第一步：neo（intel牌加速显卡驱动）下载 -- UBUNTU 20.04
本人踩坑点，不下neo只能cpu运行。
1.
sudo apt install ocl-icd-libopencl12.
mkdir neo && cd neo3.（改成国内源会快一点）
wget https://mirror.ghproxy.com/https://github.com/intel/intel-graphics-compiler/releases/download/igc-1.0.11702.1/intel-igc-core_1.0.11702.1_amd64.1.deb
wget https://mirror.ghproxy.com/https://github.com/intel/intel-graphics-compiler/releases/download/igc-1.0.11702.1/intel-igc-opencl_1.0.11702.1_amd64.deb
wget https://mirror.ghproxy.com/https://github.com/intel/compute-runtime/releases/download/22.35.24055/intel-level-zero-gpu-dbgsym_1.3.24055_amd64.ddeb
wget https://mirror.ghproxy.com/https://github.com/intel/compute-runtime/releases/download/22.35.24055/intel-level-zero-gpu_1.3.24055_amd64.deb
wget https://mirror.ghproxy.com/https://github.com/intel/compute-runtime/releases/download/22.35.24055/intel-opencl-icd_22.35.24055_amd64.deb
wget https://mirror.ghproxy.com/https://github.com/intel/compute-runtime/releases/download/22.35.24055/libigdgmm12_22.1.3_amd64.deb4.
sudo dpkg -i *.deb第二步：openvino-dev（用来转化模型，onnx-->ir）
pip install openvino-dev -i https://pypi.tuna.tsinghua.edu.cn/simple
第三步：.pt-->.onnx-->.xml/.bin
1.export.py:
.pt转.onnx
2.
mo --input_model weights/yolov5s.onnx --output_dir weightsvino.py如下：
"""
brief :Editor sztu cjh,ljl
"""
import time
import openvino.runtime as ov
import cv2
import numpy as np
import openvino.preprocess as op
from openvino import InferRequest
from collections import deque
import logging
import multiprocessing
from numba import jit
# logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(message)s')
class Colors:
# Ultralytics color palette https://ultralytics.com/
def __init__(self):
# hex = matplotlib.colors.TABLEAU_COLORS.values()
hexs = ('FF3838', 'FF9D97', 'FF701F', 'FFB21D', 'CFD231', '48F90A', '92CC17', '3DDB86', '1A9334', '00D4BB',
'2C99A8', '00C2FF', '344593', '6473FF', '0018EC', '8438FF', '520085', 'CB38FF', 'FF95C8', 'FF37C7')
self.palette = [self.hex2rgb(f'#{c}') for c in hexs]
self.n = len(self.palette)
def rgb_call(self, i, bgr=False):
c = self.palette[int(i) % self.n]
return (c[2], c[1], c[0]) if bgr else c
@staticmethod
def hex2rgb(h): # rgb order (PIL)
return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4))
@staticmethod
def rgb2hex(rgb):
'''RGB转HEX
:param rgb: RGB颜色元组，Tuple[int, int, int]
:return: int or str
'''
r, g, b = rgb
result = r + (g << 8) + (b << 16)
return hex(result)[2:]
def cap_capture(stack, source, top) -> None:
"""
:param top2: 缓冲栈容量
:param stack2: Manager.list对象,作为第二通道的视频流
:param source: 摄像头参数
:param stack: Manager.list对象,作为第一通道的视频流
:param top: 缓冲栈容量
:return: None
"""
global cap
if len(source) <= 2:
try:
cap = cv2.VideoCapture(int(source))
except Exception:
pass
else:
cap = cv2.VideoCapture(source)
while True:
ret, Frame = cap.read()
if ret:
stack.append(Frame)
if len(stack) >= top:
del stack[:]
@jit(nopython=True)
def accle_for(detections: np.ndarray, confidence: float):
detections_out = []
for prediction in detections:
_confidence_ = prediction[4].item() # 获取置信度
# 第一次进行过滤，过滤掉大部分没用信息，但是设定阈值不能太高
if _confidence_ >= confidence / 2:
detections_out.append(prediction)
return detections_out
class Vino(Colors):
def __init__(self, model_path="/home/nuc2/PycharmProjects/yolov5-master/best_ball.xml"
, weights_path="/home/nuc2/PycharmProjects/yolov5-master/best_ball.bin"
, conf_thres=0.55
, line_thickness=3
, iou_thres=0.55
, classes=list
, img_size=736
, device="GPU"
, name_porcess="Vino"):
super(Vino, self).__init__()
if img_size is None:
img_size = [736, 736]
self.confidence = conf_thres
self.line_thickness = line_thickness
self.device = device
self.iou_thres = iou_thres
self.classes = classes
self.img_size=img_size
# 在实例化的时候就开辟空间给内核激活
self.logger = logging.getLogger(name_porcess)
self.logger.info(f"{model_path, weights_path, conf_thres, line_thickness, iou_thres, device}")
self.inter = self.Core(model_path=model_path, weights_path=weights_path)
# 修改图像的亮度，brightness取值0～2 <1表示变暗 >1表示变亮
def letter_box(self, box_, img, shape_re, dw, dh, confidence_box, class_id) -> (np.ndarray, np.uint8):
color = self.rgb_call(class_id, True)
rx = img.shape[1] / (shape_re[1] - dw)
ry = img.shape[0] / (shape_re[0] - dh)
# box_ 代表着xyxy
box_[0] = int(rx * box_[0])
box_[1] = int(box_[1] * ry)
box_[2] = int(rx * box_[2]) + box_[0]
box_[3] = int(box_[3] * ry) + box_[1]
cv2.rectangle(img, (int(box_[0]), int(box_[1])), (int(box_[2]), int(box_[3])), color,
self.line_thickness) # 绘制物体框
cv2.putText(img, f"confidence:{str(int(confidence_box * 100))}%,id:{class_id}",
(int(box_[0]), int(box_[1]) - 5),
cv2.FONT_HERSHEY_SIMPLEX,
1, color, 2)
size_decrease = (int(img.shape[1] / 1), int(img.shape[0] / 1))
img_decrease = cv2.resize(img, size_decrease, interpolation=cv2.INTER_CUBIC)
return box_, img_decrease.astype(np.uint8)
def Core(self, model_path: str, weights_path: str) -> InferRequest:
try:
core = ov.Core()
# 读取用YOLOv5模型转换而来的IR模型
model = core.read_model(model_path, weights_path)
ppp = op.PrePostProcessor(model)
ppp.input().tensor().set_element_type(ov.Type.u8).set_layout(ov.Layout("NHWC")).set_color_format(
op.ColorFormat.BGR)
ppp.input().preprocess().convert_element_type(ov.Type.f32).convert_color(op.ColorFormat.RGB).scale(
[255., 255., 255.])
