Article directory
overview
This document mainly describes the platform and the method of exporting models python
using openvino
module inference .YOLOv5
IR
The document mainly includes the following contents:
openvino
module installation- Description of the model format
- The basic API interface of openvino, including
初始化
,模型加载
,模型参数获取
,模型推理
etc. - Preprocessing of image data
- Post-processing of inference results, including conversion
NMS
ofcxcywh
coordinates toxyxy
coordinates, etc. - Key method calls and parameter descriptions
- complete sample code
1. Environment deployment
pre-installation environment
(Windows) Visual Studio 2019
Anaconda3
orMiniConda3
Note: The model
openvino
must be confirmed to be used , otherwise it cannot be used.CPU
Intel
openvino
Install
pytorch 1.7.1+cu110
onnxruntime-gpu 1.7.0
conda create -n openvino python=3.8 -y
conda activate ort
pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
pip install onnxruntime-gpu==1.7.0
ONNX model conversion
The official pre-training model of YOLOv5 can be downloaded through the official link. The model format is .Download linkpt
. The official project provides a script for converting format model to format model. Project link
YOLOv5
pt
ONNX
Model export command:
python export --weights yolov5s.pt --include openvino
Note: For the installation and configuration of the environment required for exporting files to execute instructions, please refer to the official project
README
documentation, and will not repeat them here.
After the execution of the command is completed, a folder yolov5s.pt
named . The file structure is as follows:yolov5s_openvino_model
IR
yolov5s_openvino_model
├── yolov5s.bin
├── yolov5s.mapping
└── yolov5s.xml
2. OpenVINO basic API
2.1 Initialization
from openvino.runtime import Core
core = Core()
2.2 Get device information
devices = core.available_devices
for device in devices:
device_name = core.get_property(device, "FULL_DEVICE_NAME")
print(f"{
device}: {
device_name}")
Example code output:
device_name: 12th Gen Intel(R) Core(TM) i5-12400F
device_name: NVIDIA GeForce RTX 2060 SUPER (dGPU)
2.3 Loading the model
Load OpenVINO
the intermediate representation ( IR
) model file and return ExecutableNetwork
the object.
# load the openvino IR model
yolo_model_path = "weights/yolov5s_openvino_model/yolov5s.xml"
model = core.read_model(model=yolo_model_path)
compiled_model = core.compile_model(model=model, device_name="AUTO")
2.4 Get model input and output information
# 获取输入层信息
input_layer = model.inputs[0]
print(f"input_layer: {
input_layer}")
# 获取输出层信息
output_layer = model.outputs
print(f"output_layer: {
output_layer}")
Example code output:
input_layer: <Output: names[images] shape[1,3,640,640] type: f32>
output_layer: [<Output: names[output] shape[1,25200,85] type: f32>, <Output: names[345] shape[1,3,80,80,85] type: f32>, <Output: names[403] shape[1,3,40,40,85] type: f32>, <Output: names[461] shape[1,3,20,20,85] type: f32>]
It can be seen that the model has only one input layer, but has 4 output layers, which conforms to the YOLOv5
multi-layer output characteristics of the model.
During inference, we only need to focus on output
this output layer.
Furthermore, for the subsequent preprocessing and postprocessing of model inference, we need to obtain relevant information about the input and output layers of the model, including:
- The name, shape, and type of the input layer
- The name, shape, and type corresponding to the output layer
# 如果输入层只有一层,直接调用any_name获取输入层名称
input_name = input_layer.any_name
print(f"input_name: {
input_name}")
N, C, H, W = input_layer.shape
print(f"N: {
N}, C: {
C}, H: {
H}, W: {
W}")
input_dtype = input_layer.element_type
print(f"input_dtype: {
input_dtype}")
output_name = output_layer[0].any_name
print(f"output_name: {
output_name}")
output_shape = output_layer[0].shape
print(f"output_shape: {
output_shape}")
output_dtype = output_layer[0].element_type
print(f"output_dtype: {
output_dtype}")
Example output:
input_name: images
N: 1, C: 3, H: 640, W: 640
input_dtype: <Type: 'float32'>
output_name: output
output_shape: [1,25200,85]
output_dtype: <Type: 'float32'>
2.5 Model Reasoning
image = cv2.imread(str(image_filename))
