Convert and Optimize YOLOv8 real-time object detection with OpenVINO™

This Jupyter notebook can be launched on-line, opening an interactive environment in a browser window. You can also make a local installation. Choose one of the following options:

Real-time object detection is often used as a key component in computer vision systems. Applications that use real-time object detection models include video analytics, robotics, autonomous vehicles, multi-object tracking and object counting, medical image analysis, and many others.

This tutorial demonstrates step-by-step instructions on how to run and optimize PyTorch YOLOv8 with OpenVINO. We consider the steps required for object detection scenario.

The tutorial consists of the following steps: - Prepare the PyTorch model. - Download and prepare a dataset. - Validate the original model. - Convert the PyTorch model to OpenVINO IR. - Validate the converted model. - Prepare and run optimization pipeline. - Compare performance of the FP32 and quantized models. - Compare accuracy of the FP32 and quantized models. - Other optimization possibilities with OpenVINO api - Live demo

Table of contents:

• Get PyTorch model

  • Prerequisites

• Instantiate model

  • Convert model to OpenVINO IR

  • Verify model inference

  • Preprocessing

  • Postprocessing

  • Select inference device

  • Test on single image

• Check model accuracy on the dataset

  • Download the validation dataset

  • Define validation function

  • Configure Validator helper and create DataLoader

• Optimize model using NNCF Post-training Quantization API

  • Validate Quantized model inference

• Compare the Original and Quantized Models

  • Compare performance object detection models

  • Validate quantized model accuracy

• Next steps

  • Async inference pipeline

  • Integration preprocessing to model

    • Initialize PrePostProcessing API

    • Define input data format

    • Describe preprocessing steps

    • Integrating Steps into a Model

• Live demo

  • Run Live Object Detection

Get PyTorch model

Generally, PyTorch models represent an instance of the torch.nn.Module class, initialized by a state dictionary with model weights. We will use the YOLOv8 nano model (also known as yolov8n) pre-trained on a COCO dataset, which is available in this repo. Similar steps are also applicable to other YOLOv8 models. Typical steps to obtain a pre-trained model: 1. Create an instance of a model class. 2. Load a checkpoint state dict, which contains the pre-trained model weights. 3. Turn the model to evaluation for switching some operations to inference mode.

In this case, the creators of the model provide an API that enables converting the YOLOv8 model to ONNX and then to OpenVINO IR. Therefore, we do not need to do these steps manually.

Prerequisites

Install necessary packages.

%pip install -q "openvino>=2023.1.0" "nncf>=2.5.0"
%pip install -q "torch>=2.1" "torchvision>=0.16" "ultralytics==8.0.43" onnx --extra-index-url https://download.pytorch.org/whl/cpu

Import required utility functions. The lower cell will download the notebook_utils Python module from GitHub.

from pathlib import Path

# Fetch the notebook utils script from the openvino_notebooks repo
import urllib.request
urllib.request.urlretrieve(
    url='https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/main/notebooks/utils/notebook_utils.py',
    filename='notebook_utils.py'
)

from notebook_utils import download_file, VideoPlayer

Define utility functions for drawing results

from typing import Tuple, Dict
import cv2
import numpy as np
from ultralytics.yolo.utils.plotting import colors


def plot_one_box(box:np.ndarray, img:np.ndarray,
                 color:Tuple[int, int, int] = None,
                 label:str = None, line_thickness:int = 5):
    """
    Helper function for drawing single bounding box on image
    Parameters:
        x (np.ndarray): bounding box coordinates in format [x1, y1, x2, y2]
        img (no.ndarray): input image
        color (Tuple[int, int, int], *optional*, None): color in BGR format for drawing box, if not specified will be selected randomly
        label (str, *optonal*, None): box label string, if not provided will not be provided as drowing result
        line_thickness (int, *optional*, 5): thickness for box drawing lines
    """
    # Plots one bounding box on image img
    tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1  # line/font thickness
    color = color or [random.randint(0, 255) for _ in range(3)]
    c1, c2 = (int(box[0]), int(box[1])), (int(box[2]), int(box[3]))
    cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)
    if label:
        tf = max(tl - 1, 1)  # font thickness
        t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
        c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3
        cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA)  # filled
        cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA)

    return img


def draw_results(results:Dict, source_image:np.ndarray, label_map:Dict):
    """
    Helper function for drawing bounding boxes on image
    Parameters:
        image_res (np.ndarray): detection predictions in format [x1, y1, x2, y2, score, label_id]
        source_image (np.ndarray): input image for drawing
        label_map; (Dict[int, str]): label_id to class name mapping
    Returns:
        Image with boxes
    """
    boxes = results["det"]
    for idx, (*xyxy, conf, lbl) in enumerate(boxes):
        label = f'{label_map[int(lbl)]} {conf:.2f}'
        source_image = plot_one_box(xyxy, source_image, label=label, color=colors(int(lbl)), line_thickness=1)
    return source_image

# Download a test sample
IMAGE_PATH = Path('./data/coco_bike.jpg')
download_file(
    url='https://storage.openvinotoolkit.org/repositories/openvino_notebooks/data/data/image/coco_bike.jpg',
    filename=IMAGE_PATH.name,
    directory=IMAGE_PATH.parent
)

'data/coco_bike.jpg' already exists.

PosixPath('/home/ea/work/openvino_notebooks/notebooks/230-yolov8-optimization/data/coco_bike.jpg')

Instantiate model

There are several models available in the original repository, targeted for different tasks. For loading the model, required to specify a path to the model checkpoint. It can be some local path or name available on models hub (in this case model checkpoint will be downloaded automatically).

