I’ve put down dockerfile for the class. Some people asked to share it, so here it is:
For the full use you’ll also need jupyter_notebook_config.py
and your kaggle.json file, which you can get from your profile page on kaggle.com
What this file does:
- Installs software, which is used and/or mentioned in notebooks
- Install jupyterlab and some handy extensions for better experience. Jupyter notebook should also work but I never tried
- Set’s a jupyter password, so you can acces jupyterlab without token. Changes to jupyter_notebook_config.py should be made in order to enable this. Here is the explanation, what to do https://jupyter-notebook.readthedocs.io/en/latest/public_server.html#preparing-a-hashed-password
- Changes default shell to bash
- Creates another user, this way inside of docker will look more like regular setup
- Sets up conda
- Uses conda to install pytorch, fastai and other usefull packages
- runs jupyterlab by default
sudo docker image build -t fastai_v1:v3 .
I’m running this file executing
sudo docker run -p 8888:8888/tcp -v /home/user/AI:/home/user/AI --runtime=nvidia --ipc=host fastai_v1:v3
Started working on the exact same thing this morning, but yours is more complete than mine at the moment.
Had a test image up and running here: https://hub.docker.com/r/zerotosingularity/fastai_v3/
Might be good to create a PR and have an fastai container on Dockerhub?
Haven’t thought about that, will try to do so. Thanks!
Let me know if you need help on anything. Was planning on doing that in the coming days, but don’t want to steal any thunder
More time for me to work on other things
I’ve created an issue for this, will see if this is needed:
I tried building the Dockerfile but it failed…
Do you have it in a repo where I could contribute (a) PR(s)?
I would love to, but since it includes resources we shouldn’t share outside of this course, I wonder if it’s a good idea to create public repo with just this in it. And I don’t have a payed account to create private repo.
I’d like to wait for response on github issue first and think what to do after that.
In a mean time, you can just PM me or post here what failed and I’ll fix it.
I thought it would be ok to branch on github given the code is there in the open…
The build fails because I don’t have a kaggle.json file on my server and would upgrade the pip installation… Nothin earth shattering
oh, ok. I mentioned in the post that kaggle.json is needed, but it’s easy to miss. Might be a good idea to write small guide for the file, but don’t have enough time right now . Will have to wait till weekend.
It turns out course-v3 repo is indeed public. For some reason I though that we received access to it by some invite link, but I was obviously wrong. So with this in mind I created a repo for dockerfile and instructions.
No instruction right now but PRs are welcomed and I’ll try to handle check them ASAP.
Edit: probably better to just fork the original repo, here’s link for the fork https://github.com/Liberus/course-v3