Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
  • 02:06
    QuLogic pinned #21655
  • 02:06
    QuLogic reopened #21655
  • 02:06
    QuLogic commented #21655
  • 01:05
    QuLogic labeled #21850
  • 01:01
    ojeda-e synchronize #21799
  • 01:00
    ojeda-e synchronize #21799
  • 00:56
    QuLogic milestoned #21520
  • 00:29
    ericpre commented #21374
  • 00:27
    ojeda-e edited #21799
  • 00:09
    ojeda-e synchronize #21799
  • 00:01
    ericpre synchronize #20839
  • Dec 06 23:57

    tacaswell on main

    Remove unused bbox arg to _conv… Merge pull request #21867 from … (compare)

  • Dec 06 23:57
    tacaswell closed #21867
  • Dec 06 23:56
    ojeda-e commented #21871
  • Dec 06 23:54
    ojeda-e synchronize #21871
  • Dec 06 23:53
    timhoffm commented #21374
  • Dec 06 23:49
    tacaswell milestoned #21867
  • Dec 06 23:49

    tacaswell on main

    Use partialmethod for better si… Merge pull request #21868 from … (compare)

  • Dec 06 23:49
    tacaswell milestoned #21868
  • Dec 06 23:49
    tacaswell closed #21868
Thomas A Caswell
@tacaswell
how are you adding the Line to the canvas?
Eric Wieser
@eric-wieser
Ah yes if I actually remember to do that everything is fine
Thomas A Caswell
@tacaswell
:) great!
I was starting to worry we were doing very strange things inside that would make over riding draw not work as expected...
Nathan Goldbaum
@ngoldbaum
who needs symlog, I have np.abs() ;)
np.abs() + epsilon I guess
plotting quantities that have a zero point and lots of dynamic range is just terrible, not much to be done about it
Eric Wieser
@eric-wieser
How hard is it to implement using np.asinh instead of log today?
Jody Klymak
@jklymak
I like @eric-wieser the suggestion above of arcsinh(x / linregion) where linregion is the range you want the scale to be linear. Its not exactly logarithmic until about 5 x linregion, but its pretty close by 2 x linreagion.
@eric-wieser it'd take some work. You need a Transform, scale, locator/formatter and preferably a norm. But its already all there, so not a lot of thinking really
Eric Wieser
@eric-wieser
I'd have thought you could use the same locator / formatter as for log graphs
Jody Klymak
@jklymak
Possibly?
Eric Wieser
@eric-wieser
Not sure of the distinction between Transform and scale there. Either way, not something I need, so will let someone else have a go at trying that out
Jody Klymak
@jklymak
each scale has a transform
you need the lower level transform so that the drawing stack can put things in the right place
the scale co-ordinates the transform and the locators and formatters.
You do need a different locator/formatter but can probably just use the symlognorm one
Antony Lee
@anntzer
@eric-wieser you may want to have a look at lines._AxLine for a recent example of Line subclass that updates itself (effectively) at draw time
there's some funky caching with lines because they both support (ugh) dynamically changing the unit and are kind of optimized given their very widespread use
Eric Wieser
@eric-wieser
I was using that originally, but decided not to rely on unreleased matplotlib, and it didn't work in 3D anyway
Happy to contribute the 3d one to mplot3d at some point
Thomas A Caswell
@tacaswell
@anntzer mind if I (force) push commits at matplotlib/matplotlib#16206
Antony Lee
@anntzer
Go for it
Antony Lee
@anntzer
ugh, looks like #8896 got (just) merged with three full sets of baseline images in the git history, and they're not exactly lightweight either (it's 2.5M each set), would you (@tacaswell) consider editing the (admittedly now public) history to squash them away?
Eric Wieser
@eric-wieser
Ugh, that's a pain
I'd lean towards immediately rewinding history, and if no one noticed push the squashed version a few days later
Antony Lee
@anntzer
(same here)
but I'll leave it to @tacaswell to decide
Thomas A Caswell
@tacaswell
hmm, so have to decide this like now
so: con: we have not done this in recent history and it seems very worrying to do
pro: squashes 4M out of our history
we have only merged 2 PRs and 1 PR opened sense then that merge all of which are @anntzer
ok, I'm sold on rewinding history. This seems like a safe low-stakes context to see if we can do this sort of rollback
Thomas A Caswell
@tacaswell
ok, done, lets see what the fallout is

previous head: 85fd1aba15cb0bcffebf630d79d3cd1343246b98
reverted head: cb904b187370119e965e5fe811f796772647e7df

problematic commit: 62d8d65caab54dfb9daf41d0982fa58061adc6db

PRs that have to be re-done (with previous heads)
16437 :18d552d8b6172d3869b4d4791db6495b715f46cc
16206 : c4d3cd4e797ab997df834fa02ef26f3299ee9682
8896 : 0492b885f0e97ce105e59958f705b6174539c6d5

I'm working on writing it up on discourse, but duming here to get it off my computer
Thomas A Caswell
@tacaswell
doing that force-push made everysingle open PR flag as "updated"
Till Stensitzki
@Tillsten
From my view, the arcsinh alternative is much more complicated to understand and to explain than the current symlog scale. "Between -1 and 1 it's linear, logarithmic elsewhere.".
Jody Klymak
@jklymak
@Tillsten Except, thats not what symlog is doing. There is a transition between the two regions (just like arcsinh).
Till Stensitzki
@Tillsten
What wrong with that transition?
Afaik it works exactly as I expect
Till Stensitzki
@Tillsten
Because some people expect smoothness?
Jody Klymak
@jklymak
Thats fine, but it needs to be documented. i.e. its not "logarithmic elsewhere", it asymptotes to being logarithmic elsewhere.
Till Stensitzki
@Tillsten
Could you link an issues or maybe explain more what you mean by that?
I think thats basically our alogorithm
Till Stensitzki
@Tillsten
The paper only mentions the abrupt change