MLontheEdge

The overall purpose of this document is to showcase an example of Azure Machine Learning on IoT Edge Devices using Microsoft Embedded Learning Library (ELL)

ELL for Windows Devices

Acquiring Equipments

Item	Description	Est. Cost (USD)	Example Products
Raspberry Pi 3	Main device for running software	$37.00	PiDevice
Sandisk Ultra 32GB Micro SDHC	Will store the Raspbian Operating System image and software for the Raspberry Pi 3	$12.00	MicroSD
Raspberry Pi Camera	Used to capture images and video on the Pi	$6.00	PiCamera
USB Keyboard	WIRED CONNECTION: Keyboard control for the Pi	$15.00	Keyboard
USB Mouse	WIRED CONNECTION: Mouse control for the Pi	$6.00	Mouse
Ethernet Cable	WIRED CONNECTION: Network connection for the Pi	$4.00	Ethernet

Window Device Set Up

The first step is to install Microsoft ELL on your host device. In order to do so, simply follow the directions in the link provided: Microsoft ELL

Set Up for Raspberry Pi Devices

Python 3.5.3: These steps assume you are starting from an existing Raspbian install with Python 3.5.3 on it based on the Raspbian Stretch image - 2018-04-18 image
Change Hostname: It is important that your host name is changes as we will attempt to connect to it remotely. The steps on how to achieve just that is best described here: Hostname Directions
Camera Set Up: Begin by typing:

sudo raspi-config

Select 5 Interfacing Options and press Enter.
Select P1 Camera and press Enter.
Select Yes to enable the camera interface.
Load the camera module:

sudo modprobe bcm2835-v4l2

Once the picamera is plugged in, a simple test to see if you picamera is working correctly is the command raspistill:

raspistill -o image.jpg

If a preview window is opened and a new file image.jpg is saved, then the picamera is successfully installed.

Enable SSH: Begin by typing:

sudo raspi-config

Select 5 Interfacing Options and press Enter.
Select P2 SSH and press Enter
Select Yes to enable the SSH Server.
Note: : Once SSH has been enabled, be sure to change the password associated with your Raspberry Pi, if you have not done so already.
Network Connection: For this project, the majority of the network connectivity came through the attachment of an ethernet cable. However, attached are steps to connecting the Raspberry Pi to a wireless connection. Wifi Connections

Azure Storage

Azure Blob Storage:

The Raspberry Pi has a small storage capability. Therefore, it is important to save picture, videos, models, and project description on the Cloud. For this project, Azure Storage is being used. Steps on how to set up Azure Storage is linked here. Azure Storage

Note: Be sure to save and make note of your STORAGE ACCOUNT NAME and STORAGE ACCOUNT KEYS They will be needed in the Edge.py and pisetup.py script later on your Raspberry Pi as well as on the Host.py script for your Host Device.

Pi3 Folder:

The utilization of Azure Storage is required for this application. Azure allows for automatic uploads and downloads of content file. As well, it is essential for persitant updates to the current pi3 folder.
The first time the application is ran with a correct Azure Credentials, blob containers are created for use with that given account. In addition, the current version of the pi3 folder located on the Raspberry Pi is uploaded to its respective blob container.
Constant checks are being made for changes and updates that occur every 3 hours.

Important Note on the Pi3 Folder on Azure:

After the project has been ran once and the given storage containers have been made, the user can now make changes to the given model and the pi3 folder.

Using the Azure Portal or Microsoft Azure Storage Explorer, locate the edgemodels blob container.
This is where the compiled "pi3 folder" with its given model is stored. It is important that the pi3 folder is zipped before being ready to be uploaded to the given blob container.
There can only be one item in this blob container and it most be titled zippedpi3 for use on the Raspberry Pi.

Download Models

As of right now, the Microsoft ELL supports Neural Network Models that were trained with Microsoft Cognitive Toolkit(CNTK) or with Darknet. Follow the given tutorial for insights in how to download model with the ELL Library. Importing Models.

From Host Device ELL to Raspberry Pi

The next step is to now get a compressed model from the ELL Library from your laptop to your Raspberry Pi to be utilized. The directions below are modified based on a tutorial sample from the site Microsft-Ell-Image-Classification.

