code comments about suave HOT 5 OPEN

Glad the comments are (potentially) helpful :)

Good luck with the profiling - hopefully, that uncovers some easy-to-fix bottlenecks

from suave.

hi @manodeep, apologies for the delay here i lost track of this! this is very helpful feedback, thanks so much; all makes sense on a first read.

i just set up some basic speed and accuracy tests, so that when i go to address these comments I can see the difference. the scripts for those are here - i started with the most basic test of pair counting vs my estimator's version of it, will add more.

i will also check out the profiler to figure out where the current bottleneck is, and see how fixing these helps. will update here how it goes.

finally just a heads up that i'm about to change the name of my version of the repo/package!

from suave.

Meant to add earlier - will you please see if this numpy trick works for the brute-force calculation that you have? Presumably, you will need to adapt for the projection but may be some clever formulation will work for your case

from suave.

hi @manodeep, i've been trying to get that numpy trick to work - thanks for pointing me to it, the brute force calculation is hugely limiting right now. i can use the trick to compute the distances for all of the pairs, which works well. however I can't figure out how to adapt it to my formulation without hitting memory limits.

from what i understand, the trick constructs an NxN matrix of distances between the pairs, where N is the number of data points, like so:
dists = np.sqrt( -2 * np.dot(data, data.T) + np.sum(data**2, axis=1) + np.sum(data**2, axis=1)[:, np.newaxis] )
this is very fast but already seems a bit wild memory-wise so maybe i'm missing something?

anyway, then for my estimator, i need to for each pair distance return a (k,k) matrix (where k is usually 5-20, e.g. the number of bins), and then sum up these matrices over all the pairs. i can't figure out how to do this in a vectorized way, short of computing a (N,N,k,k) matrix which is ridiculous memory-wise. (more specifically, i compute a k-vector for each pair, and the (k,k) matrix is the outer product of the k-vector with itself. maybe there's a trick with this?)

i'll keep working at it but any advice appreciated!

from suave.

How about working in fixed chunks of 8/16/32 of N, N (say 8, 8 or 32, 32) and then generating the k, k matrix (so the output matrix would be [k, k, 8, 8] or [k, k, 32, 32]) and then summing over the last two dimensions. (Could be important to make the last two indices for the chunking, but I am not sure. Worth checking if the performance is better when the calculation is performed as [N, N, k, k] vs [k, k, N, N])

Stepwise:

• create the k, k matrix before all loops and zero out the matrix
• create a double for loop over i, j in chunks of 8 (or something small)
• calculate the [k, k, 8, 8] matrix for each chunk of 8, 8 in the vectorized way
• sum the last two dimensions and add to the matrix you created out of the loop
• account for the remaining pairs that are not evaluated (because of the chunks)

Would that work?

from suave.