# image.shape = (height, width, channels)
# N,C,H,W = batch size, number of channels, height, width.
N, C, H, W = input_layer.shape
# OpenCV resize expects the destination size as (width, height).
resized_image = cv2.resize(src=image, dsize=(W, H))
# resized_image.shape = (height, width, channels)
input_data = np.expand_dims(np.transpose(resized_image, (2, 0, 1)), 0).astype(np.float32)
# input_data.shape = (N, C, H, W)
# for single input models only
result = compiled_model(input_data)[output_layer]
# for multiple inputs in a list
result = compiled_model([input_data])[output_layer]
# or using a dictionary, where the key is input tensor name or index
result = compiled_model({
input_layer.any_name: input_data})[output_layer]
3. Key code
2.1 Image data preprocessing
Data preprocessing steps include resize, normalization, color channel conversion, NCWH dimension conversion, etc.
resize
Before, there was a very common trick to deal with non-square pictures, that is, to calculate the longest side of the graphic, based on this longest side, create a square, place the original graphic in the upper left corner, and fill the rest with black. The advantage of doing this is that the aspect ratio of the original graphics will not be changed, and the content of the original graphics will not be changed at the same time.
# image preprocessing, the trick is to make the frame to be a square but not twist the image
row, col, _ = frame.shape # get the row and column of the origin frame array
_max = max(row, col) # get the max value of row and column
input_image = np.zeros((_max, _max, 3), dtype=np.uint8) # create a new array with the max value
input_image[:row, :col, :] = frame # paste the original frame to make the input_image to be a square
After completing the filling of the image, continue to perform operations such as resize, normalization, and color channel conversion.
blob = cv2.dnn.blobFromImage(image, scalefactor=1 / 255.0, size=(640,640), swapRB=True, crop=False)
image
: Input image data,numpy.ndarray
the formatshape
is(H,W,C)
, and the channel order isBGR
.scalefactor
: Image data normalization coefficient, usually1/255.0
.size
: The image resize size is subject to the input requirements of the model, here is(640,640)
.swapRB
: Whether to exchange color channels, that is, convertBGR
toRGB
True
indicate exchange,False
and indicate not to exchange. Sinceopencv
the order of color channels for reading image data isBGR
, andYOLOv5
the input requirement of the model isRGB
, color channels need to be exchanged here.crop
: Whether to crop the image,False
means no cropping.
blobFromImage
The function returns a four-dimensional Mat object (NCHW dimensions order), and the shape of the data is(1,3,640,640)
2.4 Post-processing of inference results
Since there are a large number of overlapping inference results bbox
, they need to be NMS
processed, and then filtered according to each bbox
confidence level and the confidence threshold set by the user, and finally the final bbox
, corresponding category and confidence level are obtained.
2.4.1 NMS
opencv-python
Modules provide NMSBoxes
methods for NMS
processing.
cv2.dnn.NMSBoxes(bboxes, scores, score_threshold, nms_threshold, eta=None, top_k=None)
bboxes
:bbox
list, forshape
,(N,4)
forN
quantitybbox
,4
forbbox
.x,y,w,h
scores
:bbox
The corresponding confidence list,shape
is(N,1)
,N
isbbox
the quantity.score_threshold
: Confidence threshold, which is smaller than the thresholdbbox
will be filtered.nms_threshold
:NMS
threshold
NMSBoxes
The return value of the function is bbox
a list of indices, which shape
is the number.(M,)
M
bbox
2.4.2 score_threshold filtering
According to the NMS
processed bbox
index list, filter confidence less score_threshold
than bbox
.
