Bug#893656: transition: theano 0.9 -> 1.0 - please update lasagne

[Stephen Sinclair]

I was waiting for Lasagne to do a release, but I can take a look at it
soon. I'll try to do it tomorrow.


On Thu, Mar 22, 2018 at 6:41 PM, Rebecca N. Palmer 
wrote:

> lasagne - 3 test failures; fixed upstream by
>> https://github.com/Lasagne/Lasagne/pull/836
>>
> Latest upstream (37ca134; only packaging change was to drop
> remove-deprecated.patch) doesn't have these failures.  There is a warning
> that cuda_convnet is no longer available with Theano 1.0, but that has
> another dependency (pybuild2) that was never in Debian anyway.
>
> pyopencl - FTBFS for unrelated reasons (#893050); the theano-using code
>> appears to be build/test scripts (pyopencl/compyte/gen*, test*) we never
>> actually run
>>
> Builds after applying the path in #893050; 7 tests fail or crash (in
> Python 2, 6 also in Python 3), but they do this with or without Theano.
>
> sympy - build hangs at mkdir -p _build/logo (with mkdir using a full CPU
>> core?!), probably unrelated as it also happens without *-theano installed
>>
> This hang is because faketime doesn't work well with cowbuilder --login,
> i.e. nothing to do with sympy itself.  Running the testsuite with new
> Theano succeeds.
>
> Hence, it appears that updating lasagne is the only change required to
> work with Theano 1.0, but testing on CUDA-capable hardware would still be
> preferable.
>
>
-- 
debian-science-maintainers mailing list
debian-science-maintainers@lists.alioth.debian.org
http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/debian-science-maintainers

Bug#893656: transition: theano 0.9 -> 1.0 - please update lasagne

[Rebecca N. Palmer]

lasagne - 3 test failures; fixed upstream by 
https://github.com/Lasagne/Lasagne/pull/836
Latest upstream (37ca134; only packaging change was to drop 
remove-deprecated.patch) doesn't have these failures.  There is a 
warning that cuda_convnet is no longer available with Theano 1.0, but 
that has another dependency (pybuild2) that was never in Debian anyway.



pyopencl - FTBFS for unrelated reasons (#893050); the theano-using code appears 
to be build/test scripts (pyopencl/compyte/gen*, test*) we never actually run
Builds after applying the path in #893050; 7 tests fail or crash (in 
Python 2, 6 also in Python 3), but they do this with or without Theano.



sympy - build hangs at mkdir -p _build/logo (with mkdir using a full CPU 
core?!), probably unrelated as it also happens without *-theano installed
This hang is because faketime doesn't work well with cowbuilder --login, 
i.e. nothing to do with sympy itself.  Running the testsuite with new 
Theano succeeds.


Hence, it appears that updating lasagne is the only change required to 
work with Theano 1.0, but testing on CUDA-capable hardware would still 
be preferable.


--
debian-science-maintainers mailing list
debian-science-maintainers@lists.alioth.debian.org
http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/debian-science-maintainers

Bug#893656: transition: theano 0.9 -> 1.0

[Rebecca N. Palmer]

Source: theano
Severity: wishlist

(Posting here rather than on debian-release as the formal transition 
process is rarely used, and its tools mostly useless, for non-compiled 
cases.)


Theano 1.0 (currently in Salsa) contains some API-breaking changes:
http://www.deeplearning.net/software/theano/NEWS.html

The packages in main that import theano are:

brian - not tested, FTBFS for unrelated reasons (#876920), not in testing
deepnano - not tested but looks already broken, not in testing
keras - build and debci tests pass; upstream say it shouldn't be a 
problem https://github.com/keras-team/keras/issues/5209
lasagne - 3 test failures; fixed upstream by 
https://github.com/Lasagne/Lasagne/pull/836
pyopencl - FTBFS for unrelated reasons (#893050); the theano-using code 
appears to be build/test scripts (pyopencl/compyte/gen*, test*) we never 
actually run
sympy - build hangs at mkdir -p _build/logo (with mkdir using a full CPU 
core?!), probably unrelated as it also happens without *-theano installed


As one of the changes is removal of theano.sandbox.cuda (replaced by 
theano.gpuarray, with a different API) and my hardware doesn't support 
CUDA, I am unable to test this fully and testing by others would be 
welcome.  Be aware that if python(3)-nose is installed (it wasn't for 
these tests), attempting to import theano.sandbox.cuda raises SkipTest 
(?!), not ImportError: 
https://salsa.debian.org/science-team/theano/raw/master/theano/sandbox/cuda/__init__.py


--
debian-science-maintainers mailing list
debian-science-maintainers@lists.alioth.debian.org
http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/debian-science-maintainers