Revisit: Squeezing

[jo-mueller]

So...with #106 and #107 merged and released, I thought it would be a good idea to revisit this topic. Originally, the idea behind squeezing (among others) was, that many downstream processing algorithms don't squeeze data on their end, which might lead to errors due to dimension missmatches.
That being said, I think there is still a strong case for actually squeezing out singleton dimensions. And with the named axes in, we can do so without getting ourselves in trouble when trying to parse and upload annotations later.

I thought the following could be done:

1. Squeeze out (leading?) singleton dimensions and update the axis_labels.
2. In save_rois, inspect this property and process annotations accordingly. I guess for Points and for Shapes there wouldn't be much of a different, anyway, but for the Labels (arguably the most prominent use case) it would make a big difference.

[psobolewskiPhD]

Just going to link the old discussion on this: #62
Is there ever a case where we'd want singleton dimensions?
I guess it would be the case where you have a T_YX and a _ZYX? if you open those with the current system (all axes preserved) you'd have proper sliders for T and Z. But in a squeezed world, you'd get T and Z on the same slider. But squeezing resolves issues with downstream processing via other plugins. Ideally napari would handle this better, inserting a dummy dim based on Layer.axis_labels as needed.

Edit: As an alternative to always squeezing, I think we could add a menu contribution to squeeze a layer, without a widget or anything? It would need to be a dask squeeze to not just load everything in the pyramidal case I think.

Aside: Maybe squeeze should be a napari layer contextual menu feature?

[psobolewskiPhD]

As noted above, if you don't squeeze then you always have TZYX so you can also name all the viewer.dims (e.g. sliders), because the viewer will always have those.
We have access to the viewer in all cases already (widget and the omero plugin), so after loading any image we could do:
viewer.dims.axis_labels = ("t", "z", "y", "x") and this would always work and stay true for any image we load, regardless if it has just yx, zyx, tyx, tzyx, or any other combo.

Naming dims vs axes was also here #88

[jo-mueller]

As for the suggestions above, I agree that squeezing could be an opt-in or opt-out feature in order to catch different usecases. Maybe a checkbox in the OMERO browser could make it a bit more visible? In the case of a separate plugin contribution I'm afraid that it would make the image data a bit harder to use for the "average" user. I'm not sure whether non-Python people would know ad-hoc what exactly squeezing an image means.

That being said, maybe squeezing could be an opt-out feature and napari-omero would always give you squeezed image data unless you'd really need singleton dimensions for whatever reason.

[jo-mueller]

As discussed about: Maaaybe it's worthwhile to store the indeces of squeezed dimension in the metadata of downoaded images from omero. ROIs/labels/etc could then check which dimensions have been removed on load and inject them back in on upload.