Activate your Python enviroment on your host device

[Linux/macOS] source activate py36
[Windows] activate py36

Create a new directory to be copied over to the Rasberry Pi. The directory should contain should compressed CTNK or Darknets model in model.ell form. It should also contain the text file for the model classification
As before, run the wrap tool on your laptop or desktop computer, but this time specify the target platform as pi3. This tells the ELL compiler to generate machine code for the Raspberry Pi’s ARM Cortex A53 processor. This step needs to be performed in the directory to be copied to the Raspberry Pi.

python <ELL-root>/tools/wrap/wrap.py model.ell --language python --target pi3

To speed up the transfer of files to the Raspberry Pi, delete the model.ell file first before copying the folder. Now, there’s a pi3 directory that contains a CMake project that builds the Python wrapper and some helpful Python utilities. This directory should be inside a directory that also contains the model classification text.

Build Python Module:

The final step in the process is getting the pi3 folder transfered to Azure Blob Storage and the Raspberry Pi

A helper script for tranfering to Azure Blob Storage has been created; Host.py. Simply change the Azure Account Storage Name and Key in the Host.py script and run it in the same directory as the pi3 folder. This will compress the neccessary folder and make it availbale to be used by the Raspberry Pi.

python Host.py

Log in to your Raspberry Pi, find the directory you just copied from your computer, and build the python module that wraps the ELL model.

cd pi3
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make
cd ../..

You just created a Python module named model, which includes functions that report the model’s input and output dimensions and makes it possible to pass images to the model for classification.

Quick Project Run with Docker

Setting Up Docker on Raspberry Pi

Begin by updating the Raspberry Pi packages with the folloing command

sudo apt-get update && sudo apt-get upgrade

Proceed to installing Docker for the Raspberry Pi

curl -sSL https://get.docker.com | sh

Add 'Pi' user to the Docker group

sudo usermod -aG docker pi

Note: Until a Raspberry Pi reboot, the sudo prefix will be required before running any Docker commands

Verify that the Docker download was succesful in two ways

docker --version

sudo docker run armhf/hello-world

MLontheEdge for Docker A prebuilt docker image of this entire project has been made available on DockerHub. Step on how to download and run succesfully are listed below.

In the terminal, run the following docker command to pull the MLonthe Edge Docker Image

docker pull amlonedge\dockonedge:latest

Note: This is a very large image and will take approximately 15 minutes to be completely ready for use.

Once this is done, in your preferred working directory, create a env.list file. A copy of what is required is available on this repository.
The downloaded docker images requires a user Azure Blob Storage Name and Key to be entired as enviromental variables in order to run. This should be the Azure Storage accounnt where your zippedpi3 folder is located and where you would like project imgaes and video saved too.

vim env.list

AZURE_CONTAINER_NAME=yourblobstorageaccountname
AZURE_CONTAINER_KEY=yourblobstorageaccesskey

After saving the env.list file, the final step is to actually run the entire project. Run the project with the following exact command:

docker run --device=/dev/vcsm --device=/dev/vchiq --env-file env.list amlonedge/dockeronedge:latest

Once this command is ran, the project must check and update various files as neccesary. After about 30 seconds of setting up, the project is run continously. Afterwards, predictions, images and video are set to the given Azure Blob Storage Account.

Do It Yourself

Necessary Python Packages

CMake: CMake will be used on the Raspberry Pi to create python modules that can be called from given code. In order to install CMake on your Raspberry Pi, you must first be connected to the network, then open a terminal window and type:

sudo apt-get update
sudo apt-get install -y cmake

OpenBLAS: This is for fast linear algerba operations. This is highly recommended because it can significantly increase the speed of the models. Type:

sudo apt-get install -y libopenblas-dev

Curl: Curl is a command line tool used to transfer Data via URL. To install curl, type:

sudo apt-get install -y curl

OpenCV: As of this time, you will need to Build Open CV 3.3.0 on Raspbian Stretch with Python 3.5.3

