I'm very surprised that Java is that much faster than the initial implementation provided (after its been wrapped in a function). Feel like there is something non-obvious going on...
On Sunday, April 27, 2014 5:33:06 AM UTC-4, Carlos Becker wrote: > > I agree with Elliot, take a look at the performance tips. > Also, you may want to move the tic(), toc() out of the function, make sure > you compile it first, and then use @time <function calll> to time it. > > you may also get a considerable boost by using @simd in your for loops > (together with @inbounds) > Let us know how it goes ;) > > cheers. > > > El domingo, 27 de abril de 2014 09:39:03 UTC+2, Freddy Chua escribió: >> >> Alright, thanks! All these is looking very positive for Julia. >> >> On Sunday, April 27, 2014 3:36:23 PM UTC+8, Elliot Saba wrote: >>> >>> I highly suggest you read through the whole "Performance >>> Tips<http://julia.readthedocs.org/en/latest/manual/performance-tips/>" >>> page I linked to above; it has documentation on all these little features >>> and stuff. I did get a small improvement (~5%) by enabling SIMD extensions >>> on the two inner for loops, but that requires a very recent build of Julia >>> and is a somewhat experimental feature. Neat to have though. >>> -E >>> >>> >>> On Sun, Apr 27, 2014 at 12:14 AM, Freddy Chua <[email protected]> wrote: >>> >>>> wooh, this @inbounds thing is new to me... At least it does shows that >>>> Julia is comparable to Java. >>>> >>>> >>>> On Sunday, April 27, 2014 3:04:26 PM UTC+8, Elliot Saba wrote: >>>> >>>>> Since we have made sure that our for loops have the right boundaries, >>>>> we can assure the compiler that we're not going to step out of the bounds >>>>> of an array, and surround our code in the @inbounds macro. This is not >>>>> something you should do unless you're certain that you'll never try to >>>>> access memory out of bounds, but it does get the runtime down to 0.23 >>>>> seconds, which is on the same order as Java. Here's the full >>>>> code<https://gist.github.com/staticfloat/11339342>with all the >>>>> modifications made. >>>>> -E >>>>> >>>>> >>>>> On Sat, Apr 26, 2014 at 11:55 PM, Freddy Chua <[email protected]>wrote: >>>>> >>>>>> Stochastic Gradient Descent is one of the most important optimisation >>>>>> algorithm in Machine Learning. So having it perform better than Java is >>>>>> important to have more widespread adoption. >>>>>> >>>>>> >>>>>> On Sunday, April 27, 2014 2:03:28 PM UTC+8, Freddy Chua wrote: >>>>>> >>>>>>> This code takes 60+ secs to execute on my machine. The Java >>>>>>> equivalent takes only 0.2 secs!!! Please tell me how to optimise the >>>>>>> following code.begin >>>>>>> >>>>>>> begin >>>>>>> N = 10000 >>>>>>> K = 100 >>>>>>> rate = 1e-2 >>>>>>> ITERATIONS = 1 >>>>>>> >>>>>>> # generate y >>>>>>> y = rand(N) >>>>>>> >>>>>>> # generate x >>>>>>> x = rand(K, N) >>>>>>> >>>>>>> # generate w >>>>>>> w = zeros(Float64, K) >>>>>>> >>>>>>> tic() >>>>>>> for i=1:ITERATIONS >>>>>>> for n=1:N >>>>>>> y_hat = 0.0 >>>>>>> for k=1:K >>>>>>> y_hat += w[k] * x[k,n] >>>>>>> end >>>>>>> >>>>>>> for k=1:K >>>>>>> w[k] += rate * (y[n] - y_hat) * x[k,n] >>>>>>> end >>>>>>> end >>>>>>> end >>>>>>> toc() >>>>>>> end >>>>>>> >>>>>>> Sorry for repeated posting, I did so to properly indent the code.. >>>>>>> >>>>>> >>>>> >>>
