This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
Multiple numpy tests fail with numpy 1.19 #18600
Comments
|
Error Log |
|
np.delete also fails the compatibility test: |
|
Temporarily disabled this test in #18598 |
|
The failure is observed in v1.7.x branch as well, check log http://jenkins.mxnet-ci.amazon-ml.com/blue/rest/organizations/jenkins/pipelines/mxnet-validation/pipelines/centos-cpu/branches/PR-18478/runs/2/nodes/110/steps/159/log/?start=0 from PR #18478. |
7 tasks
|
@eric-haibin-lin actually we should have put constraint on the numpy version first before we can verify that the integration with the new version is functional, especially since numpy doesn't follow semantic versioning. Could you post a change to reflect this and to re-enable the test? |
|
@ciyongch should this be fixed for 1.7 release? I'd expect that users would like to use latest MXNet version with latest numpy. |
|
Hi @leezu, as the latest numpy (1.19.0) was released just for a couple of days, I think it's better but not mandatory to have this fix for 1.7 release, as numpy 1.19.0 is not fully verified with MXNet yet and it's very common to pin package A to some certain version of package B in python world. |
@ciyongch that's not true. In fact it's easy to
If we consider the numpy API experimental in MXNet 1.7, it's ok but certainly not ideal. |
|
Yes, it could be a broken issue unless there's a constraint for the numpy version for MXNet's installation. I'm not sure whether we can state numpy API as an experimental feature or not since it's already introduced in 1.6.0 release. |
|
Is it feasible to make mxnet compatible with all versions above 1.17? If not then we definitely would need to make a choice on the numpy version to support. I'm asking this because numpy doesn't follow semantic versioning. |
|
Only deprecated APIs are removed. There's no need to support the deprecated APIs in older releases. So generally this should be possible. What do you think @yzhliu |
|
Hi @leezu @szha @yzhliu , is there any updates for this topic, we're pending on this to drop the rc tag for 1.7 release. Based on the previous discussion, seems there're two options to go,
Or is it necessary to raise it on dev@ ? Thanks! |
|
As discussed here, we'll mark numpy operator as experimental in v1.7 release and probably will move forward by taking this as an known issue (some cases are broken with latest numpy version 1.19.0). Please let me know if you have any further concerns or suggestions! |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
The following four numpy op tests fails with numpy==1.19
The failure happens on the master branch CI http://jenkins.mxnet-ci.amazon-ml.com/blue/organizations/jenkins/mxnet-validation%2Fwindows-cpu/detail/master/2034/pipeline and blocks recent PRs http://jenkins.mxnet-ci.amazon-ml.com/blue/organizations/jenkins/mxnet-validation%2Fwindows-cpu/detail/PR-18598/7/pipeline
It only happens recently, and I noticed that the installed numpy version on CI has changed from
numpy-1.18.5tonumpy-1.19.0- see installation log in previous commit and master head.Also notice that the error log shows that the exceptions happen in the _np.xx functions, which is part of the official numpy APIs instead of mx.numpy APIs.
@yzhliu FYI
The text was updated successfully, but these errors were encountered: