✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: bb1f1f6. Link to the linter CI: here}

Normally, one does not call estimator_html_repr directly, but the HTML repr is shown automatically when outputting an estimator. I think there needs to be an automatic way to have the HTML repr pick up the browser's or Jupyter's preference.

Normally this is done with a prefers-color-scheme CSS media query, but it is not supported in jupyterlab yet: jupyterlab/jupyterlab#8777 In the mean time, I would be okay with a workaround that does not involve javascript.

Normally, one does not call estimator_html_repr directly, but the HTML repr is shown automatically when outputting an estimator. I think there needs to be an automatic way to have the HTML repr pick up the browser's or Jupyter's preference.

Normally this is done with a prefers-color-scheme CSS media query, but it is not supported in jupyterlab yet: jupyterlab/jupyterlab#8777 In the mean time, I would be okay with a workaround that does not involve javascript.

Yes, agree this change won't solve the dark theme problem (since as you said, this would require some detection on the browser). Instead, it adds a theming layer to the HTML renders.
I think this change can still expose some low-level methods to change the theme if they need it for report generation purposes on Jupyter. wdyt?

Also later, we can add dark theme with detection using prefers-color-scheme more easily once support on jupyterlab is ready.

@thomasjpfan how do you think about the overall approach?

It seems the main use-case of the light/dark mode switching is to match it up to the light/dark mode used by the notebook that the HTML representation is rendered into. So I think we should add this if we have a way to query the notebook viewer (JupyterLab, VS Code, others?) to find out if they are using light/dark mode. This is another level up in complexity from using CSS prefers-color-scheme because it seems JupyterLab (and probably others) let the user select light/dark mode independent of their OS/browser preference. Does someone know if prefers-color-scheme only reflects the OS/browser level preference or if JupyterLab could set it as a result of the user toggling between light/dark mode?

Overall I think adding a way for the user to set the theme for the HTML repr by hand is not that attractive because it seems it would add yet another toggle (OS level, JupyterLab level, and scikit-learn level). And all of them are not connected to each other :-/

Ok, thinking about it, it's still best to provide a way to override the settings in anycase.

For example, I may have dark mode on, but I want my diagrams to appear light-themed. In that case, we would still need a way to let user to define the theme that they want.

Maybe something like this:

For me the question is "why would you want to do that?" - this is a serious question from me as someone who never uses dark mode. So I'm a weird person because a lot of people love dark mode and/or being able to switch. However it does mean I can't imagine why you'd ever want to have the notebook in dark (light) mode but the estimator reprs in light (dark) mode. Do you have an idea?

@betatim I don't use dark mode myself either, and not aiming to solve the dark-mode problem with this MR.

However, I do use Jupyter to generate reports and diagrams. And having the ability to style the pipeline diagram will certainly be useful.
Right now, we start with DARK and LIGHT themes, but could easily extend do Solarized, low-contrast etc.
I could add a few more themes to showcase this functionality.

Put differently, this MR:

1. Can resolve Adding dark mode for pipeline diagram #26364 and similar requests to style the pipeline diagram.
2. It is not feasible and complex to add "dark mode" detection today. We can tackle this in a separate larger MR. (If it's even still wanted)

To me it seems like we need to wait for jupyterlab/jupyterlab#8777 to be able to have an automated themed' fix.

In the meantime, we could make set_config(display="diagram-dark") to explicitly set the theme to dark, otherwise use the default provided in CSS.

In the meantime, we could make set_config(display="diagram-dark") to explicitly set the theme to dark, otherwise use the default provided in CSS.

Interesting, yes set_config() may be a good place to set it than on Pipeline or make_pipeline.

Approach 1 - In Pipeline, add new optional attribute theme:

pipe = Pipeline(steps) # if `theme` not give, will fallback to current style
pipe = Pipeline(steps, theme=themes.DARK) # `theme` is set to dark style

• Pros: Seems intuitive (subjective) to add here where I'm declaring the pipeline.
• Cons: Could be argued to keep Pipeline focused on ML model params.

Approach 2 - In Config, modify existing optional attribute display:

set_config(display="diagram-dark") # current values are ("diagram", "text")

• Pros: Keeps Pipeline focused on ML model, config for display-related configs.
• Cons: Adding ("diagram-dark", "diagram", "text") now makes us stuck on just 2 types of styles? What if we want ("diagram-dark", "diagram-style-2", "diagram-style-3", "diagram", "text")

Approach 3 - In Config, add new optional attribute theme:

set_config(theme=themes.DARK) # set a new config value `theme` that defaults to `themes.LIGHT`

• Pros: Keeps Pipeline focused on ML model, config for display-related configs; theme is separate from display mode.
• Cons: Not as intuitive (subjective) to set the display of the pipeline.

I'm leaning towards Approach 1 or 3. But what do you think?

I don't think we need to support multiple themes, and it certainly doesn't make sense to add this as an argument to pipeline.

And this is only a temporary solution, in the long run, we'd depend on whatever the css info tells us.

@thomasjpfan @adrinjalali pls help review again, thank you

Looks like it can close #26364

I was thinking, I can start a next PR (and a few more in sequence) to refactor this themes section of the codebase in small and safe steps. The goal being to bring better support for CSS variables.

At some point, it may be possible to set themes on jupyter-themes, that can propagate the styles to sklearn's diagrams.

Do let me know wdyt this.

let me re-open, need to fix the branches

Reopening

@thomasjpfan @adrinjalali PR is ready for review again