Adding isin function for multidimensional arrays #8423

brsr · 2016-12-28T00:39:09Z

This resolves #8331. I had initially started the request there, and there was discussion on the mailing list, in which everyone seemed pretty happy with the idea. I hadn't realized that there was a separate in1d method for masked arrays, so this includes an isin function for masked arrays too.

endolith · 2016-12-28T05:38:34Z

I was going to submit this, too, but just using the existing function:

def in_nd(ar1, ar2, assume_unique=False, invert=False):
    """
        ...
    """
    out = np.in1d(ar1, ar2, assume_unique=False, invert=False)
    return out.reshape(ar1.shape)

endolith · 2016-12-28T05:43:46Z

numpy/lib/arraysetops.py


-    Returns a boolean array the same length as `ar1` that is True
+    Returns a boolean array the same shape as `ar1` that is True


"Returns a boolean array of the same shape"

endolith · 2016-12-28T05:45:49Z

numpy/lib/arraysetops.py

-    >>> test = np.array([0, 1, 2, 5, 0])
-    >>> states = [0, 2]
-    >>> mask = np.in1d(test, states)
+    >>> test = np.array([0, 2, 4, 6]).reshape([2, 2])


Why not

test = np.array([[0, 2], [4, 6]])

?

eric-wieser · 2016-12-29T14:51:58Z

Could probably do with better parameter names than in1 and in2. Marginally better perhaps is query and values?

endolith · 2016-12-29T15:27:18Z

Agreed, but not sure what the best names are. Conceptually, in2 is a set, not an (ordered) array

eric-wieser · 2016-12-29T18:44:15Z

item, item_set?

shoyer

It occurs to me that another possible name for this function is contains, copying the name from operator.contains. That has the arguments backwards, however, which would entail flipping the arguments here, too.

Maybe isin(query, index) for the argument names?

shoyer · 2016-12-30T00:26:38Z

numpy/lib/arraysetops.py

-        Default is False. ``np.in1d(a, b, invert=True)`` is equivalent
-        to (but is faster than) ``np.invert(in1d(a, b))``.
+        Default is False. ``np.isin(a, b, invert=True)`` is equivalent
+        to (but is faster than) ``np.invert(isin(a, b))``.

        .. versionadded:: 1.8.0


This versionadded is no longer relevant.

shoyer · 2016-12-30T00:27:36Z

numpy/lib/arraysetops.py

-    python keyword `in`, for 1-D sequences. ``in1d(a, b)`` is roughly
-    equivalent to ``np.array([item in b for item in a])``.
+    `isin` can be considered as an element-wise function version of the
+    python keyword `in`. ``in1d(ar1, ar2)`` is roughly equivalent to


should be isin, not in1d

shoyer · 2016-12-30T00:29:27Z

numpy/lib/arraysetops.py

-    equivalent to ``np.array([item in b for item in a])``.
+    `isin` can be considered as an element-wise function version of the
+    python keyword `in`. ``in1d(ar1, ar2)`` is roughly equivalent to
+    ``in1d = np.vectorize(lambda x: x in ar2); in1d(ar1)``.


I would keep the original example, which I think is a little clearer than using np.vectorize.

The original example is not valid for multidimensional arrays.

Sure, but you can still use it to describe how this function works on 1D arrays.

shoyer · 2016-12-30T00:31:40Z

numpy/lib/arraysetops.py

@@ -380,16 +381,16 @@ def setxor1d(ar1, ar2, assume_unique=False):
    flag2 = flag[1:] == flag[:-1]
    return aux[flag2]

-def in1d(ar1, ar2, assume_unique=False, invert=False):
+def isin(ar1, ar2, assume_unique=False, invert=False):


Definitely consider more meaningful names here.

brsr · 2016-12-31T00:31:51Z

Went with elements and test_elements to replace ar1 and ar2. It fits with the description in the docstring: "Test whether each element of an array is present in a second array." I felt that calling ar2 a "set" would be confusing, because the function has counterintuitive behavior if ar2 is a standard Python set.

endolith · 2016-12-31T05:26:31Z

It occurs to me that another possible name for this function is contains, copying the name from operator.contains. That has the arguments backwards, however, which would entail flipping the arguments here, too.

