BUG: cast to str_ should not convert to pure-python intermediate #9978
Conversation
ahaldane
force-pushed
the
fix_cast_to_str
branch
from
46838e9 to
703cd0b
Compare
|
Test fails due to some platforms not having |
ahaldane
force-pushed
the
fix_cast_to_str
branch
from
703cd0b to
4a1a2b4
Compare
|
Oh right.. then I will just remove the f16 test. |
| # str(np.float64(...).item()) as this would incorrectly truncate. | ||
| a = np.zeros(1, dtype='S20') | ||
| a[:] = np.array(['1.12345678901234567890'], dtype='f8') | ||
| assert_equal(a[0], b"1.1234567890123457") |
There was a problem hiding this comment.
Can you add another test for S10 too?
It would be good to record what happens when the float str is too long. I'm not sure whether letting the string truncate is OK, or if it should round.
There was a problem hiding this comment.
Also might be tidier to use a 0d array here and .astype, but up to you
There was a problem hiding this comment.
For all types (not just floats) there is a fairly well established precedent that strings get simply truncated if the dtype is too short, so I don't think we should round. Integers don't round.
Also, I'm a bit hesitant to add a test for S10 because I didn't touch any string truncation code here, and getting the truncation right seems like a separate issue, to be examined and discussed on its own. This PR's purpose is to test string casting intermediates, not to test the arguably unrelated string truncation.
|
Thanks Allan. |
Small fix for casting to string arrays. See See pandas-dev/pandas#18123.
When casting to a
string(S) array numpy currently doesdst[i] = str(src[i].item())where recall that.item()converts to a pure-python type. I think it should just dodst[i] = str(src[i]).This comes up for floating types now that we updated their
str, since nowstr(np.floatXX(...))is not the same asstr(np.floatXX(...).item(), especially in python2.