Making prediction, the model accepts a path to input image and returns list with Results class object. Results contains boxes for object detection model. Also it contains utilities for processing results, for example, plot() method for drawing.

Let us consider the examples:

models_dir = Path('./models')
models_dir.mkdir(exist_ok=True)

from PIL import Image
from ultralytics import YOLO

DET_MODEL_NAME = "yolov8n"

det_model = YOLO(models_dir / f'{DET_MODEL_NAME}.pt')
label_map = det_model.model.names

res = det_model(IMAGE_PATH)
Image.fromarray(res[0].plot()[:, :, ::-1])

2023-10-05 19:15:51.230030: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0.
2023-10-05 19:15:51.269549: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-10-05 19:15:51.909328: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
Ultralytics YOLOv8.0.43 🚀 Python-3.8.10 torch-2.0.1+cpu CPU
YOLOv8n summary (fused): 168 layers, 3151904 parameters, 0 gradients, 8.7 GFLOPs

image 1/1 /home/ea/work/openvino_notebooks/notebooks/230-yolov8-optimization/data/coco_bike.jpg: 480x640 2 bicycles, 2 cars, 1 dog, 48.7ms
Speed: 2.6ms preprocess, 48.7ms inference, 1.3ms postprocess per image at shape (1, 3, 640, 640)

Convert model to OpenVINO IR

YOLOv8 provides API for convenient model exporting to different formats including OpenVINO IR. model.export is responsible for model conversion. We need to specify the format, and additionally, we can preserve dynamic shapes in the model.

# object detection model
det_model_path = models_dir / f"{DET_MODEL_NAME}_openvino_model/{DET_MODEL_NAME}.xml"
if not det_model_path.exists():
    det_model.export(format="openvino", dynamic=True, half=False)

Verify model inference

To test model work, we create inference pipeline similar to model.predict method. The pipeline consists of preprocessing step, inference of OpenVINO model and results post-processing to get results.

Preprocessing

Model input is a tensor with the [-1, 3, -1, -1] shape in the N, C, H, W format, where * N - number of images in batch (batch size) * C - image channels * H - image height * W - image width

The model expects images in RGB channels format and normalized in [0, 1] range. Although the model supports dynamic input shape with preserving input divisibility to 32, it is recommended to use static shapes, for example, 640x640 for better efficiency. To resize images to fit model size letterbox, resize approach is used, where the aspect ratio of width and height is preserved.

To keep a specific shape, preprocessing automatically enables padding.

from typing import Tuple
from ultralytics.yolo.utils import ops
import torch
import numpy as np


def letterbox(img: np.ndarray, new_shape:Tuple[int, int] = (640, 640), color:Tuple[int, int, int] = (114, 114, 114), auto:bool = False, scale_fill:bool = False, scaleup:bool = False, stride:int = 32):
    """
    Resize image and padding for detection. Takes image as input,
    resizes image to fit into new shape with saving original aspect ratio and pads it to meet stride-multiple constraints

    Parameters:
      img (np.ndarray): image for preprocessing
      new_shape (Tuple(int, int)): image size after preprocessing in format [height, width]
      color (Tuple(int, int, int)): color for filling padded area
      auto (bool): use dynamic input size, only padding for stride constrins applied
      scale_fill (bool): scale image to fill new_shape
      scaleup (bool): allow scale image if it is lower then desired input size, can affect model accuracy
      stride (int): input padding stride
    Returns:
      img (np.ndarray): image after preprocessing
      ratio (Tuple(float, float)): hight and width scaling ratio
      padding_size (Tuple(int, int)): height and width padding size


    """
    # Resize and pad image while meeting stride-multiple constraints
    shape = img.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # only scale down, do not scale up (for better test mAP)
        r = min(r, 1.0)

    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    if auto:  # minimum rectangle
        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
    elif scale_fill:  # stretch
        dw, dh = 0.0, 0.0
        new_unpad = (new_shape[1], new_shape[0])
        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios

    dw /= 2  # divide padding into 2 sides
    dh /= 2

    if shape[::-1] != new_unpad:  # resize
        img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return img, ratio, (dw, dh)


def preprocess_image(img0: np.ndarray):
    """
    Preprocess image according to YOLOv8 input requirements.
    Takes image in np.array format, resizes it to specific size using letterbox resize and changes data layout from HWC to CHW.

    Parameters:
      img0 (np.ndarray): image for preprocessing
    Returns:
      img (np.ndarray): image after preprocessing
    """
    # resize
    img = letterbox(img0)[0]

    # Convert HWC to CHW
    img = img.transpose(2, 0, 1)
    img = np.ascontiguousarray(img)
    return img


def image_to_tensor(image:np.ndarray):
    """
    Preprocess image according to YOLOv8 input requirements.
    Takes image in np.array format, resizes it to specific size using letterbox resize and changes data layout from HWC to CHW.

    Parameters:
      img (np.ndarray): image for preprocessing
    Returns:
      input_tensor (np.ndarray): input tensor in NCHW format with float32 values in [0, 1] range
    """
    input_tensor = image.astype(np.float32)  # uint8 to fp32
    input_tensor /= 255.0  # 0 - 255 to 0.0 - 1.0

    # add batch dimension
    if input_tensor.ndim == 3:
        input_tensor = np.expand_dims(input_tensor, 0)
    return input_tensor

Postprocessing

The model output contains detection boxes candidates, it is a tensor with the [-1,84,-1] shape in the B,84,N format, where:

• B - batch size

• N - number of detection boxes

For getting the final prediction, we need to apply a non-maximum suppression algorithm and rescale box coordinates to the original image size.

