YOLO pipeline integration for deepsparse

[rahul-tuli]

The goal of this PR is to add a pipeline interface for YOLO image segmentation models. The idea is to wrap preprocessing and postprocessing steps into a callable pipeline for ease of use, and out of the box support with deepsparse.server. This also adds an installable for yolo, to install all dependencies invoke:
pip install --editable "./[yolo]"

Usage:

• Option 1, without a class names mapping

from deepsparse.pipeline import Pipeline

model_stub = "zoo:cv/detection/yolov5-s/pytorch/ultralytics/coco/pruned-aggressive_96"
task = "yolo"
input_file = "yolo-test-image.JPEG"



pipeline = Pipeline.create(
    task=task,
    model_path=stub,
    class_names="coco",
)

print("Pipeline outputs: ", pipeline(images=input_file))

output:

DeepSparse Engine, Copyright 2021-present / Neuralmagic, Inc. version: 0.11.2 (ca6e29a5) (release) (optimized) (system=avx512, binary=avx512)
Pipeline outputs:  predictions=[[[132.24066162109375, 0.04663848876953125, 269.58807373046875, 46.56146240234375, 0.4553620517253876, 2.0], [181.0561065673828, 7.16241455078125, 447.66375732421875, 612.1524047851562, 0.444583535194397, 0.0]]] boxes=[[[132.24066162109375, 0.04663848876953125, 269.58807373046875, 46.56146240234375], [181.0561065673828, 7.16241455078125, 447.66375732421875, 612.1524047851562]]] scores=[[0.4553620517253876, 0.444583535194397]] labels=[['car', 'person']]

[dbogunowicz]

lgtm, good job. I would refactor some of the code, but I assume that those pieces which I do not like are copied from the original yolo codebase, so it does not make sense to beautify it - better to have it in the original form for reference, would you agree?

[bfineran]

rebased onto feature-pipeline-main