I'll need to change the basis of some 10^7 vectors, each having
200 coordinates. So I will multiply one [200 x 200] matrix by 10^7 [200 x 1] vectors. I need it to run very fast but I need to code it fast (one day or less)
and my CUDA is poor, so I don't want to code it from scratch in CUDA or OpenCL. Maybe some existing library can do it for me? Notice that, if the solution uses GPGPU, the matrix should be transfered to the GPU only once, otherwise the performance will be poor. Could I could use OpenACC (or OpenMP, I don't know)? Is it possible to do this in a day?
I prefer open source solutions (for both convenience and ethical reasons) but I can tolerate a closed source solution, even paid (assuming it is not too costly).
This is for my dissertation.
Thank you for your attention.
You can put your vectors in a matrix, 200 * 10^7 is perhaps to much space at once depending on our system, so you can split it.
And then you use any code that is optimized for matrix matrix multiplication, like BLAS. There are many implementations on CPUs, GPUs (cuBLAS, MAGMA,...), multicores (PLASMA,...), or distributed memory.
Since you will have big matrices you vill have a better acceleration than by doing matrix vector multiplications.
You're going to multiply 10 million big vectors by a huge matrix that is the same for all of them.
It would be fastest if all possible decision-making could be compiled-out ahead of time.
In other words, there are lots of index calculations and loop testing that would be identically repeated millions of times.
This sounds like a perfect case for pre-compilation:
Write a small program that would take as input your 200x200 matrix data values, and have it print out a piece of program text defining a function capable of inputting the input vector and outputting the result vector.
It could look something like this:
void multTheMatrixByTheVector(double a[200], double b[200]){
b[0] = 0
+ a[0] * <a constant, the value of mat[0][0]>
+ a[1] * <a constant, the value of mat[1][0]>
...
+ a[199] * <a constant, the value of mat[199][0]>
;
b[1] = 0
+ a[0] * <a constant, the value of mat[0][1]>
+ a[1] * <a constant, the value of mat[1][1]>
...
+ a[199] * <a constant, the value of mat[199][1]>
;
...
b[199] = etc. etc.
}
You see, that function will be around 40000 lines long, but a decent compiler should be able to handle it.
Of course, if any of the matrix elements are zero, i.e. there's some sparsity, you can omit those lines (or let the compiler optimizer do it).
To do this on CUDA or vectorized instructions, you'd have to modify it accordingly, but that should be do-able.
When you include that function in your main program, it should be able to run about as fast as the machine can go.
It's not wasting any cycles doing index calculations, loop testing, or multiplying by empty matrix cells.
Then if it takes 10ns per multiply and add, my back-of-the envelope says it should take 400 usec per vector, or 4000 seconds overall - a little over an hour.
Related
I have an np.array of observations z where z.shape is (100000, 60). I want to efficiently calculate the 100000x100000 correlation matrix and then write to disk the coordinates and values of just those elements > 0.95 (this is a very small fraction of the total).
My brute-force version of this looks like the following but is, not surprisingly, very slow:
for i1 in range(z.shape[0]):
for i2 in range(i1+1):
r = np.corrcoef(z[i1,:],z[i2,:])[0,1]
if r > 0.95:
file.write("%6d %6d %.3f\n" % (i1,i2,r))
I realize that the correlation matrix itself could be calculated much more efficiently in one operation using np.corrcoef(z), but the memory requirement is then huge. I'm also aware that one could break up the data set into blocks and calculate bite-size subportions of the correlation matrix at one time, but programming that and keeping track of the indices seems unnecessarily complicated.
Is there another way (e.g., using memmap or pytables) that is both simple to code and doesn't put excessive demands on physical memory?
After experimenting with the memmap solution proposed by others, I found that while it was faster than my original approach (which took about 4 days on my Macbook), it still took a very long time (at least a day) -- presumably due to inefficient element-by-element writes to the outputfile. That wasn't acceptable given my need to run the calculation numerous times.
In the end, the best solution (for me) was to sign in to Amazon Web Services EC2 portal, create a virtual machine instance (starting with an Anaconda Python-equipped image) with 120+ GiB of RAM, upload the input data file, and do the calculation (using the matrix multiplication method) entirely in core memory. It completed in about two minutes!
