YOLO Models

This article describes how to prepare models from the YOLO family for integration with the Deep Learning Streamer pipeline.

Ultralytics Model Preparation

All models supported by the ultralytics/ultralytics library can be converted to OpenVINO™ IR format by using the Ultralytics exporter. DL Streamer supports many Ultralytics YOLO architectures for tasks such as zero-shot object detection, oriented object detection, segmentation, pose estimation, and more. See the Supported Models table for details.

NOTE: The instructions below are comprehensive, but for convenience, we recommend using the download_ultralytics_models.py script. It can download a YOLO model or read one from a PyTorch file and perform the required conversions automatically. See Model Conversion Scripts for more information.

If you prefer to prepare the model manually, the following minimal Python script converts an Ultralytics model stored in the yolo.pt file to IR format:

from ultralytics import YOLO

# Load a YOLO PyTorch model
model = YOLO("yolo.pt")

# Export the model
model.export(format="openvino", int8=True)  # creates 'yolo_int8_openvino_model/'

The argument passed to YOLO() can be either a local PyTorch file or an identifier for a model available from Ultralytics, such as yolo26n.pt.

The directory created by the exporter contains all files required to use the model with the gvadetect element. No further modifications are required.

Other YOLO Models

NOTE: To obtain ready-to-use versions of the models described below, we recommend using the download_public_models.sh script. See Download Public Models for details.

YOLOv7

Model preparation:

git clone https://github.com/WongKinYiu/yolov7.git
cd yolov7
python3 export.py --weights yolov7.pt --grid --dynamic-batch
# Convert the model to OpenVINO format FP32 precision
ovc yolov7.onnx --compress_to_fp16=False
# Convert the model to OpenVINO format FP16 precision
ovc yolov7.onnx

When used with gvadetect, this model requires label-file=coco_80cl.txt and model-proc=yolo-v7.json.

Older YOLOv5 Versions

YOLOv5 models trained with ultralytics/yolov5 are not compatible with the ultralytics/ultralytics library or the download_ultralytics_models.py script.

Preparing YOLOv5 7.0 from Ultralytics therefore involves two steps.

1. Convert the PyTorch model to Intel® OpenVINO™ format:

   git clone https://github.com/ultralytics/yolov5
   cd yolov5
   wget https://github.com/ultralytics/yolov5/releases/download/v7.0/yolov5s.pt
   python3 export.py --weights yolov5s.pt --include openvino --dynamic
   

2. Then reshape the model to enable a dynamic batch size while keeping the other dimensions fixed:

   from openvino.runtime import Core
   from openvino.runtime import save_model
   core = Core()
   model = core.read_model("yolov5s_openvino_model/yolov5s.xml")
   model.reshape([-1, 3, 640, 640])
   save_model(model, "yolov5s.xml")
   

YOLOX

An Intel® OpenVINO™ version of the model can be obtained from the ONNX file:

wget https://github.com/Megvii-BaseDetection/YOLOX/releases/download/0.1.1rc0/yolox_s.onnx
ovc yolox_s.onnx --compress_to_fp16=False

Model Usage

See Samples for detailed examples of Deep Learning Streamer pipelines using different YOLO models.