Finally, detection box has the [x, y, h, w, class_no_1, …, class_no_80] format, where:

• (x, y) - raw coordinates of box center

• h, w - raw height and width of the box

• class_no_1, …, class_no_80 - probability distribution over the classes.

def postprocess(
    pred_boxes:np.ndarray,
    input_hw:Tuple[int, int],
    orig_img:np.ndarray,
    min_conf_threshold:float = 0.25,
    nms_iou_threshold:float = 0.7,
    agnosting_nms:bool = False,
    max_detections:int = 300,
):
    """
    YOLOv8 model postprocessing function. Applied non maximum supression algorithm to detections and rescale boxes to original image size
    Parameters:
        pred_boxes (np.ndarray): model output prediction boxes
        input_hw (np.ndarray): preprocessed image
        orig_image (np.ndarray): image before preprocessing
        min_conf_threshold (float, *optional*, 0.25): minimal accepted confidence for object filtering
        nms_iou_threshold (float, *optional*, 0.45): minimal overlap score for removing objects duplicates in NMS
        agnostic_nms (bool, *optiona*, False): apply class agnostinc NMS approach or not
        max_detections (int, *optional*, 300):  maximum detections after NMS
    Returns:
       pred (List[Dict[str, np.ndarray]]): list of dictionary with det - detected boxes in format [x1, y1, x2, y2, score, label]
    """
    nms_kwargs = {"agnostic": agnosting_nms, "max_det":max_detections}
    preds = ops.non_max_suppression(
        torch.from_numpy(pred_boxes),
        min_conf_threshold,
        nms_iou_threshold,
        nc=80,
        **nms_kwargs
    )

    results = []
    for i, pred in enumerate(preds):
        shape = orig_img[i].shape if isinstance(orig_img, list) else orig_img.shape
        if not len(pred):
            results.append({"det": [], "segment": []})
            continue
        pred[:, :4] = ops.scale_boxes(input_hw, pred[:, :4], shape).round()
        results.append({"det": pred})

    return results

Select inference device

Select device from dropdown list for running inference using OpenVINO

import ipywidgets as widgets
import openvino as ov

core = ov.Core()

device = widgets.Dropdown(
    options=core.available_devices + ["AUTO"],
    value='AUTO',
    description='Device:',
    disabled=False,
)

device

Dropdown(description='Device:', index=2, options=('CPU', 'GPU', 'AUTO'), value='AUTO')

Test on single image

Now, once we have defined preprocessing and postprocessing steps, we are ready to check model prediction for object detection.

core = ov.Core()

det_ov_model = core.read_model(det_model_path)
if device.value != "CPU":
    det_ov_model.reshape({0: [1, 3, 640, 640]})
det_compiled_model = core.compile_model(det_ov_model, device.value)


def detect(image:np.ndarray, model:ov.Model):
    """
    OpenVINO YOLOv8 model inference function. Preprocess image, runs model inference and postprocess results using NMS.
    Parameters:
        image (np.ndarray): input image.
        model (Model): OpenVINO compiled model.
    Returns:
        detections (np.ndarray): detected boxes in format [x1, y1, x2, y2, score, label]
    """
    preprocessed_image = preprocess_image(image)
    input_tensor = image_to_tensor(preprocessed_image)
    result = model(input_tensor)
    boxes = result[model.output(0)]
    input_hw = input_tensor.shape[2:]
    detections = postprocess(pred_boxes=boxes, input_hw=input_hw, orig_img=image)
    return detections

input_image = np.array(Image.open(IMAGE_PATH))
detections = detect(input_image, det_compiled_model)[0]
image_with_boxes = draw_results(detections, input_image, label_map)

Image.fromarray(image_with_boxes)

Check model accuracy on the dataset

For comparing the optimized model result with the original, it is good to know some measurable results in terms of model accuracy on the validation dataset.

Download the validation dataset

YOLOv8 is pre-trained on the COCO dataset, so to evaluate the model accuracy we need to download it. According to the instructions provided in the YOLOv8 repo, we also need to download annotations in the format used by the author of the model, for use with the original model evaluation function.

NOTE: The initial dataset download may take a few minutes to complete. The download speed will vary depending on the quality of your internet connection.

from zipfile import ZipFile

DATA_URL = "http://images.cocodataset.org/zips/val2017.zip"
LABELS_URL = "https://github.com/ultralytics/yolov5/releases/download/v1.0/coco2017labels-segments.zip"
CFG_URL = "https://raw.githubusercontent.com/ultralytics/ultralytics/8ebe94d1e928687feaa1fee6d5668987df5e43be/ultralytics/datasets/coco.yaml"

OUT_DIR = Path('./datasets')

DATA_PATH = OUT_DIR / "val2017.zip"
LABELS_PATH = OUT_DIR / "coco2017labels-segments.zip"
CFG_PATH = OUT_DIR / "coco.yaml"

download_file(DATA_URL, DATA_PATH.name, DATA_PATH.parent)
download_file(LABELS_URL, LABELS_PATH.name, LABELS_PATH.parent)
download_file(CFG_URL, CFG_PATH.name, CFG_PATH.parent)

if not (OUT_DIR / "coco/labels").exists():
    with ZipFile(LABELS_PATH , "r") as zip_ref:
        zip_ref.extractall(OUT_DIR)
    with ZipFile(DATA_PATH , "r") as zip_ref:
        zip_ref.extractall(OUT_DIR / 'coco/images')

