Enh arbitrary scale #12818

jklymak · 2018-11-15T20:11:51Z

PR Summary

As pointed out by @ImportanceOfBeingErnest in the SecondaryAxes PR (#11859 (comment)), there is probably call for an arbitrary axis scale, so, here you go.

import matplotlib.pyplot as plt
import matplotlib.scale as mscale
import numpy as np

fig, ax = plt.subplots()
ax.plot(range(1, 1000))

def forward(x):
    return x**2

def inverse(x):
    return x**(1/2)

ax.set_xscale('function', functions=(forward, inverse))
ax.set_xlim(1, 1000)
plt.show()

PR Checklist

Has Pytest style unit tests
Code is Flake 8 compliant
New features are documented, with examples if plot related
Documentation is sphinx and numpydoc compliant
Added an entry to doc/users/next_whats_new/ if major new feature (follow instructions in README.rst there)
Documented in doc/api/api_changes.rst if API changed in a backward-incompatible way

jklymak · 2018-11-15T20:18:33Z

Note this is also a first-step in solving issues like #12665 and #12808 for colorbars with non-standard norms. Though we will have a bit of an issue getting from a norm to a transform. My guess is that the API here will need to expand to allow a Transform with an inverse as the argument as well as the two-tuple. But I'll let others comment on the basic idea first.

anntzer · 2018-11-15T21:59:19Z

lib/matplotlib/scale.py

+             self._inverse = inverse
+         else:
+             raise ValueError('arguments to ArbitraryTransform must '
+                              'be functions.')


generally, exception messages have no final dot (applies throughout)

anntzer · 2018-11-15T22:00:19Z

lib/matplotlib/scale.py

+                              'be functions.')
+
+     def transform_non_affine(self, values):
+         with np.errstate(divide='ignore', invalid='ignore'):


I think setting the errstate should be the job of the function itself? (aka. let's assume that whoever passes arbitrary functions in knows what they're doing, or document that they should)

Sounds good to me. That was leftover from LogScale (I think), and I didn't quite grok what it was doing... I'll need to make an example that does the error checking though....

well, in logscale it effectively belongs to the user-defined (meaning by us) function

anntzer · 2018-11-15T22:01:50Z

lib/matplotlib/scale.py

+
+     name = 'arbitrary'
+
+     def __init__(self, axis, functions=None):


Just don't set a default to functions, as None isn't a valid value anyways?

Wonder whether the locator and formatter classes should be arguments too, i.e. I'd write something like

def __init__(self, axis, functions, *, major_locator_class=AutoLocator, minor_locator_class=NullLocator, major_formatter_class=ScalarFormatter, minor_formatter_class=NullFormatter):

I think you're also going to run into "interesting" issues in shared-axis handling in the implementation of cla(), which assumes that you can "copy" a scale by doing

self.xaxis._scale = mscale.scale_factory( self._sharex.xaxis.get_scale(), self.xaxis)

although it's not clear why the code doesn't just do

self.xaxis._scale = self._sharex.xaxis._scale

...

Thats a good idea!

So the problem w/ passing the class instead of an instance of the class is that the Locators and Formatters need to be passed arguments, so I think it makes more sense to do

def __init__(self, axis, functions, *, major_locator=AutoLocator(), minor_locator=NullLocator(), major_formatter=ScalarFormatter(), minor_formatter=NullFormatter()):

But that won't work because you can't share a locator/formatter across multiple axes right now (as they keep a reference to their axis -- they probably shouldn't, but that's another in depth refactoring...).

Hmmm, well, I guess folks will just have to change the locators and formatters manually

anntzer · 2018-11-15T22:03:17Z

lib/matplotlib/scale.py

+            TODO
+
+         """
+         if functions is None or len(functions) < 2:


I'd probably just write forward, inverse = functions and get whatever error message you have when functions is not an iterable of len 2, rather than supplying your own error message. (After all, as of this commit passing a non-interable will also give you the builtin message.)

anntzer · 2018-11-15T22:34:55Z

looks like you got an extra commit in there

timhoffm · 2018-11-18T09:32:22Z

lib/matplotlib/scale.py

+         Parameters
+         ----------
+
+         forward: The forward function for the transform


Should be something like

forward : callable The forward function for the transform. It must have the signature:: def forward(values: array-like) -> array-like

Are there any additional constraints? E.g. must the function be monotonic? Do we need more precise type descriptions?

Changed to

forward: callable The forward function for the transform. This function must have an inverse and, for best behavior, be monotonic. It must have the signature:: def forward(values: array-like) -> array-likeThe forward function for the transform inverse: callable The inverse of the forward function. Signature as ``forward``.

timhoffm · 2018-11-18T09:36:24Z

lib/matplotlib/scale.py

+         axis: the axis for the scale
+
+         functions: (forward, inverse)
+            two-tuple of the forward and inverse functions for the scale.