For reference, the code I used was basically this:
import numpy as np
import pickle
import h5py
# read nparray, dimensions (102000, 60)
infile = open(r'file.dat', 'rb')
x = pickle.load(infile)
infile.close()
# z-normalize the data -- first compute means and standard deviations
xave = np.average(x,axis=1)
xstd = np.std(x,axis=1)
# transpose for the sake of broadcasting (doesn't seem to work otherwise!)
ztrans = x.T - xave
ztrans /= xstd
# transpose back
z = ztrans.T
# compute correlation matrix - shape = (102000, 102000)
arr = np.matmul(z, z.T)
arr /= z.shape[0]
# output to HDF5 file
with h5py.File('correlation_matrix.h5', 'w') as hf:
hf.create_dataset("correlation", data=arr)
From my rough calculations, you want a correlation matrix that has 100,000^2 elements. That takes up around 40 GB of memory, assuming floats.
That probably won't fit in computer memory, otherwise you could just use corrcoef.
There's a fancy approach based on eigenvectors that I can't find right now, and that gets into the (necessarily) complicated category...
Instead, rely on the fact that for zero mean data the covariance can be found using a dot product.
z0 = z - mean(z, 1)[:, None]
cov = dot(z0, z0.T)
cov /= z.shape[-1]
And this can be turned into the correlation by normalizing by the variances
sigma = std(z, 1)
corr = cov
corr /= sigma
corr /= sigma[:, None]
Of course memory usage is still an issue.
You can work around this with memory mapped arrays (make sure it's opened for reading and writing) and the out parameter of dot (For another example see Optimizing my large data code with little RAM)
N = z.shape[0]
arr = np.memmap('corr_memmap.dat', dtype='float32', mode='w+', shape=(N,N))
dot(z0, z0.T, out=arr)
arr /= sigma
arr /= sigma[:, None]
Then you can loop through the resulting array and find the indices with a large correlation coefficient. (You may be able to find them directly with where(arr > 0.95), but the comparison will create a very large boolean array which may or may not fit in memory).
You can use scipy.spatial.distance.pdist with metric = correlation to get all the correlations without the symmetric terms. Unfortunately this will still leave you with about 5e10 terms that will probably overflow your memory.
You could try reformulating a KDTree (which can theoretically handle cosine distance, and therefore correlation distance) to filter for higher correlations, but with 60 dimensions it's unlikely that would give you much speedup. The curse of dimensionality sucks.
You best bet is probably brute forcing blocks of data using scipy.spatial.distance.cdist(..., metric = correlation), and then keep only the high correlations in each block. Once you know how big a block your memory can handle without slowing down due to your computer's memory architecture it should be much faster than doing one at a time.
please check out deepgraph package.
https://deepgraph.readthedocs.io/en/latest/tutorials/pairwise_correlations.html
I tried on z.shape = (2500, 60) and pearsonr for 2500 * 2500. It has an extreme fast speed.
Not sure for 100000 x 100000 but worth trying.
I have implemented an algorithm that uses two other algorithms for calculating the shortest path in a graph: Dijkstra and Bellman-Ford. Based on the time complexity of the these algorithms, I can calculate the running time of my implementation, which is easy giving the code.
Now, I want to experimentally verify my calculation. Specifically, I want to plot the running time as a function of the size of the input (I am following the method described here). The problem is that I have two parameters - number of edges and number of vertices.
I have tried to fix one parameter and change the other, but this approach results in two plots - one for varying number of edges and the other for varying number of vertices.
This leads me to my question - how can I determine the order of growth based on two plots? In general, how can one experimentally determine the running time complexity of an algorithm that has more than one parameter?
It's very difficult in general.
The usual way you would experimentally gauge the running time in the single variable case is, insert a counter that increments when your data structure does a fundamental (putatively O(1)) operation, then take data for many different input sizes, and plot it on a log-log plot. That is, log T vs. log N. If the running time is of the form n^k you should see a straight line of slope k, or something approaching this. If the running time is like T(n) = n^{k log n} or something, then you should see a parabola. And if T is exponential in n you should still see exponential growth.
You can only hope to get information about the highest order term when you do this -- the low order terms get filtered out, in the sense of having less and less impact as n gets larger.
In the two variable case, you could try to do a similar approach -- essentially, take 3 dimensional data, do a log-log-log plot, and try to fit a plane to that.
However this will only really work if there's really only one leading term that dominates in most regimes.
Suppose my actual function is T(n, m) = n^4 + n^3 * m^3 + m^4.
When m = O(1), then T(n) = O(n^4).
When n = O(1), then T(n) = O(m^4).
When n = m, then T(n) = O(n^6).
In each of these regimes, "slices" along the plane of possible n,m values, a different one of the terms is the dominant term.