'datasets/val2017.zip' already exists.
'datasets/coco2017labels-segments.zip' already exists.

datasets/coco.yaml:   0%|          | 0.00/1.25k [00:00<?, ?B/s]

Define validation function

from tqdm.notebook import tqdm
from ultralytics.yolo.utils.metrics import ConfusionMatrix


def test(model:ov.Model, core:ov.Core, data_loader:torch.utils.data.DataLoader, validator, num_samples:int = None):
    """
    OpenVINO YOLOv8 model accuracy validation function. Runs model validation on dataset and returns metrics
    Parameters:
        model (Model): OpenVINO model
        data_loader (torch.utils.data.DataLoader): dataset loader
        validator: instance of validator class
        num_samples (int, *optional*, None): validate model only on specified number samples, if provided
    Returns:
        stats: (Dict[str, float]) - dictionary with aggregated accuracy metrics statistics, key is metric name, value is metric value
    """
    validator.seen = 0
    validator.jdict = []
    validator.stats = []
    validator.batch_i = 1
    validator.confusion_matrix = ConfusionMatrix(nc=validator.nc)
    model.reshape({0: [1, 3, -1, -1]})
    compiled_model = core.compile_model(model)
    for batch_i, batch in enumerate(tqdm(data_loader, total=num_samples)):
        if num_samples is not None and batch_i == num_samples:
            break
        batch = validator.preprocess(batch)
        results = compiled_model(batch["img"])
        preds = torch.from_numpy(results[compiled_model.output(0)])
        preds = validator.postprocess(preds)
        validator.update_metrics(preds, batch)
    stats = validator.get_stats()
    return stats


def print_stats(stats:np.ndarray, total_images:int, total_objects:int):
    """
    Helper function for printing accuracy statistic
    Parameters:
        stats: (Dict[str, float]) - dictionary with aggregated accuracy metrics statistics, key is metric name, value is metric value
        total_images (int) -  number of evaluated images
        total objects (int)
    Returns:
        None
    """
    print("Boxes:")
    mp, mr, map50, mean_ap = stats['metrics/precision(B)'], stats['metrics/recall(B)'], stats['metrics/mAP50(B)'], stats['metrics/mAP50-95(B)']
    # Print results
    s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Labels', 'Precision', 'Recall', 'mAP@.5', 'mAP@.5:.95')
    print(s)
    pf = '%20s' + '%12i' * 2 + '%12.3g' * 4  # print format
    print(pf % ('all', total_images, total_objects, mp, mr, map50, mean_ap))
    if 'metrics/precision(M)' in stats:
        s_mp, s_mr, s_map50, s_mean_ap = stats['metrics/precision(M)'], stats['metrics/recall(M)'], stats['metrics/mAP50(M)'], stats['metrics/mAP50-95(M)']
        # Print results
        s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Labels', 'Precision', 'Recall', 'mAP@.5', 'mAP@.5:.95')
        print(s)
        pf = '%20s' + '%12i' * 2 + '%12.3g' * 4  # print format
        print(pf % ('all', total_images, total_objects, s_mp, s_mr, s_map50, s_mean_ap))

Configure Validator helper and create DataLoader

The original model repository uses a Validator wrapper, which represents the accuracy validation pipeline. It creates dataloader and evaluation metrics and updates metrics on each data batch produced by the dataloader. Besides that, it is responsible for data preprocessing and results postprocessing. For class initialization, the configuration should be provided. We will use the default setup, but it can be replaced with some parameters overriding to test on custom data. The model has connected the ValidatorClass method, which creates a validator class instance.

from ultralytics.yolo.utils import DEFAULT_CFG
from ultralytics.yolo.cfg import get_cfg
from ultralytics.yolo.data.utils import check_det_dataset

args = get_cfg(cfg=DEFAULT_CFG)
args.data = str(CFG_PATH)

det_validator = det_model.ValidatorClass(args=args)

det_validator.data = check_det_dataset(args.data)
det_data_loader = det_validator.get_dataloader("datasets/coco", 1)

val: Scanning datasets/coco/labels/val2017.cache... 4952 images, 48 backgrounds, 0 corrupt: 100%|██████████| 5000/5000 [00:00<?, ?it/s]

det_validator.is_coco = True
det_validator.class_map = ops.coco80_to_coco91_class()
det_validator.names = det_model.model.names
det_validator.metrics.names = det_validator.names
det_validator.nc = det_model.model.model[-1].nc

After definition test function and validator creation, we are ready for getting accuracy metrics.

NOTE: Model evaluation is time consuming process and can take several minutes, depending on the hardware. For reducing calculation time, we define num_samples parameter with evaluation subset size, but in this case, accuracy can be noncomparable with originally reported by the authors of the model, due to validation subset difference. To validate the models on the full dataset set ``NUM_TEST_SAMPLES = None``.