To me, that would mean one array is a sub-array of the other, like contains([[1,2,3],[4,5,6],[7,8,9]], [[1,2],[4,5]]) would be True.

https://docs.python.org/2/reference/expressions.html#membership-test-operations calls the arguments x and s, and calls s a "collection". The first argument could be some synonym of testee or subject, but nothing seems obviously good to me.

eric-wieser · 2016-12-31T08:55:10Z

I'd argue the name should still be contains, because your intuition is inconsistent with operator.contains: operator.contains([1,2,3], [2,3]) is false

brsr · 2017-01-01T18:50:00Z

operator.contains and numpy.contains would have inconsistent behavior. Furthermore, ndarray implements __contains__, so numpy.contains and numpy.ndarray.__contains__ would have inconsistent behavior.

Compare:

>>> operator.contains([1, 2, 3], 2)
True
>>> operator.contains([1, 2, 3], [2, 3])
False
>>> operator.contains([[1, 2], [3, 4], [5, 6]], [3, 6])
False

versus

>>> np.in1d(2, [1, 2, 3])
array([ True], dtype=bool)
>>> np.in1d([2, 3], [1, 2, 3])
array([ True,  True], dtype=bool)
>>> np.in1d([3, 6], [[1, 2], [3, 4], [5, 6]])
array([ True,  True], dtype=bool)

eric-wieser · 2017-01-01T22:46:07Z

I'm assuming that you mean isin not in1d there.

Your second case is not really any more inconsistent than:

>>> 1 == [1, 2]
False
>>> 1 == np.array([1, 2])
array([ True,  False], dtype=bool)

I think that whatever we name it, differences in behaviour between np.isin and in will cause surprises, and that's not really any different to operator.contains vs np.contains causing those same surprises.

shoyer · 2017-01-01T22:59:13Z

I guess there are enough differences here that isin could make more sense. I like that the name is shorter and is consistent with pandas.

seberg · 2017-01-02T11:01:13Z

__contains__ has bad behaviour in any case :/, it would be nice to deprecate the more insane stuff there.

eric-wieser · 2017-01-02T13:25:02Z

@seberg: Is the behaviour documented anywhere? The first three google results simply tell me x.__contains__(val) means val in x without a link to what that means, and the third is a StackOverflow post saying "Don't try to use __contains__ in NumPy. It does crazy, useless shit"

seberg · 2017-01-02T13:53:24Z

Reading the code of it is the easiest. Basically, it works as long as the lhs is a single element. If it is not a single element, things likely won't make much sense.

shoyer · 2017-02-10T19:11:54Z

numpy/lib/arraysetops.py

+    Returns a boolean array of the same length as `elements` that is True
+    where an element of `elements` is in `test_elements` and False otherwise.
+
+    See `numpy.isin` for more details.


I'm a little nervous about deleting all the parameters on this method, including the notes on when they were added. Can we keep them around, even though it's slightly redundant?

shoyer

This is very close -- would love your help to get this across the finish line!

This should also be mentioned in the docs (doc/source/reference/routines.set.rst) and the release notes (doc/release/1.13.0-notes.rst).

shoyer · 2017-02-14T17:20:05Z

numpy/lib/arraysetops.py

@@ -380,106 +381,130 @@ def setxor1d(ar1, ar2, assume_unique=False):
    flag2 = flag[1:] == flag[:-1]
    return aux[flag2]

-def in1d(ar1, ar2, assume_unique=False, invert=False):
+def isin(elements, test_elements, assume_unique=False, invert=False):


It's not worth the hassle of changing argument names on an effectively deprecated function and potentially breaking some users. Let's only use the new arguments names for isin.

shoyer · 2017-02-14T17:21:55Z

numpy/lib/arraysetops.py

-    ar2 : array_like
-        The values against which to test each value of `ar1`.
+    test_elements : array_like
+        The values against which to test each value of `elements`.


Note that this argument is flattened.

shoyer · 2017-02-14T17:22:19Z

numpy/lib/arraysetops.py

-    in1d : (M,) ndarray, bool
-        The values `ar1[in1d]` are in `ar2`.
+    isin : ndarray, bool
+        The values `elements[isin]` are in `test_elements`.


