Container size

[nomandera]

Photonix is now the largest container I deploy coming in at a 1.23 GB

Adding in the the helpers

• Postgres11 283 MB
• redis 105 MB

That is a whopping 1.6GB and considering it excludes cache, models and postgress database which are volume mounted and would add another 0.8GB we are getting into chonky boi territory.

Whats our thoughts on this?

[nomandera]

Polite bump to check if Github discussions are going to be used or if general community chat is happening somewhere else.

[damianmoore]

Hi @nomandera, sorry for the delayed reply - I wasn't checking here and I've had some time off. I agree with you that Photonix is quite chunky and we should do more to slim it down. I actually was a lot bigger (if you can believe that). I did some work in October 2020 which you can read more about here: https://photonix.org/blog/face-recognition-docker-multi-stage-builds-js/ . Basically it involved using multi-stage builds, removing parts of Python modules that weren't being used, excluding all the Node.js stuff. Unfortunately that means most of the low-hanging fruit is gone already. There are still things we can do though. Most of the space is now taken up by Python libraries (notably Tensorflow and OpenCV). Here are the top packages (above 5MB):

825516	/usr/local/lib/python3.8/site-packages/
485464	/usr/local/lib/python3.8/site-packages/tensorflow
75652	/usr/local/lib/python3.8/site-packages/cv2
57068	/usr/local/lib/python3.8/site-packages/opencv_python.libs
31940	/usr/local/lib/python3.8/site-packages/numpy.libs
31812	/usr/local/lib/python3.8/site-packages/django
16052	/usr/local/lib/python3.8/site-packages/numpy
14784	/usr/local/lib/python3.8/site-packages/tensorboard_data_server
10180	/usr/local/lib/python3.8/site-packages/Pillow.libs
10060	/usr/local/lib/python3.8/site-packages/faker
9636	/usr/local/lib/python3.8/site-packages/grpc
8800	/usr/local/lib/python3.8/site-packages/pip
8056	/usr/local/lib/python3.8/site-packages/h5py
7900	/usr/local/lib/python3.8/site-packages/psycopg2
6708	/usr/local/lib/python3.8/site-packages/matplotlib
5020	/usr/local/lib/python3.8/site-packages/PIL

I think where we go from here are a few possibilities:

1. Break the AI stuff into a separate container(s) so that people who don't want that can have a much smaller image.
2. Try to remove OpenCV as a dependency as it's only used in the face recognition and I don't think it's doing anything clever.
3. Try tweaking our Tensorflow builder to strip out parts we don't use
4. Try making the ML models run on TFLite instead of Tensorflow

I'd be happy to listen to other people's views on these options or take new suggestions.