Simplify LDA input parameterization

[gokceneraslan]

I tried to simplify LDA input representation by using a simple M x V matrix of word frequencies where M and V represent number of documents and words. In the model, now instead of iterating over all words of all documents, iterations are over each element of the M x V matrix.

[bob-carpenter]

Thanks for submitting. I've been out for a while, so haven't been able to review this, but I'll get to it ASAP.

[bob-carpenter]

Please just add new models rather than replacing the existing ones.

The new implementations have very different memory properties (which will only be better in some dense cases).

[bob-carpenter]

Rather than looping to define gamma, a one-liner will do it:

gamma = count * (log(theta[i, ]) + log(phi[, j]);

The loop works for the count == 0 case (though a bit ineffciently) but presumably there aren't any zero-length documents in well-formed data sets, so I'd just write replace this whole loop with:

for (i in 1:M)
  for (j in 1:V)
    target += log_sum_exp(count * (log(theta[i, ]) + log(phi[, j]));

increment_log_prob is deprecated---this model has been around for a while without being updated.

[bob-carpenter]

Thanks for replacing <- --- that would be a good change for the original model, as well as replacing increment_log_prob with target +=.

square(sigma[m]) will be a bit more efficient, as would be sigma[m]^2.

[bob-carpenter]

Oh, and I'd suggest adding suffixes to existing model names like _counts to indicate you're taking sufficient stats rather than the raw data.

[gokceneraslan]

Oh, and I'd suggest adding suffixes to existing model names like _counts to indicate you're taking sufficient stats rather than the raw data.

You mean adding _counts to the new model? Because it's the one uses counts.

[bob-carpenter]

Anything to distinguish the way in which data is coded in the two approaches. So yes, I meant keeping <foo>.stan as is and adding <foo>_counts.stan or something similar for the sufficient stats version.

…

On Nov 17, 2018, at 6:04 PM, Gökçen Eraslan ***@***.***> wrote: Oh, and I'd suggest adding suffixes to existing model names like _counts to indicate you're taking sufficient stats rather than the raw data. You mean adding _counts to the new model? Because it's the one uses counts. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.