PaddleLite部署之树莓派3B

Author： smallsung
发布时间：July 14, 2020
7729 views
2 comments
5032 words
Categories：默认分类

1. 编译安装PaddleLite

1.2 树莓派环境配置

sudo apt update
sudo apt-get install -y --no-install-recomends gcc g++ make wget python unzip patchelf python-dev

wget https://www.cmake.org/files/v3.10/cmake-3.10.3.tar.gz
tar -zxvf cmake-3.10.3.tar.gz
cd cmake-3.10.3
./configure
make
sudo make install

export CXX=/usr/bin/g++
export CC=/usr/bin/gcc

1.3 下载Paddle-Lite源码并切换到release分支

git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite && git checkout release/v2.8

1.4 加速第三方库下载

rm -rf third-party
git submodule init
git submodule update

1.5 编译Paddle-Lite

./lite/tools/build_linux.sh --arch=armv7hf --with_python=ON --with_extra=ON --python_version=3.7

可选编译参数

--arch: (armv8|armv7|armv7hf)   arm版本，默认为armv8
--toolchain: (gcc|clang)        编译器类型，默认为gcc
--with_extra: (OFF|ON)          是否编译OCR/NLP模型相关kernel&OP，默认为OFF，只编译CV模型相关kernel&OP
--with_python: (OFF|ON)         是否编译python预测库, 默认为 OFF
--python_version: (2.7|3.5|3.7) 编译whl的Python版本，默认为 None
--with_cv: (OFF|ON)             是否编译CV相关预处理库, 默认为 OFF
--with_log: (OFF|ON)            是否输出日志信息, 默认为 ON
--with_exception: (OFF|ON)      是否在错误发生时抛出异常，默认为 OFF

注意：推荐with_cv设置为On否则会缺少部分op

2.使用PaddleLite

2.1 模型训练

本文使用PP-YOLO_MobileNetV3_small

详见PaddleDetection模型训练部分

https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.0-rc/docs/tutorials/DetectionPipeline.md

https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.0-rc/configs/ppyolo

2.2 模型导出

本文使用PP-YOLO_MobileNetV3_small

详见PaddleDetection模型导出部分

https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.0-rc/docs/advanced_tutorials/deploy/EXPORT_MODEL.md

2.3 模型调用MobileConfig/CxxConfig

2.3.1 使用opt tool优化模型

import os
import sys
# 引用Paddlelite预测库
sys.path.append('inference_lite_lib.armlinux.armv7hf/python/lib')
sys.path.append('inference_lite_lib.armlinux.armv7hf/cxx/lib')
from lite import *

# 1. 创建opt实例
opt=Opt()
# 2. 指定输入模型地址 
opt.set_model_file('./model/__model__')
opt.set_param_file('./model/__params__')
# 3. 指定转化类型： arm、x86、opencl、npu
opt.set_valid_places("arm")
# 4. 指定模型转化类型： naive_buffer、protobuf
opt.set_model_type("naive_buffer")
opt.set_quant_type('QUANT_INT8')
# 4. 输出模型地址
opt.set_optimize_out("model_int8_opt")
# 5. 执行模型优化
opt.run()

2.3.2 图片预测代码

#-*- coding: utf-8 -*-
from __future__ import print_function
import sys
import numpy as np
import cv2
import time
from PIL import Image
from PIL import ImageFont
from PIL import ImageDraw
sys.path.append('inference_lite_lib.armlinux.armv7hf/python/lib')
sys.path.append('inference_lite_lib.armlinux.armv7hf/cxx/lib')
from lite import *

font_style = ImageFont.truetype("arialuni.ttf", 25)

def read_img(im_path, resize_w,resize_h ):
  origin = Image.open(im_path)
  img = origin.resize((resize_w,resize_h), Image.BILINEAR)
  resized_img = img.copy()
  if img.mode != 'RGB':
    img = img.convert('RGB')
  img = np.array(img).astype('float32').transpose((2, 0, 1))  # HWC to CHW
  img -= 127.5
  img *= 0.007843
  img = img[np.newaxis, :]
  return origin,img


def draw_bbox_image(img, boxes, labels, save_name,scores,label_dict):
  img_width, img_height = img.size
  draw = ImageDraw.Draw(img)
  for box, label,score in zip(boxes, labels,scores):
      if(score>=0.80):
        print("label:",label_dict[int(label)])
        xmin, ymin, xmax, ymax = box[0], box[1], box[2], box[3]
        draw.rectangle((xmin, ymin, xmax, ymax), None, 'red')
        draw.text((xmin, ymin), '{},score={}'.format(label_dict[int(label)],score), (0, 255, 0),font=font_style)
  print(save_name)
  img.save(save_name)

