Re: [Chanegset]: Re: cat does not convert sparse matrices?

[Jaroslav Hajek]

On Sat, Sep 13, 2008 at 10:44 PM, David Bateman <address@hidden> wrote:
> dbateman wrote:
>>
>> In fact I think I see two bugs here
>>
>>
>> Francesco Potorti`-2 wrote:
>>>
>>> In Octave 3.0.1, I cannot catenate a full double matrix with a sparse
>>> logical matrix.  Is this normal or a bug?
>>>
>>> octave> a=zeros(3,3,0);
>>> octave> b1=zeros(3,3);
>>> octave> b2=sparse(1,1,true,3,3);
>>> octave> whos a b1 b2
>>>
>>> *** local user variables:
>>>
>>>  Prot Name        Size                     Bytes  Class
>>>  ==== ====        ====                     =====  =====   rwd a
>>> 3x3x0                        0  double
>>>   rwd b1          3x3                         72  double
>>>   rwd b2          3x3                         21  logical
>>>
>>> Total is 10 elements using 93 bytes
>>>
>>> octave> size(cat(3,a,b1))
>>> ans =
>>>
>>>   3   3   1
>>>
>>
>> Trailing singleton dimensions should be dropped, So I suspect there is a
>> function call missing in the cat function that should do this.
>>
>
> The attached patch addresses this, but
>
>>
>>
>>> octave> size(cat(3,a,b2))
>>> error: invalid conversion of NDArray to Matrix
>>> error: range error for insert
>>> error: evaluating argument list element number 1
>>>
>>
>> The sparse logical matrix should be promoted to a sparse matrix and the
>> concatenation allowed. Ok, I'll look at this wen I'm on a development
>> machine.
>
> It seems I missed the point you made by this one.. Sparse logical matrices
> are promoted to sparse double matrices before concatenation with another
> double matrix. The issue above is that a 3x3x0 is an NDArray and the Sparse
> matrix type only supports 2D arrays and so the conversion of the 3x3x0 array
> to a sparse matrices prior to the concatenation fails. It would be
> interesting to know what matlab does in this case, as I suspect it fails as
> well. If it doesn't then I suppose we need to special case for this.
>
> In fact we used to special case for all empty matrices in data.cc (do_cat)
> and just drop them from the concatenation. However, something like
> "cat(1,[],single([]))" must return a single precision empty matrix and so
> even empty matrices play their part in selecting the type of the returned
> matrix. So care would have to be taken if we did allow this not to cause
> other issues.
>
> Cheers
> David
>
>
>
>
>
>
>>
>> D.
>>
>
>
> # HG changeset patch
> # User David Bateman <address@hidden>
> # Date 1221338610 -7200
> # Node ID 71f97e500f63e0a93e9cbe7f3cd5318827d3b3f9
> # Parent  9fe441ffcbaa7b5e2f2c41d3e83b1769f92235da
> cut trailing singeltons in cat function
>
> diff --git a/src/ChangeLog b/src/ChangeLog
> --- a/src/ChangeLog
> +++ b/src/ChangeLog
> @@ -1,3 +1,7 @@ 2008-09-09  David Bateman  <address@hidden
> +2008-09-13  David Bateman  <address@hidden>
> +
> +       * data.cc (do_cat): Chopping trailing singleton dimensions.
> +
>  2008-09-09  David Bateman  <address@hidden>
>
>        * DLD-FUNCTIONS/regexp.cc (octregexp_list): Distinguish between
> diff --git a/src/data.cc b/src/data.cc
> --- a/src/data.cc
> +++ b/src/data.cc
> @@ -1802,6 +1802,10 @@ do_cat (const octave_value_list& args, s
>                                 dv_tmp (dim) : 1);
>            }
>
> +         // Reshape, chopping trailing singleton dimensions
> +         dv.chop_trailing_singletons ();
> +         tmp = tmp.reshape (dv);
> +
>          retval = tmp;
>        }
>       else
>
> _______________________________________________
> Help-octave mailing list
> address@hidden
> https://www-old.cae.wisc.edu/mailman/listinfo/help-octave
>
>

David, this does not apply cleanly to release-3-0-x. The do_cat
function seems to have non-trivially diverged between the branches. I
tracked the previous changes to changeset 7997 in John's repo (by you)
that deals mainly with single prec - apparently you fixed something
"along the way". I can't transplant 7997 and at this moment I'm not
100% sure how to adapt the patch. If you wish this bug to be fixed in
3.0.x, please adapt the patch specially for 3.0.x.

regards

-- 
RNDr. Jaroslav Hajek
computing expert
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz