tlc.integration.super_gradients.callbacks.pipeline_params¶

Utility dataclass for storing SuperGradients inference pipeline parameters.

Module Contents¶

Classes¶

Class

Description

PipelineParams

Parameters for model prediction pipelines across detection and pose models.

API¶

class PipelineParams¶

Parameters for model prediction pipelines across detection and pose models.

class_agnostic_nms: bool | None = None¶

Apply NMS across classes (True) or per-class (False).

  • Supported by: detection models exposing class-agnostic NMS.

  • Default: model-specific default or None to use model default.

conf: float | None = None¶

Confidence threshold; predictions below are discarded.

  • Supported by: detection models and pose models (YoloNASPose, DEKRPoseEstimationModel).

  • Default: model-specific default or None to use model default.

fp16: bool = True¶

Use mixed precision for inference.

  • Supported by: most pipelines.

  • Default: True.

fuse_model: bool = True¶

Fuse layers to improve performance (at the cost of memory).

  • Supported by: detection, pose, classification, segmentation.

  • Default: True.

iou: float | None = None¶

IoU threshold for NMS.

  • Supported by: detection models (e.g., YoloBase, CustomizableDetector, PPYoloE, SlidingWindowInferenceDetectionWrapper) and YoloNASPose.

  • Default: model-specific default or None to use model default.

max_predictions: int | None = None¶

Max detections returned per image after NMS.

  • Supported by: detection models exposing full NMS controls (CustomizableDetector, PPYoloE, SlidingWindowInferenceDetectionWrapper).

  • Default: model-specific default or None to use model default.

multi_label_per_box: bool | None = None¶

Allow multiple class labels per anchor.

  • Supported by: detection models that support multi-label decoding (e.g., PPYoloE, some CustomizableDetector heads).

  • Default: model-specific default or None to use model default.

nms_top_k: int | None = None¶

Top-K candidates considered before NMS.

  • Supported by: detection models exposing full NMS controls (CustomizableDetector, PPYoloE, SlidingWindowInferenceDetectionWrapper).

  • Default: model-specific default or None to use model default.

post_nms_max_predictions: int | None = None¶

Top-K predictions retained after NMS in pose decoding.

  • Supported by: YoloNASPose.

  • Default: model-specific default or None to use model default.

pre_nms_max_predictions: int | None = None¶

Top-K candidates kept before NMS in pose decoding.

  • Supported by: YoloNASPose.

  • Default: model-specific default or None to use model default.

skip_image_resizing: bool = False¶

Skip image resizing in the processor (stride-constrained auto-padding).

  • Supported by: detection, pose, classification, segmentation.

  • Default: False.