2.4.3 bbox coordinate conversion and restoration
YOLOv5
bbox
The coordinates output by the model are in cxcywh
a format that needs to be converted to xyxy
a format. In addition, since the image has been manipulated before resize
, the coordinates need to be bbox
restored to the size of the original image.
The conversion method is as follows:
# 获取原始图片的尺寸(填充后)
image_width, image_height, _ = input_image.shape
# 计算缩放比
x_factor = image_width / INPUT_WIDTH # 640
y_factor = image_height / INPUT_HEIGHT # 640
# 将cxcywh坐标转换为xyxy坐标
x1 = int((x - w / 2) * x_factor)
y1 = int((y - h / 2) * y_factor)
w = int(w * x_factor)
h = int(h * y_factor)
x2 = x1 + w
y2 = y1 + h
x1
, y1
, x2
, y2
are the coordinates bbox
of xyxy
.
4. Sample code
There are two source codes, one of which is splicing and calling functions, which is more convenient for debugging, and the other is packaged into classes, which is convenient for integration into other projects.
3.1 Unpackaged
"""
this file is to demonstrate how to use openvino to do inference with yolov5 model exported from onnx to openvino format
"""
from typing import List
import cv2
import numpy as np
import time
from pathlib import Path
from openvino.runtime import Core
def build_model(model_path: str) -> cv2.dnn_Net:
"""
build the model with opencv dnn module
Args:
model_path: the path of the model, the model should be in onnx format
Returns:
the model object
"""
# load the model
core = Core()
model = core.read_model(model_path)
for device in core.available_devices:
print(device)
compiled_model = core.compile_model(model=model, device_name="AUTO")
# output_layer = compiled_model.output(0)
return compiled_model
def inference(image: np.ndarray, model: cv2.dnn_Net) -> np.ndarray:
"""
inference the model with the input image
Args:
image: the input image in numpy array format, the shape should be (height, width, channel),
the color channels should be in GBR order, like the original opencv image format
model: the model object
Returns:
the output data of the model, the shape should be (1, 25200, nc+5), nc is the number of classes
"""
# image preprocessing, include resize, normalization, channel swap like BGR to RGB, and convert to blob format
# get a 4-dimensional Mat with NCHW dimensions order.
blob = cv2.dnn.blobFromImage(image, 1 / 255.0, (INPUT_WIDTH, INPUT_HEIGHT), swapRB=True, crop=False)
output_layer = model.output(0)
outs = model([blob])[output_layer]
start = time.perf_counter()
# inference
# outs = model.forward()
end = time.perf_counter()
# print("inference time: ", end - start)
# the shape of the output data is (1, 25200, nc+5), nc is the number of classes
return outs
def xywh_to_xyxy(bbox_xywh, image_width, image_height):
"""
Convert bounding box coordinates from (center_x, center_y, width, height) to (x_min, y_min, x_max, y_max) format.
Parameters:
bbox_xywh (list or tuple): Bounding box coordinates in (center_x, center_y, width, height) format.
image_width (int): Width of the image.
image_height (int): Height of the image.
Returns:
tuple: Bounding box coordinates in (x_min, y_min, x_max, y_max) format.
