Hi guys, I hope you are doing well!

I have been working on a new plugin which relies a lot on sklearn-decomposition, but I am aiming it to be a more practical decomposition plugin, for example, introducing label layers and time plots.

Since I intend on building it on top of the structure of this one, I wanted to know your opinion about it to make a decision. What would be more interesting for you? @jdeschamps @guiwitz @psobolewskiPhD

1. Push changes here and modify this one (including its name)
2. Keep this one as it is and fork it to work on a separate project

In case of the second option, of course you are welcome to join if you like, just send me a message later 😃

Best, Marcelo

I mentioned this in the PR that's been merged and closed: https://github.com/jdeschamps/napari-sklearn-decomposition/pull/2 So posting as an issue so we don't lose track of it totally. Basically, if you try to change the decomposition method from the default (PCA at the moment) then the image dropdown is greyed out. Going back to PCA also doesn't work. But delete and reopen the sample Faces and the new choice works. So I guess when the new widget is appended it doesn't inherit the layers list and that list needs to be modified to trigger the magicgui magic that makes the list.

The plugin implements 3 different decompositions. Applying colormaps to the output decompositions makes them easier to interpret (and stand out more vs. the data). PCA and ICA return both positive and negative values, so it makes sense to visualize using a diverging colormap. napari implements PiYG, which is colorblind-safe, so that one is used by default. NMF specifically does not return negative values, so viridis a uniform, colorblind-safe colormap is used by default. Also, the layers are named explicitly.

Adding tests, to address https://github.com/jdeschamps/napari-sklearn-decomposition/issues/6 Hopefully with CI fixed, they will run! Currently it's just the most basic tests:

1. that sample data is loaded correctly
2. that a widget is created (Note: I commented out the example tests because they fail, but for now leaving them as a reference.)

This is needed according to @talley, see: https://napari.zulipchat.com/#narrow/stream/309872-plugins/topic/important.20info.20for.20npe2.20cookiecutter.20users

I was reading the code and playing with the plugin to learn a bit more. Made some small tweaks:

1. set default N_components to 6. It was zero, which made nothing happen on-run. 6 matches the example
2. Set the viewer to first slice upon output. The layer sliders are linked, so when the source is in the middle slice (200) then the output is missing.
3. cleaned up some unused imports
4. tried to fix NMF This I failed. I noticed that when changing the mode the subsequent ones the image input was greyed out. I swapped the default from PCA to NMF and confirmed that NMF now works with defaults (I made minor tweak). But the issue remains: somehow the image is only passed to the default, "manual" call of _on_choice_change. I tried a few things, but couldn't figure it out—sorry!

Fix for Issue https://github.com/jdeschamps/napari-sklearn-decomposition/issues/4 where changing the decomposition method from the default (PCA at the moment) made the image dropdown non-functional: greyed out. As @jdeschamps suggested the fix is to just reset_choices and connect to the event, based on: https://github.com/pattonw/napari-affinities/blob/7d9eab9100daf607ce7351f8717bff37ad71acf0/src/napari_affinities/widget.py#L61

Ok, it works now. napari>File>Sample Data>napari-sklearn ....

The returned data is wrong though: it's still the ravelled, rather than a stack. see: https://scikit-learn.org/stable/modules/generated/sklearn.datasets.fetch_olivetti_faces.html#sklearn.datasets.fetch_olivetti_faces