detectionmetrics.models package

Submodules

detectionmetrics.models.model module

detectionmetrics.models.onnx module

class detectionmetrics.models.onnx.OnnxImageSegmentationModel(model, model_type, ontology_fname, model_cfg, model_fname)

Bases: ImageSegmentationModel

detectionmetrics.models.tensorflow module

class detectionmetrics.models.tensorflow.ImageSegmentationTensorflowDataset(dataset: ImageSegmentationDataset, resize: Tuple[int, int] | None = None, crop: Tuple[int, int] | None = None, batch_size: int = 1, splits: List[str] = ['test'], lut_ontology: dict | None = None, normalization: dict | None = None, keep_aspect: bool = False)

Bases: object

Dataset for image segmentation Tensorflow models

Parameters:

  • dataset (ImageSegmentationDataset) – Image segmentation dataset

  • resize (Optional[Tuple[int, int]], optional) – Target size for resizing images, defaults to None

  • crop (Optional[Tuple[int, int]], optional) – Target size for center cropping images, defaults to None

  • batch_size (int, optional) – Batch size, defaults to 1

  • splits (str, optional) – Splits to be used from the dataset, defaults to [“test”]

  • lut_ontology (dict, optional) – LUT to transform label classes, defaults to None

  • normalization (dict, optional) – Parameters for normalizing input images, defaults to None

  • keep_aspect (bool, optional) – Whether to keep aspect ratio when resizing images. If true, resize to match smaller sides size and crop center. Defaults to False

load_data(idx: str, images_fnames: List[str], labels_fnames: List[str]) → Tuple[tensorflow.Tensor, tensorflow.Tensor]

Function for loading data for each dataset sample

Parameters:

  • idx (str) – Sample index

  • images_fnames (List[str]) – List containing all image filenames

  • labels_fnames (List[str]) – List containing all corresponding label filenames

Returns:

Image and label tensor pairs

Return type:

Tuple[tf.Tensor, tf.Tensor]

read_image(fname: str, label=False) → tensorflow.Tensor

Read a single image or label

Parameters:

  • fname (str) – Input image or label filename

  • label (bool, optional) – Whether the input data is a label or not, defaults to False

Returns:

Tensorflow tensor containing read image or label

Return type:

tf.Tensor

class detectionmetrics.models.tensorflow.TensorflowImageSegmentationModel(model: str | tensorflow.Module | tensorflow.keras.Model, model_cfg: str, ontology_fname: str)

Bases: ImageSegmentationModel

Image segmentation model for Tensorflow framework

Parameters:

  • model (Union[str, torch.nn.Module]) – Either the filename of a Tensorflow model in SavedModel format or the model already loaded into an arbitrary Tensorflow or Keras model.

  • model_cfg (str) – JSON file containing model configuration

  • ontology_fname (str) – JSON file containing model output ontology

eval(dataset: ImageSegmentationDataset, split: str | List[str] = 'test', ontology_translation: str | None = None, predictions_outdir: str | None = None, results_per_sample: bool = False) → DataFrame

Perform evaluation for an image segmentation dataset

Parameters:

  • dataset (ImageSegmentationDataset) – Image segmentation dataset for which the evaluation will be performed

  • split (str | List[str], optional) – Split to be used from the dataset, defaults to “test”

  • ontology_translation (str, optional) – JSON file containing translation between dataset and model output ontologies

  • predictions_outdir (Optional[str], optional) – Directory to save predictions per sample, defaults to None. If None, predictions are not saved.

  • results_per_sample (bool, optional) – Whether to store results per sample or not, defaults to False. If True, predictions_outdir must be provided.

Returns:

DataFrame containing evaluation results

Return type:

pd.DataFrame

get_computational_cost(image_size: Tuple[int] | None = None, runs: int = 30, warm_up_runs: int = 5) → dict

Get different metrics related to the computational cost of the model

Parameters:

  • image_size (Tuple[int], optional) – Image size used for inference

  • runs (int, optional) – Number of runs to measure inference time, defaults to 30

  • warm_up_runs (int, optional) – Number of warm-up runs, defaults to 5

Returns:

Dictionary containing computational cost information

inference(image: Image) → Image

Perform inference for a single image

Parameters:

image (Image.Image) – PIL image

Returns:

segmenation result as PIL image

Return type:

Image.Image

detectionmetrics.models.tensorflow.crop_center(image: tensorflow.Tensor, width: int, height: int) → tensorflow.Tensor

Crop tensorflow image center to target size

Parameters:

  • image (tf.Tensor) – Input image tensor

  • width (int) – Target width for cropping

  • height (int) – Target width for cropping

Returns:

Cropped image tensor

Return type:

tf.Tensor

detectionmetrics.models.tensorflow.get_computational_cost(model: tensorflow.Module, dummy_input: tensorflow.Tensor, model_fname: str | None = None, runs: int = 30, warm_up_runs: int = 5) → dict

Get different metrics related to the computational cost of the model

Parameters:

  • model (tf.Module) – Loaded TensorFlow SavedModel

  • dummy_input (tf.Tensor) – Dummy input data for the model

  • model_fname (Optional[str], optional) – Model filename used to measure model size, defaults to None

  • runs (int, optional) – Number of runs to measure inference time, defaults to 30

  • warm_up_runs (int, optional) – Number of warm-up runs, defaults to 5

Returns:

DataFrame containing computational cost information

Return type:

pd.DataFrame

detectionmetrics.models.tensorflow.resize_image(image: tensorflow.Tensor, method: str, width: int | None = None, height: int | None = None, closest_divisor: int = 16) → tensorflow.Tensor

Resize tensorflow image to target size. If only one dimension is provided, the aspect ratio is preserved.

Parameters:

  • image (tf.Tensor) – Input image tensor

  • method (str) – Resizing method (e.g. bilinear, nearest)

  • width (Optional[int], optional) – Target width for resizing

  • height (Optional[int], optional) – Target height for resizing

  • closest_divisor (int, optional) – Closest divisor for the target size, defaults to 16. Only applies to the dimension not provided.

Returns:

Resized image tensor

Return type:

tf.Tensor

detectionmetrics.models.torch module

Module contents