I have to admit that I had to construct a new specimen to be sure I
understood the distinction between sample1 and sample2. sample1 is
random in each 2x2 region; sample1 chooses the same element in all 2x2
regions, randomly if its x argument is random.
]specimen =: ,./4 8$i. 2 2
0 1 0 1 0 1 0 1
2 3 2 3 2 3 2 3
0 1 0 1 0 1 0 1
2 3 2 3 2 3 2 3
0 1 0 1 0 1 0 1
2 3 2 3 2 3 2 3
0 1 0 1 0 1 0 1
2 3 2 3 2 3 2 3
sample1=: 3 : 'wind (?@4: { ,);._3"2 y'
sample1^:2 specimen
2 3
1 1
sample1^:2 specimen
2 1
3 0
sample2=: 4 : 'wind x&{@,;._3"2 y'
(?@4: sample2 ])^:2 specimen
3 3
3 3
(?@4: sample2 ])^:2 specimen
0 0
0 0
On Wed, Jan 2, 2013 at 10:59 AM, Raul Miller <[email protected]> wrote:
> On Wed, Jan 2, 2013 at 10:47 AM, Brian Schott <[email protected]> wrote:
>> Regarding "Why?", I was training a neural net to recognize 13x13
>> images and I wanted to create slightly imprecise images to improve the
>> net, so I generated randomized compressions. Does that explain it?
>
> With this statement of purpose, other valid choices seem to include:
>
> 1. Pick a random pixel for each choice from 2 by 2 pixel region.
>
> sample1=: 3 : 'wind (?@4: { ,);._3"2 y'
> sample1^:2 specimen
>
> 2. Move the random choice out of sample
>
> sample2=: 4 : 'wind x&{@,;._3"2 y'
> (?@4: sample2 ])^:2 specimen
>
> so I still do not know why you made the choice you did, nor whether it
> was the right choice for your purposes.
>
>> Now my neural net is trained and I want to just compress 52x52 images
>> to 13x13 and am waffling on whether the filtering process now should
>> be similarly random or should be based on something like the "median
>> filter" value of the 4x4 subregion like Cliff Reiter does. Having
>> looked at the matrix and vector C routines supplied by Apple, which do
>> not offer a median filter-like option, I am leaning toward the type of
>> compression I mentioned in a previous message in this thread named
>> "vDSP_vcmprs". I am still trying to figure out its use, though, and
>> hoping that it use of "stride" will enable me to compress in
>> checkboard sub regions of 2x2 or 4x4. As to the choice of which item
>> to select in each region, I am still waffling on whether or not to
>> choose randomly in the spirit of statistical random sampling (not as
>> before to generate different samples).
>
> Stride sounds analogous. But if it can perform precisely this
> mechanism you would need to use it twice and one of the applications
> would need to be a stride over rows instead of a stride over pixels,
> and I do not know if vcmprs supports that nor if there's another
> mechanism for doing so. I'd have to spend more time than I feel
> comfortable with, I think, to discover that.
>
Yes, I will have to play with stride myself. Thanks for your help.
> Thanks,
>
> --
> Raul
--
(B=)
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
