Object recognition without AVX

[el-fredo]

I know that Immich needs a CPU with AVX support to use the machine learning features. Due to the huge increase in energy costs, I decided to go with a very energy efficient home server, which is powered by an Intel Pentium Silver J5040. Unfortunately, this CPU does not support AVX.

Then I found this thread: Machine learning on CPUs without AVX: howto
It explains how to use the whole TensorFlow thing on CPUs without AVX. Unfortunately, this tutorial is for professionals, because you have to create containers and build / compile TensorFlow (as far as I understand).

Do you think there might be a simpler solution in the future? Could this method perhaps be built directly into Immich? So I mean that the object detection in Immich is working with every CPU?

I would also offer myself as a test subject if someone could take me by the hand. Then I would put together a step-by-step guide for beginners. However, I would also like to mention that I have Immich running under Unraid. Maybe that makes things even more complicated?

[bertmelis]

I wrote the "tutorial" and I am by far not a professional. Besides, further down that thread you can find a precompiled tfjs. I created it on a J4105.

Where do you need help? I can guide you through the process and improve the documentation while I do this.

[el-fredo]

Thank you so much for offering your help. Unfortunately, I have no idea about Docker and Docker-Compose. From that you are a pro for me ;-) In Unraid I am only using the community appstore to install containers.

For an absolute beginner like me, it would probably be easier to use your precompiled images. I think our CPUs are almost identical. So as the first step, I download your libtensorflow.tar.gz from your Google Drive. Then I need to install this as a container in Unraid. I do that with Docker Compose, right?

My first question is, where to store the downloaded file and do I have to unpack it?

Second question: Should I use the docker-compose.yml I already used to install Immich on Unraid, and just modify the part:

immich-machine-learning:
    image: altran1502/immich-machine-learning:release

Or should I create a new docker-compose.yml, just for this new machine learning image?

[bertmelis]

Well, I don't know a thing about Unraid. Can you ash into your server? Or do you have a terminal on your server GUI?

[el-fredo]

Yes, Unraid provides a terminal. I can also open a terminal in each installed container.

[el-fredo]

I have come one step further:

Using the command docker import /mnt/user/download/libtensorflow.tar.gz immich-ml-new:latest in the Unraid terminal, I was able to import your image.

So when I type docker image ls, I see all installed images, and also the new immich-ml-new:latest (that's the name I have chosen).

The next step now is to create the Docker container, right? And this is where I get stuck. Can you tell me what the Docker command should be?

For simplicity, I tried to create the container using Portainer. There I edited the existing container immich-immich-machine-learning-1, but with your image instead of the original one. Unfortunately, I get the error 400.

[el-fredo]

Okay, great.

I had briefly tested Photoprism, here the object detection worked perfectly (with my CPU without AVX support). But I prefer Immich, because it is exactly what I have been searching for.

The background why I startet this thread: In Unraid I have to pull a total of 7 containers to get Immich running. And I just thought it would be easy to replace the original ML-container with one that does not require AVX. So users can choose which one they pull, depending on their CPU.

In any case, thanks for your efforts 👍🏼

[bertmelis]

I'll still create the non-avx version but I'll build tensorflow manually. I only have different priorities now. Have a few days?

[bertmelis]

@el-fredo Would this be easier?
https://github.com/bertmelis/immich-machine-learning-no-avx

[el-fredo]

It is so much easier! Thank you so much for your effort.

All I had to do was change the single line in my docker-compose.yml and click on "Update Stack". That was everything I had to do. It pulled your container, which took a while. I restarted Immich, went to Administration -> Jobs and clicked on "Detect Objects" -> Run. Now my server is busy. Very busy... More than 24,000 jobs to do, the CPU load of all four cores is at 99 - 100% 😏

In the smartphone app, I already see the "Things" section when I tap Search. A few objects are already listed there, so it seems to work.

Thank you very much again!

[bertmelis]

It's indeed quite a large image because it add another layer on top of an already large image. It's quick and dirty but it works.

[alextran1502]

@bertmelis Thank you for the support! I appreciate your continuous responses to this topic.

[bertmelis]

With pleasure. I now published my very first image. Learning by doing.