<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<font size="-1"><font face="Arial">Dear Matteo,<br>
thank you for kindly replying!<br>
<br>
Yes, I do have the latest versions of joblib (0.11) and sklearn
(0.19.1), see at the bottom of this email.<br>
<br>
The problem seems independent of either running in jupyter, or
evoking ipython or python directly in the console.<br>
<br>
I am now wondering whether there may be something wrong in my
snippet.<br>
When first running your gist, I encountered an:<br>
<br>
UnboundLocalError: local variable 'best_clf' referenced
before assignment<br>
<br>
which I solved by moving the best_clf declaration a few lines
down:<br>
<br>
-----------------------------------------<br>
#best_clfs = {} #moved down 7 lines<br>
confusion = ConfusionMatrix()<br>
verbose(1, "Estimating error using nested CV for model
selection")<br>
partitioner = partitionerCD<br>
splitter = Splitter('partitions')<br>
tms = Parallel(n_jobs=2)(delayed(_run_one_partition)(isplit,
partitions) <br>
for isplit, partitions in
enumerate(partitionerCD.generate(fds)))<br>
best_clfs = {}<br>
for tm in tms:<br>
confusion += tm.ca.stats<br>
best_clfs[tm.measure.descr] =
best_clfs.get(tm.measure.descr, 0) + 1<br>
-----------------------------------------<br>
<br>
But now, running the snippet in ipython/python specifically for
the SVM parallelization issue, I saw the error message popping
up again:<br>
<br>
UnboundLocalError: local variable 'best_clf' referenced
before assignment<br>
<br>
May this be the culprit? As a reminder, the full snippet I am
using is included in my previous email.<br>
<br>
Thank you and very best wishes,<br>
Marco<br>
<br>
<br>
In [21]: mvpa2.wtf(exclude=['runtime','process']) ##other
possible arguments (['sources', <br>
Out[21]: <br>
Current date: 2017-11-24 21:23<br>
PyMVPA:<br>
Version: 2.6.3<br>
Hash: 9c07e8827819aaa79ff15d2db10c420a876d7785<br>
Path:
/usr/lib/python2.7/dist-packages/mvpa2/__init__.pyc<br>
Version control (GIT):<br>
GIT information could not be obtained due
"/usr/lib/python2.7/dist-packages/mvpa2/.. is not under GIT"<br>
SYSTEM:<br>
OS: posix Linux 4.13.0-1-amd64 #1 SMP Debian
4.13.4-2 (2017-10-15)<br>
Distribution: debian/buster/sid<br>
EXTERNALS:<br>
Present: atlas_fsl, cPickle, ctypes, good
scipy.stats.rv_continuous._reduce_func(floc,fscale), good
scipy.stats.rv_discrete.ppf, griddata, gzip, h5py, hdf5,
ipython, joblib, liblapack.so, libsvm, libsvm verbosity control,
lxml, matplotlib, mdp, mdp ge 2.4, mock, nibabel, nose, numpy,
numpy_correct_unique, pprocess, pylab, pylab plottable, pywt,
pywt wp reconstruct, reportlab, running ipython env, scipy, skl,
statsmodels<br>
Absent: afni-3dinfo, atlas_pymvpa, cran-energy, datalad,
elasticnet, glmnet, good scipy.stats.rdist, hcluster, lars,
mass, nipy, nipy.neurospin, numpydoc, openopt, pywt wp
reconstruct fixed, rpy2, scipy.weave, sg ge 0.6.4, sg ge 0.6.5,
sg_fixedcachesize, shogun, shogun.krr, shogun.lightsvm,
shogun.mpd, shogun.svmocas, shogun.svrlight, weave<br>
Versions of critical externals:<br>
ctypes : 1.1.0<br>
h5py : 2.7.1<br>
hdf5 : 1.10.0<br>
ipython : 5.5.0<br>
joblib : 0.11<br>
lxml : 4.1.0<br>
matplotlib : 2.0.0<br>
mdp : 3.5<br>
mock : 2.0.0<br>
nibabel : 2.3.0dev<br>
numpy : 1.13.1<br>
pprocess : 0.5<br>
pywt : 0.5.1<br>
reportlab : 3.4.0<br>
scipy : 0.19.1<br>
skl : 0.19.1<br>
Matplotlib backend: TkAgg<br>
<br>
<br>
</font></font>
<blockquote type="cite"><font size="-1"><font face="Arial">On
24/11/2017 17:32, Matteo Visconti di Oleggio Castello wrote:<br>