NUM_TEST_SAMPLES = 300

fp_det_stats = test(det_ov_model, core, det_data_loader, det_validator, num_samples=NUM_TEST_SAMPLES)

0%|          | 0/300 [00:00<?, ?it/s]

print_stats(fp_det_stats, det_validator.seen, det_validator.nt_per_class.sum())

Boxes:
               Class      Images      Labels   Precision      Recall      mAP@.5  mAP@.5:.95
                 all         300        2145       0.594       0.543       0.579       0.417

print_stats reports the following list of accuracy metrics:

• Precision is the degree of exactness of the model in identifying only relevant objects.

• Recall measures the ability of the model to detect all ground truths objects.

• mAP@t - mean average precision, represented as area under the Precision-Recall curve aggregated over all classes in the dataset, where t is the Intersection Over Union (IOU) threshold, degree of overlapping between ground truth and predicted objects. Therefore, mAP@.5 indicates that mean average precision is calculated at 0.5 IOU threshold, mAP@.5:.95 - is calculated on range IOU thresholds from 0.5 to 0.95 with step 0.05.

Optimize model using NNCF Post-training Quantization API

NNCF provides a suite of advanced algorithms for Neural Networks inference optimization in OpenVINO with minimal accuracy drop. We will use 8-bit quantization in post-training mode (without the fine-tuning pipeline) to optimize YOLOv8.

The optimization process contains the following steps:

1. Create a Dataset for quantization.

2. Run nncf.quantize for getting an optimized model.

3. Serialize OpenVINO IR model, using the openvino.runtime.serialize function.

Reuse validation dataloader in accuracy testing for quantization. For that, it should be wrapped into the nncf.Dataset object and define a transformation function for getting only input tensors.

import nncf  # noqa: F811
from typing import Dict


def transform_fn(data_item:Dict):
    """
    Quantization transform function. Extracts and preprocess input data from dataloader item for quantization.
    Parameters:
       data_item: Dict with data item produced by DataLoader during iteration
    Returns:
        input_tensor: Input data for quantization
    """
    input_tensor = det_validator.preprocess(data_item)['img'].numpy()
    return input_tensor


quantization_dataset = nncf.Dataset(det_data_loader, transform_fn)

INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino

The nncf.quantize function provides an interface for model quantization. It requires an instance of the OpenVINO Model and quantization dataset. Optionally, some additional parameters for the configuration quantization process (number of samples for quantization, preset, ignored scope, etc.) can be provided. YOLOv8 model contains non-ReLU activation functions, which require asymmetric quantization of activations. To achieve a better result, we will use a mixed quantization preset. It provides symmetric quantization of weights and asymmetric quantization of activations. For more accurate results, we should keep the operation in the postprocessing subgraph in floating point precision, using the ignored_scope parameter.

NOTE: Model post-training quantization is time-consuming process. Be patient, it can take several minutes depending on your hardware.

ignored_scope = nncf.IgnoredScope(
    types=["Multiply", "Subtract", "Sigmoid"],  # ignore operations
    names=[
        "/model.22/dfl/conv/Conv",           # in the post-processing subgraph
        "/model.22/Add",
        "/model.22/Add_1",
        "/model.22/Add_2",
        "/model.22/Add_3",
        "/model.22/Add_4",
        "/model.22/Add_5",
        "/model.22/Add_6",
        "/model.22/Add_7",
        "/model.22/Add_8",
        "/model.22/Add_9",
        "/model.22/Add_10"
    ]
)


# Detection model
quantized_det_model = nncf.quantize(
    det_ov_model,
    quantization_dataset,
    preset=nncf.QuantizationPreset.MIXED,
    ignored_scope=ignored_scope
)

INFO:nncf:12 ignored nodes was found by name in the NNCFGraph
INFO:nncf:9 ignored nodes was found by types in the NNCFGraph
INFO:nncf:Not adding activation input quantizer for operation: 128 /model.22/Sigmoid
INFO:nncf:Not adding activation input quantizer for operation: 156 /model.22/dfl/conv/Conv
INFO:nncf:Not adding activation input quantizer for operation: 178 /model.22/Sub
INFO:nncf:Not adding activation input quantizer for operation: 179 /model.22/Add_10
INFO:nncf:Not adding activation input quantizer for operation: 193 /model.22/Sub_1
INFO:nncf:Not adding activation input quantizer for operation: 218 /model.22/Mul_5

Statistics collection: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 300/300 [00:31<00:00,  9.54it/s]
Applying Fast Bias correction: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 63/63 [00:03<00:00, 18.94it/s]

from openvino.runtime import serialize
int8_model_det_path = models_dir / f'{DET_MODEL_NAME}_openvino_int8_model/{DET_MODEL_NAME}.xml'
print(f"Quantized detection model will be saved to {int8_model_det_path}")
serialize(quantized_det_model, str(int8_model_det_path))

Quantized detection model will be saved to models/yolov8n_openvino_int8_model/yolov8n.xml

Validate Quantized model inference

nncf.quantize returns the OpenVINO Model class instance, which is suitable for loading on a device for making predictions. INT8 model input data and output result formats have no difference from the floating point model representation. Therefore, we can reuse the same detect function defined above for getting the INT8 model result on the image.

device

Dropdown(description='Device:', index=2, options=('CPU', 'GPU', 'AUTO'), value='AUTO')

if device.value != "CPU":
    quantized_det_model.reshape({0: [1, 3, 640, 640]})
quantized_det_compiled_model = core.compile_model(quantized_det_model, device.value)
input_image = np.array(Image.open(IMAGE_PATH))
detections = detect(input_image, quantized_det_compiled_model)[0]
image_with_boxes = draw_results(detections, input_image, label_map)

Image.fromarray(image_with_boxes)

Compare the Original and Quantized Models

Compare performance object detection models

Finally, use the OpenVINO Benchmark Tool to measure the inference performance of the FP32 and INT8 models.