Note that this has the same shape as elements.

shoyer · 2017-02-14T17:25:57Z

numpy/lib/arraysetops.py

-    # Ravel both arrays, behavior for the first array could be different
-    ar1 = np.asarray(ar1).ravel()
-    ar2 = np.asarray(ar2).ravel()
+    # Ravel both arrays, will reshape result later to match elements


Even though isin should replace most uses of in1d, I think this code is actually easier to understand when a separate helper function is used for the 1D functionality, in the style of unique/_unique1d in this file. So I would keep the core logic in in1d.

brsr · 2017-02-15T14:36:59Z

So it sounds like we want to leave in1d as it is in 1.12, and the new function isin should just reshape the result of in1d?

shoyer · 2017-02-15T15:24:20Z

@brsr You're certainly welcome to do any internal clean-up that is warranted, but from a public API perspective I would indeed keep in1d mostly changed. I would certainly search for references to in1d in the documentation and replace most of those with isin as the preferred solution, but we want to keep in1d around in an intact state to maximize backwards compatibility.

brsr · 2017-02-16T15:44:20Z

I believe I've caught all the instances of in1d in the docs. The docs for where actually demonstrate the exact use case that I wanted this function for. Let me know what else needs tweaking.

homu · 2017-02-22T16:18:43Z

☔ The latest upstream changes (presumably #8446) made this pull request unmergeable. Please resolve the merge conflicts.

brsr · 2017-04-10T13:52:52Z

@eric-wieser, I'm not following your reasoning. Using the example from the docs, isin is equivalent to np.array([item in b for item in a]) if a and b are 1-D arrays. Expressing that in terms of the names currently used in isin would be something like np.array([element in test_elements for element in elements]). elements is plural because the function tests each singular element in elements.

eric-wieser · 2017-04-10T14:01:14Z

elements is plural because the function tests each singular element in elements

Right, but in numpy by default everything is plural, so that's not really helpful.

For instance, np.add(a, b):

When passed 0d arrays, does (a + b),
When passed 1d arrays, does [(ai + bi) for ai, bi in zip(a, b)]
When passed 2d arrays, does [[(aij + bij) for aij, bij in zip(ai, bi)] for ai, bi in zip(a, b) ]
...

It makes the most sense to document this as "calculates a + b, broadcasting over a and b"

So for np.isin(a, bs):

When passed 0d a, does (a in bs),
When passed 1d a, does [(ai in bs) for ai in a]
When passed 2d a, does [[(aij in bs) for aij in ai] for ai in a ]
...

Again, we have a base operation, in parentheses, and then a bunch of numpy broadcasting happening.

Let's be consistent here - we should always describe the function in terms of its lowest-dimensional generalization. So it makes the most sense to document this as "calculates a in bs, broadcasting over a alone"

eric-wieser · 2017-04-10T14:12:56Z

Also, we already bikeshedded the name, but...

The more I look at this PR, the harder I find it not to read np.isin as i sin, as python sets a precedent with __iadd__, and I'm used to seeing np.sin.

I know we've picked it for pandas compatibility, but as it's not a method of ndarray, this doesn't help with duck typing.

How do people feel about np.is_in?

shoyer · 2017-04-10T16:09:45Z

How do people feel about np.is_in?

I can see your point, but it looks very stilted to me, so I still prefer isin, which looks like a shortened version of the builtin name isinstance. In context, I don't think I would be confused, because the usage is so different: np.isin returns a boolean array typically used for indexing, rather than floats like np.sin.

eric-wieser · 2017-04-10T17:12:42Z

numpy/lib/arraysetops.py

@@ -516,7 +516,7 @@ def isin(elements, test_elements, assume_unique=False, invert=False):
    Returns
    -------
    isin : ndarray, bool
-        Has the same shape as `elements`. The values `elements[isin]`
+        Has the same shape as `element`. The values `elements[isin]`


Missed an s here

eric-wieser · 2017-04-10T17:14:17Z

numpy/lib/arraysetops.py

+
+    If `test_elements` is a set (or other non-sequence collection) it will
+    be converted to an object array with one element, rather than an array
+    of the values contained in `test_elements`. Converting the set to


Probably worth pointing out that this is just the np.array constructor doing its thing, and is not something weird this function chose to do. Mentioning "array constructor" somewhere should do the job

eric-wieser · 2017-04-10T17:14:55Z

numpy/lib/arraysetops.py

-    elements = np.array(elements)
-    return in1d(elements, test_elements, assume_unique=assume_unique,
-                invert=invert).reshape(elements.shape)
+    element = np.array(element)


Should be asarray, or asanyarray if you think this will worked for masked arrays

eric-wieser · 2017-04-10T17:15:35Z

numpy/ma/extras.py

-    elements = ma.array(elements)
-    return in1d(elements, test_elements, assume_unique=assume_unique,
-                invert=invert).reshape(elements.shape)
+    element = ma.array(element)


Again, asarray or asanyarray

eric-wieser · 2017-04-10T17:17:25Z

numpy/ma/tests/test_extras.py

+        ec[0, 0, 1] = True
+        ec[0, 2, 3] = True
+        c = isin(a, b)
+        assert_array_equal(c, ec)


Can we have a test here for type(c)? Should it be MaskedArray or array?