<br>
Hi Marco,<br>
<br>
some ideas in random order<br>
<br>
- what version of sklearn/joblib are you using? I would make
sure to use<br>
the latest version (0.11), perhaps not importing it from
sklearn (unless<br>
you have the latest sklearn version, 0.19.1)<br>
- are you running the code in a jupyter notebook? There might
be some<br>
issues with that (see
<a class="moz-txt-link-freetext" href="https://github.com/joblib/joblib/issues/174">https://github.com/joblib/joblib/issues/174</a>). As a<br>
test you might try to convert your notebook to a script and
then run it<br>
<br>
</font></font><br>
<br>
<div class="moz-cite-prefix">On 23/11/2017 12:07, marco tettamanti
wrote:<br>
</div>
<blockquote type="cite"
cite="mid:77c2f8e5-6d2c-68c8-ef14-a430601be7a0@gmail.com"> <font
size="-1"><font face="Arial">Dear Matteo (and others),<br>
sorry, I am again asking for your help!<br>
<br>
I have experimented with the analysis of my dataset using an
adaptation of your joblib-based gist.<br>
As I wrote before, it works perfectly, but not with some
classifiers: SVM classifiers always cause the code to
terminate with an error.<br>
<br>
If I set:<br>
myclassif=clfswh['!gnpp','!skl','svm'] #Note that
'gnnp' and 'skl' were excluded for independent reasons<br>
the code runs through without errors.<br>
<br>
However, with:<br>
myclassif=clfswh['!gnpp','!skl']<br>
I get the following error:<br>
MaybeEncodingError: Error sending result:
'[TransferMeasure(measure=SVM(svm_impl='C_SVC',
kernel=LinearLSKernel(), weight=[], probability=1,<br>
weight_label=[]),
splitter=Splitter(space='partitions'),
postproc=BinaryFxNode(space='targets'),
enable_ca=['stats'])]'. Reason: 'TypeError("can't<br>
pickle SwigPyObject objects",)'<br>
<br>
After googling for what may cause this particular error, I
have found that the situation improves slightly (i.e. more
splits executed, sometimes even all splits) by importing the
following:<br>
import os<br>
from sklearn.externals.joblib import Parallel,
delayed<br>
from sklearn.externals.joblib.parallel import
parallel_backend<br>
and then specifying just before 'Parallel(n_jobs=2)': <br>
with parallel_backend('threading'):<br>
However, also in this case, the code invariably terminates
with a long error message (I only report an extract, but in
case I can send the whole error message):<br>
<type 'str'>: (<type
'exceptions.UnicodeEncodeError'>,
UnicodeEncodeError('ascii',<br>
u'JoblibAttributeError\n___________________________________________________________________________\nMultiprocessing<br>
exception:\n...........................................................................\n/usr/lib/python2.7/runpy.py
in<br>
_run_module_as_main(mod_name=\'ipykernel_launcher\',
alter_argv=1)\n 169 pkg_name =
mod_name.rpartition(\'.\')[0]\n 170<br>
main_globals = sys.modules["__main__"].__dict__\n
171 if alter_argv:\n 172 sys.argv[0] =
fname\n 173 return _run_code(code,<br>
main_globals, None,\n--> 174<br>
<br>
</font></font><br>
I think I have sort of understood that the problem is due to
some failure in pickling the parallelized jobs, but I have no
clues if and how it can be solved.<br>
Do you have any suggestions?<br>
<br>
Thank you and very best wishes,<br>
Marco<br>
<br>
p.s. This is again the full code:<br>
<br>
########## * ##########<br>
##########<br>
<br>
PyMVPA:<br>
Version: 2.6.3<br>
Hash: 9c07e8827819aaa79ff15d2db10c420a876d7785<br>
Path:
/usr/lib/python2.7/dist-packages/mvpa2/__init__.pyc<br>
