MAINT: suggested speedups #1
Conversation
| for d in self.data: | ||
| self.distros.append(norm(loc=d[0], scale=d[1])) | ||
| # truncnorm takes care of each gaussian contribution integrating | ||
| # to unity, etc. |
There was a problem hiding this comment.
truncnorm a much better way of handling a clipped normal distribution, and it should be fairly fast.
|
|
||
| return _pdf | ||
|
|
||
| arrs = [d.pdf(x) for d in self.distros] |
There was a problem hiding this comment.
Removing a double nested loop should be a hell of a lot faster. Each of the d.pdf calls is vectorised.
| arrs = [d.cdf(x) for d in self.distros] | ||
| return np.sum(arrs, axis=0) / len(self.data) | ||
|
|
||
| def _ppf_single(self, q): |
There was a problem hiding this comment.
This approach to calculating ppf is taken from scipy.stats. If it works well there, it should work well here. Of course I may have ignored improved bracketing that you've already worked out for these distributions.
There was a problem hiding this comment.
q is normally used as the argument for ppf (at least in scipy.stats). Changing the argument name to q would benefit me (at least), and possibly others who are interested in contributing.
| args=[v]).root | ||
| vfun = np.vectorize(self._ppf_single, otypes='d') | ||
| _ppf = vfun(np.atleast_1d(q)) | ||
|
|
||
| if len(_ppf) == 1: |
There was a problem hiding this comment.
Probably not a good idea to have this, you probably need if _ppf.size == 1
>>> x = np.random.random(size=(1, 10))
>>> len(x)
1
I was browsing PyPI and came across the package. It's great to know that there's a community building up around refnx. I knew that @arm61 was working on stuff around this area, I guess this is to do with that.
The project infrastructure looks great. Automated testing, pypi, etc.
I've made this PR to suggest some possible speedups via removal of loops, vectorisation, etc. It's not currently passing tests, possibly because the PDFs are being calculated differently. That should only require minor tweaks though.