Tensorflow item recognition
Leveraging Google’s Tensorflow Machine Learning libraries for item recognition in images is fantastically easy to get going. The below Dockerfile will setup a container with everything required and allow the user to feed a URL to a file for classification:
Dockerfile:
Download raw from here: https://pastebin.com/raw/mdJ225vp
FROM ubuntu:16.04
RUN apt-get update && apt-get install -y openssh-server
RUN mkdir /var/run/sshd
RUN echo ‘root:tensorflow’ | chpasswd
RUN sed -i ‘s/PermitRootLogin prohibit-password/PermitRootLogin yes/’ /etc/ssh/sshd_config
# SSH login fix. Otherwise user is kicked off after login
RUN sed ‘s@session\s*required\s*pam_loginuid.so@session optional pam_loginuid.so@g’ -i /etc/pam.d/sshd
ENV NOTVISIBLE “in users profile”
RUN echo “export VISIBLE=now” » /etc/profile
EXPOSE 22
CMD [“/usr/sbin/sshd”, “-D”]
RUN apt-get install -y python-pip git
RUN pip install numpy tensorflow
RUN git clone https://github.com/jonas-werner/tf_item_recognition.git
Save the above into a file called “Dockerfile”.
Enter the directory where the Dockerfile is saved and build the Docker image:
docker build -t tf-image_recognition .
Verify the Docker image:
jonas@continuity:~/CODE/tf-image_recognition$ docker image ls
REPOSITORY TAG IMAGE ID CREATED SIZE
tf-image_recognition latest cd28f91139a2 6 minutes ago 1.07GB
Run the image. We’ll expose SSH on port 22 on the container as 2222 on the host:
docker run -it -d -P -p 2222:22 tf-image_recognition
Verify the local Docker gateway IP using the container ID (81f13360885f in this case – use “docker ps” to find out):
jonas@continuity:~/CODE/tf-image_recognition$ docker inspect 81f13360885f | grep Gateway
“Gateway”: “172.17.0.1”,
“IPv6Gateway”: “”,
“Gateway”: “172.17.0.1”,
“IPv6Gateway”: “”,
SSH and execute the image classification script (password: “tensorflow”):
ssh root@172.17.0.1 -p 2222 “wget https://jonamiki.com/wp-content/uploads/2013/07/wpid-IMAG0438_1.jpg; python /tf_item_recognition/classify_image.py –image_file ~/*.jpg; rm ~/*.jpg”
This is the image we’ve pulled down:
And this is the classification result:
motor scooter, scooter (score = 0.89661)
moped (score = 0.04356)
disk brake, disc brake (score = 0.00290)
crash helmet (score = 0.00289)
snowmobile (score = 0.00216)
Not too bad 🙂 Tensorflow accurately detects that the image contains a scooter, a crash helmet and even sees the disk brake on the scooter! Try with any image URL to see what Tensorflow will classify your image as. Have fun!