Version control (GIT):<br>
GIT information could not be obtained due
"/usr/lib/python2.7/dist-packages/mvpa2/.. is not under GIT"<br>
SYSTEM:<br>
OS: posix Linux 4.13.0-1-amd64 #1 SMP Debian
4.13.4-2 (2017-10-15)<br>
<br>
<br>
print fds.summary()<br>
Dataset: 36x534@float32, <sa:
chunks,targets,time_coords,time_indices>, <fa:
voxel_indices>, <a:
imgaffine,imghdr,imgtype,mapper,voxel_dim,voxel_eldim><br>
stats: mean=0.548448 std=1.40906 var=1.98546 min=-5.41163
max=9.88639<br>
No details due to large number of targets or chunks. Increase
maxc and maxt if desired<br>
Summary for targets across chunks<br>
targets mean std min max #chunks<br>
C 0.5 0.5 0 1 18<br>
D 0.5 0.5 0 1 18<br>
<br>
<br>
#Evaluate prevalent best classifier with nested crossvalidation<br>
verbose.level = 5<br>
<br>
partitionerCD = ChainNode([NFoldPartitioner(cvtype=2,
attr='chunks'), Sifter([('partitions', 2), ('targets', ['C',
'D'])])], space='partitions')<br>
# training partitions<br>
for fds_ in partitionerCD.generate(fds): <br>
training = fds[fds_.sa.partitions == 1]<br>
#print list(zip(training.sa.chunks, training.sa.targets))<br>
# testing partitions<br>
for fds_ in partitionerCD.generate(fds): <br>
testing = fds[fds_.sa.partitions == 2]<br>
#print list(zip(testing.sa.chunks, testing.sa.targets))<br>
<br>
#Helper function (partitionerCD recursively acting on dstrain,
rather than on fds):<br>
def select_best_clf(dstrain_, clfs):<br>
"""Select best model according to CVTE<br>
Helper function which we will use twice -- once for proper
nested<br>
cross-validation, and once to see how big an optimistic bias
due<br>
to model selection could be if we simply provide an entire
dataset.<br>
Parameters<br>
----------<br>
dstrain_ : Dataset<br>
clfs : list of Classifiers<br>
Which classifiers to explore<br>
Returns<br>
-------<br>
best_clf, best_error<br>
"""<br>
best_error = None<br>
for clf in clfs:<br>
cv = CrossValidation(clf, partitionerCD)<br>
# unfortunately we don't have ability to reassign clf
atm<br>
# cv.transerror.clf = clf<br>
try:<br>
error = np.mean(cv(dstrain_))<br>
except LearnerError, e:<br>
# skip the classifier if data was not appropriate
and it<br>
# failed to learn/predict at all<br>
continue<br>
if best_error is None or error < best_error:<br>
best_clf = clf<br>
best_error = error<br>
verbose(4, "Classifier %s cv error=%.2f" % (clf.descr,
error))<br>
verbose(3, "Selected the best out of %i classifiers %s with
error %.2f"<br>
% (len(clfs), best_clf.descr, best_error))<br>
return best_clf, best_error<br>
<br>
# This function will run all classifiers for one single
partitions<br>
myclassif=clfswh['!gnpp','!skl'][5:6] #Testing a single SVM
classifier<br>
def _run_one_partition(isplit, partitions,
classifiers=myclassif): #see §§<br>
verbose(2, "Processing split #%i" % isplit)<br>
dstrain, dstest = list(splitter.generate(partitions))<br>
best_clf, best_error = select_best_clf(dstrain, classifiers)<br>
# now that we have the best classifier, lets assess its
transfer<br>
# to the testing dataset while training on entire training<br>
tm = TransferMeasure(best_clf,
splitter,postproc=BinaryFxNode(mean_mismatch_error,space='targets'),
enable_ca=['stats'])<br>
tm(partitions)<br>
return tm<br>
<br>
#import os<br>
#from sklearn.externals.joblib import Parallel, delayed<br>
#from sklearn.externals.joblib.parallel import parallel_backend<br>
<br>
# Parallel estimate error using nested CV for model selection<br>
confusion = ConfusionMatrix()<br>
