Francesc Alted: > A Friday 22 October 2010 20:54:21 braingateway escrigué: > >> Gaetan de Menten : >> >>> On Fri, Oct 22, 2010 at 10:53, Francesc Alted <fal...@pytables.org> >>> > wrote: > >>>> A Friday 22 October 2010 10:45:13 Francesc Alted escrigué: >>>> >>>>> A Friday 22 October 2010 10:19:50 Gaetan de Menten escrigué: >>>>> >>>>>> Hello, >>>>>> >>>>>> I guess you are pretty busy, but I would really appreciate your >>>>>> opinion about those two tickets (it's been 3 months without >>>>>> reply). I'm willing to make (other) patches, tests or whatever >>>>>> you want to get those changes included. Especially 287 is a >>>>>> burden for me since I have to distribute my patch along with my >>>>>> program and explain to my users how to patch their copy of >>>>>> PyTables... So I would like for it to be either included in >>>>>> some form, or rejected altogether so that I would work around >>>>>> it. >>>>>> >>>>> Yeah, too many things to do lately... Anyway, I plan to release >>>>> 2.2.1 next week (let's see). I've had a quick look into #287 and >>>>> seems good to me, so I plan to add it to the forthcoming release. >>>>> Which is the other you are referring to? >>>>> >>>> Er... you are referring to #286 indeed. I'll try to think about >>>> this (of course, a patch would help me think faster :-) >>>> >>> Thanks for the quick reply. I'll produce a patch for #286... in >>> time. That one is more a nitpick than anything, so I'll come back >>> to it in a few weeks now that I know you are possibly >>> interested... >>> >> Hi Gaëtan, >> >> I am just about to ask for the similar feature. Without this feature, >> I have to recreate the Expr object every time I try to change the >> start and end position, which is very annoying, although it will not >> affect the performance a lot. And the eval() can only be called >> once, other wise it is going to give wrong result. >> > > Now that I think, I ended implementing this behaviour because I needed a > way to pass a lot of parameters to the computational kernel. For > passing such parameters, I implemented the next methods for `Expr`: > > * setInputsRange(start=None, stop=None, step=None) > * setOutput(out, append_mode=False) > * setOutputRange(start=None, stop=None, step=None) > > Do you think that it is better to be able to pass these params to > `eval()` instead? Hmm, I'm not that sure... > > >> I try to create an expr object with certain expression, then I am >> going to change the input range in a loop, and repeatedly call >> expr.eval(). I found out this is impossible, only the first time >> calling expr.eval() will work, outputing (10,5) array. Then it >> always ouputs the same (5,) array, unless I recreate expr again. Is >> this an expected behavior? >> > > Recreating the expression (via `Expr` constructor) is very cheap in > terms of time. Why do you not like doing this? > > I know it does not affect the performance:-) It is just make the programming style more consitent;-p I should say I am already quite happy about the Expr behavior now. But you know, users are always greedy :-D >> By the way, is there any way to set the input range for several >> different dimensions? For example, if I get (4e12,100) array: A, and >> I probably want to evaluate expression only on A[x:y,2:5]. I think >> this will be very useful feature for large arrays. >> > > It should be pretty useful indeed. But, for doing this I'm afraid that > we'd need to implement something similar to NumPy views, but for > PyTables containers (this is how numexpr deals with the problem at > least). Unfortunately I don't think this is going to happen anytime > soon, so you will need to explicitly *copy* the interesting chunks of > the array prior to perform computations. > OK, I think I have to live with that then:-/ Anyway, I did not noticed much performance-worsening with an explicity *copy*.
Thanks a lot, LittleBigBrain ------------------------------------------------------------------------------ Nokia and AT&T present the 2010 Calling All Innovators-North America contest Create new apps & games for the Nokia N8 for consumers in U.S. and Canada $10 million total in prizes - $4M cash, 500 devices, nearly $6M in marketing Develop with Nokia Qt SDK, Web Runtime, or Java and Publish to Ovi Store http://p.sf.net/sfu/nokia-dev2dev _______________________________________________ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