NOTE: For more accurate performance, it is recommended to run benchmark_app in a terminal/command prompt after closing other applications. Run benchmark_app -m <model_path> -d CPU -shape "<input_shape>" to benchmark async inference on CPU on specific input data shape for one minute. Change CPU to GPU to benchmark on GPU. Run benchmark_app --help to see an overview of all command-line options.

device

Dropdown(description='Device:', index=2, options=('CPU', 'GPU', 'AUTO'), value='AUTO')

# Inference FP32 model (OpenVINO IR)
!benchmark_app -m $det_model_path -d $device.value -api async -shape "[1,3,640,640]"

[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ WARNING ] Default duration 120 seconds is used for unknown device AUTO
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2023.2.0-12690-0ee0b4d9561
[ INFO ]
[ INFO ] Device info:
[ INFO ] AUTO
[ INFO ] Build ................................. 2023.2.0-12690-0ee0b4d9561
[ INFO ]
[ INFO ]
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(AUTO) performance hint will be set to PerformanceMode.THROUGHPUT.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 16.72 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     images (node: images) : f32 / [...] / [?,3,?,?]
[ INFO ] Model outputs:
[ INFO ]     output0 (node: output0) : f32 / [...] / [?,84,?]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[ INFO ] Reshaping model: 'images': [1,3,640,640]
[ INFO ] Reshape model took 11.83 ms
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     images (node: images) : u8 / [N,C,H,W] / [1,3,640,640]
[ INFO ] Model outputs:
[ INFO ]     output0 (node: output0) : f32 / [...] / [1,84,8400]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 424.62 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: torch_jit
[ INFO ]   EXECUTION_DEVICES: ['CPU']
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 12
[ INFO ]   MULTI_DEVICE_PRIORITIES: CPU
[ INFO ]   CPU:
[ INFO ]     AFFINITY: Affinity.CORE
[ INFO ]     CPU_DENORMALS_OPTIMIZATION: False
[ INFO ]     CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE: 1.0
[ INFO ]     ENABLE_CPU_PINNING: True
[ INFO ]     ENABLE_HYPER_THREADING: True
[ INFO ]     EXECUTION_DEVICES: ['CPU']
[ INFO ]     EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
[ INFO ]     INFERENCE_NUM_THREADS: 36
[ INFO ]     INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]     NETWORK_NAME: torch_jit
[ INFO ]     NUM_STREAMS: 12
[ INFO ]     OPTIMAL_NUMBER_OF_INFER_REQUESTS: 12
[ INFO ]     PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]     PERFORMANCE_HINT_NUM_REQUESTS: 0
[ INFO ]     PERF_COUNT: False
[ INFO ]     SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
[ INFO ]   MODEL_PRIORITY: Priority.MEDIUM
[ INFO ]   LOADED_FROM_CACHE: False
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'images'!. This input will be filled with random values!
[ INFO ] Fill input 'images' with random values
[Step 10/11] Measuring performance (Start inference asynchronously, 12 inference requests, limits: 120000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 34.66 ms
[Step 11/11] Dumping statistics report
[ INFO ] Execution Devices:['CPU']
[ INFO ] Count:            15024 iterations
[ INFO ] Duration:         120272.02 ms
[ INFO ] Latency:
[ INFO ]    Median:        77.25 ms
[ INFO ]    Average:       95.83 ms
[ INFO ]    Min:           60.23 ms
[ INFO ]    Max:           270.56 ms
[ INFO ] Throughput:   124.92 FPS

# Inference INT8 model (OpenVINO IR)
!benchmark_app -m $int8_model_det_path -d $device.value -api async -shape "[1,3,640,640]" -t 15

[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2023.2.0-12690-0ee0b4d9561
[ INFO ]
[ INFO ] Device info:
[ INFO ] AUTO
[ INFO ] Build ................................. 2023.2.0-12690-0ee0b4d9561
[ INFO ]
[ INFO ]
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(AUTO) performance hint will be set to PerformanceMode.THROUGHPUT.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 73.93 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     images (node: images) : f32 / [...] / [1,3,?,?]
[ INFO ] Model outputs:
[ INFO ]     output0 (node: output0) : f32 / [...] / [1,84,21..]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[ INFO ] Reshaping model: 'images': [1,3,640,640]
[ INFO ] Reshape model took 56.37 ms
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     images (node: images) : u8 / [N,C,H,W] / [1,3,640,640]
[ INFO ] Model outputs:
[ INFO ]     output0 (node: output0) : f32 / [...] / [1,84,8400]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 1742.92 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: torch_jit
[ INFO ]   EXECUTION_DEVICES: ['CPU']
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 18
[ INFO ]   MULTI_DEVICE_PRIORITIES: CPU
[ INFO ]   CPU:
[ INFO ]     AFFINITY: Affinity.CORE
[ INFO ]     CPU_DENORMALS_OPTIMIZATION: False
[ INFO ]     CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE: 1.0
[ INFO ]     ENABLE_CPU_PINNING: True
[ INFO ]     ENABLE_HYPER_THREADING: True
[ INFO ]     EXECUTION_DEVICES: ['CPU']
[ INFO ]     EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
[ INFO ]     INFERENCE_NUM_THREADS: 36
[ INFO ]     INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]     NETWORK_NAME: torch_jit
[ INFO ]     NUM_STREAMS: 18
[ INFO ]     OPTIMAL_NUMBER_OF_INFER_REQUESTS: 18
[ INFO ]     PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]     PERFORMANCE_HINT_NUM_REQUESTS: 0
[ INFO ]     PERF_COUNT: False
[ INFO ]     SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
[ INFO ]   MODEL_PRIORITY: Priority.MEDIUM
[ INFO ]   LOADED_FROM_CACHE: False
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'images'!. This input will be filled with random values!
[ INFO ] Fill input 'images' with random values
[Step 10/11] Measuring performance (Start inference asynchronously, 18 inference requests, limits: 15000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 58.19 ms
[Step 11/11] Dumping statistics report
[ INFO ] Execution Devices:['CPU']
[ INFO ] Count:            3150 iterations
[ INFO ] Duration:         15116.07 ms
[ INFO ] Latency:
[ INFO ]    Median:        79.96 ms
[ INFO ]    Average:       85.97 ms
[ INFO ]    Min:           56.29 ms
[ INFO ]    Max:           154.33 ms
[ INFO ] Throughput:   208.39 FPS