# 配置config

config = MobileConfig()
config.set_model_from_file('./model_int8_opt.nb')

#config = CxxConfig()
#config.set_model_file('./model/__model__') # 指定模型文件路径 
#config.set_param_file('./model/__params__') # 指定参数文件路径
#places = [Place(TargetType.ARM, PrecisionType.FP32)]
#config.set_valid_places(places)

# 创建predictor
predictor = create_paddle_predictor(config)
print('Predict Start')
time_start=time.time()
image_path='./1.jpg'
output_path='./1.predict.jpg'
# 设置输入

input_tensor1 = predictor.get_input(0)
height, width = 320, 320
input_tensor1.resize([1, 3, height, width])
origin,img=read_img(image_path, height, width)
data = img.flatten()
input_tensor1.set_float_data(data)

origin_width, origin_height = origin.size
input_tensor2 = predictor.get_input(1)
input_tensor2.resize([1, 2])
input_tensor2.set_int32_data([origin_height, origin_width])

# 运行
predictor.run()

# 获取输出
output_tensor = predictor.get_output(0)
print(output_tensor.float_data())
print('Time Cost{}'.format(time.time()-time_start))
print('Predict End')

if output_tensor.shape()[1] != 6:
  print("No object found in {}".format(image_path))
  origin.save(output_path)

bboxes=np.array(list(output_tensor.float_data()), copy = False).reshape(-1,6)

#删除score<0.8的bbox
del_idx=[]
for i in range(len(bboxes)):
 if bboxes[i][1]<0.8:
  del_idx.append(i)
bboxes=np.delete(bboxes, del_idx, axis=0) 

labels = bboxes[:, 0].astype('int32')
scores = bboxes[:, 1].astype('float32')
boxes = bboxes[:, 2:].astype('float32')
  
print("label:{}".format(labels))
print("score:{}".format(scores))
print("boxes:{}".format(boxes))

label_dict={ 0:'meat', 1:'veg'}
draw_bbox_image(origin, boxes, labels, output_path,scores,label_dict)

2.3.3 视频预测代码

# coding=UTF-8
import cv2
import threading
import os
import sys
import time
import numpy as np

sys.path.append('inference_lite_lib.armlinux.armv7hf/python/lib')
sys.path.append('inference_lite_lib.armlinux.armv7hf/cxx/lib')
from lite import *

from PIL import Image
from PIL import ImageFont
from PIL import ImageDraw

font_style = ImageFont.truetype("arialuni.ttf", 25)

class Cameara:
    def __init__(self, src=0):
        self.stream = cv2.VideoCapture(src)
        self.stream.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
        self.stream.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
        (self.grabbed, self.frame) = self.stream.read()
        self.stopped = False

    def start(self):
        threading.Thread(target=self.update, args=()).start()

    def update(self):
        while True:
            if self.stopped:
                return
            (self.grabbed, self.frame) = self.stream.read()

    def read(self):
        return self.frame

    def stop(self):
        self.stopped = True

camera = Cameara("/dev/video0")

def data_process(frame, resize_w,resize_h):
  origin = frame.copy()
  frame = frame.resize((resize_w,resize_h), Image.BILINEAR)
  resize = frame.copy()
  if frame.mode != 'RGB':
    frame = frame.convert('RGB')
  frame = np.array(frame).astype('float32').transpose((2, 0, 1))  # HWC to CHW
  frame -= 127.5
  frame *= 0.007843
  frame = frame[np.newaxis, :]
  return origin,resize,frame

def draw_bbox_image(img, boxes, labels,scores,label_dict,color):
  img_width, img_height = img.size
  draw = ImageDraw.Draw(img)
  for box, label,score in zip(boxes, labels,scores):
     print("label:",label_dict[int(label)])
     xmin, ymin, xmax, ymax = box[0], box[1], box[2], box[3]
     draw.rectangle((xmin, ymin, xmax, ymax), None, color)
     draw.text((xmin, ymin), '{},score={}'.format(label_dict[int(label)],score),color,font=font_style)
  return img



if __name__ == "__main__":
  config = MobileConfig()
  config.set_model_from_file('./model_int8_opt.nb')
  # 创建predictor
  predictor = create_paddle_predictor(config)
  camera.start()
  cv2.namedWindow('image', cv2.WINDOW_NORMAL)

