larry-discuss team mailing list archive

[Keith Goodman <kwgoodman@xxxxxxxxx>]

On Mon, Feb 8, 2010 at 6:39 AM, Keith Goodman <kwgoodman@xxxxxxxxx> wrote:
> On Sun, Feb 7, 2010 at 10:15 PM,  <josef.pktd@xxxxxxxxx> wrote:
>> On Mon, Feb 8, 2010 at 12:51 AM, Keith Goodman <kwgoodman@xxxxxxxxx> wrote:
>>> On Sun, Feb 7, 2010 at 9:03 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>>> On Sun, Feb 7, 2010 at 11:32 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>>>> On Sun, Feb 7, 2010 at 11:17 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>>>>> On Sun, Feb 7, 2010 at 11:15 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>>>>>> On Sun, Feb 7, 2010 at 11:08 PM, Keith Goodman <kwgoodman@xxxxxxxxx> wrote:
>>>>>>>> On Sun, Feb 7, 2010 at 8:06 PM, Keith Goodman <kwgoodman@xxxxxxxxx> wrote:
>>>>>>>>> On Sun, Feb 7, 2010 at 7:53 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>>>>>>>>> On Sun, Feb 7, 2010 at 10:43 PM, Keith Goodman <kwgoodman@xxxxxxxxx> wrote:
>>>>>>>>>>> On Sun, Feb 7, 2010 at 7:40 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>>>>>>>>>>> On Sun, Feb 7, 2010 at 10:23 PM, Keith Goodman <kwgoodman@xxxxxxxxx> wrote:
>>>>>>>>>>>>> On Sun, Feb 7, 2010 at 7:11 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>>> If you want to go this way, you could have lixs :  lix with strings,
>>>>>>>>>>>>>> and lixr : lix with repr , ... ?
>>>>>>>>>>>>>> lix would be safe  because it uses the labels directly, the other ones
>>>>>>>>>>>>>> are only recommended in restricted cases where the string (or other)
>>>>>>>>>>>>>> representation makes sense.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Too confusing with two I think.
>>>>>>>>>>>>>
>>>>>>>>>>>>>> if you want to allow slicing i.e. `:`  then I think lix would need to
>>>>>>>>>>>>>> use its __getitem__ instead of __call__ or __init__
>>>>>>>>>>>>>> lar1[1:4, lix['msft', google], lixs['2010-02-01':]   or lar1[1:4,
>>>>>>>>>>>>>> lix[['msft', google]], lixs['2010-02-01':] ?
>>>>>>>>>>>>>
>>>>>>>>>>>>> lar1[lix(date)]: will work. So will lar[lix(date1):lix(date2)]
>>>>>>>>>>>>>
>>>>>>>>>>>>> Allowing slices slows things down because now we have to look instead
>>>>>>>>>>>>> each slice object (slice.start, slice.stop) for lix object and if
>>>>>>>>>>>>> found convert them.
>>>>>>>>>>>>
>>>>>>>>>>>> I wouldn't mix lix and regular array slices or indices for the same
>>>>>>>>>>>> axis. I would restrict it to if either one of (slice.start,
>>>>>>>>>>>> slice.stop) are lix then both are interpreted as labels.
>>>>>>>>>>>>
>>>>>>>>>>>> either lar[lix[date1:date2],:]   or   lar[lix(date1):lix(date2), :]
>>>>>>>>>>>> but not lar[lix(date1):-3]
>>>>>>>>>>>
>>>>>>>>>>> How come? Seems handy to me.
>>>>>>>>>>
>>>>>>>>>> Just a feeling, mixing oranges and apples, (If I have one label, I
>>>>>>>>>> expect also to have the other, or I have neither in the other case)
>>>>>>>>>>
>>>>>>>>>> It would be useful to write a babyclass with the different versions of
>>>>>>>>>> getitem. It's easier to see what's going on and to experiment than
>>>>>>>>>> using the already more complicated getitem of larry.
>>>>>>>>>>
>>>>>>>>>> Josef
>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>> I don't know if lix should hold some code for conversion from label
>>>>>>>>>>>>>> representation to indices, or be just an identifier for use in
>>>>>>>>>>>>>> larry.__getitem__
>>>>>>>>>>>>>
>>>>>>>>>>>>> Yeah, a method getindex that takes label as input, perhaps. larry's
>>>>>>>>>>>>> getitem is already getting harry, so keeping the code in lix instead
>>>>>>>>>>>>> of adding yet more code to getitem might be a good idea. Or make a
>>>>>>>>>>>>> function in util.misc like a did for string indexing.
>>>>>>>>>
>>>>>>>>> I'm thinking of allowing only one label element in lix. But allow
>>>>>>>>> mixing label and integers for slicing.
>>>>>>>>>
>>>>>>>>> I tried allowing multiple labels per axis in the first indexing by
>>>>>>>>> labels blueprint. It's hard. So as a first step, only allow one. For
>>>>>>>>> example:
>>>>>>>>>
>>>>>>>>> lar[lix(date1):lix(date2)]
>>>>>>>>> lar[:lix(date2)]
>>>>>>>>> lar[lix(date1):-1]
>>>>>>>>
>>>>>>>> BTW, all this can already be done of course:
>>>>>>>>
>>>>>>>> lar[lar.labelindex(date1,0):lar.labelindex(date2,0)]
>>>>>>>> lar[:lar.labelindex(date2,0)]
>>>>>>>> lar[lar.labelindex(date1,0):-1]
>>>>>>>
>>>>>>> the usecase I had initially in mind was a list of labels