So there's no way to determine the function just from taking some points with fixed m, and some points with fixed n. If you did that, you wouldn't get the right answer for n = m -- you wouldn't be able to discover "middle" leading terms like that.
I would recommend that the best way to predict asymptotic growth when you have lots of variables / complicated data structures, is with a pencil and piece of paper, and do traditional algorithmic analysis. Or possibly, a hybrid approach. Try to break the question of efficiency into different parts -- if you can split the question up into a sum or product of a few different functions, maybe some of them you can determine in the abstract, and some you can estimate experimentally.
Luckily two input parameters is still easy to visualize in a 3D scatter plot (3rd dimension is the measured running time), and you can check if it looks like a plane (in log-log-log scale) or if it is curved. Naturally random variations in measurements plays a role here as well.
In Matlab I typically calculate a least-squares solution to two-variable function like this (just concatenates different powers and combinations of x and y horizontally, .* is an element-wise product):
x = log(parameter_x);
y = log(parameter_y);
% Find a least-squares fit
p = [x.^2, x.*y, y.^2, x, y, ones(length(x),1)] \ log(time)
Then this can be used to estimate running times for larger problem instances, ideally those would be confirmed experimentally to know that the fitted model works.
This approach works also for higher dimensions but gets tedious to generate, maybe there is a more general way to achieve that and this is just a work-around for my lack of knowledge.
I was going to write my own explanation but it wouldn't be any better than this.
Scipy (http://www.scipy.org/) offers two KD Tree classes; the KDTree and the cKDTree.
The cKDTree is much faster, but is less customizable and query-able than the KDTree (as far as I can tell from the docs).
Here is my problem:
I have a list of 3million 2 dimensional (X,Y) points. I need to return all of the points within a distance of X units from every point.
With the KDtree, there is an option to do just this: KDtree.query_ball_tree() It generates a list of lists of all the points within X units from every other point. HOWEVER: this list is enormous and quickly fills up my virtual memory (about 744 million items long).
Potential solution #1: Is there a way to parse this list into a text file as it is writing?
Potential solution #2: I have tried using a for loop (for every point in the list) and then finding that single point's neighbors within X units by employing: KDtree.query_ball_point(). HOWEVER: this takes forever as it needs to run the query millions of times. Is there a cKDTree equivalent of this KDTree tool?
Potential solution #3: Beats me, anyone else have any ideas?
From scipy 0.12 on, both KD Tree classes have feature parity. Quoting its announcement:
cKDTree feature-complete
Cython version of KDTree, cKDTree, is now feature-complete. Most
operations (construction, query, query_ball_point, query_pairs,
count_neighbors and sparse_distance_matrix) are between 200 and 1000
times faster in cKDTree than in KDTree. With very minor caveats,
cKDTree has exactly the same interface as KDTree, and can be used as a
drop-in replacement.
Try using KDTree.query_ball_point instead. It takes a single point, or array of points, and produces the points within a given distance of the input point(s).
You can use this function to perform batch queries. Give it, say, 100000 points at a time, and then write the results out to a file. Something like this:
BATCH_SIZE = 100000
for i in xrange(0, len(pts), BATCH_SIZE):
neighbours = tree.query_ball_point(pts[i:i+BATCH_SIZE], X)
# write neighbours to a file...
I am computing a similarity matrix based on Euclidean distance in MATLAB. My code is as follows:
for i=1:N % M,N is the size of the matrix x for whose elements I am computing similarity matrix
for j=1:N
D(i,j) = sqrt(sum(x(:,i)-x(:,j)).^2)); % D is the similarity matrix
end
end
Can any help with optimizing this = reducing the for loops as my matrix x is of dimension 256x30000.
Thanks a lot!
--Aditya
The function to do so in matlab is called pdist. Unfortunately it is painfully slow and doesnt take Matlabs vectorization abilities into account.
The following is code I wrote for a project. Let me know what kind of speed up you get.
Qx=repmat(dot(x,x,2),1,size(x,1));
D=sqrt(Qx+Qx'-2*x*x');
Note though that this will only work if your data points are in the rows and your dimensions the columns. So for example lets say I have 256 data points and 100000 dimensions then on my mac using x=rand(256,100000) and the above code produces a 256x256 matrix in about half a second.
There's probably a better way to do it, but the first thing I noticed was that you could cut the runtime in half by exploiting the symmetry D(i,j)==D(i,j)
You can also use the function norm(x(:,i)-x(:,j),2)
I think this is what you're looking for.