Also, a test of what happens if you call np.isin instead of np.ma.isin might be interesting as well.

eric-wieser · 2017-04-10T17:17:38Z

numpy/lib/tests/test_arraysetops.py

+        ec[1, 2, 2] = True
+        c = isin(a, b)
+        assert_array_equal(c, ec)
+


Needs a test for 0d a here

eric-wieser · 2017-04-10T17:25:21Z

you might be able to simplify your tests with something like:

isin_slow = np.vectorize(lambda a, b: a in b, excluded={'b'})

# do this for a bunch of things, now that you don't need to find the result by hand
assert_equal(isin_slow(a, b), isin(a, b))

eric-wieser · 2017-04-12T20:46:39Z

numpy/lib/tests/test_arraysetops.py

+        #both args
+        assert_isin_equal(5, 6)
+        #zero-d array-like as...
+        x = []


[] is 1d with shape (0,), not 0d with shape ().

I'm thinking testing things like np.isin(somearr, np.array(1)), or np.isin(np.array(1), somearr)

eric-wieser · 2017-04-12T20:50:16Z

numpy/ma/tests/test_extras.py

+        a = array(a, mask=mask)
+        values = [0, 10, 20, 30,  1,  3, 11, 22, 33]
+        mask   = [0,  1,  0,  1,  0,  1,  0,  1,  0]
+        b = array(values, mask=mask)


Or maybe even just

array(data=[0, 10, 20, 30, 1, 3, 11, 22, 33], mask=[0, 1, 0, 1, 0, 1, 0, 1, 0])

Better already with the alignment

eric-wieser · 2017-04-12T23:46:23Z

numpy/ma/extras.py

 def in1d(ar1, ar2, assume_unique=False, invert=False):
    """
    Test whether each element of an array is also present in a second
    array.

    The output is always a masked array. See `numpy.in1d` for more details.

+    This function has been deprecated: use `isin` instead.
+
+    Deprecated since version 1.13.0.


Should use the .. deprecated:: 1.13.0 syntax. Look around at other files to see how things are worded around it. Some files use .. note as well, I think

For what it's worth, numpy/HOWTO_DOCUMENT.rst.txt says something different about deprecation warnings.

eric-wieser · 2017-04-12T23:47:04Z

numpy/lib/arraysetops.py

 def in1d(ar1, ar2, assume_unique=False, invert=False):
    """
    Test whether each element of a 1-D array is also present in a second array.

    Returns a boolean array the same length as `ar1` that is True
    where an element of `ar1` is in `ar2` and False otherwise.

+    This function has been deprecated: use `isin` instead.
+
+    Deprecated since version 1.13.0.


eric-wieser · 2017-04-12T23:50:48Z

numpy/lib/tests/test_arraysetops.py

+        c = [(9, 8), (7, 6)]
+        d = (9, 7)
+        assert_isin_equal(c, d)
+        #zero-d scalar as:


These tests are good, but tests for 0d arrays are also important, and usually catch more bugs — so it would be good to have those too.

Related questions:

should isin(1, 1):

return np.array(True)

return np.bool_(True)

raise TypeError, because 1 is not iterable before coercion to a 0d array

should isin(1, np.array(1)):

return np.array(True)

return np.bool_(True)

etc

Also, I think the comments labelling which arg is a scalar are a bit much. This'd be more readable with just the section headings, and some space between ssections

eric-wieser · 2017-04-12T23:54:12Z

numpy/lib/arraysetops.py

@@ -411,6 +417,8 @@ def in1d(ar1, ar2, assume_unique=False, invert=False):

    See Also
    --------
+    isin                  : Version of this function that preserves the
+                            shape of ar1.
    numpy.lib.arraysetops : Module with a number of other functions for


"See also: The module you're looking at right now" is a little weird. @brsr, you don't need to worry about it in this PR though.

eric-wieser · 2017-04-13T01:08:57Z

numpy/lib/arraysetops.py

+        can speed up the calculation.  Default is False.
+    invert : bool, optional
+        If True, the values in the returned array are inverted (that is,
+        False where an element of `ar1` is in `ar2` and True otherwise).


