample: Path) → Tensor

Loads an image into the format necessary for inference by a Detectron2 model.

rods_to_csv(prediction: DetectionResult, image: Path | str | ndarray, classes: dict | None = None, output_dir: Path | str = Path(), *, filename: str = 'extracted_rods.csv', **kwargs) → None

Extract rod enpoint positions from detected object masks and save them to *.csv files.

For each detected object, two enpoints are determined from the detected segmentation mask. These endpoints are then saved in *.csv format described by DEFAULT_COLUMNS. The data is either saved into a new file, given by filename and output_dir or integrated into this file, if it already exists.

Hint

This function is intended to be used as a saving function with detect().

Parameters:

• prediction (DetectionResult) – Prediction output of a Detectron2 network. It can also be given as prediction["instances"] as detectron2.structures.Instances or dict, as long as the resulting dict contains at least the same keys as DetectionResult.

• image (Union[Path, str, np.ndarray]) – (Path to ) the image that prediction was created from.

• classes (dict, optional) –

  _description_

  By default None.

• output_dir (Union[Path, str], optional) –

  Path to a folder the output file will be written to.

  By default Path().

• filename (str, optional) –

  Name of the *.csv file the rod position data should be saved in. Already existing data for an image in this file might get overwritten.

  By default "extracted_rods.csv".

• **kwargs (dict, optional) –

  The following keywords are used to determine the used frame-camera combination of the image used to create the prediction. This allows the proper saving when a dataset format is given to the detect() function instead of a list of files.

  cam1_name : str

  cam2_name : str

  frames : Iterable[int]

  dataset_format : str

  The following keyword arguments are passed to rod_endpoints():

  method: Literal[“simple”, “advanced”]

  expected_particles : Union[int, Dict[int, int], None]

See also

replace_missing_rods(), add_points(), rod_endpoints(), csv_extract_colors()

rods_to_mat(prediction: DetectionResult, image: Path | str, classes: dict | None = None, output_dir: Path | str = Path(), *_, **kwargs) → None

Extract rod enpoint positions from detected object masks and save them to *.mat files.

The generated *.mat contain one variable rod_data_links with each rod being represented by a Point1 and Point2.

Hint

This function is intended to be used as a saving function with detect().

Parameters:

• prediction (DetectionResult) – Prediction output of a Detectron2 network. It can also be given as prediction["instances"] as detectron2.structures.Instances or dict, as long as the resulting dict contains at least the same keys as DetectionResult.

• image (Union[Path, str]) – Path to image that prediction was created from.

• classes (dict, optional) –

  Dictionary of classes detectable by the model with

  {key} -> Index of class in the model

  {value} -> Name of the class

  By default None.

• output_dir (Union[Path, str]) –

  Path to a folder the output file will be written to.

  By default Path().

• **kwargs (dict, optional) –

  Keywords, that are propagated to rod_endpoints():

  method: Literal[“simple”, “advanced”]

  expected_particles : Union[int, Dict[int, int], None]

See also

rod_endpoints()