A pedagogical implementation of panel apps served up on a remote machine.
See the full app here.
I spent a day figuring out how to make this happen at work, and decided to spend an evening consolidating my knowledge.
git clone https://github.com/ericmjl/minimal-panel-appThe first version of the app was coded up entirely on an iPad,
using a combination of blink
Juno,
and nano on my home remote server
(which is nothing more than a converted gaming tower).
Web app development in Python is now doable and we can use modern tablets as a thin client!
Deploying the HIV drug resistance model to Heroku was challenging because I had to watch out for memory and storage usage. There are 8 models to make predictions on, and loading all of them together causes memory overload on Heroku's free tier.
I got around this by pickling the models individually, and only loading them when needed. I also minimized disk usage by using gzip when pickling the files.
There are multiple "apps" that are being hosted by a single Panel server here. Each "app" is basically one Jupyter notebook. In each notebook, I define a self-contained, hostable unit that an end-user can interact with. One of them is the homepage, written using Panel's tooling just to prove the point, but the others are actual user-facing interfaces that provide a way to interact with either data or a machine learning model.