These names are out of date. I'd probably ditch this wording all together, and explain it in terms of not in.

eric-wieser · 2017-04-13T01:10:46Z

numpy/lib/arraysetops.py

+    array([[ True, False],
+           [ False, True]], dtype=bool)
+    >>> element[mask]
+    array([0, 6])"""


nit: needs an extra newline in this docstring

eric-wieser · 2017-04-13T01:11:03Z

numpy/lib/arraysetops.py

+    it will be converted to an object array with one element, rather than an
+    array of the values contained in `test_elements`. This is a consequence of
+    the `array` constructor's way of handling non-sequence collections.
+    Converting the set to a list usually gives the desired behavior.


This might be a good thing to show an example of, for people skipping the notes

eric-wieser · 2017-04-13T01:13:10Z

numpy/lib/info.py

@@ -147,6 +147,8 @@
 setxor1d         Set exclusive-or of 1D arrays with unique elements.
 in1d             Test whether elements in a 1D array are also present in
                 another array.
+isin             Test whether elements in an array are also present in
+                 another array, preserving shape of first array.


Maybe "test whether each element of one array is present anywhere within another"?

No strong feelings here

eric-wieser · 2017-04-13T01:47:51Z

numpy/lib/info.py

@@ -147,8 +147,8 @@
 setxor1d         Set exclusive-or of 1D arrays with unique elements.
 in1d             Test whether elements in a 1D array are also present in
                 another array.
-isin             Test whether elements in an array are also present in
-                 another array, preserving shape of first array.
+isin             Test whether each element of one (possibly multidimensional) 


Why just possibly? I think you can put ND before both of these arrays, to make it explicit compared to its surrounding 1D functions

eric-wieser

Looks pretty good to me now - nice job!

remove 1D from section description, since they're not all 1D. match order in the list at top of arraysetops.py, where all the ND functions are first and all the 1D ones are after.

homu · 2017-04-28T02:29:36Z

☔ The latest upstream changes (presumably #8996) made this pull request unmergeable. Please resolve the merge conflicts.

eric-wieser · 2017-04-28T15:22:51Z

doc/release/1.13.0-notes.rst

@@ -22,6 +22,8 @@ Dropped Support
 Deprecations
 ============

+* ``in1d`` is now deprecated in favor of ``isin``.


I don't think we should mark things as deprecated without also raising a DeprecationWarning.

I think it would be best to not deprecate in1d in this PR, so that merging this otherwise nice work can happen in time for 1.13.

You can always make a new PR that adds deprecation once this is merged

This may be my fault: I may have said something like to suggest that we should deprecate in1d, but I was being sloppy with my language. I don't think we can (or should) ever raise a DeprecationWarning. in1d works just fine, albeit not with as nice an API as isin, and is used all over the place, so there is absolutely no good reason to remove it.

What we can and should do is discourage using in1d in favor of isin for new code. Instead of saying in1d is deprecated, its docstring should say something like: "We recommend using isin instead of in1d for new code".

eric-wieser · 2017-04-28T15:23:36Z

Will need a rebase at some point, but if you're happy for all the commits to be squashed into one, then I can do that for you via the github UI

brsr · 2017-04-28T18:20:02Z

@eric-wieser , squashing is fine with me.

eric-wieser · 2017-05-05T00:04:13Z

Thanks @brsr!

This fixes numpygh-8331 Also update the docs for arraysetops to remove the outdated "1D" from the description, which was already incorrect for np.unique.

