Travis is unhappy (again) #1773
Comments
|
Isn't that the failure because of #1770 in combination with multinomial sampling? Do you mean this one? https://travis-ci.org/pymc-devs/pymc3/jobs/200522344 |
|
Wrong copy&paste. |
|
My last PR (#1775) also fails with these errors. Anyway I can help? |
|
@aloctavodia Not sure what's behind this. Probably a theano update. |
|
@nouiz any idea what recent theano change might have triggered that? |
|
I'm not able to reproduce the first error with the master of Theano. Which Theano flags get used? |
|
Also can't reproduce locally. Seems like we need the docker. @ColCarroll do you have that set up? |
|
My docker is still on 3.5.2, but has newest theano. That particular test
passes fine (new warning, though!):
```
Check effective sample size is equal to number of samples when initializing
with MAP ... Optimization terminated successfully.
Current function value: 4.594693
Iterations: 0
Function evaluations: 1
Gradient evaluations: 1
WARNING (theano.tensor.blas): We did not found a dynamic library into the
library_dir of the library we use for blas. If you use ATLAS, make sure to
compile it with dynamics library.
100%|█████| 100/100 [00:00<00:00, 628.66it/s]
ok
```
…On Mon, Feb 13, 2017 at 4:13 PM Thomas Wiecki ***@***.***> wrote:
Also can't reproduce locally. Seems like we need the docker. @ColCarroll
<https://github.com/ColCarroll> do you have that set up?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1773 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ACMHEJmj8up-7oc9886CCRpU4ZEQtVPNks5rcMdygaJpZM4L-KLh>
.
|
|
I'm totally stumped by this. No one can reproduce it, even in the docker clone of the travis environment. @ColCarroll Did you mention you tried to roll back to an older theano version but it didn't help? |
|
Sigh, I'm an idiot and was running a test that was non-problematic. Can now reproduce -- I'm looking for a theano SHA where it works (if someone wants to share |
|
Fails at least back to 0.8.2 -- I don't think this is a theano problem. I can debug more later, but this reproduces the error pretty quick if someone wants to take a look: |
|
It also doesn't seem to be a pymc3 problem -- I think. Or maybe it's an interaction between a pymc3 change and a theano change? |
|
What happens if you roll both back. |
|
Still fails with
Theano 0.8.2,
pymc3@a8b4d705
a8b4d70 Thomas Wiecki Fri Feb 3 19:54:40 2017 +0100
Merge pull request #1753 from pymc-devs/fix_types_lkj
…On Wed, Feb 15, 2017 at 9:29 AM Thomas Wiecki ***@***.***> wrote:
What happens if you roll both back.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1773 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ACMHEEoBMe9Dg3lKdXfLPz3ZFSi7jjA2ks5rcwvYgaJpZM4L-KLh>
.
|
|
Hrm, a travis issue?! |
|
Did the version of some other dependency change recently? From the place where I think the error is generated it looks like some array does not have the expected rank. It's not using the git version of numpy, does it? Because there was this numpy/numpy@d9d266a recently. |
|
My docker image had pretty old versions of all libraries while it was failing, then I updated all. It looks like scipy just had a new version pushed to conda. Tried raising/lowering scipy and numpy (a few shas related to the change you mentioned) with no effect. If someone has tests passing, could you share a gist with your |
|
I just looked into the code in |
|
It's the same problem in all test failures: Each time a full matrix is passed to Metropolis, which is passed on to |
|
@aseyboldt Good catch! I suppose fixing it is more critical than figuring out why. |
|
True, but I'd still like to know :-) |
Probably a theano master problem?
The text was updated successfully, but these errors were encountered: