plot_simulated_data.py takes one hour to finish about nilearn HOT 18 CLOSED

3822.42 seconds in my machine. Mostly because of Searchlight.

In any case this makes it really painful to build the doc and to the user it
looks as if the example was broken (there is no output whatsoever for one
hour).

OK, the fact that there is no output needs to be changed.

That said, we need real examples on nisl, not only toys. This means that
we will have to cope with long-running doc builds (and yes searchlight is
slow and useless).

from nilearn.

3822.42 seconds in my machine. Mostly because of Searchlight.

On my machine, it takes 2 minutes (without any cache).

In any case this makes it really painful to build the doc and to the user it looks as if the example was broken (there is no output whatsoever for one hour).

Normally, the progression is shown like that:
Job #1, processed 190/1435 voxels(13.24%, 76 seconds remaining)

So there must be another problem...

from nilearn.

Normally, the progression is shown like that:
Job #1, processed 190/1435 voxels(13.24%, 76 seconds remaining)

So there must be another problem...

It could be that the output is captured by the gen_rst.py. There is a
'Tee' object that is supposed to sort this out, but... Need investigation

from nilearn.

I'm in the process of changing the Searchlight class API (see PR #65). I'll have a look on this problem at the same time.

from nilearn.

It could be that the output is captured by the gen_rst.py. There is a 'Tee' object that is supposed to sort this out, but...

I remember adressing this problem by outputing the progress in the standard error. Maybe something has been broken since I've done that.

from nilearn.

The reason why there's no output in the console, is because verbose=1 should be passed to Searchlight, which is not the case in plot_simulated_data.py.

from nilearn.

On my machine, with one core only, the searchlight part takes 475s.

from nilearn.

I'm looking into this. It seems to depend on my Python version, maybe it's a problem on my side.

from nilearn.

After some discussion with Gael, adding estimator=svm.SVR(kernel="linear") to the SearchLight initialization parameters partly solves the problem (this was a bug). The computation time is reduced, and the result is way better.

from nilearn.

thanks @pgervais I'll try it out

from nilearn.

After some discussion with Gael, adding estimator=svm.SVR(kernel="linear") to the SearchLight initialization parameters partly solves the problem (this was a bug). The computation time is reduced, and the result is way better.

Estimated computation time on my machine is ~50 000 seconds.

from nilearn.

Estimated computation time on my machine is ~50 000 seconds.

Are you sure that you are talking about the plot_simulated_data example?

from nilearn.

Are you sure that you are talking about the plot_simulated_data example?

No, I thought that we were talking about plot_haxby_searchlight...

from nilearn.

The bug is only in plot_simulated_data.py, not plot_haxby_searchlight.py, because one is a regression, and the other a classification task.

from nilearn.

Are you sure that you are talking about the plot_simulated_data example?

No, I thought that we were talking about plot_haxby_searchlight...

So the take home message is that if you put an SVC in a regression pb, it
is very slow, and the same for an SVR in a classification pb. I believe
that it is because of the low explained variance, that makes the
optimization pb hard (probably because of the large number of support
vectors).

from nilearn.

The bug is only in plot_simulated_data.py, not plot_haxby_searchlight.py, because one is a regression, and the other a classification task.

That's what was puzzling me ;)

from nilearn.

THANKS!! works much better now, it went down to 159s which is really reasonable. I promise to have the buildbot working soon.

from nilearn.

THANKS!! works much better now, it went down to 159s which is really
reasonable.

Well, thanks for pointing it out. It was useful.

I promise to have the buildbot working soon.

Great!

G

from nilearn.