Installation on Raspberry Pi OS? #25
Comments
Yes, I'm actually using it on a Raspberry Pi 4 with 8GB. Because it's ARM, simply including pytorch it in the requirements doesn't work for some reason. I installed pytorch from this custom wheel: https://github.com/ljk53/pytorch-rpi/blob/master/torch-1.9.0a0%2Bgitd69c22d-cp39-cp39-linux_aarch64.whl. You can also build it from scratch, but it takes unnecessarily long.
Let me know how the setup goes. |
Ah, I'd forgotten to mention which Pi I'm using: I have a Pi 400 because it's currently so difficult to get your hands on a normal Pi 4. Therefore I only have 4GB RAM. Today I installed Ubuntu 20.10 and also increased the Swap memory in case it's about the RAM. Trying to install Pytorch from https://github.com/ljk53/pytorch-rpi once again failed, but the wheel from https://github.com/Qengineering/PyTorch-Raspberry-Pi-64-OS apparently worked (also PyTorch 1.9.0). The conceptarium installation also seems to have worked, although I had to manually install quite a few dependencies. Now the backend is running, and streamlit with the browser interface is also working. But I cannot login with localhost:8000 + a new token, and there is no "knowledge" folder created anywhere on my system. This is the log in the backend terminal:
There's also an error in the frontend, but I'm assuming the problem is in the backend again? What I've tried is simply paste the "knowledge" folder from the conceptarium on my Mac in the home directory and access that, but I get exactly the same results. |
Update: I've got it now! My imported knowledge folder should have gone into the base conceptarium directory, not the main user folder as was the case with the Docker version. Still: Shouldn't it have created such a folder for me when I first try to log in? First impression is that it's slower, but perfectly usable. I'll try to set up remote access in the next few days. |
I totally get that, I bought one a couple weeks ago in part for the same reason! I don't plan on using it for a few months, but was worried it'd also go out of stock by then.
Would it be possible to
Huh, yeah it should be created if it's not there, I'll open up a new issue for that and try to replicate it.
Awesome! |
In fact they were listed in the requirements.txt, but in my case still not installed automatically. Here are the ones that could then be installed with a simple
I also had to manually install sentence transformers 2.1.0. 2.0.0 did not work. |
One other thing: I'd installed streamlit manually before starting the installation, because that was the requirement that hadn't worked at all on the 32-bit OS. So that could be another one that's not installed directly. If I remember correctly streamlit and orjson had to be manually installed for the Docker instance as well. |
Oh maybe the first pytorch issues in requirements.txt led to those other packages not being tackled at all? I see they're all in their respective requirements, and the docker image is built via github actions to bake those dependencies in the image. Anyway, closing this for now. |
I'd first tried to install the conceptarium on the preinstalled 32-bit Pi OS, which failed because I could not get Streamlit to run.
Then I've tried it with the new 64-bit version of Pi OS. Streamlit works now, but the Pytorch-version linked in the readme could not be installed.
Have you managed to install the conceptarium on Raspberry Pi OS? If not, which OS have you been using, so that I can try that out?
The text was updated successfully, but these errors were encountered: