HaroldMua/tfserving-record

Multi-Camera Live Object Detection with Tensorflow Serving

如何使用异步进程获取IP摄像头的视频流

ImageZMQ 可以实现，多个sender(Raspberry Pi)的视频流发送至 reciever(PC)。

例如，可在多个Raspberry Pi运行以下代码，发送视频流：

from imutils.video import VideoStream
import imagezmq
import cv2
import time

cap = VideoStream(0).start()

sender = imagezmq.ImageSender(connect_to='tcp://192.168.11.216:5555')  # change to IP address and port of server thread
cam_id = 'Camera 1'  # this name will be displayed on the corresponding camera stream
time.sleep(2)

id = 0
while True:
    frame = cap.read()
    sender.send_image(cam_id, frame)

    id += 1
    print("frame id: %d" % id)

在PC端运行以下代码，接收视频流：

import cv2
import imagezmq

image_hub = imagezmq.ImageHub()

while True:  # show streamed images until Ctrl-C
    rpi_name, image = image_hub.recv_image()
    cv2.imshow(rpi_name, image)  # 1 window for each RPi

    cv2.waitKey(1)
    image_hub.send_reply(b'OK')

代码特点

• 通过定义config.ini,读取Tensorflow Serving的URL配置信息
• flask视频流的实现参考flask-video-streaming, 该部分代码是python类之静态方法、类方法的优秀实现

Record the walkthrough use of tfserving.

• Tensorflow Serving shows you how to use TensorFlow Serving with Docker(CPU/GPU).
• serving_basic shows you how to use TensorFlow Serving components to export a trained TensorFlow model(as SavedModel format) and use a Docker serving image to easily load the model for serving.
• convert_model_to_TFserving

Deploy a Tensorflow model wiht TF Serving

• Tensorflow object detection API
• Deployment of TensorFlow models into production with TensorFlow Serving, Docker, Kubernetes and Microsoft Azure
• How to deploy an object detection model with tensorflow serving
• Bravo!!! Deploy tensorflow object detection model in less than 5 minutes

Deploy a Pytorch model wiht TF Serving

• yolov5 model downloads

使用TensorFlow Serving镜像的两种方式

• 使用tensorflow/serving基础镜像，将模型文件挂载在对应文件夹
• 使用基于tensorflow/serving基础镜像的自定义镜像，将模型文件放置在镜像中

Reference

• 微信公众号-NLP奇幻之旅，系列文章：利用TFserving实现模型部署及预测

• 微信公众号-Python中文社区，利用TFserving部署深度学习模型

• 通过keras将ptorch .pth模型转tensorflow .pb模型

• ptorch2keras

• pytorch-onnx-tensorflow-pb

• pytorch保存网络结构及参数

• tesorflow三种模型的加载和保存方法

• pytorch加载本地和官方预训练模型

• yolov5-pytorch -> onnx -> tf savemodel

• tensorflow-yolov3 serving

• pytorch2tf

• Deploying Object Detection Model with TensorFlow Serving

• tf-chpt-2-pb

• camera-feed-object-detector-tf-serve

• darknet-tensorflow-serving

• tensorflow_serving_examples

• html5播放rtsp方案

github projects:

• opencv-flask

  • 将目标检测程序生成为系统的动态链接库文件
  • 用flask调用opencv和yolo-lite

• yolov5-flask-vue

• https://github.com/lbeaucourt/Object-detection

• object-detection

  • Blog

• https://github.com/LanderMoerkerke/camera-feed-object-detector-tf-serve

• flask-video-streaming

  • Blog

• flask-opencv-streaming

• opencv-stream-video-to-web

• live-video-stream

• multi-camera-live-object-tracking

Client

Tips

TODO

将coco label从pkl文件读写并保存

client端的代码逻辑是不错的：

• 提取视频帧有一个守护进程
• 保存检测的视频帧有一个守护进程
• 主程序执行预测处理

优化：

• tf serving需要warm up?
• 应选择挂载保存在本地的模型的方式？

收获：

• 在feed.py中，学习使用了锁，线程