Pass Cholesky to multivariate gaussian

[jucor]

Issue by jucor from Friday Nov 01, 2013 at 18:30 GMT
Originally opened as google-deepmind/torch-randomkit#5

---

[jucor]

Comment by jucor from Saturday Nov 09, 2013 at 08:15 GMT

---

The more I think of it, the more I want it :-)
Either as a separate function, or an optional options table as the last argument. That complicates a bit the argument passing, but that should be doable, since the covariance is never a table.

[jucor]

Comment by d11 from Monday Nov 11, 2013 at 10:36 GMT

---

Hmm, if we went with the latter, it'd look like the following, right?

multivariateGaussianRand( n | resultTensor, mu, { cholesky = myChol })

another option could be

multivariateGaussianRand( n | resultTensor, mu, cholesky, true )

where the last argument is a flag indicating that the covariance matrix is already decomposed…
I'm not sure what I favour for this, really… which is more torch-y? Has anyone given a preference about how to call it?

On 9 Nov 2013, at 08:15, Julien Cornebise notifications@github.com wrote:

The more I think of it, the more I want it :-)
Either as a separate function, or an optional options table as the last argument.

—
Reply to this email directly or view it on GitHub.

[jucor]

Comment by d11 from Monday Nov 11, 2013 at 10:39 GMT

---

On second thoughts, the 'flag' option doesn't work so well if we want to add support for passing the precision matrix too.

[jucor]

Comment by jucor from Monday Nov 11, 2013 at 10:39 GMT

---

I was thinking a hybrid of both:

multivariateGaussianRand( [n | resultTensor,] mu, covOrCholesky [, options])

where options is a table with

• cholesky = true or false (default false) indicates whether we passed a cholesky or a variance
• and any other setting that we may see fit to extend in the future, e.g. returnCholesky

[jucor]

Comment by d11 from Monday Nov 11, 2013 at 10:40 GMT

---

I see - yes, that could work quite well I think.

On 11 November 2013 10:39, Julien Cornebise notifications@github.comwrote:

I was thinking a hybrid of both:

multivariateGaussianRand( [n | resultTensor,] mu, covOrCholesky [, options])

where options is a table with

• cholesky = true or false (default false) indicates whether we passed
  a cholesky or a variance
• and any other setting that we may see fit to extend in the future,
  e.g. returnCholesky

—
Reply to this email directly or view it on GitHubhttps://github.com/google-deepmind/torch-randomkit/issues/5#issuecomment-28188896
.

[jucor]

Comment by jucor from Monday Nov 11, 2013 at 10:41 GMT

---

Yes, exactly: so we can have

multivariateGaussianRand( [n | resultTensor,] mu, covOrCholeskyOrPrecisionOrCholprecision [, options])

where options would have a flag for cholesky and a flag for precision -- thus allowing to pass the inverse of the Cholesky (although precision and cholesky-precision is more useful for logpdf rather than for the sampling, but at least we'd have a unified interface).

[jucor]

@d11 : done :) But I discovered #10 while doing this one...

[d11]

Nice! Thanks! We should put a note in the docs, if you didn't already

[jucor]

I'm doing it right now ;-)

On 18 Nov 2013, at 18:05, Dan Horgan notifications@github.com wrote:

Nice! Thanks! We should put a note in the docs, if you didn't already

—
Reply to this email directly or view it on GitHub.

[jucor]

And done: http://jucor.github.io/torch-distributions/#toc_13

On 18 Nov 2013, at 18:05, Julien Cornebise julien@deepmind.com wrote:

I'm doing it right now ;-)

On 18 Nov 2013, at 18:05, Dan Horgan notifications@github.com wrote:

Nice! Thanks! We should put a note in the docs, if you didn't already

—
Reply to this email directly or view it on GitHub.

[d11]

Awesome :)
On 18 Nov 2013, at 18:09, Julien Cornebise notifications@github.com wrote:

And done: http://jucor.github.io/torch-distributions/#toc_13

On 18 Nov 2013, at 18:05, Julien Cornebise julien@deepmind.com wrote:

I'm doing it right now ;-)

On 18 Nov 2013, at 18:05, Dan Horgan notifications@github.com wrote:

Nice! Thanks! We should put a note in the docs, if you didn't already

—
Reply to this email directly or view it on GitHub.

—
Reply to this email directly or view it on GitHub.