[MRG] Release 0.20.1 #12383
[MRG] Release 0.20.1 #12383
Conversation
…2156) Follow-up on scikit-learn#12152 And added comment why transformer_weights is not passed through, see scikit-learn#11183 (review) for more discussion
…an abstractmethod (scikit-learn#12182) * _run_search raises NotImplementedError instead of being and abstractmethod * add error message * test for a BaseSearchCV child w/o a _run_search * make the test python2 compatible, still in 0.20 zone. * specify cv in tests not to trigger the related FutureWarning * PEP8
…) with negative integer values (scikit-learn#12180) * Fix Issue scikit-learn#12179 OneHotEncoder "only non-negative integers" message should suggest using categories='auto' * Fix Issue scikit-learn#12179 OneHotEncoder "only non-negative integers" message should suggest using categories='auto' * Fix Issue scikit-learn#12179 OneHotEncoder "only non-negative integers" message should suggest using categories='auto' * Fixes scikit-learn#12180 Modify the error message * Fix the spacing
…C and LinearSVR (scikit-learn#12185) * Added Tip for tuning C parameter in LinearSVC/SVR * Added reference about performance improvement * Clearer explanation about large C values * Update svm.rst
|
I looked into it and here is what I got so far. In the CircleCI build calling import numpy as np
n = 56
array = np.repeat(np.nan, n*n).reshape(n, n)
np.linalg.slogdet(array) # Segmentation fault on CircleCI machineNot sure what is the reason behind this ... suggestions welcome otherwise I'll keep debugging. The problematic script is gdb details |
|
It's very strange for this to crop up all of a sudden... Have you tried
clearing the conda instance from the cache?
It's hilarious that it's plot_stock_market again causing headaches...
|
|
Interesting problem. I suppose adding Also FWIW all tests pass on conda forge using this PR conda-forge/scikit-learn-feedstock#80 |
Not sure how I would do that to be honest. From a quick look at .circle/config.yml it does not look like we are caching the conda install but maybe I am missing something. What I tried:
Depends on your sense of humour I guess, but yeah I totally get your point ;-) |
|
You hadn't mentioned the corrupted numpy-base before had you? It certainly looks suspicious! |
Looking at it a bit more it may be a red herring not sure ... I was able to delete the cached packages ( I have tested that adding |
|
Overall it feels like a packaging issue upstream with numpy and MKL. Maybe we can open an issue at numpy with your minimal example from #12383 (comment) and use nomkl (which would also make builds faster)? |
Has anyone tried and manage to reproduce the crash locally or is this just an issue on the circle ci host? The stock market example works with my numpy with openblas installed from pypi. |
|
I can't reproduce the small example #12383 (comment) using the same docker image as in CircleCI. Have not tried to run the full build though, so versions might differ slightly.. |
The thing is I only managed to reproduce the error on the CircleCI machine ... I haven't tried reusing the same docker image to reproduce locally. I'll try first using nomkl in the CircleCI build. |
A PR with |
|
I have a PR with nomkl and |
|
Such a great thing to wake up to lol. Why are there NaN in that array anyway? |
|
For the record, the installed versions of numpy that are failing are |
|
I won't be able to now. Possibly some issue with the default conda channel cf #12636 (comment) It also fails on master. |
Good morning! Good question about the NaNs, not sure.
Which cache do you have in mind and how would you clear it? I would have thought that there were no cache for the miniconda install but maybe I am missing something. |
|
Looks like there is indeed no miniconda cache. Your ssh was down, so I restarted it and I'll try to have a look, but it'll be a bit slow. |
|
Seems unrelated to the numpy version btw. Even crashes with numpy 1.13 |
I saw the same thing: I tried with a few different numpy version and the segmentation was still happening. |
|
but doing |
It did not fix it for the python 2 circle ci build in #12636 (first commit of that PR). |
|
Sorry edited: I didn't use conda-forge, I did |
|
you said
but that's not happening for me using your branch. Is that expected? |
|
honestly I have no idea and will just cherry-pick. |
|
You're right. It shows the commits that are not included, or cannot be
detected as identical to something included (not sure how this works
actually). The first step in that successive rebase -i is to mark the long
head of the list as drop.
|
|
It included some that you picked before, though. That was confusing to me. |
|
Hmm, so Circle CI fails in the PR and passes in 0.20.X? So many mysterious things. |
This is gathering commits for 0.20.1. See https://github.com/scikit-learn/scikit-learn/milestone/27 for remaining work. From master == f5ef674, I have performed the following interactive rebase:
(Devs take note: it's really nice if the first line of a commit message makes it clear what component it relates to, as well as whether it's DOC/FIX/TST, etc. Feel free to rename PRs or edit the commit message upon squash-and-merge, although making sure not to remove the PR number.)