[WIP] MAINT: get rid of np.int #5032
Conversation
np.int is a confusing synonym for __builtin__.int; it maps to C long, the size of which is platform-dependent. The NumPy devs are speaking of deprecating np.int, and all uses of it in SciPy have been replaced. Notably, this makes sample_without_replacement 64-bit clean; previously it used np.int_t, which is a C long (32 bits wide on Windows). See also http://github.com/numpy/numpy#6103.
|
Is |
|
Are you referring to a specific part of the PR, or in general? |
|
In general. Or |
|
For indices, |
|
I think we removed all appearances of float in favour of float64. |
|
I am also wondering why not |
$ git grep 'np\.float[^0-9]' -- '*.py' | wc -l
111 |
|
meh. Ok. maybe not. but that is just a single command to fix ;) |
|
the question is whether we want to use numpy types or python types, right? |
|
I'm in favor of using |
|
It's a bit hard to know whether something will be used as index, right? I guess we don't generally pass integers around, though, apart from the cross-validation module, where they will be used as indices. |
| samples_range = np.linspace(2, 1000, 4).astype(np.int) | ||
| features_range = np.linspace(2, 1000, 4).astype(np.int) | ||
| samples_range = np.linspace(2, 1000, 4).astype(int) | ||
| features_range = np.linspace(2, 1000, 4).astype(int) |
There was a problem hiding this comment.
These should all be linspace(..., dtype=int) to avoid a copy
|
too many conflicts. If we want to do this we should start from scratch. |
np.intis a confusing synonym for__builtin__.int; it maps to Clong, the size of which is platform-dependent and not necessarily the same as that of Cint. The NumPy devs are speaking of deprecatingnp.int, and all uses of it in SciPy have just been removed (scipy/scipy@10ca036).Don't merge yet: these are not all occurrences, and
np.floatneeds the same treatment.See also numpy/numpy#6103.