  label_dict={ 0:'meat', 1:'veg'}

  while True:
      frame = camera.read()
  
      frame = Image.fromarray(cv2.cvtColor(frame,cv2.COLOR_BGR2RGB))

      input_tensor1 = predictor.get_input(0)
      height, width = 320, 320
      input_tensor1.resize([1, 3, height, width])
      origin,resize,img=data_process(frame,320,320)
      data = img.flatten()
      input_tensor1.set_float_data(data)

      origin_width, origin_height = origin.size
      input_tensor2 = predictor.get_input(1)
      input_tensor2.resize([1, 2])
      input_tensor2.set_int32_data([origin_height, origin_width])

      print('Predict Start')
      time_start=time.time()

      predictor.run()

      output_tensor = predictor.get_output(0)
      print(output_tensor.float_data())
      print('Time Cost{}'.format(time.time()-time_start))
      print('Predict End')

      if output_tensor.shape()[1] != 6:
        print("No object found")
        continue

      bboxes=np.array(list(output_tensor.float_data()), copy = False).reshape(-1,6)

      del_idx=[]
      for i in range(len(bboxes)):
        if bboxes[i][2]<0.8:
          del_idx.append(i)
      bboxes=np.delete(bboxes, del_idx, axis=0) 


      labels = bboxes[:, 0].astype('int32')
      scores = bboxes[:, 1].astype('float32')
      boxes = bboxes[:, 2:].astype('float32')
  
      print("label:{}".format(labels))
      print("score:{}".format(scores))
      print("boxes:{}".format(boxes))

      img=draw_bbox_image(origin, boxes, labels,scores,label_dict,color)
      img = cv2.cvtColor(np.asarray(img),cv2.COLOR_RGB2BGR)
      cv2.imshow('image',img)

      if cv2.waitKey(1) & 0xFF == ord("q"):
          break
  cv2.destroyAllWindows()

附编译好的paddlelite 2.8-rc库

https://wwa.lanzous.com/ifoitm4dv9e

参考文档

https://paddle-lite.readthedocs.io/zh/latest/user_guides/source_compile.html
https://paddle-lite.readthedocs.io/zh/latest/user_guides/Compile/Linux.html

Last modification：May 8, 2021

如果觉得我的文章对你有用，请随意赞赏

2 comments

pplite
April 7th, 2021 at 10:18 am

博主您好，请问在树莓派上编译之后，运行predictor.run()会出现段错误，是什么原因呢？

Reply
1. smallsung
  April 22nd, 2021 at 07:53 pm
  
  @pplite
  
  你具体出现的错误是什么，可以贴出来。
  
  Reply

PaddleLite部署之树莓派3B

smallsung • 2020 年 07 月 14 日

<h1>1. 编译安装PaddleLite</h1><h2>1.2 树莓派环境配置</h2><pre><code>sudo apt update
sudo apt-get install -y --no-install-recomends gcc g++ make wget python unzip patchelf python-dev

wget https://www.cmake.org/files/v3.10/cmake-3.10.3.tar.gz
tar -zxvf cmake-3.10.3.tar.gz
cd cmake-3.10.3
./configure
make
sudo make install

