Summary

• Updated deepsparse support for yolov8
• Small similar changes to the ones made for sparseml's yolov8 update
• process_engine_outputs has a slight change. Previously providing nm for number of masks, which is no longer used by the yolov8 non_max_suppression function.

Testing:

• Tested locally using both sparsezoo and yolov8 models for both segmentation and detection
• Workflows tested: pipeline creation and forward pass, annotation, eval and benchmarking

sparsezoo models tested:

1. detection: "zoo:cv/detection/yolov8-m/pytorch/ultralytics/coco/base_quant-none"
2. segmentation: "zoo:cv/segmentation/yolov8-n/pytorch/ultralytics/coco/base-none"

yolov8 models tested:

1. detection: yolov8n.pt
2. segmentation: yolov8n-seg.pt

Pipeline creation + forward pass:

Detection

from deepsparse import Pipeline
model_path = "yolov8n.onnx" 
images = ["basilica.jpg"]
yolo_pipeline = Pipeline.create(
    task="yolov8",
    model_path=model_path,
)
pipeline_outputs = yolo_pipeline(images=images)

Segmentation:

from deepsparse import Pipeline
model_path = "yolov8n-seg.onnx" 
images = ["basilica.jpg"]
yolo_pipeline = Pipeline.create(
    task="yolov8",
    subtask="segmentation",
    model_path=model_path,
)

pipeline_outputs = yolo_pipeline(images=images)

Annotation:

Detection:
deepsparse.yolov8.annotate --source basilica.jpg --model_filepath "yolov8n.onnx"
Segmentation:
deepsparse.yolov8.annotate --source basilica.jpg --model_filepath "yolov8n-seg.onnx" --subtask segmentation

Results verified for outputs produced by both workflows above

Eval

Detection:
deepsparse.yolov8.eval --model_path yolov8n.onnx
Segementation:
deepsparse.yolov8.eval --model_path yolov8n-seg.onnx --subtask segmentation