Thanks to visit codestin.com
Credit goes to github.com

Skip to content

einsum broadcast regression (with optimize=True) #10343

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.

Sign up for GitHub

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

Already on GitHub? Sign in to your account

Jump to bottom
Closed
d70-t opened this issue Jan 8, 2018 · 8 comments · Fixed by #10352
Closed

einsum broadcast regression (with optimize=True) #10343

d70-t opened this issue Jan 8, 2018 · 8 comments · Fixed by #10352
Labels
06 - Regression component: numpy._core

Comments

@d70-t
Copy link

d70-t commented Jan 8, 2018

In numpy 1.13.3, it was possible to execute the following snippet without errors, while in 1.14.0 this happens:

In [1]: import numpy as np
In [2]: a = np.ones((10,2))
In [3]: b = np.ones((1,2))
In [4]: np.einsum('t...i,ti->t...', a, b)
Traceback (most recent call last):
  File "<ipython-input-4-fa62d1d882f9>", line 1, in <module>
    np.einsum('t...i,ti->t...', a, b)
  File "/usr/local/lib/python2.7/dist-packages/numpy/core/einsumfunc.py", line 1087, in einsum
    einsum_call=True)
  File "/usr/local/lib/python2.7/dist-packages/numpy/core/einsumfunc.py", line 710, in einsum_path
    "not match previous terms.", char, tnum)
ValueError: ("Size of label '%s' for operand %d does not match previous terms.", 't', 1)

However optimize=False solves the problem:

In [5]: np.einsum('t...i,ti->t...', a, b, optimize=False)
Out[5]: array([2., 2., 2., 2., 2., 2., 2., 2., 2., 2.])

Is this intended behavior and the user is now responsible to explicitly disable optimization or is this a bug?
@seberg seberg added this to the 1.14.1 release milestone Jan 8, 2018
@charris
Copy link
Member

charris commented Jan 8, 2018

Both? optimize=True was made the default, but it shouldn't be causing errors.

@dgasmith Another one ;)

@eric-wieser
Copy link
Member

Seems we failed to format the error message correctly either

@mhvk mhvk mentioned this issue Jan 8, 2018
Closed
@mhvk
Copy link
Contributor

mhvk commented Jan 8, 2018

This hit at least one user of astropy (though not our tests, annoyingly!). The following breaks (but works with optimize=False):

np.einsum('...ij,j...->i...', np.eye(3)[np.newaxis], np.array([[1., 1.], [2., 2.], [3., 3.]]))

This was referenced Jan 9, 2018
Merged
Merged
@mhvk mhvk mentioned this issue Jan 10, 2018
Closed
@mhvk
Copy link
Contributor

mhvk commented Jan 10, 2018

The fix in #10352 seems correct. Note also that the current optimization choice makes no sense: for 2 argument, optimization should not be attempted; see #10357.

@mhvk
Copy link
Contributor

mhvk commented Jan 10, 2018

A very simple fix for the slowdown is in #10359 - I think we should consider reverting the change in default of optimize, though. In the end, einsum is very likely used mostly with small numbers of arrays, for which the additional cost is not worth it. It seems more likely that users with long expressions will read the docstring and realize there is an optimize keyword than that users with short ones will realize the default is bad for their case.

joostvanzwieten added a commit to joostvanzwieten/nutils that referenced this issue Jan 12, 2018
@joostvanzwieten
exclude numpy 1.14.0
a060809 
`einsum` in this version of numpy does not support single expansion.  See
[numpy#10343].

[numpy#10343]: numpy/numpy#10343
joostvanzwieten added a commit to joostvanzwieten/nutils that referenced this issue Jan 12, 2018
@joostvanzwieten
disable einsum optimization
d34fbc4 
This circumvents a bug in numpy 1.14.0: `einsum` does not support single
expansion.  See [numpy#10343].

[numpy#10343]: numpy/numpy#10343
joostvanzwieten added a commit to joostvanzwieten/nutils that referenced this issue Jan 12, 2018
@joostvanzwieten
disable einsum optimization
64015df 
This circumvents a bug in numpy 1.14.0: `einsum` with the now default
`optimize=True` does not support single expansion (see [numpy#10343]).

[numpy#10343]: numpy/numpy#10343
joostvanzwieten added a commit to joostvanzwieten/nutils that referenced this issue Jan 12, 2018
@joostvanzwieten
disable einsum optimization
5142455 
This circumvents a bug in numpy 1.14.0: `einsum` with the now default
`optimize=True` does not support singleton expansion (see [numpy#10343]).

[numpy#10343]: numpy/numpy#10343
@mhvk
Copy link
Contributor

mhvk commented Jan 14, 2018

@charris - this issue is biting lots of astropy users now (astropy/astropy#7051 (comment)) as it screws up one of our coordinate transformations. Those are all with 3x3 matrices, so pending not just a solution but also a better understanding of impacts of performance, I would indeed suggest to revert the default for 1.14.1. Obviously, we also need #10352 and especially its tests so this doesn't recur! (We missed it in astropy in part because for python3 our transformations use matmul -- guess we should test our last python2 version also with numpy-dev...)

charris added a commit to charris/numpy that referenced this issue Jan 14, 2018
@charris
BUG: Make the einsum optimize default to False.
b3fc900 
There have been bugs reported with einsum with the `optimize=True`
default, see numpy#10343. This PR makes the default `False` for the 1.14.1
release. Because optimizing the contractions is not a trivial problem,
this will allow us to pursue fixes in the longer time frame of the 1.15
release.
@charris charris mentioned this issue Jan 14, 2018
Merged
@charris
Copy link
Member

charris commented Jan 14, 2018

@mhvk See #10403.

@mhvk
Copy link
Contributor

mhvk commented Jan 14, 2018

OK, good, that gives more time.

mattjj added a commit to HIPS/autograd that referenced this issue Feb 1, 2018
@mattjj
use einsum optimize=False to avoid upstream numpy bug
55b9202 
numpy/numpy#10343
@jluttine jluttine mentioned this issue Feb 8, 2018
Closed
@charris charris mentioned this issue Feb 9, 2018
Merged
@oseiskar oseiskar mentioned this issue Feb 16, 2018
Closed
@dgasmith dgasmith mentioned this issue Feb 21, 2018
Closed
VikingScientist pushed a commit to VikingScientist/nutils that referenced this issue Mar 3, 2018
@joostvanzwieten @VikingScientist
disable einsum optimization
4db9942 
This circumvents a bug in numpy 1.14.0: `einsum` with the now default
`optimize=True` does not support singleton expansion (see [numpy#10343]).

[numpy#10343]: numpy/numpy#10343
NileGraddis pushed a commit to AllenInstitute/datacube that referenced this issue May 11, 2018
downgrade numpy (waiting on numpy/numpy#10343)
e664c10
@cerebis cerebis mentioned this issue May 16, 2018
Merged
@toslunar toslunar mentioned this issue Jun 4, 2018
Merged
janeknowsbest77 added a commit to janeknowsbest77/autograd that referenced this issue Jul 25, 2024
@janeknowsbest77 @mattjj
use einsum optimize=False to avoid upstream numpy bug
bd8a432 
numpy/numpy#10343
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
06 - Regression component: numpy._core
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

BUG: Fix einsum optimize logic for singleton dimensions larsoner/numpy
5 participants
@seberg @charris @eric-wieser @mhvk @d70-t