Validate quantized model accuracy

As we can see, there is no significant difference between INT8 and float model result in a single image test. To understand how quantization influences model prediction precision, we can compare model accuracy on a dataset.

int8_det_stats = test(quantized_det_model, core, det_data_loader, det_validator, num_samples=NUM_TEST_SAMPLES)

0%|          | 0/300 [00:00<?, ?it/s]

print("FP32 model accuracy")
print_stats(fp_det_stats, det_validator.seen, det_validator.nt_per_class.sum())

print("INT8 model accuracy")
print_stats(int8_det_stats, det_validator.seen, det_validator.nt_per_class.sum())

FP32 model accuracy
Boxes:
               Class      Images      Labels   Precision      Recall      mAP@.5  mAP@.5:.95
                 all         300        2145       0.594       0.543       0.579       0.417
INT8 model accuracy
Boxes:
               Class      Images      Labels   Precision      Recall      mAP@.5  mAP@.5:.95
                 all         300        2145       0.623       0.517       0.572       0.406

Great! Looks like accuracy was changed, but not significantly and it meets passing criteria.

Next steps

This section contains suggestions on how to additionally improve the performance of your application using OpenVINO.

Async inference pipeline

The key advantage of the Async API is that when a device is busy with inference, the application can perform other tasks in parallel (for example, populating inputs or scheduling other requests) rather than wait for the current inference to complete first. To understand how to perform async inference using openvino, refer to Async API tutorial

Integration preprocessing to model

Preprocessing API enables making preprocessing a part of the model reducing application code and dependency on additional image processing libraries. The main advantage of Preprocessing API is that preprocessing steps will be integrated into the execution graph and will be performed on a selected device (CPU/GPU etc.) rather than always being executed on CPU as part of an application. This will improve selected device utilization.

For more information, refer to the overview of Preprocessing API.

For example, we can integrate converting input data layout and normalization defined in image_to_tensor function.

The integration process consists of the following steps: 1. Initialize a PrePostProcessing object. 2. Define the input data format. 3. Describe preprocessing steps. 4. Integrating Steps into a Model.

Initialize PrePostProcessing API

The openvino.preprocess.PrePostProcessor class enables specifying preprocessing and postprocessing steps for a model.

from openvino.preprocess import PrePostProcessor

ppp = PrePostProcessor(quantized_det_model)

Define input data format

To address particular input of a model/preprocessor, the input(input_id) method, where input_id is a positional index or input tensor name for input in model.inputs, if a model has a single input, input_id can be omitted. After reading the image from the disc, it contains U8 pixels in the [0, 255] range and is stored in the NHWC layout. To perform a preprocessing conversion, we should provide this to the tensor description.

ppp.input(0).tensor().set_shape([1, 640, 640, 3]).set_element_type(ov.Type.u8).set_layout(ov.Layout('NHWC'))
pass

To perform layout conversion, we also should provide information about layout expected by model

Describe preprocessing steps

Our preprocessing function contains the following steps: * Convert the data type from U8 to FP32. * Convert the data layout from NHWC to NCHW format. * Normalize each pixel by dividing on scale factor 255.

ppp.input(input_id).preprocess() is used for defining a sequence of preprocessing steps:

ppp.input(0).preprocess().convert_element_type(ov.Type.f32).convert_layout(ov.Layout('NCHW')).scale([255., 255., 255.])

print(ppp)

Input "images":
    User's input tensor: [1,640,640,3], [N,H,W,C], u8
    Model's expected tensor: [1,3,?,?], [N,C,H,W], f32
    Pre-processing steps (3):
      convert type (f32): ([1,640,640,3], [N,H,W,C], u8) -> ([1,640,640,3], [N,H,W,C], f32)
      convert layout [N,C,H,W]: ([1,640,640,3], [N,H,W,C], f32) -> ([1,3,640,640], [N,C,H,W], f32)
      scale (255,255,255): ([1,3,640,640], [N,C,H,W], f32) -> ([1,3,640,640], [N,C,H,W], f32)

Integrating Steps into a Model

Once the preprocessing steps have been finished, the model can be finally built. Additionally, we can save a completed model to OpenVINO IR, using openvino.runtime.serialize.

quantized_model_with_preprocess = ppp.build()
serialize(quantized_model_with_preprocess, str(int8_model_det_path.with_name(f"{DET_MODEL_NAME}_with_preprocess.xml")))

