12:50 p.m.
Dear all,
I am facing the following simple problem for which I only find a loop
solution but loops are (generally?) slow compared to vectorized
approaches (if possible).
<hansl>
matrix Y = seq(1,10)' # original realizations
SEL = ceil(muniform(10,1)*10) # selection vector (select i-th row
from Y)
eval Y ~ SEL
matrix X = zeros(10,1) # put re-sampled Y values here
# I want to avoid this annoying loop...
loop i=1..rows(SEL) -q
catch X[i]=A[SEL[i]] # i-th position Y value
endloop
</hansl>
In MATLAB this works simply by: X = Y(SEL)
Is there any way to vectorize this task?? Thanks in advance.
Artur
1:25 p.m.
Am 19.10.2017 um 14:50 schrieb Artur T.:
I am facing the following simple problem
Are you sure you don't simply need the built-in resample() function?
for which I only find a loop
solution but loops are (generally?) slow compared to vectorized
approaches (if possible).
Actually, in my experience loops aren't so slow in hansl. There have
been cases where I invested some effort to replace a loop with some
matrix-oriented tricks and didn't get any speed gains. But of course it
depends on the particular case.
In MATLAB this works simply by: X = Y(SEL)
Have you tried Y[SEL] (square brackets, as with all indexation in hansl)?
cheers,
sven
1:35 p.m.
Am 19.10.2017 um 14:50 schrieb Artur T.:
> I am facing the following simple problem
Are you sure you don't simply need the built-in resample() function?
> for which I only find a loop solution but loops are (generally?) slow
> compared to vectorized approaches (if possible).
Actually, in my experience loops aren't so slow in hansl. There have
been cases where I invested some effort to replace a loop with some
matrix-oriented tricks and didn't get any speed gains. But of course
it depends on the particular case.
> In MATLAB this works simply by: X = Y(SEL)
Have you tried Y[SEL] (square brackets, as with all indexation in hansl)?
cheers,
sven
Oh my god, this was so trivial. Thank you, Sven! I totally forgot about
this solution... :-/ In the example below the gain in speed is enormous
(26x faster)
<hansl>
set verbose off
scalar T = 100
matrix Y = seq(1,T)' # original realizations
set stopwatch
loop j=1..1000 -q
SEL = ceil(muniform(T,1)*10) # selection vector (select i-th
row from Y)
matrix X = zeros(T,1) # put re-sampled Y values here
loop i=1..rows(SEL) -q
X[i]=Y[SEL[i]] # i-th position Y value
endloop
endloop
scalar T1 = $stopwatch
printf "Loop version = %.3f sec.\n", $stopwatch
set stopwatch
loop j=1..1000 -q
SEL = ceil(muniform(T,1)*10) # selection vector (select i-th
row from Y)
matrix X = zeros(T,1) # put re-sampled Y values here
X = Y[SEL]
endloop
scalar T2 = $stopwatch
printf "Vectorized version = %.3f sec.\n", $stopwatch
eval T1/T2
</hanslY
Artur
1:40 p.m.
Am 19.10.2017 um 15:35 schrieb Artur T.:
Am 19.10.2017 um 14:50 schrieb Artur T.:
> Are you sure you don't simply need the built-in resample()
function?
> Have you tried Y[SEL] (square brackets, as with all indexation in
hansl)?
>
> cheers,
> sven
Oh my god, this was so trivial. Thank you, Sven! I totally forgot about
this solution... :-/ In the example below the gain in speed is enormous
(26x faster)
OK, still it looks as if you just want to resample(). It would also be
interesting how fast that is in comparison.
-s
1:44 p.m.
On 19.10.2017 15:40, Sven Schreiber wrote:
Am 19.10.2017 um 15:35 schrieb Artur T.:
> Am 19.10.2017 um 14:50 schrieb Artur T.:
>> Are you sure you don't simply need the built-in resample() function?
>> Have you tried Y[SEL] (square brackets, as with all indexation in
>> hansl)?
