numpy-discussion

[Numpy-discussion] Further comments

[Edward C. J. <edc...@er...>]

Should the variable "type" be used in numarray? It is an important 
function in Python.

---------------------------------------

There needs to be a function or method that returns the number of 
elements in an array.

def Elements(array):
     """Number of elements in an array.

     This version is slow.
     """
     return numarray.multiply.reduce(array.shape)

---------------------------------------

I write code using both PIL and numarray. PIL uses strings for modes and
numarray uses (optionally) strings as typecodes. This causes problems.
One fix is to emit a DeprecationWarning when string typecodes are used. 
Two functions are needed: StringTypeWarningOn and StringTypeWarningOff. 
The default should be to ignore this warning.

In my code I use the following workarounds:

def SameType(x, y):
     """Are the two input the same object of NumericType?"""
     if isinstance(x, NumericType) and isinstance(y, NumericType) \
            and x == y:
         return True
     return False

def IsTypeInList(typecode, seq):
     """Is a NumericType object in a list of NumericType objects?"""
     if not isinstance(typecode, NumericType):
         return False
     for item in seq:
         if isinstance(item, NumericType) and typecode == item:
             return True
     return False

---------------------------------------

The following function is useful for downsizing arrays. I suggest that
it should be a ufunc method. This is how I have used reduceat in Numeric.

def blockreduce(array, blocksizes, ufunc):
     """Apply ufunc.reduce to blocks in an array."""
     dims = len(array.shape)
     if type(blocksizes) is IntType:
         blocksizes = dims * [blocksizes]
     if len(blocksizes) != dims:
         raise TypeError, 'blocksizes must be same length as shape'
     for i in range(dims):
         if array.shape[i] % blocksizes[i] != 0:
             raise ValueError, 'blocksizes must exactly divide ' \
                'the corresponding array dimension'
     for i in range(dims):
         array = array.copy()
         newshape = (array.shape[0] / blocksizes[i], blocksizes[i]) + \
                       array.shape[1:]
         array.shape = newshape
         array = ufunc.reduce(array, 1)
         dims = len(array.shape)
         # (0,1,2,3) --> (1,2,3,0)
         perm =  tuple(range(1, dims)) + (0,)
         array = numarray.transpose(array, perm)
     return array

RE: [Numpy-discussion] Further comments

[Perry G. <pe...@st...>]

Edward Jones writes: 
> 
> Should the variable "type" be used in numarray? It is an important 
> function in Python.
> 
We pondered this for a while. Yes, it conflicts with the built 
in function (and so that function is aliased within the numarray
code to a different name). Do you see this causing problems for 
you? Normally there is no conflict in using it as a keyword name
for function or method arguments, but if you can show practical
cases where it is a problem, we may reconsider. It seemed to us
that type was the best name for the keyword, particularly since
we were discouraging people from thinging of typecodes.

> ---------------------------------------
> 
> There needs to be a function or method that returns the number of 
> elements in an array.
> 
> def Elements(array):
>      """Number of elements in an array.
> 
>      This version is slow.
>      """
>      return numarray.multiply.reduce(array.shape)
> 
> ---------------------------------------
> 
I don't disagree. It would be nice to have a consensus for what the
function name (and perhaps an array attribute?) should be. I believe
IDL uses nelements (which I find more suggestive than elements), but
perhaps others have better names.

> I write code using both PIL and numarray. PIL uses strings for modes and
> numarray uses (optionally) strings as typecodes. This causes problems.
> One fix is to emit a DeprecationWarning when string typecodes are used. 
> Two functions are needed: StringTypeWarningOn and StringTypeWarningOff. 
> The default should be to ignore this warning.
> 
I'm not sure I understand. Can you give me an example of problem
code or usage? It sounds like you are trying to test the types
of  PIL and numarray objects in a generic sense. But I'd understand
better if you could show an example.

> 
> The following function is useful for downsizing arrays. I suggest that
> it should be a ufunc method. This is how I have used reduceat in Numeric.
> 
> def blockreduce(array, blocksizes, ufunc):
>      """Apply ufunc.reduce to blocks in an array."""
>      dims = len(array.shape)
>      if type(blocksizes) is IntType:
>          blocksizes = dims * [blocksizes]
>      if len(blocksizes) != dims:
>          raise TypeError, 'blocksizes must be same length as shape'
>      for i in range(dims):
>          if array.shape[i] % blocksizes[i] != 0:
>              raise ValueError, 'blocksizes must exactly divide ' \
>                 'the corresponding array dimension'
>      for i in range(dims):
>          array = array.copy()
>          newshape = (array.shape[0] / blocksizes[i], blocksizes[i]) + \
>                        array.shape[1:]
>          array.shape = newshape
>          array = ufunc.reduce(array, 1)
>          dims = len(array.shape)
>          # (0,1,2,3) --> (1,2,3,0)
>          perm =  tuple(range(1, dims)) + (0,)
>          array = numarray.transpose(array, perm)
>      return array
> 
We certainly have frequent need for a binning function (i.e.,
the equivalent of a blockreduce for add). Do others see this
as a generally useful extension for all binary ufuncs?
(As a side comment, the requirement that the dimensions be
evenly divisible is often a pain; one question is to ask whether
this must be a requirement or not, or whether there is a way to
disable this requirement, i.e., permit the last block to be smaller
than usual?)

Thanks, Perry Greenfield 

Re: [Numpy-discussion] Further comments

[Magnus L. H. <ma...@he...>]

Perry Greenfield <pe...@st...>:
>
> > There needs to be a function or method that returns the number of 
> > elements in an array.
> > 
> > def Elements(array):
> >      """Number of elements in an array.
> > 
> >      This version is slow.
> >      """
> >      return numarray.multiply.reduce(array.shape)

len(ravel(array)) may be a faster in some cases, it seems. (Or
ravel(array).shape[0], for that matter).

BTW: Using the name "array" here seems just as bad as the use of
"type" in numarray... Just a thought :]

[snip]
> > The following function is useful for downsizing arrays. I suggest that
> > it should be a ufunc method. This is how I have used reduceat in Numeric.
> > 
> > def blockreduce(array, blocksizes, ufunc):
> >      """Apply ufunc.reduce to blocks in an array."""
[snip]
> We certainly have frequent need for a binning function (i.e.,
> the equivalent of a blockreduce for add). Do others see this
> as a generally useful extension for all binary ufuncs?

Not sure if I fully understand this -- but in my work on sequences and
time series, I certainly need to do stuff like sums and averages over
fixed-sized blocks of one-dimensional arrays... (Discretization.) But
there are easier ways of doing that, I suppose (with reshape or
indices, for example).

-- 
Magnus Lie Hetland
http://hetland.org