cc @behzadnouri

Hmm, although this is maybe more an issue with my numpy installation?
As I don't really see a reason why np.bincount would fail when getting int64 values (the docs says it needs integers and not floats, but does not say anything about integer size)

Possibly an issue with bincount itself: numpy/numpy#4366

since ids are compressed, they cannot exceed memory address space. so i think it should be safe to cast on pandas side, i.e.: ids[ids != -1].astype(np.intp).

with copy=False arg: ids[ids != -1].astype(np.intp, copy=False)

I setup a vm for linux-32

these break for the same reasons as well.
https://github.com/pydata/pandas/blob/master/pandas/core/groupby.py#L2580
https://github.com/pydata/pandas/blob/master/pandas/core/groupby.py#L1815

both of them u added int64 not me. xref #10988

def group_info(self): indeed requires to return int64_t values but not def nunique. Either case, the right way to do it would be to leave internal computations alone, but only change the output as in:

diff --git a/pandas/core/groupby.py b/pandas/core/groupby.py
index e72f7c6..06b1105 100644
--- a/pandas/core/groupby.py
+++ b/pandas/core/groupby.py
@@ -1808,15 +1808,17 @@ class BinGrouper(BaseGrouper):
     @cache_readonly
     def group_info(self):
         ngroups = self.ngroups
-        obs_group_ids = np.arange(ngroups, dtype='int64')
+        obs_group_ids = np.arange(ngroups)
         rep = np.diff(np.r_[0, self.bins])

         if ngroups == len(self.bins):
-            comp_ids = np.repeat(np.arange(ngroups, dtype='int64'), rep)
+            comp_ids = np.repeat(np.arange(ngroups), rep)
         else:
-            comp_ids = np.repeat(np.r_[-1, np.arange(ngroups, dtype='int64')], rep)
+            comp_ids = np.repeat(np.r_[-1, np.arange(ngroups)], rep)

-        return comp_ids, obs_group_ids, ngroups
+        return comp_ids.astype('int64', copy=False),  \
+               obs_group_ids.astype('int64', copy=False),  \
+               ngroups

     @cache_readonly
     def ngroups(self):
@@ -2565,8 +2567,8 @@ class SeriesGroupBy(GroupBy):

         # group boundries are where group ids change
         # unique observations are where sorted values change
-        idx = com._ensure_int64(np.r_[0, 1 + np.nonzero(ids[1:] != ids[:-1])[0]])
-        inc = com._ensure_int64(np.r_[1, val[1:] != val[:-1]])
+        idx = np.r_[0, 1 + np.nonzero(ids[1:] != ids[:-1])[0]]
+        inc = np.r_[1, val[1:] != val[:-1]]

         # 1st item of each group is a new unique observation
         mask = isnull(val)
@@ -2577,7 +2579,7 @@ class SeriesGroupBy(GroupBy):
             inc[mask & np.r_[False, mask[:-1]]] = 0
             inc[idx] = 1

-        out = np.add.reduceat(inc, idx)
+        out = np.add.reduceat(inc, idx).astype('int64', copy=False)
         return Series(out if ids[0] != -1 else out[1:],
                       index=self.grouper.result_index,
                       name=self.name)

I changed them because they didn't work on window. I will try your diff and let u know.

Hello,

I have issue with minlength which seems to be fix with this solved issue... Are we going to have pandas 0.17.1 version tag soon? It really unreliable without it to pass to 0.17.x which seems to bring many benefit despite this ".size() minlength" issue...

And if not is using master consider "safe" (read as stable as an official tagged commit)?

Thanks

Sorry, I think I am wrong... This bug fix seems to be include in 0.17.0

But I have this error : <type 'exceptions.ValueError'> minlength must be positive

File "/usr/local/lib/python2.7/dist-packages/pandas/core/groupby.py", line 820, in size
return self.grouper.size()
File "/usr/local/lib/python2.7/dist-packages/pandas/core/groupby.py", line 1383, in size
out = np.bincount(ids[ids != -1], minlength=ngroup)
ValueError: minlength must be positive

Which seems related...

I just do :

grouped.size()

Which were working just fine prior to upgrade from 0.15.0 to 0.17.0

Could it be related?

Downgrading to 0.16.2 solve my issue...

:(

you would have to show a copy-pastable example along with pd.show_versions()

I will try to reproduce with dummy data, because I can't include the actual data... Give me a couples of days.

Thanks