[SciPy-dev] Note on changeset 5878: Add old_behavior arg for convolve2d and correlate2d

[David Cournapeau]

Pearu Peterson wrote:
> Hi,
>
> Currently scipy.signal.convolve gives the following warning by default:
>
> """
> Current default behavior of convolve and correlate functions is deprecated.
>
> Convolve and corelate currently swap their arguments if the second argument
> has dimensions larger than the first one, and the mode is relative to
> the input
> with the largest dimension. The new behavior is to never swap the
> inputs, which
> is what most people expects, and is how correlation is usually defined.
>
> You can control the behavior with the old_behavior flag - the flag will
> disappear in scipy 0.9.0, and the functions will then implement the new
> behavior only.
> """
>
> The problem right now is that it is difficult to write a stable
> application that would not give this warning.
> Using old_behavior=False argument means that when upgrading to scipy
> 0.9.0, the application will become broken because the kw argument will
> be removed. In my application the old and new behaviors are identical.
>
> My suggestion is to raise the deprecation warning only when the old
> and the new behavior are in conflict.

Yes, that's a good idea. I will try to fix this today, ping me if I forgot

>  For example, consider the diff below.
>
> What do you think?
>
> Pearu
>
> $ svn diff signaltools.py
> Index: signaltools.py
> ===================================================================
> --- signaltools.py	(revision 5953)
> +++ signaltools.py	(working copy)
> @@ -100,10 +100,11 @@
>      val = _valfrommode(mode)
>
>      if old_behavior:
> -        warnings.warn(DeprecationWarning(_SWAP_INPUTS_DEPRECATION_MSG))
>          if np.iscomplexobj(in2):
> +            # hmm, how is this related to deprecation warning?
>              in2 = in2.conjugate()

There is a doc problem here: the new behavior is not only to never swap
inputs, but also to take the conjugate of the second argument (that is
the usual definition of convolution).

cheers,

David