Windows 10 使用 OpenVINO

• categories
• >
• NeuralNetwork
• >
• OpenVINO

---

参考

• Release Notes for Intel® Distribution of OpenVINO™ Toolkit 2020.3 LTS
• Introduction to Intel® Deep Learning Deployment Toolkit
• Data Structures
• (OPTIONAL) EXPORTING A MODEL FROM PYTORCH TO ONNX AND RUNNING IT USING ONNX RUNTIME
• Use the Model Downloader and Model Optimizer for the Intel® Distribution of OpenVINO™ Toolkit on Raspberry Pi*
• Install OpenVINO™ toolkit for Raspbian* OS

环境

• 转换平台: Windows 10
• Python版本: Python 3.7.
• 部署平台: Raspberry Pi 3B+
• 已有模型: pytorch
• 目标模型: onnx

安装OpenVINO toolkit至默认位置。我的默认位置: Introduction to Intel® Deep Learning Deployment Toolkit。

模型转换

pytorch -> onnx

用于单张图片判断代码和模型转换代码。

import torch
from torchvision import models
import torch.nn as nn
import os
import torchvision.transforms as transforms
import time
import sys
from PIL import Image

import torch.utils.model_zoo as model_zoo

model_name = 'C:\\Users\\milk\\Desktop\\model.pth'
classes = ['food', 'other', 'recycle', 'refuse']

print(classes)

model = models.densenet201(pretrained=True)

model.classifier = nn.Sequential(nn.Linear(1920, 256),
                                 nn.ReLU(),
                                 nn.Dropout(0.2),
                                 nn.Linear(256, len(classes)),
                                 nn.LogSoftmax(dim=1))

model.eval()
map_location = lambda storage, loc: storage
model.load_state_dict(torch.load(model_name, map_location=map_location))

transformations = transforms.Compose([
                                      transforms.ToTensor(),
                                      transforms.Normalize(
                                          [0.485, 0.456, 0.406],
                                          [0.229, 0.224, 0.225])
                                      ])

image = Image.open('ju.png')
image = image.convert('RGB')
image = image.resize((256, 256))
tick = time.time()
image = transformations(image)
image = torch.autograd.Variable(image[None, ...])
print("image shape: ", image.shape)
outputs = model(image)

print(time.time() - tick)
predict = outputs.max(1, keepdim=True)[1]
print("predict:\t", classes[predict])
print("time cost:\t", time.time() - tick)


torch_out = model(image)
torch.onnx.export(
  model, 
  image, 
  'C:\\users\\milk\\Desktop\\deploy\\onnx_model.onnx',
  export_params=True,
  output_names=['ashbin'])
print("finish")

模型优化

设置环境变量

环境变量文件: setvars.bat

PS C:\Program Files (x86)\IntelSWTools\openvino\bin> .\setupvars.bat
Python 3.7.9
[setupvars.bat] OpenVINO environment initialized

安装依赖库

路径: c:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\model_optimizer

路径下文件:

PS C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\model_optimizer> ls


    目录: C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\model_optimizer


Mode                LastWriteTime         Length Name
----                -------------         ------ ----
d-----        2020/12/3     11:54                extensions
d-----        2020/12/3     11:29                install_prerequisites
d-----        2020/12/3     11:54                mo
-a----         2020/7/2     17:28            999 mo.py
-a----         2020/7/2     17:28            932 mo_caffe.py
-a----         2020/7/2     17:28            932 mo_kaldi.py
-a----         2020/7/2     17:28            932 mo_mxnet.py
-a----         2020/7/2     17:28            929 mo_onnx.py
-a----         2020/7/2     17:28            923 mo_tf.py
-a----         2020/7/2     17:28            136 requirements.txt
-a----         2020/7/2     17:28             85 requirements_caffe.txt
-a----         2020/7/2     17:28             69 requirements_kaldi.txt
-a----         2020/7/2     17:28             90 requirements_mxnet.txt
-a----         2020/7/2     17:28             81 requirements_onnx.txt
-a----         2020/7/2     17:28             87 requirements_tf.txt
-a----         2020/7/2     17:28             41 version.txt

安装全部依赖requirements.txt或仅安装onnx支持依赖requirements_onnx.txt

pip install -r requirements_onnx.txt

转换模型

查看帮助:

python mo_onnx.py --help

执行转换:

