Hi all, I mentioned it before in the chat that I also was busy with getting an Azure environment up and running with the example model that makes use of the webhook functionality provided by numerai. Guess what, version 1 is available for you!
The github repo with everything needed to get it up and running is:
There is already some documentation available but I intend to improve that more after my holiday (which starts tomorrow). The highlights of Azure Compute are:
- Docker containers: it makes use of Docker Hub for storing the docker private container that contains the inference code. Example python code (that is based on the example model from the team) and powershell code are in the repo
- Azure Container Instance + Logic App: it makes use of these 2 resources to have the webhook functionality working, every week the container will be started, and after uploading the predictions it will terminate to save cost. For the example model (38 features minimal set) it is enough to have 2.5 Gb ram reseverd in Azure Container Instance, which means <<1 dollar per month. Ofcourse, depending on what you want to do more Ram may need to be reserved (or GPU instead of CPU but for inference I think most of us dont need that power).
- Fully automated powershell scripts in the repo to provision the Azure resources, you only need to fill in required information like subscription, credentials, etc…
And since it uses docker containers it is also quite easy to first run it locally for testing, or later switch to another cloud provider for example. I intend to get the Oracle version also working after the holidays, to see if the example models can also run on their free tier (24Gb ram, 4 Ampere cores 24x7x31days, 10Gb storage for free).
Ofcourse I am open to any improvements and such, and ofcourse you can tweak all the code to include stuff like GPU, multiple models, etc. First I am off for a couple of days and relax a bit Have a nice new year!