Link tranform

[francisco-dlp]

Add transform argument to link that takes a tuple with the forward and backward transform.

[minrk]

cc @SylvainCorlay for review

[SylvainCorlay]

Thanks.

To provide some background, one reason why we only added it to the directional case originally is that one needs to ensure that the two tranforms are inverses of each other to avoid infinite bouncing behavior.

[francisco-dlp]

Thanks for the feedback. Why would two transforms that are not inverses of each other cause bouncing behaviour?

I think that this is an useful feature. For example, say that I define a span selector and I want to fix its width without changing the class definition. With transform it is as simple as:

import traitlets

class SpanSelector(traitlets.HasTraits):
    x1 = traitlets.Float()
    x2 = traitlets.Float()

ss = SpanSelector()
fix_width = traitlets.link((ss, "x1"), (ss, "x2"), (lambda x: x + 10, lambda x: x - 10))

Of course things may not work as expected if the transforms are not inverses, but I think that it is fair to assume that the users know what they're doing.

I have merged the current master into the branch to solve the merge conflicts.

[maartenbreddels]

I think this is a useful feature, and it would be nice to have this in ipywidgets as well. Can we check in _update_source and _update_target if the transformations are actually the inverse? And raise an Exception is they differ.

[francisco-dlp]

I think that checking if the transformations are the inverse of each other is not trivial e.g. they may be the inverse of each other only in a certain domain.

[rmorshea]

@francisco-dlp because we check link.updating I don't think we have to worry about bouncing -
the link will always suppress change notifications while it's "busy". However, this also means that it would be possible to break the consistency of the transform relationship:

class MyClass(HasTraits):
    i = Int()
    j = Int()
    
    @observe("j")
    def another_update(self, change):
        self.i = change.new * 2

mc = MyClass()

trans = (lambda x: x**2, lambda x: x**(0.5))
l = link((mc, "i"), (mc, "j"), trans)

mc.i = 2

>>> mc.i
8
>>> mc.j
4

Inconsistent states will result when a consequence of a link's update changes either the source or the target trait. To resolve this issue we could raise if, after the update context is closed, the transform's relationship has been broken.

[SylvainCorlay]

I think that checking if the transformations are the inverse of each other is not trivial e.g. they may be the inverse of each other only in a certain domain.

That is why we did not implement bidirectional transform originally!

If this was to be added, it would be the responsibility of the user to provide inverses of the transforms

[francisco-dlp]

@rmorshea, I think that the issue that you mention is unrelated to this pull request e.g.:

class MyClass(HasTraits):
    i = Int()
    j = Int()
    
    @observe("j")
    def another_update(self, change):
        self.i = change.new * 2

mc = MyClass()
l = link((mc, "i"), (mc, "j"))

mc.i = 2

>>> mc.i
4
>>> mc.j
2

@SylvainCorlay, there is of course a danger of things not working as expected if the arguments passed to a function are wrong or inconsistent, but this is true in general. Hence, I don't see why this should be an issue for adding an useful feature in this case.

[maartenbreddels]

If it doesn't cause bouncing around, and we just say it's the responsibly of the user/programmer to make sure the transformations are fine I don't see a reason not to implement it.

[rmorshea]

@francisco-dlp, true - inconsistency was possible before - however if we decided that we should raise when it's encountered, that would impact how we implement transforms.

[francisco-dlp]

@rmorshea, I have modified the code to raise a TraitError when the link doesn't work as intended. Thanks for pointing this out, as I think that it'll help prevent some hard to debug issues.

[francisco-dlp]

If anybody would like to use this feature while waiting for this PR to get merged (or not) you could try link_traits. The purpose of the package is to implement the dlink and link functions of traitlets for enthought traits. It can be used to link traits to traits, traits to traitlets and (redundantly) traitlets to traitlets.