>>>>>>> lar1[:, ['msft', 'google', 'f2', 'f3'], ['open', 'close']].diff(3).log().diff(0)
>>>>>>
>>>>>> although it doesn't make economic sense
>>>>>>
>>>>>>>
>>>>>>> labelindex allows only single label
>>>>>>>
>>>>>>> Josef
>>>>>>>
>>>>>
>>>>> something like this would do what I had in mind:
>>>>>
>>>>> class lix2(object):
>>>>>    def __init__(self, label):
>>>>>        if type(label) != list:
>>>>>            raise TypeError, 'label must be a list'
>>>>>        self.label = label
>>>>>
>>>>> class A(object):
>>>>>
>>>>>    def __init__(self, data, label):
>>>>>        if type(label) != list:
>>>>>            raise TypeError, 'label must be a list'
>>>>>        self.label = label
>>>>>        self.data = data
>>>>>
>>>>>    def __getitem__(self, ind):
>>>>>        if isinstance(ind, lix2):
>>>>>            idx = map(self.label.index, ind.label)
>>>>>            return self.data[idx]
>>>>>
>>>>> aa = A(np.arange(10), 'a b c d e f g h i j'.split())
>>>>> print aa[lix2(['a', 'b', 'i'])]
>>>>>
>>>>>>>> aa.label
>>>>> ['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j']
>>>>>>>> aa.data
>>>>> array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>>>>>>> aa[lix2(['a', 'b', 'i'])]
>>>>> array([0, 1, 8])
>>>>>
>>>>
>>>>
>>>> some examples with my latest version with lix3, see attachment
>>>>
>>>>>>> aa = B(np.arange(10), 'a b c d e f g h i j'.split())
>>>>>>> print aa[lix2('a', 'b', 'i')]
>>>> [0 1 8]
>>>>>>> print aa[lix3('a', 'b', 'i')]
>>>> [0 1 8]
>>>>>>> print aa[lix3['a':'d']]
>>>> [0 1 2]
>>>>>>> print aa[lix3['a':'h']]
>>>> [0 1 2 3 4 5 6]
>>>>>>> print aa[lix3['c':'h']]
>>>> [2 3 4 5 6]
>>>>>>> print aa[lix3['c':-2]]
>>>> [2 3 4 5 6 7]
>>>>>>> print aa[lix3[2:'e']]
>>>> [2 3]
>>>>>>> print aa[lix3['c':'h':2]]
>>>> [2 4 6]
>>>>>>> print aa[lix3['c':]]
>>>> [2 3 4 5 6 7 8 9]
>>>>>>> print aa[lix3[:'c']]
>>>> [0 1]
>>>>>>>
>>>
>>> The problem is that these lists, when used to index into lar.x, will
>>> do fancy indexing. Is there a way to convert the list to something
>>> that doesn't do fancy indexing? Either that, or I'd have to add
>>> support for fancy indexing to larry.__getitem__. Something I'd like to
>>> do but a big project.
>>
>> I forgot, I needed to check a few examples, using array with correctly
>> broadcasted indices seems to work. I don't remember where/when we
>> discussed it, but *rectangular* indexing shouldn't be very difficult.
>>
>> If this kind of examples  lar[np.array([0,1,2])[:,None],[1,3]] works
>> correctly, the main work would be to add the None for the additional
>> axes, looks doable with enough tests.
>>
>>>>> lar = la.larry(np.ones((3,4)))
>>>>> lar[[0,1,2],[1,3]]
>> Traceback (most recent call last):
>>  File "C:\Josef\eclipsegworkspace\larry-josef\larry-josef\la\deflarry.py",
>> line 1384, in __getitem__
>>    x = self.x[index]
>> ValueError: shape mismatch: objects cannot be broadcast to a single shape
>>>>> lar.x[[0,1,2],[1,3]]
>> Traceback (most recent call last):
>> ValueError: shape mismatch: objects cannot be broadcast to a single shape
>>>>> lar[np.array([0,1,2])[:,None],[1,3]]
>> label_0
>>    0
>>    1
>>    2
>> label_1
>>    1
>>    3
>> x
>> array([[ 1.,  1.],
>>       [ 1.,  1.],
>>       [ 1.,  1.]])
>>>>> lar[[0,1,2],[1,3,0]]
>> Traceback (most recent call last):
>>  File "C:\Josef\eclipsegworkspace\larry-josef\larry-josef\la\deflarry.py",
>> line 1411, in __getitem__
>>    return larry(x, label)
>>  File "C:\Josef\eclipsegworkspace\larry-josef\larry-josef\la\deflarry.py",
>> line 85, in __init__
>>    if x.shape[i] != nlabel:
>> IndexError: tuple index out of range
>
> I don't think I want lar[index] to give different results from
> lar.x[index]. My goal is for larry and arrays to behave the same way
> where feasible. But I could borrow an ideas from the first blueprint:
>
> lar.lix[index]
>
> where index would only contain labels. Inside the lix method I would
> convert the labels to indices and then return lar[index_converted]. It
> would have the advantage of not slowing down an already slow
> larry.__getitem__. And the user would not have to wrap labels in a
> class. One downside is that you cannot do
>
> lar['price']['aapl'][date]
>
> instead it would be
>
> lar.lix['price'].lix['appl'].lix['date']
>
> Or
>
> lar['price', 'aapl', date]

Oh, and I forgot one of the main point, lix would use rectangular
indexing. So it would convert multiple lists to rectangular indexing
before passing index_converted to larry.__getitem__.