Bug#898126: theano: memory leak

2018-05-11 Thread Rebecca N. Palmer

Control: severity -1 normal

This is not a new problem (I'm not sure if it's got worse per test, or 
just got noticed because the test suite got longer), so removing the 
migration block.


The largest (but not only) leaks are
test_pickle_big_fusion (theano.tensor.tests.test_opt.test_fusion)
test_const4 (theano.tensor.tests.test_opt.test_local_subtensor_merge)
test_none_slice (theano.tensor.tests.test_opt.test_local_subtensor_merge)
test_shape_i_const (theano.tensor.tests.test_subtensor.T_subtensor)

Weirdly, memory use increases several times more during one of these 
tests when it is part of a full testsuite run, than during the same test 
when it is part of a shorter (e.g. just these few tests) run.


In a short run, the leaked memory belongs to the main test process (not 
a subprocess) and gc.collect() doesn't free it, but I haven't tried 
these in a full run.




Bug#898126: theano: memory leak

2018-05-07 Thread Rebecca N. Palmer

Package: python-theano
Version: 1.0.1+dfsg-2
Severity: serious

(Filing this as RC to give myself time to investigate: I may downgrade 
it later.)


Memory usage increases over theano's tests, to ~6GB by the end of the 
python2 set (6981 tests) in Debian sid amd64, and enough to fail the 
Ubuntu arm64 and ppc64el builds (LP#1769672).