DEV Community

Kenichiro Nakamura
Kenichiro Nakamura

Posted on

Export Custom Vision model to Raspberry Pi 3: Issue and fix

#customvision #raspberrypi3 #docker

I play with Custom Vision Service export capability to docker file for RP3 today. There are several things I need to do to make it work, at least as of today.

I believe the issue will be addressed soon, so please refer to this information only when you have issue.

I don't explain what is Custom Vision or how to export model. Please refer to official document.

Perform image classification at the edge with Custom Vision Service

Error

When I build the image, I encounter following error at step 5/9.

Step 5/9 : RUN pip install flask pillow --index-url 'https://www.piwheels.org/simple' ---> Running in 686a1e4aa7bc Looking in indexes: https://www.piwheels.org/simple Collecting flask Downloading https://www.piwheels.org/simple/flask/Flask-1.1.2-py2.py3-none-any.whl (94 kB) ERROR: Could not find a version that satisfies the requirement pillow (from versions: none) ERROR: No matching distribution found for pillow

Dockerfile

Generated Dockerfile is below. If this is different from yours, then your issue maybe different. Key points are:

  • It's from python:3.7-slim
  • It misses libgl1-mesa-glx module on the first apt install. 
FROM python:3.7-slim RUN apt update && apt install -y libjpeg62-turbo libopenjp2-7 libtiff5 libatlas-base-dev RUN pip install absl-py six protobuf wrapt gast astor termcolor keras_applications keras_preprocessing --no-deps RUN pip install numpy==1.16 tensorflow==1.13.1 --extra-index-url 'https://www.piwheels.org/simple' --no-deps RUN pip install flask pillow --index-url 'https://www.piwheels.org/simple' # By default, we run manual image resizing to maintain parity with CVS webservice prediction results. # If parity is not required, you can enable faster image resizing by uncommenting the following lines. # RUN echo "deb http://security.debian.org/debian-security jessie/updates main" >> /etc/apt/sources.list & apt update -y # RUN apt install -y zlib1g-dev libjpeg-dev gcc libglib2.0-bin libsm6 libxext6 libxrender1 libjasper-dev libpng16-16 libopenexr23 libgstreamer1.0-0 libavcodec58 libavformat58 libswscale5 libqtgui4 libqt4-test libqtcore4 # RUN pip install opencv-python --extra-index-url 'https://www.piwheels.org/simple' COPY app /app # Expose the port EXPOSE 80 # Set the working directory WORKDIR /app # Run the flask server for the endpoints CMD python -u app.py

Fix

I updated the Dockerfile as below.

FROM balenalib/raspberrypi3-debian-python:3.7 RUN apt update && apt install -y libjpeg62-turbo libopenjp2-7 libtiff5 libatlas-base-dev libgl1-mesa-glx RUN pip install absl-py six protobuf wrapt gast astor termcolor keras_applications keras_preprocessing --no-deps RUN pip install numpy==1.16 tensorflow==1.13.1 --extra-index-url 'https://www.piwheels.org/simple' --no-deps RUN pip install flask pillow --index-url 'https://www.piwheels.org/simple' # By default, we run manual image resizing to maintain parity with CVS webservice prediction results. # If parity is not required, you can enable faster image resizing by uncommenting the following lines. # RUN echo "deb http://security.debian.org/debian-security jessie/updates main" >> /etc/apt/sources.list & apt update -y # RUN apt install -y zlib1g-dev libjpeg-dev gcc libglib2.0-bin libsm6 libxext6 libxrender1 libjasper-dev libpng16-16 libopenexr23 libgstreamer1.0-0 libavcodec58 libavformat58 libswscale5 libqtgui4 libqt4-test libqtcore4 # RUN pip install opencv-python --extra-index-url 'https://www.piwheels.org/simple' COPY app /app # Expose the port EXPOSE 80 # Set the working directory WORKDIR /app # Run the flask server for the endpoints CMD python -u app.py

Reference

I use the same base image from Custom Vision + Azure IoT Edge on a Raspberry Pi 3.

I also added "libgl1-mesa-glx" module as there is issue

Additional Info

README has section Build ARM container on x64 machine which says:

Export "ARM" Dockerfile from customvision.ai. Then build it using docker buildx command.

docker buildx build --platform linux/arm/v7 -t <your image name> --load .

I am using Windows 10 Docker Desktop v19.03.8.

I tested by running the command by enabling experimental feature which works fine, but I also try without enable experimental feature and simply run docker build which works just fine. So I am not 100% sure if we need buildx right now.

Top comments (0)