export CXX=/usr/bin/g++
export CC=/usr/bin/gcc</code></pre><h2>1.3 下载Paddle-Lite源码 并切换到release分支</h2><pre><code>git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite &amp;&amp; git checkout release/v2.8</code></pre><h2>1.4 加速第三方库下载</h2><pre><code>rm -rf third-party
git submodule init
git submodule update</code></pre><h2>1.5 编译Paddle-Lite</h2><pre><code>./lite/tools/build_linux.sh --arch=armv7hf --with_python=ON --with_extra=ON --python_version=3.7</code></pre><p>可选编译参数</p><pre><code>--arch: (armv8|armv7|armv7hf)   arm版本，默认为armv8
--toolchain: (gcc|clang)        编译器类型，默认为gcc
--with_extra: (OFF|ON)          是否编译OCR/NLP模型相关kernel&amp;OP，默认为OFF，只编译CV模型相关kernel&amp;OP
--with_python: (OFF|ON)         是否编译python预测库, 默认为 OFF
--python_version: (2.7|3.5|3.7) 编译whl的Python版本，默认为 None
--with_cv: (OFF|ON)             是否编译CV相关预处理库, 默认为 OFF
--with_log: (OFF|ON)            是否输出日志信息, 默认为 ON
--with_exception: (OFF|ON)      是否在错误发生时抛出异常，默认为 OFF </code></pre><p>注意：推荐with_cv设置为On否则会缺少部分op</p><h1>2.使用PaddleLite</h1><h2>2.1 模型训练</h2><p>本文使用PP-YOLO_MobileNetV3_small</p><p>详见PaddleDetection模型训练部分</p><p><span class="external-link"><a class="no-external-link" href="https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.0-rc/docs/tutorials/DetectionPipeline.md" target="_blank"><i data-feather="external-link"></i>https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.0-rc/docs/tutorials/DetectionPipeline.md</a></span></p><p><span class="external-link"><a class="no-external-link" href="https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.0-rc/configs/ppyolo" target="_blank"><i data-feather="external-link"></i>https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.0-rc/configs/ppyolo</a></span></p><h2>2.2 模型导出</h2><p>本文使用PP-YOLO_MobileNetV3_small</p><p>详见PaddleDetection模型导出部分</p><p><span class="external-link"><a class="no-external-link" href="https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.0-rc/docs/advanced_tutorials/deploy/EXPORT_MODEL.md" target="_blank"><i data-feather="external-link"></i>https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.0-rc/docs/advanced_tutorials/deploy/EXPORT_MODEL.md</a></span></p><h2>2.3 模型调用MobileConfig/CxxConfig</h2><h3>2.3.1 使用opt tool优化模型</h3><pre><code class="lang-python">import os
import sys
# 引用Paddlelite预测库
sys.path.append(&#039;inference_lite_lib.armlinux.armv7hf/python/lib&#039;)
sys.path.append(&#039;inference_lite_lib.armlinux.armv7hf/cxx/lib&#039;)
from lite import *

# 1. 创建opt实例
opt=Opt()
# 2. 指定输入模型地址 
opt.set_model_file(&#039;./model/__model__&#039;)
opt.set_param_file(&#039;./model/__params__&#039;)
# 3. 指定转化类型： arm、x86、opencl、npu
opt.set_valid_places(&quot;arm&quot;)
# 4. 指定模型转化类型： naive_buffer、protobuf
opt.set_model_type(&quot;naive_buffer&quot;)
opt.set_quant_type(&#039;QUANT_INT8&#039;)
# 4. 输出模型地址
opt.set_optimize_out(&quot;model_int8_opt&quot;)
# 5. 执行模型优化
opt.run()</code></pre><h3>2.3.2 图片预测代码</h3><pre><code class="lang-python">#-*- coding: utf-8 -*-
from __future__ import print_function
import sys
import numpy as np
import cv2
import time
from PIL import Image
from PIL import ImageFont
from PIL import ImageDraw
sys.path.append(&#039;inference_lite_lib.armlinux.armv7hf/python/lib&#039;)
sys.path.append(&#039;inference_lite_lib.armlinux.armv7hf/cxx/lib&#039;)
from lite import *

font_style = ImageFont.truetype(&quot;arialuni.ttf&quot;, 25)

def read_img(im_path, resize_w,resize_h ):
  origin = Image.open(im_path)
  img = origin.resize((resize_w,resize_h), Image.BILINEAR)
  resized_img = img.copy()
  if img.mode != &#039;RGB&#039;:
    img = img.convert(&#039;RGB&#039;)
  img = np.array(img).astype(&#039;float32&#039;).transpose((2, 0, 1))  # HWC to CHW
  img -= 127.5
  img *= 0.007843
  img = img[np.newaxis, :]
  return origin,img

def draw_bbox_image(img, boxes, labels, save_name,scores,label_dict):
  img_width, img_height = img.size
  draw = ImageDraw.Draw(img)
  for box, label,score in zip(boxes, labels,scores):
      if(score&gt;=0.80):
        print(&quot;label:&quot;,label_dict[int(label)])
        xmin, ymin, xmax, ymax = box[0], box[1], box[2], box[3]
        draw.rectangle((xmin, ymin, xmax, ymax), None, &#039;red&#039;)
        draw.text((xmin, ymin), &#039;{},score={}&#039;.format(label_dict[int(label)],score), (0, 255, 0),font=font_style)
  print(save_name)
  img.save(save_name)

# 配置config

config = MobileConfig()
config.set_model_from_file(&#039;./model_int8_opt.nb&#039;)

#config = CxxConfig()
#config.set_model_file(&#039;./model/__model__&#039;) # 指定模型文件路径 
#config.set_param_file(&#039;./model/__params__&#039;) # 指定参数文件路径
#places = [Place(TargetType.ARM, PrecisionType.FP32)]
#config.set_valid_places(places)

