基于raspberry pi 3B+的OPENVINO实践之三
OPENVINO模型优化
OPENVINO开发环境部署
参考1：https://software.intel.com/en-us/openvino-toolkit/choose-download/free-download-linux
参考2：https://software.intel.com/en-us/articles/OpenVINO-ModelOptimizer
这一部分是使用virtualbox + ubuntu来实现的。
模型来源
https://github.com/movidius/ncappzoo.git
这个repo的模型本来是为测试NCSDK准备的，但是NCS2不支持，拿来主义，OPENVINO也可以使用。
测试一：Gender net
修改deploy.prototxt:
input_dim: 1
mv deploy.prototxt gender_net.prototxt
python3 model_optimizer/mo.py --input_model models/GenderNet/gender_net.caffemodel --output_dir .
生成三个文件，但是我们只需要.xml、.bin文件。
import cv2
import imutils
from imutils.video.pivideostream import PiVideoStream
from imutils.video import FPS
import time
stat = r'../gender/stat.txt'
xml = r'../gender/gender_net.xml'
file = r'../gender/gender_net.bin'
with open(stat,'r') as f:
lines = f.readlines()
mean = lines[0].strip().split(" ")
std = lines[1].strip().split(" ")
mean = [eval(x) for x in mean]
std = [eval(x) for x in std]
print(mean)
print(std)
net = cv2.dnn.readNet(xml, file)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_MYRIAD)
winName = 'gender test'
cv2.namedWindow(winName, cv2.WINDOW_AUTOSIZE)
print("[INFO] sampling THREADED frames from `picamera` module...")
vs = PiVideoStream().start()
time.sleep(2.0)
fps = FPS().start()
# loop over some frames...this time using the threaded stream
while True:
# grab the frame from the threaded video stream and resize it
# to have a maximum width of 400 pixels
frame = vs.read()
img = frame.copy()
img[:,:,0] = (img[:,:,0] - mean[0]) * std[0]
img[:,:,1] = (img[:,:,1] - mean[1]) * std[1]
img[:,:,2] = (img[:,:,2] - mean[2]) * std[2]
image = cv2.UMat(imutils.resize(img, width=227))
blob = cv2.dnn.blobFromImage(image, size=(227,227), ddepth=cv2.CV_8U)
net.setInput(blob)
out = net.forward()
if out[0,0] > out[0,1]:
cv2.putText(frame,"MAN detected",(10,10),cv2.FONT_HERSHEY_COMPLEX,0.5,(0,0,255),0)
#cv2.putText(frame,"MAN detected,[INFO] elasped time: {:.2f} FPS: {:.2f}".format(fps.elapsed(),fps.fps()),(10,10),cv2.FONT_HERSHEY_COMPLEX,6,(0,0,255),25)
else:
cv2.putText(frame,"Female detected",(10,10),cv2.FONT_HERSHEY_COMPLEX,0.5,(0,0,255),0)
cv2.imshow(winName,frame)
key = cv2.waitKey(1) & 0xFF
# if the `q` key was pressed, break from the loop
if key == ord("q"):
break
# update the FPS counter
fps.update()
# stop the timer and display FPS information
fps.stop()
print("[INFO] elasped time: {:.2f}".format(fps.elapsed()))
print("[INFO] approx. FPS: {:.2f}".format(fps.fps()))
# do a bit of cleanup
cv2.destroyAllWindows()
vs.stop()
- 测试结果可能与上位机训练的结果不一致，可能的原因：
** 模型转换过程丢失精度？
** 没有正确运行，比如图片预处理。这个与DEMO一致。
** 训练集合可能不是针对黄种人，造成数据缺口。这个是最可能的原因。 - 可能的改进措施：
** fine tuning
测试二：AgeNet
同上，结果相似，非常不准确。
Any idea?