larry-discuss team mailing list archive

[Keith Goodman <kwgoodman@xxxxxxxxx>]

On Thu, Feb 4, 2010 at 8:04 AM,  <josef.pktd@xxxxxxxxx> wrote:
> On Thu, Feb 4, 2010 at 10:33 AM, Keith Goodman <kwgoodman@xxxxxxxxx> wrote:
>> On Wed, Feb 3, 2010 at 7:04 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>> On Wed, Feb 3, 2010 at 9:54 PM,  <josef.pktd@xxxxxxxxx> wrote:
>>>> On Wed, Feb 3, 2010 at 9:16 PM, Keith Goodman <kwgoodman@xxxxxxxxx> wrote:
>>>>> On Wed, Feb 3, 2010 at 6:06 PM, joep <josef.pktd@xxxxxxxxx> wrote:
>>>>>> Blueprint changed by joep:
>>>>>>
>>>>>> Whiteboard set to:
>>>>>>
>>>>>> assert_larry
>>>>>> http://bazaar.launchpad.net/~kwgoodman/larry/trunk/annotate/head%3A/la/tests/deflarry_nose_test.py#L49
>>>>>>
>>>>>> and the method check_function in test class
>>>>>>
>>>>>> http://bazaar.launchpad.net/~kwgoodman/larry/trunk/annotate/head%3A/la/tests/deflarry_nose_test.py#L128
>>>>>>
>>>>>> already contain the extracted boiler plate assert function to compare a
>>>>>> larry with desired data matrix x and labels
>>>>>>
>>>>>> the later has a view keyword to choose whether to verify nocopy or
>>>>>> noreference
>>>>>>
>>>>>> used in nosetests for larry methods
>>>>>
>>>>> I moved this to the larry-discuss list since it is hard to discuss it
>>>>> through the whiteboard on the blueprint.
>>>>>
>>>>> Yes, let's start with those functions and then prettify them.
>>>>>
>>>>> What should the signature be?
>>>>>
>>>>> The current signature is:
>>>>>
>>>>> assert_larry(opname, la, t, lab, msgstr)
>>>>>
>>>>> How about changing that to the same signature as np.testing.assert_equal? So:
>>>>>
>>>>> assert_equal(actual, desired, err_msg='', verbose=True)
>>>>>
>>>>> Then we don't have to separate the data and the label. And instead of
>>>>> nancode we can use the numpy nan aware assert in numpy 1.4.
>>>>>
>>>>> Oops...dinner time!
>>>>
>>>> I'm just browsing the code and adding some notes.
>>>>
>>>> Yes, matching the numpy signature for assert_equal and
>>>> assert_almost_equal is a good idea.
>>>> If you require numpy 1.4 for the test suite, then most of the boiler
>>>> plate is gone (nan handling)
>>>>
>>>> I think that assert_larry_equal is equivalent to
>>>> assert_equal(la1.x, la2.x)
>>>> assert_equal(la1.label, la2.label)
>>>>
>>>> slicing_test.py/test_slicing and test_morph use directly an assert_
>>>> which could be replaced by np.testing.assert_equal
>>>
>>> just another comment:
>>>
>>> there are 3 patterns in the test suite corresponding to the previous comments
>>>
>>> * python unittest with boilerplate
>>> * numpy 1.3 nosetests with explicit nan handling
>>> * numpy 1.4 nosetests where numpy.testing assert do the nan handling
>>>
>>> In the 4th option an explicit function assert_larry_xxx is not really
>>> necessary, and the test for x and labels and nocopy/noreference could
>>> also be "yielded" directly from the test function/method.
>>
>> Instead of making many unit tests out of one call to larry's
>> assert_equal, which would occur if we used yield, I think it is better
>> for the whole thing be one unit test. That would mean that we'd have
>> to wrap calls to np.testing.assert_equal in try...except blocks,
>> collect any error messages, and raise an AssertionError at the end of
>> the function if needed.
>
> several asserts don't have to be yielded, if you want them to be only
> one unittest that fails at the first assertion error, e.g.
> def test_movingsum32(self)  in deflarry_nose_test.py

I think it is better for debugging if all the info is printed out when
the test fails. For example, if a test fails on the label, Id like to
know if it passed on the array.

>>
>> As for signature, how about
>>
>> assert_equal(actual, desired, msg='', dtype=True, noreference=True,
>> nocopy=False, verbose=True)
>
> this doesn't work, since noreference and nocopy also need the original larry,
> the signature of check function is

Good point.

An alternative to passing in the original and the actual is to pass in
the original and the function that modifies the original to produce
the actual. Then two larrys are always passed in. But that sounds
messy.

> check_function(self, t, label, p, orig, view=False)
>
> the signature could be
> assert_larry_equal(actual, desired, msg='', dtype=True, original=None,
> noreference=True,
> nocopy=False, verbose=True)

Should we go with the signature above?

> but for noreference=True check an original has to be included
>
> with
> assert_larry_equal(actual, desired, original, msg='', dtype=True,
> noreference=True,
> nocopy=False, verbose=True)
>
> the original would have to be passed in even if both noreference and
> nocopy are False
>
>
>> So by default the dtype would be compared. Sometimes you expect the
>> dtype to change so maybe an option would be to pass in the dtype for
>> the "desired" larry.
>>
>> I think that would cover the most common use cases.
>
> Yes, I think so for the comparison of two larrys
>
> slicing tests e.g. test_slicing, would need a new function for
> noreference, nocopy that verifies e.g. that a slice is really a view.
> (I don't know what the numpy tests for view versus copy are for fancy
> slicing/indexing)
>
> Josef
>
>>
>> _______________________________________________
>> Mailing list: https://launchpad.net/~larry-discuss
>> Post to     : larry-discuss@xxxxxxxxxxxxxxxxxxx
>> Unsubscribe : https://launchpad.net/~larry-discuss
>> More help   : https://help.launchpad.net/ListHelp
>>
>
> _______________________________________________
> Mailing list: https://launchpad.net/~larry-discuss
> Post to     : larry-discuss@xxxxxxxxxxxxxxxxxxx
> Unsubscribe : https://launchpad.net/~larry-discuss
> More help   : https://help.launchpad.net/ListHelp
>