ObjectDetectionDataFilter Class

ml_debugger.data_filter.object_detection.object_detection_torch_data_filter.ObjectDetectionTorchDataFilter

Bases: CommonObjectDetectionDataFilter, ObjectDetectionTorchLogger

DataFilter for object detection tasks using PyTorch models.

Combines CommonObjectDetectionDataFilter (query, filter, aggregation) with ObjectDetectionTorchLogger (NMS, _parse_and_save_io_data, bbox hashing) to provide real-time filtering and Active Learning query support.

__init__(model, model_name, version_name, result_name=None, n_epoch='latest', filter_config=None, target_layers=None, additional_fields=None, auto_sync=False, force_table_recreate=False, api_endpoint=None, api_key=None, n_class=None, score_thresh=None, iou_thresh=None, max_detections_per_image=None)

Initialize object detection data filter.

Parameters:

Name	Type	Description	Default
model	Module	PyTorch model to trace.	required
model_name	str	Name of the ML model.	required
version_name	str	Version identifier for the ML model.	required
result_name	Optional[str]	The name of the existing evaluation result to retrieve.	None
n_epoch	Union[str, Optional[int]]	Filter option for n_epoch value.	'latest'
filter_config	Optional[Union[BBoxStrategy, Dict[str, Any]]]	BBox-level aggregation strategy for real-time filtering. Controls bbox selection, aggregation, and image-level threshold. Can be a BBoxStrategy instance, a dict, or None (defaults).	None
target_layers	Optional[Dict[str, str]]	Mapping of layer aliases to module paths.	None
additional_fields	Optional[List[dict]]	Extra fields for database schema.	None
auto_sync	bool	Enable background syncing of logged data.	False
force_table_recreate	bool	Whether to drop and recreate existing tables.	False
api_endpoint	Optional[str]	URL of the service API for data upload.	None
api_key	Optional[str]	API key for authenticating with the service.	None
n_class	Optional[int]	Number of classes. Auto-detected from model if None.	None
score_thresh	Optional[float]	Minimum score threshold for user-visible NMS decisions.	None
iou_thresh	Optional[float]	IoU threshold for user-visible NMS decisions.	None
max_detections_per_image	Optional[int]	Maximum detections per image for user-visible output.	None

__call__(model_input, input_ids, dataset_type='pool', **kwargs)

Invoke the data filter on a single inference, recording I/O data.

Parameters:

Name	Type	Description	Default
model_input	Any	Input data for the model inference.	required
input_ids	List[str]	Identifiers of each input data. Must not contain duplicates.	required
dataset_type	str	Identifier of input dataset. (e.g. 'pool')	'pool'
**kwargs	Any	Additional keyword arguments for parsing and saving I/O data.	{}

Returns:

Type	Description
Tuple[Any, List[Optional[bool]]]	Tuple of (model_output, filter_flags): - model_output: Raw model output. - filter_flags: List of booleans indicating if each input matches filter condition (True = matches / should be extracted, False = does not match, None = no filter configured or no bboxes).

Raises:

Type	Description
ValueError	If input_ids contains duplicates.

get_hooked_features(layer_name)

Retrieve the captured output for a given layer alias.

Parameters:

Name	Type	Description	Default
layer_name	str	Alias of the layer whose activation was captured.	required

Returns:

Name	Type	Description
Any	Any	Activation data stored for the specified layer.

Raises:

Type	Description
KeyError	If no activation has been captured for layer_name.

export(output_path=None)

Export extracted features into a ZIP archive.

Uses the internal n_epoch resolved during validator setup to filter records, consistent with upload() and wait_for_save().

Parameters:

Name	Type	Description	Default
output_path	Optional[str]	Path or directory for saving the ZIP file. If no .zip extension, the default filename is appended. Defaults to cwd.	None

Returns:

Type	Description
Optional[Path]	Path to the created ZIP file, or None on non-primary distributed ranks.

wait_for_save(interval=3)

upload()

query(n_data, strategy, dataset_type='pool', type_cast=None)

Sort and query dataset based on strategy with image-level aggregation.

Parameters:

Name	Type	Description	Default
n_data	int	Maximum number of images to query.	required
strategy	Union[str, BBoxStrategy, Dict[str, Any]]	Query strategy. - 'high_error_proba': Per-image max of error_proba, sorted descending. Returns images where at least one bbox has high error. - 'low_error_proba': Per-image min of error_proba, sorted ascending. Returns images where at least one bbox has low error. - BBoxStrategy: Full control over bbox selection, aggregation, target column, and sort order. - dict: Validated as BBoxStrategy. Unknown keys are rejected, unspecified fields use defaults. Example: {'target_column': 'det_error_proba', 'aggregation': 'mean', 'top_n': 3}	required
dataset_type	str	Filter of input dataset (e.g. 'pool').	'pool'
type_cast	Optional[type]	Type for casting input_id (e.g. int).	None

Returns:

Type	Description
List[str]	List of input_ids of queried images.

Raises:

Type	Description
ValueError	If strategy is invalid or no data found.

get_image_scores(strategy='high_error_proba', dataset_type='pool')

Get image-level error scores sorted by the given strategy.

Aggregates per-bbox error probabilities to image-level scores using the specified strategy, and returns them as a sorted DataFrame.

Parameters:

Name	Type	Description	Default
strategy	Union[str, BBoxStrategy, Dict[str, Any]]	Scoring strategy. - 'high_error_proba': Sort by error_proba descending (default). - 'low_error_proba': Sort by error_proba ascending. - BBoxStrategy: Full control over bbox selection, aggregation, target column, and sort order. - dict: Validated as BBoxStrategy.	'high_error_proba'
dataset_type	str	Filter of input dataset (e.g. 'pool').	'pool'

Returns:

Type	Description
DataFrame	DataFrame with columns [input_id, error_proba, pred_score],
DataFrame	sorted by error_proba according to the strategy's sort order.

Raises:

Type	Description
ValueError	If strategy is invalid or no data found.