"""
center_x, center_y, width, height = bbox_xywh
x_min = max(0, int(center_x - width / 2))
y_min = max(0, int(center_y - height / 2))
x_max = min(image_width - 1, int(center_x + width / 2))
y_max = min(image_height - 1, int(center_y + height / 2))
return x_min, y_min, x_max, y_max
def wrap_detection(
input_image: np.ndarray,
output_data: np.ndarray,
labels: List[str],
confidence_threshold: float = 0.6
) -> (List[int], List[float], List[List[int]]):
# the shape of the output_data is (25200,5+nc),
# the first 5 elements are [x, y, w, h, confidence], the rest are prediction scores of each class
image_width, image_height, _ = input_image.shape
x_factor = image_width / INPUT_WIDTH
y_factor = image_height / INPUT_HEIGHT
# transform the output_data[:, 0:4] from (x, y, w, h) to (x_min, y_min, x_max, y_max)
# output_data[:, 0:4] = np.apply_along_axis(xywh_to_xyxy, 1, output_data[:, 0:4], image_width, image_height)
indices = cv2.dnn.NMSBoxes(output_data[:, 0:4].tolist(), output_data[:, 4].tolist(), 0.6, 0.4)
# print(indices)
raw_boxes = output_data[:, 0:4][indices]
raw_confidences = output_data[:, 4][indices]
raw_class_prediction_probabilities = output_data[:, 5:][indices]
criteria = raw_confidences > confidence_threshold
raw_class_prediction_probabilities = raw_class_prediction_probabilities[criteria]
raw_boxes = raw_boxes[criteria]
raw_confidences = raw_confidences[criteria]
bounding_boxes, confidences, class_ids = [], [], []
for class_prediction_probability, box, confidence in zip(raw_class_prediction_probabilities, raw_boxes,
raw_confidences):
# find the least and most probable classes' indices and their probabilities
# min_val, max_val, min_loc, mac_loc = cv2.minMaxLoc(class_prediction_probability)
most_probable_class_index = np.argmax(class_prediction_probability)
label = labels[most_probable_class_index]
confidence = float(confidence)
x, y, w, h = box
left = int((x - 0.5 * w) * x_factor)
top = int((y - 0.5 * h) * y_factor)
width = int(w * x_factor)
height = int(h * y_factor)
bounding_box = [left, top, width, height]
bounding_boxes.append(bounding_box)
confidences.append(confidence)
class_ids.append(most_probable_class_index)
return class_ids, confidences, bounding_boxes
coco_class_names = ["person", "bicycle", "car", "motorcycle", "airplane", "bus", "train", "truck", "boat",
"traffic light", "fire hydrant", "stop sign", "parking meter", "bench", "bird", "cat",
"dog", "horse", "sheep", "cow", "elephant", "bear", "zebra", "giraffe", "backpack",
"umbrella", "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard", "sports ball",
"kite", "baseball bat", "baseball glove", "skateboard", "surfboard", "tennis racket",
"bottle", "wine glass", "cup", "fork", "knife", "spoon", "bowl", "banana", "apple",
"sandwich", "orange", "broccoli", "carrot", "hot dog", "pizza", "donut", "cake", "chair",
"couch", "potted plant", "bed", "dining table", "toilet", "tv", "laptop", "mouse",
"remote", "keyboard", "cell phone", "microwave", "oven", "toaster", "sink",
"refrigerator", "book", "clock", "vase", "scissors", "teddy bear", "hair drier",
"toothbrush"]
# generate different colors for coco classes
colors = np.random.uniform(0, 255, size=(len(coco_class_names), 3))
INPUT_WIDTH = 640
INPUT_HEIGHT = 640
CONFIDENCE_THRESHOLD = 0.7
NMS_THRESHOLD = 0.45
def video_detector(video_src):
cap = cv2.VideoCapture(video_src)
# 3. inference and show the result in a loop
while cap.isOpened():
success, frame = cap.read()
start = time.perf_counter()
if not success:
break
# image preprocessing, the trick is to make the frame to be a square but not twist the image
row, col, _ = frame.shape # get the row and column of the origin frame array
_max = max(row, col) # get the max value of row and column
input_image = np.zeros((_max, _max, 3), dtype=np.uint8) # create a new array with the max value