>>
>> cheers,
>> sven
>
> Oh my god, this was so trivial. Thank you, Sven! I totally forgot
> about this solution... :-/ In the example below the gain in speed is
> enormous (26x faster)
OK, still it looks as if you just want to resample(). It would also be
interesting how fast that is in comparison.
I am programming a stationary bootstrap where the block length varies
randomly. This is why I avoid resample(). As resample() is written in C
(I assume) it should be faster, I guess...
Artur
2:05 p.m.
Am 19.10.2017 um 15:44 schrieb Artur T.:
I am programming a stationary bootstrap where the block length varies
randomly. This is why I avoid resample(). As resample() is written in C
(I assume) it should be faster, I guess...
Then perhaps you should look at the "SB" contributed package (by F. Di
Iorio, S. Fachin).
If it doesn't do what you need it would still be useful to know.
cheers,
sven
2:26 p.m.
Am 19.10.2017 um 15:44 schrieb Artur T.:
> I am programming a stationary bootstrap where the block length varies
> randomly. This is why I avoid resample(). As resample() is written in
> C (I assume) it should be faster, I guess...
Then perhaps you should look at the "SB" contributed package (by F. Di
Iorio, S. Fachin).
If it doesn't do what you need it would still be useful to know.
cheers,
sven
Good idea, I did the horse race (see the attached file). The output is
almost identical for 20000 MC iterations and using the SB's sample
script. My SBAT() function beats SB() substantially.
<OUTPUT>
*******************************************************************************
SB() took = 10.134 sec.
estimated coefficients 4.47 1.29 -2.40
95% bootstrap confidence intervals
Lower limits 3.06 1.08 -2.91
Upper limits 5.77 1.52 -1.89
*******************************************************************************
*******************************************************************************
SBAT() took = 6.445 sec.
estimated coefficients 4.47 1.29 -2.40
95% bootstrap confidence intervals
Lower limits 3.23 1.10 -2.86
Upper limits 5.63 1.49 -1.94
*******************************************************************************
</OUTPUT>
Artur
8:51 a.m.
@Stefano: If you are willing, Stefano, and conditionally that my code
works correct, we could update or extend your version of the SB-package
by the function I programmed. What do you think, Stefano
Best,
Artur
On 19.10.2017 16:26, Artur T. wrote:
Am 19.10.2017 um 15:44 schrieb Artur T.:
>
>> I am programming a stationary bootstrap where the block length
>> varies randomly. This is why I avoid resample(). As resample() is
>> written in C (I assume) it should be faster, I guess...
>
> Then perhaps you should look at the "SB" contributed package (by F.
> Di Iorio, S. Fachin).
> If it doesn't do what you need it would still be useful to know.
>
> cheers,
> sven
Good idea, I did the horse race (see the attached file). The output is
almost identical for 20000 MC iterations and using the SB's sample
script. My SBAT() function beats SB() substantially.
<OUTPUT>
*******************************************************************************
SB() took = 10.134 sec.
estimated coefficients 4.47 1.29 -2.40
95% bootstrap confidence intervals
Lower limits 3.06 1.08 -2.91
Upper limits 5.77 1.52 -1.89
*******************************************************************************
*******************************************************************************
SBAT() took = 6.445 sec.
estimated coefficients 4.47 1.29 -2.40
95% bootstrap confidence intervals
Lower limits 3.23 1.10 -2.86
Upper limits 5.63 1.49 -1.94
*******************************************************************************
</OUTPUT>
Artur
9:15 a.m.
Am 20.10.2017 um 10:51 schrieb Artur T.:
@Stefano: If you are willing, Stefano, and conditionally that my code
works correct, we could update or extend your version of the SB-package
by the function I programmed. What do you think, Stefano
Excellent idea (I'm not Stefano though :-),
I was going to suggest the same thing.
cheers,
sven
2373
days inactive
2374
days old
8 comments
2 participants
participants (2)
-
Artur T. -
Sven Schreiber