verbose(1, "Estimating error using nested CV for model
selection")<br>
partitioner = partitionerCD<br>
splitter = Splitter('partitions')<br>
# Here we are using joblib Parallel to parallelize each
partition<br>
# Set n_jobs to the number of available cores (or how many you
want to use)<br>
#with parallel_backend('threading'):<br>
# tms =
Parallel(n_jobs=2)(delayed(_run_one_partition)(isplit,
partitions) <br>
tms = Parallel(n_jobs=2)(delayed(_run_one_partition)(isplit,
partitions)<br>
for isplit, partitions in
enumerate(partitionerCD.generate(fds)))<br>
# Parallel retuns a list with the results of each parallel loop,
so we need to<br>
# unravel it to get the confusion matrix<br>
best_clfs = {}<br>
for tm in tms:<br>
confusion += tm.ca.stats<br>
best_clfs[tm.measure.descr] =
best_clfs.get(tm.measure.descr, 0) + 1<br>
<br>
##########<br>
########## * ##########<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<div class="moz-cite-prefix">On 13/11/2017 09:12, marco
tettamanti wrote:<br>
</div>
<blockquote type="cite"
cite="mid:b63d0b4c-525b-b1d0-54b5-45c765778976@gmail.com"> <font
size="-1"><font face="Arial">Dear Matteo,<br>
grazie mille, this is precisely the kind of thing I was
looking for: it works like charm!<br>
Ciao,<br>
Marco<br>
<br>
</font></font>
<blockquote type="cite"><font size="-1"><font face="Arial">On
11/11/2017 21:44, Matteo Visconti di Oleggio Castello
wrote:<br>
<br>
Hi Marco,<br>
<br>
in your case, I would then recommend looking into joblib
to parallelize<br>
your for loops (<a class="moz-txt-link-freetext"
href="https://pythonhosted.org/joblib/parallel.html">https://pythonhosted.org/joblib/parallel.html</a>).<br>
<br>
As an example, here's a gist containing part of the
PyMVPA's nested_cv<br>
example where I parallelized the loop across partitions.
I feel this is<br>
what you might want to do in your case, since you have a
lot more folds.<br>
<br>
Here's the gist:<br>
<a class="moz-txt-link-freetext"
href="https://gist.github.com/mvdoc/0c2574079dfde78ea649e7dc0a3feab0">https://gist.github.com/mvdoc/0c2574079dfde78ea649e7dc0a3feab0</a><br>
</font></font><br>
<br>
<div class="moz-cite-prefix">On 10/11/2017 21:13, marco
tettamanti wrote:<br>
</div>
<blockquote type="cite"
cite="mid:f898a361-4187-9f95-23f2-d9cb994cace9@gmail.com">
Dear Matteo,<br>
thank you for the willingness to look into my code.<br>
<br>
This is taken almost verbatim from <a
class="moz-txt-link-freetext"
href="http://dev.pymvpa.org/examples/nested_cv.html">http://dev.pymvpa.org/examples/nested_cv.html</a>,
except for the leave-one-pair-out partitioning, and a
slight reduction in the number of classifiers (in the
original example, they are around 45).<br>
<br>
Any help or suggestion would be greatly appreciated!<br>
All the best,<br>
Marco<br>
<br>
<br>
########## * ##########<br>
##########<br>
<br>
PyMVPA:<br>
Version: 2.6.3<br>
Hash: 9c07e8827819aaa79ff15d2db10c420a876d7785<br>
Path:
/usr/lib/python2.7/dist-packages/mvpa2/__init__.pyc<br>
Version control (GIT):<br>
GIT information could not be obtained due
"/usr/lib/python2.7/dist-packages/mvpa2/.. is not under
GIT"<br>
SYSTEM:<br>
OS: posix Linux 4.13.0-1-amd64 #1 SMP Debian
4.13.4-2 (2017-10-15)<br>
<br>
<br>
print fds.summary()<br>
Dataset: 36x534@float32, <sa:
chunks,targets,time_coords,time_indices>, <fa:
voxel_indices>, <a:
imgaffine,imghdr,imgtype,mapper,voxel_dim,voxel_eldim><br>
stats: mean=0.548448 std=1.40906 var=1.98546 min=-5.41163
max=9.88639<br>
No details due to large number of targets or chunks.
