[Numpy-discussion] A few comments (MDP conversion)
Pietro Berkes
berkes at gatsby.ucl.ac.uk
Thu Jul 13 09:11:00 EDT 2006
Dear numpys,
a couple of weeks ago Tiziano and I completed the conversion of our data
processing library MDP to numpy. We collected a few ideas and
questions:
- we found the convertcode.py module quite useful to perform a first,
low-level conversion. We had some problem when 'typecode' was
used as a keyword argument because the parser only converts
'typecode=' to 'dtype=' and skips the cases where there is an
additional space character before the equal sign ('typecode =').
Other names that might be easily converted are 'ArrayType' and
'NewAxis'.
- some functions changed the columns/rows conventions ('cov', for
example). It would be really helpful to explicitly write this in the
list of necessary changes in the documentation. It could be nice
to have the file 'convertcode.py' to print a warning every time one
of this functions is used in the code.
- the linalg functions svd, eig, inv, pinv, diag, and possibly
others perform an upcast, e.g. from 'f' to 'd'. This is apparently
because the litelapack module only wraps double precision routines.
Wouldn't it be more consistent to cast the results back to the
numerical type of the input? Otherwise, the user always has to take
care of the casting, which makes the use of single precision
arrays quite cumbersome.
- we found a bug in the 'eig' function in the case the solution
is complex: in (linalg.py, line 269) 'Complex' should be 'w.dtype'
We wondered whether it would be possible to obtain SVN write access in
orderto be able to fix this kind of issues by ourselves in the future.
We could also contribute docstrings for some of the functions.
In general, we found the conversion to numpy quite straightforward
and would like to thank you all for the great work!
Cheers,
Pietro Berkes and Tiziano Zito
http://mdp-toolkit.sourceforge.net/
More information about the NumPy-Discussion
mailing list