On Tuesday, September 3, 2013 10:03:00 PM UTC-7, Mikera wrote:
>
>
>
> On Wednesday, 4 September 2013 12:37:33 UTC+8, Brian Craft wrote:
>>
>>
>>
>> On Tuesday, September 3, 2013 9:14:30 PM UTC-7, Mikera wrote:
>>>
>>> On Wednesday, 4 September 2013 10:00:42 UTC+8, Brian Craft wrote:
>>>
>>>> I'm loading data files of about 1-2G, which are composed of a bunch of
>>>> numeric data blocks. I need to store the data blocks w/o storing
>>>> duplicates. They arrive as vectors of floats, and are stored as primitive
>>>> byte arrays.
>>>>
>>>> I first tried memoizing the function that saves a block (returning an
>>>> id), with the core memoize function. This failed because every block
>>>> became
>>>> a different key in the memoization, regardless of the content. It looks
>>>> like clojure treats variables referencing primitive arrays as equal only
>>>> if
>>>> they refer to the same array. Note:
>>>>
>>>> cavm.core=> ({[1 2 3] "foo"} [1 2 3])
>>>> "foo"
>>>> cavm.core=> ({(float-array [1 2 3]) "foo"} (float-array [1 2 3]))
>>>> nil
>>>> cavm.core=> (let [a (float-array [1 2 3])] ({a "foo"} a))
>>>> "foo"
>>>>
>>>>
>>>>
>>>> I next tried memoizing over the vector of floats, however performance
>>>> became pathologically slow, and the process threw an OOM. I'm guessing
>>>> this
>>>> is due to the memory requirements of a clojure vector of floats vs. a
>>>> primitive array of bytes holding the same data. Is there an easy way to
>>>> compare the storage requirements?
>>>>
>>>> Any suggestions on how better to handle this?
>>>>
>>>
>>> You may want to use the :ndarray-float array implementation in the
>>> latest version of core.matrix.
>>>
>>> This is effectively a wrapper over a raw Java float array: so your
>>> storage requirement should be close to the size of the raw byte data
>>> (assuming the data blocks are large enough that the size of the wrapper is
>>> negligible)
>>>
>>
>> Ah, interesting.
>>
>> > *matrix-implementation*
>> :vectorz
>> > ({(matrix [1 2 3 4]) "foo"} (matrix [1 2 3 4]))
>> "foo"
>>
>> I don't otherwise need core.matrix at this point in the loader, but this
>> is convenient. Why does that work?
>>
>
> That works because Vectorz (the underlying Java lib) has a sane
> implementation of .equals and .hashCode. It's pretty fast as well, though
> it is still O(n) since it doesn't do hashcode caching.
>
> Note that the :vectorz implementation uses 8-byte doubles rather than
> 4-byte floats though - so if you really need single precision to keep the
> overall memory usage down then it might not be the best choice. I
> personally never use 4-byte floats because the numerical errors soon become
> problematic, but YMMV.
>
>
It looks like I can get this working for byte arrays as follows.
(deftype BAHashable [ba]
Object
(equals [f g] (java.util.Arrays/equals ba (.ba g)))
(hashCode [f] (java.util.Arrays/hashCode ba)))
({(BAHashable. (byte-array (map byte [1 2 3]))) "foo"} (BAHashable.
(byte-array (map byte [1 2 3]))))
"foo"
I'm less certain of whether this is a good idea.
--
--
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to [email protected]
Note that posts from new members are moderated - please be patient with your
first post.
To unsubscribe from this group, send email to
[email protected]
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
---
You received this message because you are subscribed to the Google Groups
"Clojure" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/groups/opt_out.
