FEAT Add skops space creator app #307
Conversation
Description This is the code for a streamlit app whose goal it is to let users easily build HF spaces for sklearn models. In particular, it offers a GUI to interface with our skops.card.Card class, giving users a wysiwyg experience. The app is currently hosted on my personal account: https://huggingface.co/spaces/BenjaminB/skops-model-card-creator2 We want to move it to skops and update it automatically. Comments I have adopted the CI to create this space using skops-ci. If the space itself is faulty, the CI job will still finish successfully, but at least one can check it and do some manual testing. Adding automatic tests will prove quite difficult, because, AFAICT, streamlit doesn't have anything like a testing client. Automatic testing would require some heavy lifting from the likes of Selenium or Playwright. It is possible to write tests for the parts of the code that are streamlit-agnostic, e.g. in tasks.py. However, that's the least brittle part of the code, so the additional value is low. Right now, there is no job that automatically builds this space and pushes it to the sklearn orga. I assume for that, we should create a workflow similar to clean-skops-user.yml that runs regularly and that uses a secret token so that random people cannot override the app. I had some strange issues with pre-commit and mypy. I don't want mypy to check the streamlit app -- it is a bit hacky in some places, which results in mypy errors. Therefore, I added an exclusion rule to pyproject.toml, which is indeed respected when I run mypy manually. However, pre-commit will stil check the repo, even though it's told to use the pyproject.toml. Thus I had to exclude the space code explicitly in the pre-commit-config.yaml. If anyone knows what the issue is, please let me know.
Setup fails otherwise, as it thinks that spaces could also be part of the package.
|
So this is now importing the token. It's odd cause the effective change is adding |
Even though the install steps are shared, it doesn't make sense to put the space deployment job into the testing workflow, since the latter runs multiple times because of the test matrix. The secret for deploying on the HF scikit-learn orga is not set yet, so this is WIP.
Forgot to add it with last commit ...
|
@skops-dev/maintainers I made all changes as discussed here and with Adrin on Slack. Whether the "main" deployment works, we will only see once we merge to main :) The test space is deployed here: https://huggingface.co/spaces/skops-ci/skops-space-creator-b5559d5e-667a-4498-a0b2-a8df12cf5ee5 Before going further, some points:
|
|
There's a warning on the bottom of the space you link here:
To me it's more of a scikit-learn model card creator/editor, or skops model card manager or something. It's not creating a space, it's creating a model card.
We can keep that for another PR. |
That's what I mentioned about flakiness of streamlit, refreshing should remove the error, and everything works despite of it. I add the button with that key only once, I really don't have an idea how the key is duplicated (would the button not appear twice then??).
Although I agree that this the main purpose, it's also not quite all of it. First of all, you can also create a model repo, not just the model card, second, it technically can also be used to create/edit any kind of model card. Also, I think having "skops" appear prominently would be good. |
|
@BenjaminBossan should we move this forward? |
Yes, what is the next step? |
|
You have some items here: #307 (comment)
|
Most notably, the new card class stores the specific section types like PlotSection in its data. This allows us to more directly modify those sections, e.g. when a user updates a plot file. Also use the new get_toc method. Another change was to add a specific task for updating the title of a plot, instead of having a big task both for title and content with many if...else. This is cleaner.
|
Hmm, I get |
Would a factory reset on the space help? but this is odd since they're unique spaces anyway. Hmm... |
Exactly. I reran the whole workflow, the new space has the same issue. |
|
Something in the manifest/setup must be flaky, we're seeing too many of these issues. |
Possibly, but when I was debugging that stuff, I couldn't identify anything (e.g. I tried removing MANIFEST completely and starting from an almost empty |
|
Should we merge this, have it deployed, and then try to fix it? |
We could try that and just don't advertise this feature until it's fixed. |
Description
This is the code for a streamlit app whose goal it is to let users easily build HF spaces for sklearn models. In particular, it offers a GUI to interface with our
skops.card.Cardclass, giving users a wysiwyg experience.The app is currently hosted on my personal account:
https://huggingface.co/spaces/BenjaminB/skops-model-card-creator2
We want to move it to skops and update it automatically.
Comments
I have adopted the CI to create this space using
skops-ci. If the space itself is faulty, the CI job will still finish successfully, but at least one can check it and do some manual testing.Adding automatic tests will prove quite difficult, because, AFAICT, streamlit doesn't have anything like a testing client. Automatic testing would require some heavy lifting from the likes of Selenium or Playwright.
It is possible to write tests for the parts of the code that are streamlit-agnostic, e.g. in
tasks.py. However, that's the least brittle part of the code, so the additional value is low.I had some strange issues with pre-commit and mypy. I don't want mypy to check the streamlit app -- it is a bit hacky in some places, which results in mypy errors. Therefore, I added an exclusion rule to
pyproject.toml, which is indeed respected when I run mypy manually. However, pre-commit will still check the space, even though it's told to use thepyproject.toml. Thus I had to exclude the space code explicitly in the.pre-commit-config.yaml. If anyone knows what the issue is, please let me know.TODO
Right now, there is no job that automatically builds this space and pushes it to the sklearn orga. I assume for that, we should create a workflow similar to
clean-skops-user.ymlthat runs regularly and that uses a secret token so that random people cannot override the app.