ppp.input().model().set_layout(ov.Layout("NCHW"))
ppp.output(0).tensor().set_element_type(ov.Type.f32)
_model = ppp.build()
# 加载模型，可用CPU or GPU
compilemodel = core.compile_model(model, self.device)
# 推理结果
results = compilemodel.create_infer_request()
self.logger.info("Success to Start")
return results
except:
self.logger.info("Failed to Start")
def resized(self, image, new_shape):
old_size = image.shape[:2]
# 记录新形状和原生图像矩形形状的比率
ratio = float(new_shape[-1] / max(old_size))
# 新尺寸
new_size = tuple([int(x * ratio) for x in old_size])
image = cv2.resize(image, (new_size[1], new_size[0]))
dw = new_shape[1] - new_size[1]
dh = new_shape[0] - new_size[0]
color = [100, 100, 100]
new_im = cv2.copyMakeBorder(image, 0, dh, 0, dw, cv2.BORDER_CONSTANT, value=color)
return new_im, dw, dh
def run(self, img) -> (np.uint8, deque, deque):
# 尺寸处理,变成一个贴合矩形
img_re, dw, dh = self.resized(img, (self.img_size, self.img_size))
shape_re = img_re.shape
# 获得输入张量
input_tensor = np.expand_dims(img_re, 0)
# 输入到推理引擎
self.inter.infer({0: input_tensor})
# # 获得推理结果
output = self.inter.get_output_tensor(0)
# # 获得检测数据
detections = output.data[0]
# # 使用deque，不使用list
boxes = deque()
class_ids = deque()
confidences_deque = deque()
# 以下是将输出的队列
det_out = deque()
detections = accle_for(detections, self.confidence)
for prediction in detections:
_confidence_ = prediction[4].item() # 获取置信度
# 第一次进行过滤，过滤掉大部分没用信息，但是设定阈值不能太高
# if _confidence_ >= self.confidence / 2:
classes_scores = prediction[5:]
t, w, e, max_indx = cv2.minMaxLoc(classes_scores)
class_id = max_indx[1]
if classes_scores[class_id] > .25:
confidences_deque.append(_confidence_)
class_ids.append(class_id)
# 得到有用信息
x, y, w, h = prediction[0].item(), prediction[1].item(), prediction[2].item(), prediction[3].item()
xmin = x - (w / 2)
ymin = y - (h / 2)
box = np.array([xmin, ymin, w, h])
boxes.append(box)
# 第二次进行筛选
indexes = cv2.dnn.NMSBoxes(boxes, confidences_deque, self.confidence, self.iou_thres)
detections = deque()
for i in indexes:
j = i.item()
if self.classes:
if class_ids[j] in self.classes:
detections.append({"ci": class_ids[j], "cf": confidences_deque[j], "bx": boxes[j]})
else:
detections.append({"ci": class_ids[j], "cf": confidences_deque[j], "bx": boxes[j]})
for detection in detections:
box = detection["bx"]
classId = detection["ci"]
confidence = detection["cf"]
if confidence < self.confidence:
break
box_, img = self.letter_box(box, img, shape_re, dw, dh, confidence, classId)
# 返回值存在4个,分别是box,置信度,目标物的id,和代表色
det_out.append((box_, confidence, classId))
return img.astype(np.uint8), det_out
if __name__ == '__main__':
model_path, weights_path = "/home/nuc2/PycharmProjects/yolov5-master/best_bucket.xml", "/home/nuc2/PycharmProjects/yolov5-master/best_bucket.bin"
confidence = 0.6
vino1 = Vino(model_path, weights_path, confidence)
# cap = cv2.VideoCapture(0)
Cap_data = multiprocessing.Manager().list()
p1 = multiprocessing.Process(target=cap_capture,
args=(Cap_data, "0", 50))
p1.start()
while True:
if len(Cap_data) != 0:
start = time.time()
# ret,frame=cap.read()
frame = Cap_data.pop()
frame, det = vino1.run(frame)
for xyxy, conf, cls in reversed(det):
# print(vino1.rgb2hex(vino1.rgb_call(cls)))
mid_pos = [int((xyxy[0] + xyxy[2]) / 2),
int((xyxy[1] + xyxy[3]) / 2)]
cv2.circle(frame, mid_pos, 2, (0, 255, 0), 2)
end = time.time()
print(end - start, "ms")
fps = 1 / (end - start)
cv2.putText(frame, f"{round(fps, 1)}", (50, 50), cv2.FONT_HERSHEY_COMPLEX, 1, (255, 0, 0), 2)
cv2.imshow("frame", frame)
if cv2.waitKey(1) & 0xFF == 27:
break