See ArbitraryTransform for details on the functions.

Actually this is more user facing so I changed this to have the more complete info as well. A bit repetitive, but...

lib/matplotlib/scale.py

timhoffm · 2018-11-18T09:39:12Z

lib/matplotlib/scale.py

@@ -545,6 +623,7 @@ def limit_range_for_scale(self, vmin, vmax, minpos):
    'log':    LogScale,
    'symlog': SymmetricalLogScale,
    'logit':  LogitScale,
+    'arbitrary': ArbitraryScale,


It's a bit misleading that LogTransformBase can have an ArbitraryScale.

Not following this comment...

Thanks for the suggestions @timhoffm This was the only one I didn't understand...

Sorry, I misread the code. Never mind.

lib/matplotlib/tests/test_scale.py

tacaswell · 2018-11-19T20:32:37Z

Only did a quick read, but 👍 to the functionality.

anntzer · 2018-11-19T21:14:43Z

Minor preference for naming this FuncScale/set_scale("func", ...) consistently with FuncFormatter, and also quite a bit shorter than Arbitrary :)

jklymak · 2018-11-25T16:49:05Z

OK< name changed to FuncScale and FuncTransform. We also now call axis.set_scale('function') rather than axis.set_scale('arbitrary').

This needs to be squashed before merging because I have an extra image in there, but want to keep the old commit in case someone wants me to go back to Arbitrary....

anntzer · 2018-11-26T10:15:54Z

lib/matplotlib/scale.py

+     is_separable = True
+     has_inverse = True
+
+     def __init__(self, forward, inverse):


Is there a reason why FuncTransform takes the two functions as separate args but FuncScale takes the pair as single arg? (Perhaps there is, didn't think more than that about it.)

Just because the kwarg gets passed in as a tuple, whereas to me it makes more sense for the transform to take them separately. But I don't feel very strongly about it.

ImportanceOfBeingErnest · 2018-12-10T09:51:43Z

lib/matplotlib/scale.py

+         axis.set_minor_formatter(NullFormatter())
+         # update the minor locator for x and y axis based on rcParams
+         if rcParams['xtick.minor.visible']:
+             axis.set_minor_locator(AutoMinorLocator())


Can you incoorporate the fix from #12938 here as well?

Sorry for the delay - Done!

anntzer · 2019-01-03T21:56:08Z

examples/scales/scales.py

+        return np.rad2deg(
+                np.ma.log(np.abs(np.ma.tan(masked) + 1.0 / np.ma.cos(masked))))
+    else:
+        return np.rad2deg(np.log(np.abs(np.tan(a) + 1.0 / np.cos(a))))


Can you just rely on this returning nan for invalid values? (possibly wrap in with np.errstate(invalid="ignore") to silence warnings)
At least the test below returns nan when square-rooting negative values.

This is just an example, right? More sophisticated users (than me) can add any error catching they want if they want to implement this...

My point is that

# Function Mercator transform def forward(a): a = np.deg2rad(a) return np.rad2deg(np.log(np.abs(np.tan(a) + 1 / np.cos(a))))

is much shorter, and basically works just as well (AFAICT...)

Sure, probably. I just stole the code from the custom_scale.py example to show that it was directly transferable. But I simplified in the latest commit.

anntzer · 2019-01-16T18:00:39Z

lib/matplotlib/tests/test_scale.py

+        good = x >= 0
+        y = np.full_like(x, np.NaN)
+        y[good] = x[good]**(1/2)
+        return y


I guess in the same vein something as simple as

def forward(x): return x**(1/2)

also works (apparently it does in the scales.py example above...)? In which case you can probably even just pass them as lambdas below...

Sure, they can be lambdas, but I don't find that easier to read than just defining the functions.... But main point taken...

dstansby · 2019-01-16T21:12:29Z

My immediate thought on this is what happens if inverse isn't actually the inverse of forward?

jklymak · 2019-01-16T21:16:38Z

My immediate thought on this is what happens if inverse isn't actually the inverse of forward?

Garbage in, garbage out?

anntzer · 2019-01-17T00:30:52Z

lib/matplotlib/tests/test_scale.py

@@ -1,6 +1,8 @@
 from matplotlib.testing.decorators import image_comparison
 import matplotlib.pyplot as plt
 from matplotlib.scale import Log10Transform, InvertedLog10Transform
+import matplotlib.ticker as mticker


the import is not actually needed?

anntzer · 2019-01-17T00:31:32Z

lib/matplotlib/scale.py

+            Both functions must have the signature::
+
+               def forward(values: array-like) ->
+                       array-likeThe forward function for the transform


something wrong with the formatting

oooops. Should be

Both functions must have the signature:: def forward(values: array-like) returns array-like

Actually

def forward(values: array-like) -> array-like

looks best to me (it's nearly normal type annotation syntax), but not insisting on that.

anntzer · 2019-01-17T00:48:24Z

lib/matplotlib/scale.py

+             It must have the signature::
+
+                def forward(values: array-like) ->
+                        array-likeThe forward function for the transform


Goops, sorry, should have checked the whole thing

anntzer · 2019-01-17T10:42:17Z

lib/matplotlib/scale.py

+                def forward(values: array-like) -> array-like
+
+         inverse: callable
+            The inverse of the forward function.  Signature as ``forward``.


