RBF epsilon (changes 0.15 => 0.16)
Hi everyone, In our work on dipy, we've run into a small issue moving from 0.15 to 0.16: https://github.com/nipy/dipy/issues/688 The gist of this is that we are using RBF to interpolate a signal measured on the unit sphere. To do this, we have designed a custom 'norm' function based on the angle between measurements (rather than the default Euclidean). Code is here: https://github.com/nipy/dipy/blob/master/dipy/core/sphere.py#L500 This seemed to work fine with 0.15, but the changes to the default epsilon introduced here: https://github.com/scipy/scipy/pull/4643 seem to be throwing us off. I've taken a look at the discussion on that PR and the issue that led to this work, and it does make sense, but I wanted to get some feedback to see how we should proceed. Setting a larger default epsilon in our code (effectively undoing this change) would be one way to go, but I want to make sure that it wouldn't be an obvious blunder, and that there isn't some issue on scipy that we should be thinking about. Thanks! Ariel
Hi Ariel, Off-the-cuff, I'd think that the value of epsilon is best managed on the user side. [I can't say anything intelligent about the issue in dipy, so I'd rather say nothing about it]. In general, it seems that scipy's rbf implementation could use some work, and here's one issue which summarizes some suggestions for it: https://github.com/scipy/scipy/issues/4790 There is also a PR, https://github.com/scipy/scipy/pull/5070 which was closed by the author, and I'm not sure what's the status of it. As things stand, chances are that we'll be tempted to break backwards compat for scipy.interpolate.Rbf. Since you're using it, your suggestions are most welcome (and patches are even more so!) Cheers, Evgeni On Thu, Aug 27, 2015 at 1:11 AM, Ariel Rokem <arokem@gmail.com> wrote:
Hi everyone,
In our work on dipy, we've run into a small issue moving from 0.15 to 0.16:
https://github.com/nipy/dipy/issues/688
The gist of this is that we are using RBF to interpolate a signal measured on the unit sphere. To do this, we have designed a custom 'norm' function based on the angle between measurements (rather than the default Euclidean). Code is here:
https://github.com/nipy/dipy/blob/master/dipy/core/sphere.py#L500
This seemed to work fine with 0.15, but the changes to the default epsilon introduced here: https://github.com/scipy/scipy/pull/4643 seem to be throwing us off.
I've taken a look at the discussion on that PR and the issue that led to this work, and it does make sense, but I wanted to get some feedback to see how we should proceed. Setting a larger default epsilon in our code (effectively undoing this change) would be one way to go, but I want to make sure that it wouldn't be an obvious blunder, and that there isn't some issue on scipy that we should be thinking about.
Thanks!
Ariel
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev
Hi Evgeni - thanks for the feedback! On Fri, Aug 28, 2015 at 11:33 AM, Evgeni Burovski < evgeny.burovskiy@gmail.com> wrote:
Hi Ariel,
Off-the-cuff, I'd think that the value of epsilon is best managed on the user side. [I can't say anything intelligent about the issue in dipy, so I'd rather say nothing about it].
work, and here's one issue which summarizes some suggestions for it:
https://github.com/scipy/scipy/issues/4790
There is also a PR,
https://github.com/scipy/scipy/pull/5070
which was closed by the author, and I'm not sure what's the status of it.
As things stand, chances are that we'll be tempted to break backwards compat for scipy.interpolate.Rbf. Since you're using it, your suggestions are most welcome (and patches are even more so!)
OK - I will read up on this topic (I see there are some suggestions on
That was my intuition. Just wanted to make sure we're not doing anything crazy. In general, it seems that scipy's rbf implementation could use some that issue), and try to see whether I can be helpful. Thanks again, Ariel
Cheers,
Evgeni
On Thu, Aug 27, 2015 at 1:11 AM, Ariel Rokem <arokem@gmail.com> wrote:
Hi everyone,
In our work on dipy, we've run into a small issue moving from 0.15 to 0.16:
https://github.com/nipy/dipy/issues/688
The gist of this is that we are using RBF to interpolate a signal measured on the unit sphere. To do this, we have designed a custom 'norm' function based on the angle between measurements (rather than the default Euclidean). Code is here:
https://github.com/nipy/dipy/blob/master/dipy/core/sphere.py#L500
This seemed to work fine with 0.15, but the changes to the default epsilon introduced here: https://github.com/scipy/scipy/pull/4643 seem to be throwing us off.
I've taken a look at the discussion on that PR and the issue that led to this work, and it does make sense, but I wanted to get some feedback to see how we should proceed. Setting a larger default epsilon in our code (effectively undoing this change) would be one way to go, but I want to make sure that it wouldn't be an obvious blunder, and that there isn't some issue on scipy that we should be thinking about.
Thanks!
Ariel
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org http://mail.scipy.org/mailman/listinfo/scipy-dev
participants (2)
-
Ariel Rokem -
Evgeni Burovski