input_image[:row, :col, :] = frame # paste the original frame to make the input_image to be a square
# inference
output_data = inference(input_image, net) # the shape of output_data is (1, 25200, 85)
# define coco dataset class names dictionary
# 4. wrap the detection result
class_ids, confidences, boxes = wrap_detection(input_image, output_data[0], coco_class_names)
# wrap_detection(input_image, output_data[0], coco_class_names) ##
# 5. draw the detection result on the frame
for (class_id, confidence, box) in zip(class_ids, confidences, boxes):
color = colors[int(class_id) % len(colors)]
label = coco_class_names[int(class_id)]
# cv2.rectangle(frame, box, color, 2)
# print(type(box), box[0], box[1], box[2], box[3], box)
xmin, ymin, width, height = box
cv2.rectangle(frame, (xmin, ymin), (xmin + width, ymin + height), color, 2)
# cv2.rectangle(frame, box, color, 2)
# cv2.rectangle(frame, [box[0], box[1], box[2], box[3]], color, thickness=2)
# cv2.rectangle(frame, (box[0], box[1] - 20), (box[0] + 100, box[1]), color, -1)
cv2.putText(frame, str(label), (box[0], box[1] - 5), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 0), 2)
finish = time.perf_counter()
FPS = round(1.0 / (finish - start), 2)
cv2.putText(frame, str(FPS), (10, 50), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 0), 2)
# 6. show the frame
cv2.imshow("frame", frame)
# 7. press 'q' to exit
if cv2.waitKey(1) == ord('q'):
break
# 8. release the capture and destroy all windows
cap.release()
cv2.destroyAllWindows()
if __name__ == '__main__':
# there are 4 steps to use opencv dnn module to inference onnx model exported by yolov5 and show the result
# 1. load the model
model_path = Path("weights/yolov5s_openvino_model/yolov5s.xml")
# model_path = Path("weights/POT_INT8_openvino_model/yolov5s_int8.xml")
net = build_model(str(model_path))
# 2. load the video capture
video_source = 0
# video_source = 'rtsp://admin:[email protected]:554/h264/ch1/main/av_stream'
video_detector(video_source)
exit(0)
3.2 Encapsulated into a class call
import cv2
import numpy as np
from pathlib import Path
import time
from typing import List
from glob import glob
from openvino.runtime import Core
class YoloV5OpenvinoInference:
def __init__(self, model_path: str,
imgsize: int = 640,
labels: List[str] = None,
score_threshold: float = 0.6,
nms_threshold: float = 0.45):
self.load_model(model_path)
self.imgsize = imgsize
self.score_threshold = score_threshold
self.nms_threshold = nms_threshold
self.coco_class_names = ["person", "bicycle", "car", "motorcycle", "airplane", "bus", "train", "truck", "boat",
"traffic light", "fire hydrant", "stop sign", "parking meter", "bench", "bird", "cat",
"dog", "horse", "sheep", "cow", "elephant", "bear", "zebra", "giraffe", "backpack",
"umbrella", "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard",
"sports ball",
"kite", "baseball bat", "baseball glove", "skateboard", "surfboard", "tennis racket",
"bottle", "wine glass", "cup", "fork", "knife", "spoon", "bowl", "banana", "apple",
"sandwich", "orange", "broccoli", "carrot", "hot dog", "pizza", "donut", "cake",
"chair",
"couch", "potted plant", "bed", "dining table", "toilet", "tv", "laptop", "mouse",
"remote", "keyboard", "cell phone", "microwave", "oven", "toaster", "sink",
"refrigerator", "book", "clock", "vase", "scissors", "teddy bear", "hair drier",
"toothbrush"]
if labels is None or len(labels) == 0:
self.labels = self.coco_class_names
else:
self.labels = labels
self.colors = np.random.uniform(0, 255, size=(len(self.labels), 3))
self.x_factor = 1
self.y_factor = 1
self.xyxy = True
def load_model(self, model_path: str) -> None:
if not Path(model_path).exists():
raise FileNotFoundError(f"model file {
model_path} not found")