Increase maxc and maxt if desired<br>
Summary for targets across chunks<br>
targets mean std min max #chunks<br>
C 0.5 0.5 0 1 18<br>
D 0.5 0.5 0 1 18<br>
<br>
<br>
#Evaluate prevalent best classifier with nested
crossvalidation<br>
verbose.level = 5<br>
<br>
partitionerCD = ChainNode([NFoldPartitioner(cvtype=2,
attr='chunks'), Sifter([('partitions', 2), ('targets',
['C', 'D'])])], space='partitions')<br>
# training partitions<br>
for fds_ in partitionerCD.generate(fds): <br>
training = fds[fds_.sa.partitions == 1]<br>
#print list(zip(training.sa.chunks,
training.sa.targets))<br>
# testing partitions<br>
for fds_ in partitionerCD.generate(fds): <br>
testing = fds[fds_.sa.partitions == 2]<br>
#print list(zip(testing.sa.chunks,
testing.sa.targets))<br>
<br>
#Helper function (partitionerCD recursively acting on
dstrain, rather than on fds):<br>
def select_best_clf(dstrain_, clfs):<br>
"""Select best model according to CVTE<br>
Helper function which we will use twice -- once for
proper nested<br>
cross-validation, and once to see how big an
optimistic bias due<br>
to model selection could be if we simply provide an
entire dataset.<br>
Parameters<br>
----------<br>
dstrain_ : Dataset<br>
clfs : list of Classifiers<br>
Which classifiers to explore<br>
Returns<br>
-------<br>
best_clf, best_error<br>
"""<br>
best_error = None<br>
for clf in clfs:<br>
cv = CrossValidation(clf, partitionerCD)<br>
# unfortunately we don't have ability to reassign
clf atm<br>
# cv.transerror.clf = clf<br>
try:<br>
error = np.mean(cv(dstrain_))<br>
except LearnerError, e:<br>
# skip the classifier if data was not
appropriate and it<br>
# failed to learn/predict at all<br>
continue<br>
if best_error is None or error < best_error:<br>
best_clf = clf<br>
best_error = error<br>
verbose(4, "Classifier %s cv error=%.2f" %
(clf.descr, error))<br>
verbose(3, "Selected the best out of %i classifiers %s
with error %.2f"<br>
% (len(clfs), best_clf.descr, best_error))<br>
return best_clf, best_error<br>
<br>
#Estimate error using nested CV for model selection:<br>
best_clfs = {}<br>
confusion = ConfusionMatrix()<br>
verbose(1, "Estimating error using nested CV for model
selection")<br>
partitioner = partitionerCD<br>
splitter = Splitter('partitions')<br>
for isplit, partitions in
enumerate(partitionerCD.generate(fds)):<br>
verbose(2, "Processing split #%i" % isplit)<br>
dstrain, dstest = list(splitter.generate(partitions))<br>
best_clf, best_error = select_best_clf(dstrain,
clfswh['!gnpp','!skl'])<br>
best_clfs[best_clf.descr] =
best_clfs.get(best_clf.descr, 0) + 1<br>
# now that we have the best classifier, lets assess
its transfer<br>
# to the testing dataset while training on entire
training<br>
tm = TransferMeasure(best_clf, splitter,<br>
postproc=BinaryFxNode(mean_mismatch_error,
space='targets'), enable_ca=['stats'])<br>
tm(partitions)<br>
confusion += tm.ca.stats<br>
<br>
##########<br>
########## * ##########<br>
<br>
<br>
<br>
<br>
<br>
<br>
<blockquote type="cite"><font size="-1"><font face="Arial">On
10/11/2017 15:43, Matteo Visconti di Oleggio
Castello wrote:<br>
<br>
What do you mean with "cycling over approx 40
different classifiers"? Are<br>
you testing different classifiers? If that's the
case, a possibility is to<br>
create a script that takes as argument the type of
classifiers and runs the<br>
classification across all folds. In that way you can
submit 40 jobs and<br>
parallelize across classifiers.<br>
<br>
If that's not the case, because the folds are
independent and deterministic<br>
I would create a script that performs the
classification on blocks of folds<br>
(say fold 1 to 30, 31, to 60, etc...), and then
submit different jobs, so<br>
to parallelize there.<br>
<br>
I think that if you send a snippet of the code
you're using it can be more<br>
evident which are good points for parallelization.<br>
</font></font><br>
<br>
<div class="moz-cite-prefix">On 10/11/2017 09:57, marco
tettamanti wrote:<br>
</div>
<blockquote type="cite"
cite="mid:9a2fb439-554c-360c-6aef-4c0693dc1a1a@gmail.com">
<pre wrap="">Dear Matteo and Nick,
thank you for your responses.