# 创建predictor
predictor = create_paddle_predictor(config)
print(&#039;Predict Start&#039;)
time_start=time.time()
image_path=&#039;./1.jpg&#039;
output_path=&#039;./1.predict.jpg&#039;
# 设置输入

input_tensor1 = predictor.get_input(0)
height, width = 320, 320
input_tensor1.resize([1, 3, height, width])
origin,img=read_img(image_path, height, width)
data = img.flatten()
input_tensor1.set_float_data(data)

origin_width, origin_height = origin.size
input_tensor2 = predictor.get_input(1)
input_tensor2.resize([1, 2])
input_tensor2.set_int32_data([origin_height, origin_width])

# 运行
predictor.run()

# 获取输出
output_tensor = predictor.get_output(0)
print(output_tensor.float_data())
print(&#039;Time Cost{}&#039;.format(time.time()-time_start))
print(&#039;Predict End&#039;)

if output_tensor.shape()[1] != 6:
  print(&quot;No object found in {}&quot;.format(image_path))
  origin.save(output_path)

bboxes=np.array(list(output_tensor.float_data()), copy = False).reshape(-1,6)

#删除score&lt;0.8的bbox
del_idx=[]
for i in range(len(bboxes)):
 if bboxes[i][1]&lt;0.8:
  del_idx.append(i)
bboxes=np.delete(bboxes, del_idx, axis=0)

labels = bboxes[:, 0].astype(&#039;int32&#039;)
scores = bboxes[:, 1].astype(&#039;float32&#039;)
boxes = bboxes[:, 2:].astype(&#039;float32&#039;)
  
print(&quot;label:{}&quot;.format(labels))
print(&quot;score:{}&quot;.format(scores))
print(&quot;boxes:{}&quot;.format(boxes))

label_dict={ 0:&#039;meat&#039;, 1:&#039;veg&#039;}
draw_bbox_image(origin, boxes, labels, output_path,scores,label_dict)</code></pre><h3>2.3.3 视频预测代码</h3><pre><code class="lang-python"># coding=UTF-8
import cv2
import threading
import os
import sys
import time
import numpy as np

sys.path.append(&#039;inference_lite_lib.armlinux.armv7hf/python/lib&#039;)
sys.path.append(&#039;inference_lite_lib.armlinux.armv7hf/cxx/lib&#039;)
from lite import *

from PIL import Image
from PIL import ImageFont
from PIL import ImageDraw

font_style = ImageFont.truetype(&quot;arialuni.ttf&quot;, 25)

class Cameara:
    def __init__(self, src=0):
        self.stream = cv2.VideoCapture(src)
        self.stream.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
        self.stream.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
        (self.grabbed, self.frame) = self.stream.read()
        self.stopped = False

def start(self):
        threading.Thread(target=self.update, args=()).start()

def update(self):
        while True:
            if self.stopped:
                return
            (self.grabbed, self.frame) = self.stream.read()

def read(self):
        return self.frame

def stop(self):
        self.stopped = True

camera = Cameara(&quot;/dev/video0&quot;)

def data_process(frame, resize_w,resize_h):
  origin = frame.copy()
  frame = frame.resize((resize_w,resize_h), Image.BILINEAR)
  resize = frame.copy()
  if frame.mode != &#039;RGB&#039;:
    frame = frame.convert(&#039;RGB&#039;)
  frame = np.array(frame).astype(&#039;float32&#039;).transpose((2, 0, 1))  # HWC to CHW
  frame -= 127.5
  frame *= 0.007843
  frame = frame[np.newaxis, :]
  return origin,resize,frame

def draw_bbox_image(img, boxes, labels,scores,label_dict,color):
  img_width, img_height = img.size
  draw = ImageDraw.Draw(img)
  for box, label,score in zip(boxes, labels,scores):
     print(&quot;label:&quot;,label_dict[int(label)])
     xmin, ymin, xmax, ymax = box[0], box[1], box[2], box[3]
     draw.rectangle((xmin, ymin, xmax, ymax), None, color)
     draw.text((xmin, ymin), &#039;{},score={}&#039;.format(label_dict[int(label)],score),color,font=font_style)
  return img

if __name__ == &quot;__main__&quot;:
  config = MobileConfig()
  config.set_model_from_file(&#039;./model_int8_opt.nb&#039;)
  # 创建predictor
  predictor = create_paddle_predictor(config)
  camera.start()
  cv2.namedWindow(&#039;image&#039;, cv2.WINDOW_NORMAL)