It is crucial that this is done on Raspbian Stretch with Python 3.5.3. A link to download and format for Raspbian Stretch is provided here. Strech Image
Make sure that the Raspbian Stretch OS is up to date:

sudo apt-get update
sudo apt-get upgrade -y

Ensure that after the update, Python 3 is running Python 3.5.3:

python3 --version

Install some pre-requisities for the OpenCV Download:

sudo apt-get install -y build-essential cmake pkg-config
sudo apt-get install -y libjpeg-dev libtiff5-dev libjasper-dev libpng12-dev libdc1394-22-dev
sudo apt-get install -y libavcodec-dev libavformat-dev libswscale-dev libv4l-dev
sudo apt-get install -y libxvidcore-dev libx264-dev
sudo apt-get install -y libgtk2.0-dev libgtk-3-dev
sudo apt-get install -y libatlas-base-dev gfortran
sudo apt-get install -y python3-dev

Ensure that after the install, we running the correct pip module version:

python3 -m pip --version

Verify that the result of that command is a pip version 9.0.1 or greater

pip 9.0.1 from /usr/lib/python3/dist-packages (python 3.5)

Install Jinja2:

python3 -m install jinja2

Install Numpy:

python3 -m pip install numpy

Download and Unzip the OpenCV Source:

cd ~
wget -O opencv.zip 'https://github.com/Itseez/opencv/archive/3.3.0.zip'
unzip opencv.zip

Switch into folder and download the Opencv_contrib repo:

cd opencv-3.3.0
wget -O opencv_contrib.zip 'https://github.com/Itseez/opencv_contrib/archive/3.3.0.zip'
unzip opencv_contrib.zip

Create and go into the following build directory

cd ~/opencv-3.3.0
mkdir build
cd build

Run cmake with the following commands and DON'T FORGET THE '..' AT THE END :

cmake -D CMAKE_BUILD_TYPE=RELEASE \
  -D CMAKE_INSTALL_PREFIX=/usr/local \
  -D OPENCV_EXTRA_MODULES_PATH=../opencv_contrib-3.3.0/modules \
  -D BUILD_EXAMPLES=OFF \
  -D BUILD_TESTS=OFF \
  -D BUILD_PERF_TESTS=OFF \
  -D BUILD_opencv_python2=0 \
  -D PYTHON2_EXECUTABLE= \
  -D PYTHON2_INCLUDE_DIR= \
  -D PYTHON2_LIBRARY= \
  -D PYTHON2_NUMPY_INCLUDE_DIRS= \
  -D PYTHON2_PACKAGES_PATH= \
  -D BUILD_opencv_python3=1 \
  -D PYTHON3_EXECUTABLE=/usr/bin/python3 \
  -D PYTHON3_INCLUDE_DIR=/usr/include/python3.5 \
  -D PYTHON3_LIBRARY=/usr/lib/arm-linux-gnueabihf/libpython3.5m.so \
  -D PYTHON3_NUMPY_INCLUDE_DIRS=/home/pi/.local/lib/python3.5/site-packages/numpy/core/include \
  -D PYTHON3_PACKAGES_PATH=/home/pi/.local/lib/python3.5/site-packages \
  ..

Once that is complete, Run make:

Note: make will take approximately 2 hours to complete.

make

Run make install:

sudo make install

Test that OpenCV was installed correctly:

python3 -c "import cv2; print(cv2.__version__)"

Remove the unneccesay zip file and directory

rm ~/opencv.zip
rm -rf ~/opencv-3.3.0

Running Applications

The steps below show how the given Azure ML on Edge Project is deployed

Clone or Download the given repository:

git clone https://github.com/CatalystCode/MLontheEdge.git

Switch in the MLontheEdge and Raspi Folders

cd MLontheEdge/Raspi/

Export the Azure Blob Storage Name and Storage Key as enviromental Variable

$EDITOR ~/.profile
#ADD the following line to the bottom of the file:
export AZURE_CONTAINER_NAME=yourblobstoragename
export AZURE_CONTAINER_KEY=yourblobstoragekey

Run pisetup.py in order to set up the entire project. This script checks for the model version as well as make it ready for use.

python3 pisetup.py

Run the Edge.py script adapted specially for the Raspberry Pi

python3 Edge.py

While the script is running, a camera preview window will be opened allow you to see what the picamera sees. The scripts takes a picture every 5 seconds and returns what the model thinks it sees in that picture.
If the model and python script recognize an object, a video is captured of the 10 seconds before that moment and 15 seconds after the given moment and then saved it to an Azure Blob Storage account.

CatalystCode / MLontheEdge