This paragraph is indented one character less than the one for forward.

Also the type annotations actually also use a space before the colon (see rst syntax for definition lists, and numpy/numpydoc#78).

Likewise for the docstring below.

API: add ability to set formatter and locators from scale init FIX: rename to FuncScale DOC: add whats new FIX: simplify Mercator transform TST: simplify test

jklymak · 2019-01-21T17:51:18Z

ping on this one - I think I've addressed all review suggestions.... Thanks!

anntzer · 2019-01-21T18:15:15Z

The 3.7 failure is spurious.

jklymak · 2019-01-22T18:35:07Z

Thanks a lot for reviewing and merging!

jklymak added this to the v3.1 milestone Nov 15, 2018

jklymak added the status: work in progress label Nov 15, 2018

jklymak mentioned this pull request Nov 15, 2018

ENH: add secondary x/y axis #11859

Merged

5 tasks

anntzer reviewed Nov 15, 2018

View reviewed changes

anntzer mentioned this pull request Nov 15, 2018

Interesting things happen when calling cla() on shared axes with log scales #12819

Closed

jklymak force-pushed the enh-arbitrary-scale branch 3 times, most recently from bf1154f to d403e65 Compare November 18, 2018 02:51

jklymak added topic: ticks axis labels and removed status: work in progress labels Nov 18, 2018

jklymak force-pushed the enh-arbitrary-scale branch from d403e65 to 1e2d23e Compare November 18, 2018 02:54

timhoffm reviewed Nov 18, 2018

View reviewed changes

jklymak force-pushed the enh-arbitrary-scale branch from 1e2d23e to 8548e47 Compare November 18, 2018 16:20

timhoffm approved these changes Nov 18, 2018

View reviewed changes

anntzer reviewed Nov 26, 2018

View reviewed changes

ImportanceOfBeingErnest reviewed Dec 10, 2018

View reviewed changes

jklymak force-pushed the enh-arbitrary-scale branch from 99055d4 to c9984fe Compare January 3, 2019 17:49

anntzer reviewed Jan 3, 2019

View reviewed changes

jklymak force-pushed the enh-arbitrary-scale branch from 5ce3637 to 2ca2865 Compare January 16, 2019 17:07

anntzer reviewed Jan 16, 2019

View reviewed changes

jklymak force-pushed the enh-arbitrary-scale branch from 2ca2865 to 6220a18 Compare January 16, 2019 21:23

anntzer reviewed Jan 17, 2019

View reviewed changes

jklymak force-pushed the enh-arbitrary-scale branch 2 times, most recently from 19a2475 to bb16ae4 Compare January 17, 2019 00:36

anntzer reviewed Jan 17, 2019

View reviewed changes

jklymak force-pushed the enh-arbitrary-scale branch from bb16ae4 to bc91f3e Compare January 17, 2019 00:49

anntzer reviewed Jan 17, 2019

View reviewed changes

ENH: add arbitrary-scale functionality

9c74f77

API: add ability to set formatter and locators from scale init FIX: rename to FuncScale DOC: add whats new FIX: simplify Mercator transform TST: simplify test

jklymak force-pushed the enh-arbitrary-scale branch from bc91f3e to 9c74f77 Compare January 17, 2019 17:29

jklymak added the status: needs review label Jan 21, 2019

anntzer merged commit c22847b into matplotlib:master Jan 21, 2019

anntzer removed the status: needs review label Jan 21, 2019

jklymak deleted the enh-arbitrary-scale branch January 22, 2019 18:35

story645 mentioned this pull request May 17, 2023

Create Arbitrary scale transforms #12361

Closed

Uh oh!

Enh arbitrary scale #12818

Enh arbitrary scale #12818

Uh oh!

Conversation

jklymak commented Nov 15, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Summary

PR Checklist

Uh oh!

jklymak commented Nov 15, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anntzer Nov 15, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anntzer commented Nov 15, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

timhoffm Nov 18, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tacaswell commented Nov 19, 2018

Uh oh!

anntzer commented Nov 19, 2018

Uh oh!

jklymak commented Nov 25, 2018

Uh oh!

anntzer Nov 26, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anntzer Jan 3, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

jklymak commented Nov 15, 2018 •

edited

Loading

anntzer Nov 15, 2018 •

edited

Loading

timhoffm Nov 18, 2018 •

edited

Loading

anntzer Nov 26, 2018 •

edited

Loading

anntzer Jan 3, 2019 •

edited

Loading