I take the occasion to ask some follow-up questions, because I am struggling to
make pymvpa2 computations faster and more efficient.
I often find myself in the situation of giving up with a particular analysis,
because it is going to take far more time that I can bear (weeks, months!). This
happens particularly with searchlight permutation testing (gnbsearchlight is
much faster, but does not support pprocess), and nested cross-validation.
As for the latter, for example, I recently wanted to run nested cross-validation
in a sample of 18 patients and 18 controls (1 image x subject), training the
classifiers to discriminate patients from controls in a leave-one-pair-out
partitioning scheme. This yields 18*18=324 folds. For a small ROI of 36 voxels,
cycling over approx 40 different classifiers takes about 2 hours for each fold
on a decent PowerEdge T430 Dell server with 128GB RAM. This means approx. 27
days for all 324 folds!
The same server is equipped with 32 CPUs. With full parallelization, the same
analysis may be completed in less than one day. This is the reason of my
interest and questions about parallelization.
Is there anything that you experts do in such situations to speed up or make the
computation more efficient?
Thank you again and best wishes,
Marco
</pre>
<blockquote type="cite">
<pre wrap="">On 10/11/2017 10:07, Nick Oosterhof wrote:
There have been some plans / minor attempts for using parallelisation more
parallel, but as far as I know we only support pprocces, and only for (1)
searchlight; (2) surface-based voxel selection; and (3) hyperalignment. I
do remember that parallelisation of other functions was challenging due to
some getting the conditional attributes set right, but this is long time
ago.
</pre>
<blockquote type="cite">
<pre wrap="">On 09/11/2017 18:35, Matteo Visconti di Oleggio Castello wrote:
Hi Marco,
AFAIK, there is no support for parallelization at the level of
cross-validation. Usually for a small ROI (such a searchlight) and with
standard CV schemes, the process is quite fast, and the bottleneck is
really the number of searchlights to be computed (for which parallelization
exists).
In my experience, we tend to parallelize at the level of individual
participants; for example we might set up a searchlight analysis with
however n_procs you can have, and then submit one such job for every
participant to a cluster (using either torque or condor).
HTH,
Matteo
On 09/11/2017 10:08, marco tettamanti wrote:
</pre>
<blockquote type="cite">
<pre wrap="">Dear all,
forgive me if this has already been asked in the past, but I was wondering
whether there has been any development meanwhile.
Are there any chances that one can generally apply parallel computing (multiple
CPUs or clusters) with pymvpa2, in addition to what is already implemented for
searchlight (pprocess)? That is, also for general cross-validation, nested
cross-validation, permutation testing, RFE, etc.?
Has anyone had succesful experience with parallelization schemes such as
ipyparallel, condor or else?
Thank you and best wishes!
Marco
</pre>
</blockquote>
</blockquote>
</blockquote>
</blockquote>
<br>
<pre class="moz-signature" cols="80">--
Marco Tettamanti, Ph.D.
Nuclear Medicine Department & Division of Neuroscience
IRCCS San Raffaele Scientific Institute
Via Olgettina 58
I-20132 Milano, Italy
Phone ++39-02-26434888
Fax ++39-02-26434892
Email: <a class="moz-txt-link-abbreviated" href="mailto:tettamanti.marco@hsr.it">tettamanti.marco@hsr.it</a>
Skype: mtettamanti
<a class="moz-txt-link-freetext" href="http://scholar.google.it/citations?user=x4qQl4AAAAAJ">http://scholar.google.it/citations?user=x4qQl4AAAAAJ</a></pre>
</blockquote>
<br>
</blockquote>
</blockquote>
<br>
</blockquote>
</blockquote>
</blockquote>
<br>
</body>
</html>