The model with integrated preprocessing is ready for loading to a device. Now, we can skip these preprocessing steps in detect function:

def detect_without_preprocess(image:np.ndarray, model:ov.Model):
    """
    OpenVINO YOLOv8 model with integrated preprocessing inference function. Preprocess image, runs model inference and postprocess results using NMS.
    Parameters:
        image (np.ndarray): input image.
        model (Model): OpenVINO compiled model.
    Returns:
        detections (np.ndarray): detected boxes in format [x1, y1, x2, y2, score, label]
    """
    output_layer = model.output(0)
    img = letterbox(image)[0]
    input_tensor = np.expand_dims(img, 0)
    input_hw = img.shape[:2]
    result = model(input_tensor)[output_layer]
    detections = postprocess(result, input_hw, image)
    return detections


compiled_model = core.compile_model(quantized_model_with_preprocess, device.value)
input_image = np.array(Image.open(IMAGE_PATH))
detections = detect_without_preprocess(input_image, compiled_model)[0]
image_with_boxes = draw_results(detections, input_image, label_map)

Image.fromarray(image_with_boxes)

Live demo

The following code runs model inference on a video:

import collections
import time
from IPython import display


# Main processing function to run object detection.
def run_object_detection(source=0, flip=False, use_popup=False, skip_first_frames=0, model=det_model, device=device.value):
    player = None
    if device != "CPU":
        model.reshape({0: [1, 3, 640, 640]})
    compiled_model = core.compile_model(model, device)
    try:
        # Create a video player to play with target fps.
        player = VideoPlayer(
            source=source, flip=flip, fps=30, skip_first_frames=skip_first_frames
        )
        # Start capturing.
        player.start()
        if use_popup:
            title = "Press ESC to Exit"
            cv2.namedWindow(
                winname=title, flags=cv2.WINDOW_GUI_NORMAL | cv2.WINDOW_AUTOSIZE
            )

        processing_times = collections.deque()
        while True:
            # Grab the frame.
            frame = player.next()
            if frame is None:
                print("Source ended")
                break
            # If the frame is larger than full HD, reduce size to improve the performance.
            scale = 1280 / max(frame.shape)
            if scale < 1:
                frame = cv2.resize(
                    src=frame,
                    dsize=None,
                    fx=scale,
                    fy=scale,
                    interpolation=cv2.INTER_AREA,
                )
            # Get the results.
            input_image = np.array(frame)

            start_time = time.time()
            # model expects RGB image, while video capturing in BGR
            detections = detect(input_image[:, :, ::-1], compiled_model)[0]
            stop_time = time.time()

            image_with_boxes = draw_results(detections, input_image, label_map)
            frame = image_with_boxes

            processing_times.append(stop_time - start_time)
            # Use processing times from last 200 frames.
            if len(processing_times) > 200:
                processing_times.popleft()

            _, f_width = frame.shape[:2]
            # Mean processing time [ms].
            processing_time = np.mean(processing_times) * 1000
            fps = 1000 / processing_time
            cv2.putText(
                img=frame,
                text=f"Inference time: {processing_time:.1f}ms ({fps:.1f} FPS)",
                org=(20, 40),
                fontFace=cv2.FONT_HERSHEY_COMPLEX,
                fontScale=f_width / 1000,
                color=(0, 0, 255),
                thickness=1,
                lineType=cv2.LINE_AA,
            )
            # Use this workaround if there is flickering.
            if use_popup:
                cv2.imshow(winname=title, mat=frame)
                key = cv2.waitKey(1)
                # escape = 27
                if key == 27:
                    break
            else:
                # Encode numpy array to jpg.
                _, encoded_img = cv2.imencode(
                    ext=".jpg", img=frame, params=[cv2.IMWRITE_JPEG_QUALITY, 100]
                )
                # Create an IPython image.
                i = display.Image(data=encoded_img)
                # Display the image in this notebook.
                display.clear_output(wait=True)
                display.display(i)
    # ctrl-c
    except KeyboardInterrupt:
        print("Interrupted")
    # any different error
    except RuntimeError as e:
        print(e)
    finally:
        if player is not None:
            # Stop capturing.
            player.stop()
        if use_popup:
            cv2.destroyAllWindows()

Run Live Object Detection

Use a webcam as the video input. By default, the primary webcam is set with source=0. If you have multiple webcams, each one will be assigned a consecutive number starting at 0. Set flip=True when using a front-facing camera. Some web browsers, especially Mozilla Firefox, may cause flickering. If you experience flickering, set use_popup=True.

NOTE: To use this notebook with a webcam, you need to run the notebook on a computer with a webcam. If you run the notebook on a remote server (for example, in Binder or Google Colab service), the webcam will not work. By default, the lower cell will run model inference on a video file. If you want to try live inference on your webcam set WEBCAM_INFERENCE = True

Run the object detection:

WEBCAM_INFERENCE = False

if WEBCAM_INFERENCE:
    VIDEO_SOURCE = 0  # Webcam
else:
    VIDEO_SOURCE = 'https://storage.openvinotoolkit.org/repositories/openvino_notebooks/data/data/video/people.mp4'

device

Dropdown(description='Device:', index=2, options=('CPU', 'GPU', 'AUTO'), value='AUTO')

run_object_detection(source=VIDEO_SOURCE, flip=True, use_popup=False, model=det_ov_model, device=device.value)

Source ended