Adding isin function for multidimensional arrays

17faf5a

endolith approved these changes Dec 28, 2016

View reviewed changes

charris added 01 - Enhancement component: numpy.lib labels Dec 29, 2016

shoyer reviewed Dec 30, 2016

View reviewed changes

Fix comments on pull request

34a40da

shoyer reviewed Feb 10, 2017

View reviewed changes

shoyer reviewed Feb 14, 2017

View reviewed changes

shoyer mentioned this pull request Feb 14, 2017

Add isin method to Dataset and DataArray pydata/xarray#1268

Closed

brsr added 3 commits February 15, 2017 12:50

keep in1d mostly the same, changes in isin now

2b4a81b

screwed up the whitespace

db5e5fd

docs, convert elements to array in case it isn't already

f63cf31

Merge branch 'master' into master

a9bce34

Renaming elements to element

552a193

eric-wieser reviewed Apr 10, 2017

View reviewed changes

eric-wieser's comments

fa0b0be

eric-wieser reviewed Apr 12, 2017

View reviewed changes

Fixes to tests

0ff6be4

eric-wieser reviewed Apr 12, 2017

View reviewed changes

brsr added 3 commits April 12, 2017 20:48

Docstrings, further expanding test

545df63

spacing

41e5b0b

Actual zero-d array

521d517

eric-wieser reviewed Apr 13, 2017

View reviewed changes

More docstring changes

8805bbb

eric-wieser reviewed Apr 13, 2017

View reviewed changes

eric-wieser approved these changes Apr 13, 2017

View reviewed changes

brsr added 2 commits April 13, 2017 10:28

clean up function listing

4d3f67c

remove 1D from section description, since they're not all 1D. match order in the list at top of arraysetops.py, where all the ND functions are first and all the 1D ones are after.

Update 1.13.0-notes.rst

0395f39

Merge branch 'master' into master

3d809a6

eric-wieser reviewed Apr 28, 2017

View reviewed changes

discouraged, not deprecated

d22cafc

eric-wieser merged commit 69b0c42 into numpy:master May 5, 2017


		Returns a boolean array the same length as `ar1` that is True
		Returns a boolean array the same shape as `ar1` that is True

Adding isin function for multidimensional arrays #8423

Adding isin function for multidimensional arrays #8423

Conversation

brsr commented Dec 28, 2016

endolith commented Dec 28, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eric-wieser commented Dec 29, 2016

endolith commented Dec 29, 2016

eric-wieser commented Dec 29, 2016

shoyer left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

brsr commented Dec 31, 2016

endolith commented Dec 31, 2016

eric-wieser commented Dec 31, 2016

brsr commented Jan 1, 2017 • edited Loading

eric-wieser commented Jan 1, 2017 • edited Loading

shoyer commented Jan 1, 2017

seberg commented Jan 2, 2017

eric-wieser commented Jan 2, 2017 • edited Loading

seberg commented Jan 2, 2017

Choose a reason for hiding this comment

shoyer left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

brsr commented Feb 15, 2017 • edited Loading

shoyer commented Feb 15, 2017

brsr commented Feb 16, 2017

homu commented Feb 22, 2017

brsr commented Apr 10, 2017

eric-wieser commented Apr 10, 2017 • edited Loading

eric-wieser commented Apr 10, 2017 • edited Loading

shoyer commented Apr 10, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eric-wieser Apr 10, 2017 • edited Loading

Choose a reason for hiding this comment

eric-wieser Apr 10, 2017 • edited Loading

Choose a reason for hiding this comment

eric-wieser commented Apr 10, 2017 • edited Loading

Choose a reason for hiding this comment

eric-wieser Apr 12, 2017 • edited Loading

Choose a reason for hiding this comment

eric-wieser Apr 12, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eric-wieser Apr 12, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eric-wieser Apr 13, 2017 • edited Loading

Choose a reason for hiding this comment

eric-wieser Apr 13, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eric-wieser left a comment

Choose a reason for hiding this comment

homu commented Apr 28, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eric-wieser commented Apr 28, 2017

brsr commented Apr 28, 2017

eric-wieser commented May 5, 2017

brsr commented Jan 1, 2017 •

edited

Loading

eric-wieser commented Jan 1, 2017 •

edited

Loading

eric-wieser commented Jan 2, 2017 •

edited

Loading

brsr commented Feb 15, 2017 •

edited

Loading

eric-wieser commented Apr 10, 2017 •

edited

Loading

eric-wieser commented Apr 10, 2017 •

edited

Loading

eric-wieser Apr 10, 2017 •

edited

Loading

eric-wieser Apr 10, 2017 •

edited

Loading

eric-wieser commented Apr 10, 2017 •

edited

Loading

eric-wieser Apr 12, 2017 •

edited

Loading

eric-wieser Apr 12, 2017 •

edited

Loading

eric-wieser Apr 12, 2017 •

edited

Loading

eric-wieser Apr 13, 2017 •

edited

Loading

eric-wieser Apr 13, 2017 •

edited

Loading