label_dict={ 0:&#039;meat&#039;, 1:&#039;veg&#039;}

while True:
      frame = camera.read()
  
      frame = Image.fromarray(cv2.cvtColor(frame,cv2.COLOR_BGR2RGB))

input_tensor1 = predictor.get_input(0)
      height, width = 320, 320
      input_tensor1.resize([1, 3, height, width])
      origin,resize,img=data_process(frame,320,320)
      data = img.flatten()
      input_tensor1.set_float_data(data)

origin_width, origin_height = origin.size
      input_tensor2 = predictor.get_input(1)
      input_tensor2.resize([1, 2])
      input_tensor2.set_int32_data([origin_height, origin_width])

print(&#039;Predict Start&#039;)
      time_start=time.time()

predictor.run()

output_tensor = predictor.get_output(0)
      print(output_tensor.float_data())
      print(&#039;Time Cost{}&#039;.format(time.time()-time_start))
      print(&#039;Predict End&#039;)

if output_tensor.shape()[1] != 6:
        print(&quot;No object found&quot;)
        continue

bboxes=np.array(list(output_tensor.float_data()), copy = False).reshape(-1,6)

del_idx=[]
      for i in range(len(bboxes)):
        if bboxes[i][2]&lt;0.8:
          del_idx.append(i)
      bboxes=np.delete(bboxes, del_idx, axis=0)

labels = bboxes[:, 0].astype(&#039;int32&#039;)
      scores = bboxes[:, 1].astype(&#039;float32&#039;)
      boxes = bboxes[:, 2:].astype(&#039;float32&#039;)
  
      print(&quot;label:{}&quot;.format(labels))
      print(&quot;score:{}&quot;.format(scores))
      print(&quot;boxes:{}&quot;.format(boxes))

img=draw_bbox_image(origin, boxes, labels,scores,label_dict,color)
      img = cv2.cvtColor(np.asarray(img),cv2.COLOR_RGB2BGR)
      cv2.imshow(&#039;image&#039;,img)

if cv2.waitKey(1) &amp; 0xFF == ord(&quot;q&quot;):
          break
  cv2.destroyAllWindows()

</code></pre><p>附编译好的paddlelite 2.8-rc库</p><p><code>https://wwa.lanzous.com/ifoitm4dv9e</code></p><p>参考文档</p><pre><code>https://paddle-lite.readthedocs.io/zh/latest/user_guides/source_compile.html
https://paddle-lite.readthedocs.io/zh/latest/user_guides/Compile/Linux.html</code></pre>

PaddleLite部署之树莓派3B

1. 编译安装PaddleLite

1.2 树莓派环境配置

1.3 下载Paddle-Lite源码并切换到release分支

1.4 加速第三方库下载

1.5 编译Paddle-Lite

2.使用PaddleLite

2.1 模型训练

2.2 模型导出

2.3 模型调用MobileConfig/CxxConfig

2.3.1 使用opt tool优化模型

2.3.2 图片预测代码

2.3.3 视频预测代码

2 comments

Leave a Comment Cancel reply
使用cookie技术保留您的个人信息以便您下次快速评论，继续评论表示您已同意该条款

AHPU大数据与人工智能协会2020年简介

Ubuntu环境下深度学习环境配置不完全指北CUDA+cuDNN+NCCL

PaddleLite部署之树莓派3B

面向监考业务的人证核实移动终端与系统

编译支持Vulkan移动端加速的Pytorch

编译支持Vulkan移动端加速的Pytorch

Nvidia Jetson Nano环境配置

面向监考业务的人证核实移动终端与系统

AHPU大数据与人工智能协会2020年简介

ubuntu引导快速修复

PaddleLite部署之树莓派3B

1. 编译安装PaddleLite

1.2 树莓派环境配置

1.3 下载Paddle-Lite源码 并切换到release分支

1.4 加速第三方库下载

1.5 编译Paddle-Lite

2.使用PaddleLite

2.1 模型训练

2.2 模型导出

2.3 模型调用MobileConfig/CxxConfig

2.3.1 使用opt tool优化模型

2.3.2 图片预测代码

2.3.3 视频预测代码

2 comments

Leave a Comment Cancel reply 使用cookie技术保留您的个人信息以便您下次快速评论，继续评论表示您已同意该条款

PaddleLite部署之树莓派3B

1.3 下载Paddle-Lite源码并切换到release分支

Leave a Comment Cancel reply
使用cookie技术保留您的个人信息以便您下次快速评论，继续评论表示您已同意该条款