NotificationsYou must be signed in to change notification settings
Fork11.9k
Star31k

BUG: Fix norm type promotion#10667

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to ourterms of service andprivacy statement. We’ll occasionally send you account related emails.

Already on GitHub?Sign in to your account

Jump to bottom

Closed

kmaehashi wants to merge1 commit intonumpy:masterfromkmaehashi:fix-norm-type-promotion

Closed

BUG: Fix norm type promotion#10667

kmaehashi wants to merge1 commit intonumpy:masterfromkmaehashi:fix-norm-type-promotion

+2 −2

Conversation

Copy link

Member

kmaehashi commentedFeb 26, 2018

I testednp.linalg.norm with the following code, but it seems the type is not preserved correctly.

importnumpyasnpfordtypein [np.float16,np.float32,np.float64]:a=np.ones((2,),dtype=dtype)ret=np.linalg.norm(a,3,None,False)ifnot (dtype==ret.dtype):print('expected',dtype,'but got',ret.dtype)

Without this fix (NumPy 1.14.1):

expected <class 'numpy.float16'> but got float64expected <class 'numpy.float32'> but got float64

With NumPy 1.14.1 + this fix, nothing should be printed.

Related to#10368 andcupy/cupy#875 (comment)

fix norm type promotion

ea4bb4e

kmaehashi mentioned this pull request

Feb 26, 2018

Fix to preserve dtype of input array in cupy.linalg.normcupy/cupy#875

Merged

Copy link

Member

eric-wieser commentedFeb 26, 2018•
edited
Loading

~~This doesn't affect the result dtype at all, AFAICT~~

**= operates in-place, so should not result in a type promotion (edit: except on scalars, where it does not)

Copy link

Member

eric-wieser commentedFeb 26, 2018•
edited
Loading

Edit: I'm wrong.

Thanks to#10374,np.reciprocal is unusable in generic code. You can usenp.true_divide(1, ord, dtype=...) here instead

I think that there's a deeper problem here with numpy type promotion, and howscalar op= val gives a different promotion toarr op= val

Copy link

MemberAuthor

kmaehashi commentedFeb 26, 2018•
edited
Loading

Ah, I see.

x=np.ndarray(1,dtype=np.float32)x**=1.0/2.0x.dtype# => float32x=np.float32(1)x**=1.0/2.0x.dtype# => float64

Thanks to#10374, np.reciprocal is unusable in generic code.

I think the code in this PR is generic (i.e., works with both scalar and array).
Do you mean such use ofnp.reciprocal is discouraged (even withdtype=ret.dtype)?

importnumpyasnpprint('scalar',np.linalg.norm(np.ones((2,),dtype=np.float32),3,0,False).dtype)print('array',np.linalg.norm(np.ones((2,2),dtype=np.float32),3,0,False).dtype)