<2023-03-02>

## Getting feedback

Not only think about end users, but also embassadors, since they can lead to new users.

## In Potential Scope

- Model Cards
- has the potential to be adopted by upstream sklearn
- the app to create one, easier inspection of the models added to the model cards
- there are annoying issues with model cards that make it a bit tricky to work with them and create them
- we think this is somewhat high priority, and the annoyances need to be fixed, otherwise users will be deterred from using it.
- Persistence
- skops
- it's in a good shape, but needs many improvements, such as better inspection of the file, better support for c-extension types, extendability, etc
- there are many optimization potentials on the format's speed and size
- this is high priority if we want to push it more aggressively. For instance, it would need to be more stable for a place like sagemaker to potentially support it.
- ONNX
- it kinda works, as long as you have a single sklearn estimator.
- todo would be, better tools to check if user's model is suppoted, and to better support complex estimators such as pipelines and column transformers
- it would also need much better docs and workflow for people to implement what's needed for a custom estimator
- we haven't started this, it has potential to bring users, but it'd be a bit of work.
- Serving
- we do serving right now, but it's very slow, and half the time gives a timeout
- has many issues with different dtypes, etc
- could support a better way than sending/receiving json
- improve inference performance
- pretty much all of the work here would be in api-inference-community repo
- priority is not clear since there isn't much usage on the backend side. But it's more of a chicken and egg problem. Some improvements would be nice.
- Documenting on the skops side, how to serve using fastapi would be nice.

## Usage

It's been pretty stable around 50-100 downloads a day from pypi.

We've been having continuous engagement from community members on the repo.

The scikit-learn core team is excited about the project.

Ideas on increasing usage:
- outreach: we're giving talks around
- PyCon Italy is the next scheduled one
- Did one in PyData Berlin meetup
- Did another talk with Linux Foundation's ML Security Committee
- usage/conversion:
- the model card app would make people happy probably
- we need clearer communication on why people should use skops format, people are still confused as what the difference between `.skops` and `.onnx` is, for example.
- many people know of ONNX, convenience methods for that could bring them here.
- people always look for how to deploy their models online, if we have something related to that, it'd bring more people here. But this is a very large topic and whole bunch of companies/teams work on it.
- we could try to convince a place like sagemaker to support `.skops` format. (the format would need to be more mature for that)