Docker / container environment

[Ontopic]

Lovely package! Not getting the traction is deserves though (yet!).

I would love to run this in a container (preferably Docker). I've less experience myself with Python's dev-tooling, so containerising "Tensorflow, PyTorch, Sklearn, Matplotlib, Numpy, and Hugging Face Transformers. You should install TensorFlow and PyTorch following the instructions from their respective websites" seems like a daunting task.

Should be easy to do though with some knowledge of Docker, Python and ai-framework-landscape brought together. From there extensions to API's, UI's and Apps can be implemented Quick as well ;)

I'm thinking a dressed down / pre-fab Dockerfile with https://github.com/ufoym/deepo as a starting point could do the trick, but interested in hearing what geekjr and others may have tried.

Thanks again!

p.s. Please do more videos. I never was more convinced of giving a repo a try by a video than yours. No hype talk, no fake simple, no pretend awesomest new patented blabla. Nah, just the clear truth. Must be in your timing. Props

[geekjr]

Hello,
I will look into this further. I think being able to run QuickAI in a Docker container will improve the setup process for a lot of people.

[geekjr]

Docker support is now added! Thanks @Ontopic

[Ontopic]

Thank you! Really helpful to get people to give it a go without having the need to go through the entire setup, defeats the purpose of how quickly it allows you to test training ;)

Maybe next video with Docker? Really a start from scratch ;)

I might start a separate discussion for this, but is there a "plan" for models in the future? F.e. GPT-J, which is now ready for HuggingFace

[Ontopic]

Guess you are, just saw your commit history. Hit me up if there's any way I can assist.

[geekjr]

GPTJ is something that I have tried to add before. The only issue is that the gpu I have(2060 Super) can't run it without OOM, which means that I won't be able to test it before release.

[geekjr]

I added a GPTJ function in the main branch,

[geekjr]

That function is very useful as it would be needed for GPTJ fine tuning support. However, due to the VRAM limitation, I would be unable to test the code, which is why I am not working on it(I was able to test the new GPTJ function in Colab on one of the free 16GB VRAM GPUS, but I don't think that is not enough for fine tuning).

[Ontopic]

I was just looking at vast.ai myself. You on any discord channel?

[Ontopic]

Also see huggingface/transformers#13022, I think transforms is still in-flux for the upcoming hours, then probably should be easier.

[geekjr]

I see that PR is adding support of GPTJ. I mean even after that, i think ill still need more VRAM for training.

[geekjr]

Vast ai seems good rates, but I would still rather not use it.