core.services.model.utils

utils

Functions:

Name	Description
build_model_impl	Instantiate and initialize a model object within the current Picsellia pipeline context.
evaluate_model_impl	Run evaluation of model predictions using the appropriate evaluation strategy based on inference type.

build_model_impl(context, model_cls, pretrained_weights_name=None, trained_weights_name=None, config_name=None, exported_weights_name=None)

Instantiate and initialize a model object within the current Picsellia pipeline context.

This function supports both training and processing contexts and uses the appropriate model version to build the model. It also downloads any associated model weights (pretrained, trained, exported).

Parameters:

Name	Type	Description	Default
context	PicselliaProcessingContext | PicselliaTrainingContext | LocalProcessingContext | LocalTrainingContext	The current pipeline context, which must be either a training or processing context.	required
model_cls	type[TModel]	The model class to instantiate.	required
pretrained_weights_name	str	The name of the pretrained weights to download.	None
trained_weights_name	str	The name of the trained weights to download.	None
config_name	str	The name of the model configuration file.	None
exported_weights_name	str	The name of the exported weights for inference.	None

Returns:

Name	Type	Description
TModel	TModel	An instance of the initialized model with weights downloaded.

evaluate_model_impl(context, picsellia_predictions, inference_type, assets, output_dir, training_labelmap=None)

Run evaluation of model predictions using the appropriate evaluation strategy based on inference type.

This function leverages Picsellia's ModelEvaluator to: - Compare predictions to ground truth. - Compute classification metrics for classification tasks. - Compute COCO metrics for detection and segmentation tasks.

Parameters:

Name	Type	Description	Default
context	PicselliaProcessingContext | PicselliaTrainingContext | LocalProcessingContext | LocalTrainingContext	The current pipeline context, expected to contain an experiment.	required
picsellia_predictions	list[PicselliaClassificationPrediction] | list[PicselliaRectanglePrediction] | list[PicselliaPolygonPrediction] | list[PicselliaOCRPrediction]	List of predictions generated by the model, matching the inference type.	required
inference_type	InferenceType	The type of model inference performed (e.g., classification, detection).	required
assets	list[Asset] | MultiAsset	Ground truth assets against which predictions are evaluated.	required
output_dir	str	The directory where evaluation metrics and results will be written.	required
training_labelmap	dict[str, str] | None	Optional mapping of training labels for evaluation alignment.	None