将目标目录下的onnx_model.onnx转换为.bin和.xml文件。（时间大概一分钟）

PS C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\model_optimizer> python mo_onnx.py --log_level INFO --input_model C:\Users\milk\Desktop\deploy\onnx_model.onnx --output_dir C:\Users\milk\Desktop\deploy\onnx_model

成功输出结果:

[ SUCCESS ] Generated IR version 10 model.
[ SUCCESS ] XML file: C:\Users\milk\Desktop\deploy\onnx_model\onnx_model.xml
[ SUCCESS ] BIN file: C:\Users\milk\Desktop\deploy\onnx_model\onnx_model.bin
[ SUCCESS ] Total execution time: 46.46 seconds.
It's been a while, check for a new version of Intel(R) Distribution of OpenVINO(TM) toolkit here https://software.intel.com/en-us/openvino-toolkit/choose-download?cid=&source=upgrade&content=2020_3_LTS or on the GitHub*

onnx_model目录文件:

onnx_mode.bin
onnx_model.mapping
onnx_model.xml

文件说明:

• onnx_model.bin：训练后的数据文件。包含权重和偏差二进制数据。
• onnx_model.mapping：映射文件。
• onnx_model.xml：拓扑文件。描述网络拓扑。

使用模型

在树莓派上使用。需要根据官方文档配置环境。

判断单张图片。

from openvino.inference_engine import IECore
import time
classes = ['food', 'other', 'recycle', 'refuse']

ie = IECore()

print("read network")
tick = time.time()
net = ie.read_network('onnx_model/onnx_model.xml', 'onnx_model/onnx_model.bin')

input_blob = next(iter(net.input_info))
output_blob = next(iter(net.outputs))
net.batch_size = 1

print("load network")
exec_net = ie.load_network(network=net, device_name='MYRIAD')
print("time: ", time.time() - tick)

import torchvision.transforms as transforms
transformations = transforms.Compose([
                                      transforms.ToTensor(),
                                      transforms.Normalize(
                                          [0.485, 0.456, 0.406],
                                          [0.229, 0.224, 0.225])
                                      ])
from PIL import Image
import torch
def predict(filename):
    img = Image.open(filename)
    img = img.resize((256, 256))
    img = img.convert('RGB')
    img = transformations(img)
    img = torch.autograd.Variable(img[None, ...])
    res = exec_net.infer(inputs={input_blob: img})
    res = res[output_blob]
    return classes[res.argmax()]

tick = time.time()
print(predict('test/ju.png'))
print("time: ", time.time() - tick)

摄像机判断。（键盘q退出，空格键判断）

from openvino.inference_engine import IECore
import time
classes = ['food', 'other', 'recycle', 'refuse']

ie = IECore()

print("read network")
tick = time.time()
net = ie.read_network('onnx_model/onnx_model.xml', 'onnx_model/onnx_model.bin')

input_blob = next(iter(net.input_info))
output_blob = next(iter(net.outputs))
net.batch_size = 1

print("load network")
exec_net = ie.load_network(network=net, device_name='MYRIAD')
print("time: ", time.time() - tick)

import torchvision.transforms as transforms
transformations = transforms.Compose([
                                      transforms.ToTensor(),
                                      transforms.Normalize(
                                          [0.485, 0.456, 0.406],
                                          [0.229, 0.224, 0.225])
                                      ])

from PIL import Image
import torch
import cv2
def predict(frame):
    img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
    img = img.resize((256, 256))
    img = img.convert('RGB')
    img = transformations(img)
    img = torch.autograd.Variable(img[None, ...])
    res = exec_net.infer(inputs={input_blob: img})
    res = res[output_blob]
    return classes[res.argmax()]

capture = cv2.VideoCapture(0)

while True:
    _, frame = capture.read()
    cv2.imshow('garbage', frame)
    if cv2.waitKey(1) & 0xFF == ord('q'):
        break
    elif cv2.waitKey(1) & 0xFF == ord(' '):
        print("========== predict start ==========")
        tick = time.time()
        print(predict(frame))
        print('time: ', time.time() - tick)


capture.release()
cv2.destroyAllWindows()

效果

---

• NerualNetwork

comment:

• Valine
• LiveRe
• ChangYan