D=zeros(N);
jIndx=repmat(1:N,N,1);iIndx=jIndx'; %'# fix SO's syntax highlighting
D(:)=sqrt(sum((x(iIndx(:),:)-x(jIndx(:),:)).^2,2));
Here, I have assumed that the distance vector, x is initalized as an NxM array, where M is the number of dimensions of the system and N is the number of points. So if your ordering is different, you'll have to make changes accordingly.
To start with, you are computing twice as much as you need to here, because D will be symmetric. You don't need to calculate the (i,j) entry and the (j,i) entry separately. Change your inner loop to for j=1:i, and add in the body of that loop D(j,i)=D(i,j);
After that, there's really not much redundancy left in what that code does, so your only room left for improvement is to parallelize it: if you have the Parallel Computing Toolbox, convert your outer loop to a parfor and before you run it, say matlabpool(n), where n is the number of threads to use.
Greetings. I'm trying to approximate the function
Log10[x^k0 + k1], where .21 < k0 < 21, 0 < k1 < ~2000, and x is integer < 2^14.
k0 & k1 are constant. For practical purposes, you can assume k0 = 2.12, k1 = 2660. The desired accuracy is 5*10^-4 relative error.
This function is virtually identical to Log[x], except near 0, where it differs a lot.
I already have came up with a SIMD implementation that is ~1.15x faster than a simple lookup table, but would like to improve it if possible, which I think is very hard due to lack of efficient instructions.
My SIMD implementation uses 16bit fixed point arithmetic to evaluate a 3rd degree polynomial (I use least squares fit). The polynomial uses different coefficients for different input ranges. There are 8 ranges, and range i spans (64)2^i to (64)2^(i + 1).
The rational behind this is the derivatives of Log[x] drop rapidly with x, meaning a polynomial will fit it more accurately since polynomials are an exact fit for functions that have a derivative of 0 beyond a certain order.
SIMD table lookups are done very efficiently with a single _mm_shuffle_epi8(). I use SSE's float to int conversion to get the exponent and significand used for the fixed point approximation. I also software pipelined the loop to get ~1.25x speedup, so further code optimizations are probably unlikely.
What I'm asking is if there's a more efficient approximation at a higher level?
For example:
Can this function be decomposed into functions with a limited domain like
log2((2^x) * significand) = x + log2(significand)
hence eliminating the need to deal with different ranges (table lookups). The main problem I think is adding the k1 term kills all those nice log properties that we know and love, making it not possible. Or is it?
Iterative method? don't think so because the Newton method for log[x] is already a complicated expression
Exploiting locality of neighboring pixels? - if the range of the 8 inputs fall in the same approximation range, then I can look up a single coefficient, instead of looking up separate coefficients for each element. Thus, I can use this as a fast common case, and use a slower, general code path when it isn't. But for my data, the range needs to be ~2000 before this property hold 70% of the time, which doesn't seem to make this method competitive.
Please, give me some opinion, especially if you're an applied mathematician, even if you say it can't be done. Thanks.
You should be able to improve on least-squares fitting by using Chebyshev approximation. (The idea is, you're looking for the approximation whose worst-case deviation in a range is least; least-squares instead looks for the one whose summed squared difference is least.) I would guess this doesn't make a huge difference for your problem, but I'm not sure -- hopefully it could reduce the number of ranges you need to split into, somewhat.
If there's already a fast implementation of log(x), maybe compute P(x) * log(x) where P(x) is a polynomial chosen by Chebyshev approximation. (Instead of trying to do the whole function as a polynomial approx -- to need less range-reduction.)
I'm an amateur here -- just dipping my toe in as there aren't a lot of answers already.
One observation:
You can find an expression for how large x needs to be as a function of k0 and k1, such that the term x^k0 dominates k1 enough for the approximation:
x^k0 +k1 ~= x^k0, allowing you to approximately evaluate the function as
k0*Log(x).
This would take care of all x's above some value.
I recently read how the sRGB model compresses physical tri stimulus values into stored RGB values.
It basically is very similar to the function I try to approximate, except that it's defined piece wise:
k0 x, x < 0.0031308
k1 x^0.417 - k2 otherwise
I was told the constant addition in Log[x^k0 + k1] was to make the beginning of the function more linear. But that can easily be achieved with a piece wise approximation. That would make the approximation a lot more "uniform" - with only 2 approximation ranges. This should be cheaper to compute due to no longer needing to compute an approximation range index (integer log) and doing SIMD coefficient lookup.
For now, I conclude this will be the best approach, even though it doesn't approximate the function precisely. The hard part will be proposing this change and convincing people to use it.