BUG: Fixes GH9311 groupby on datetime64 #9345
Merged
+1,391
−743
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| @@ -1504,7 +1504,11 @@ def aggregate(self, values, how, axis=0): | |||
| raise NotImplementedError | |||
| out_shape = (self.ngroups,) + values.shape[1:] | |||
|
|
|||
| if is_numeric_dtype(values.dtype): | |||
| if is_datetime: | |||
There was a problem hiding this comment.
What does this do to operations like groupby('foo').mean() for datetime columns? Currently these raise DataError: No numeric types to aggregate but perhaps they shouldn't...
There was a problem hiding this comment.
mean of datetime64 is a can of worms. It can be done, but completely separate from this.
There was a problem hiding this comment.
Agreed... wanted to make sure to understand if that was changing that behavior for groupby aggregations.
There was a problem hiding this comment.
yep, I think that should be cause before here though (iow there is a test for it now that asserts that it fails)
|
This looks like some nice fixes but also looks likely under-tested -- you've changed a lot of functions and only added tests for one of them. |
| @@ -1487,7 +1487,7 @@ def wrapper(*args, **kwargs): | |||
| (how, dtype_str)) | |||
| return func, dtype_str | |||
|
|
|||
| def aggregate(self, values, how, axis=0): | |||
| def aggregate(self, values, how, axis=0, is_datetime=False): | |||
|
|
|||
There was a problem hiding this comment.
you can't put a flag like this here, its very odd. There are already introspection determinations in the grouper objects.
jreback
added
Bug
Dtype Conversions
|
I would probably scale this back to only change operations without any possibility of precision/overflow issues -- I think that's Also, it is almost impossible to write too many tests :). |
|
Awesome. Thanks for the feedback. I'll have updates, and more tests in here shortly. |
|
@shoyer or @jreback - I'm very close to having the above fixes ready on this PR. I do need a bit of guidance on one thing though. Now that the Cython functions are using iNaT for nan values on the integer functions, they are returning the integer value (-9223372036854775808) back in the ndarray, as I would expect. In order to get it working so that na functions will work (like fillna), I need to replace those values back with something that will be evaluated as na. Is there a helper function somewhere to do that? |
|
@iwschris iNaT -> NaT happens automatically when you cast back to datetime64. You can use |
|
Awesome. Thanks! |
chrisbyboston
force-pushed
the
groupby_nano_int
branch
from
aff1443
to
2219ef7
Compare
|
@shoyer -- Pretty sure this is ready for review again. I was able to keep the safe casting turned on in internals.py and I validated that vbench is still looking pretty good. Here's the output from vbench: |
| @@ -4,6 +4,11 @@ | |||
| # or we get a bootstrapping problem | |||
| from StringIO import StringIO | |||
|
|
|||
| MAX_INT8 = 127 | |||
There was a problem hiding this comment.
use the dtype properties instead of hard coding these, e.g., https://github.com/pydata/pandas/blob/484f66814a32324248e3d03203ea991ed896fd7a/pandas/core/common.py#L52
|
OK, this looks much more reasonable to me. I'm slightly troubled by the |
| @@ -85,6 +87,8 @@ | |||
| _dataframe_apply_whitelist = \ | |||
| _common_apply_whitelist | frozenset(['dtypes', 'corrwith']) | |||
|
|
|||
| _non_arithmetic_agg = ('first', 'last', 'min', 'max', 'nth', 'count') | |||
There was a problem hiding this comment.
This should probably be a frozenset
|
@shoyer Thanks for looking at it so quickly. I'll make the changes suggested, and I'll use a profiler to figure out why |
|
Alright, I found it. Here's the relevant function: def _try_coerce_result(self, result):
""" reverse of try_coerce_args """
if isinstance(result, np.ndarray):
if result.dtype == 'i8':
result = tslib.array_to_datetime(
result.astype(object).ravel()).reshape(result.shape)
elif result.dtype.kind in ['i', 'f', 'O']:
result = result.astype('M8[ns]', casting='safe')
elif isinstance(result, (np.integer, np.datetime64)):
result = lib.Timestamp(result)
return resultWe aren't even hitting the What's more, this block... if result.dtype == 'i8':
result = tslib.array_to_datetime(
result.astype(object).ravel()).reshape(result.shape)...I don't believe is necessary any more, as I think it was covering for this error in numpy 1.6. I'm going to clean this function up and make sure all the tests are passing and that vbench looks better, and I'll push up my changes. |
|
@iwschris Interesting -- sounds like a good plan to me! |
|
@iwschris The point of that coercion was to handle the cases where a returned input was non-i8. E.g. an op was applied to a datetime64 (say mean in a groubpy) that returned a float / object. So it prob wasn't hit very much (maybe not tested at all). |
chrisbyboston
force-pushed
the
groupby_nano_int
branch
from
2219ef7
to
0acdaab
Compare
|
@shoyer I think I've made all the changes that you requested, and the groupby performance is fixed. Let me know if you see anything else that needs to be changed. |
chrisbyboston
force-pushed
the
groupby_nano_int
branch
from
d33dae7
to
01e6526
Compare
|
Cool. I think that last push takes care of everything mentioned up to this point. Anything else? |
|
just FYI The reason we always want to take a view on a M8/m8 object is that it does NOT copy. so these DO need to be treated separately (from a regular int64) |
chrisbyboston
force-pushed
the
groupby_nano_int
branch
from
01e6526
to
9fff246
Compare
|
I see. Good info. Newest stuff now pushed. |
|
@jreback interesting -- didn't realize that about view vs astype for datetime64 |
|
@iwschris can you just run another vbench vs master (you can limit it with |
|
You bet. Should have it in a few minutes. |
|
@jreback looks like objects weren't accounted for in that group of |
|
@iwschris right, yeh that should be the else. |
chrisbyboston
force-pushed
the
groupby_nano_int
branch
from
9fff246
to
dd89607
Compare
datetime64 columns were changing at the nano-second scale when applying a groupby aggregator.
chrisbyboston
force-pushed
the
groupby_nano_int
branch
from
dd89607
to
5f6cbf8
Compare
|
|
@jreback - vbench output posted, and the modification for Let me know if there is anything else. |
|
vbench looks pretty good to me -- some nice speedups for grouped aggregations! |
|
the vbenchs are due to other PR's actually. Here is what I get @iwschris I think you were benching against an older version. in any event it looks fine. ping when green |
|
Good to know for future PR's. It just takes a fetch and a rebase, right? |
|
yep, then I run with
if you originally checked out with (or you can set the tracking branch, e.g. |
|
Yep, our vbench's match now. |
|
@jreback it's green. |
jreback
added a commit
that referenced
this pull request
Feb 14, 2015
BUG: Fixes GH9311 groupby on datetime64
|
Thanks for working with me on it! I'll take a peek at #4095. |
|
Indeed, really nicely done |
jreback
added a commit
that referenced
this pull request
Sep 3, 2015
nickeubank
pushed a commit
to nickeubank/pandas
that referenced
this pull request
Sep 29, 2015
…already fixed in master)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
datetime64 columns were changing at the nano-second scale when
applying a groupby aggregator.
closes #9311
closes #6620