self.core = Core()
self.loaded_model = self.core.read_model(model_path)
self.compiled_model = self.core.compile_model(model=self.loaded_model, device_name="AUTO")
self.output_layer = self.loaded_model.output(0)
def preprocess(self, image: np.ndarray) -> np.ndarray:
row, col, _ = image.shape
_max = max(row, col)
input_image = np.zeros((_max, _max, 3), dtype=np.uint8)
input_image[:row, :col, :] = image
image_width, image_height, _ = input_image.shape
self.x_factor = image_width / self.imgsize
self.y_factor = image_height / self.imgsize
return input_image
def inference(self, image: np.ndarray) -> np.ndarray:
image = self.preprocess(image)
blob = cv2.dnn.blobFromImage(image, 1 / 255.0, (self.imgsize, self.imgsize), swapRB=True, crop=False)
start = time.perf_counter()
outs = self.compiled_model([blob])
# TODO
# 此处由于model.output(0)返回和推理输出的数据类型可能不一致(openvino.runtime.ConstOutput和openvino.runtime.Output)
# 导致取出输出层数据报错,暂时先这样处理
output_layer = [item for item in outs if item.any_name == "output"][0]
outs = outs[output_layer]
# outs = self.compiled_model([blob])['output']
end = time.perf_counter()
print("inference time: ", end - start)
return outs[0]
def wrap_detection(self, result: np.ndarray, to_xyxy: bool = True) -> (List[int], List[float], List[List[int]]):
# using NMS algorithm to filter out the overlapping bounding boxes
indices = cv2.dnn.NMSBoxes(result[:, 0:4].tolist(), result[:, 4].tolist(), 0.6, 0.4)
# get the real data after filtering
result = result[indices]
# filter the bounding boxes with confidence lower than the threshold
result = result[result[:, 4] > self.score_threshold]
bounding_boxes, confidences, classes = [], [], []
for item in result:
box = item[0:4]
confidence = float(item[4])
class_prediction_probability = item[5:]
most_probable_class_index = np.argmax(class_prediction_probability)
#
x, y, w, h = box
left = int((x - 0.5 * w) * self.x_factor)
top = int((y - 0.5 * h) * self.y_factor)
width = int(w * self.x_factor)
height = int(h * self.y_factor)
if to_xyxy:
self.xyxy = True
bounding_box = [left, top, left + width, top + height]
else:
self.xyxy = False
bounding_box = [left, top, width, height]
bounding_boxes.append(bounding_box)
confidences.append(confidence)
classes.append(most_probable_class_index)
return classes, confidences, bounding_boxes
def detect(self, image: np.ndarray, visualize=True) -> np.ndarray:
img = self.preprocess(image)
result = self.inference(img)
class_ids, confidences, boxes = self.wrap_detection(result, to_xyxy=True)
if visualize:
for (class_id, confidence, box) in zip(class_ids, confidences, boxes):
color = yolo_v5.colors[int(class_id) % len(yolo_v5.colors)]
label = yolo_v5.coco_class_names[int(class_id)]
if self.xyxy:
cv2.rectangle(image, (box[0], box[1]), (box[2], box[3]), color, 2)
else:
cv2.rectangle(image, box, color, 2)
cv2.rectangle(image, (box[0], box[1] - 20), (box[0] + 100, box[1]), color, -1)
cv2.putText(image, str(label), (box[0], box[1] - 5), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 0), 2)
return image
def video_detector(video_src):
cap = cv2.VideoCapture(video_src)
while cap.isOpened():
success, frame = cap.read()
if not success:
break
frame = yolo_v5.detect(frame)
cv2.imshow("frame", frame)
if cv2.waitKey(1) == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
def image_detector(image_src):
image = cv2.imread(image_src)
image = yolo_v5.detect(image)
cv2.imshow("image", image)
cv2.waitKey(0)
cv2.destroyAllWindows()
if __name__ == '__main__':
model_path = "weights/yolov5s_openvino_model/yolov5s.xml"
yolo_v5 = YoloV5OpenvinoInference(model_path=model_path)
video_source = 0
# video_source = 'rtsp://admin:[email protected]:554/h264/ch1/main/av_stream'
video_detector(video_source)
# image_path = "data/images/bus.jpg"
# image_detector(image_path)
exit(0)