-
Notifications
You must be signed in to change notification settings - Fork 81
/
Copy pathperformance.html
654 lines (602 loc) · 56.9 KB
/
performance.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN"
"http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8" />
<title>How to optimize for speed — scikit-learn 0.21.3 documentation</title>
<!-- htmltitle is before nature.css - we use this hack to load bootstrap first -->
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<link rel="stylesheet" href="../_static/css/bootstrap.min.css" media="screen" />
<link rel="stylesheet" href="../_static/css/bootstrap-responsive.css"/>
<link rel="stylesheet" href="../_static/nature.css" type="text/css" />
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/gallery.css" type="text/css" />
<script type="text/javascript">
var DOCUMENTATION_OPTIONS = {
URL_ROOT: '../',
VERSION: '0.21.3',
COLLAPSE_INDEX: false,
FILE_SUFFIX: '.html',
HAS_SOURCE: true,
SOURCELINK_SUFFIX: '.txt'
};
</script>
<script type="text/javascript" src="../_static/jquery.js"></script>
<script type="text/javascript" src="../_static/underscore.js"></script>
<script type="text/javascript" src="../_static/doctools.js"></script>
<script type="text/javascript" src="../_static/js/copybutton.js"></script>
<script type="text/javascript" src="../_static/js/extra.js"></script>
<script type="text/javascript" src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.0/MathJax.js?config=TeX-AMS_SVG"></script>
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="author" title="About these documents" href="../about.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="Advanced installation instructions" href="advanced_installation.html" />
<link rel="prev" title="Utilities for Developers" href="utilities.html" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<script src="../_static/js/bootstrap.min.js" type="text/javascript"></script>
<script>
VERSION_SUBDIR = (function(groups) {
return groups ? groups[1] : null;
})(location.href.match(/^https?:\/\/scikit-learn.org\/([^\/]+)/));
</script>
<link rel="canonical" href="https://scikit-learn.org/stable/developers/performance.html" />
<script type="text/javascript">
$("div.buttonNext, div.buttonPrevious").hover(
function () {
$(this).css('background-color', '#FF9C34');
},
function () {
$(this).css('background-color', '#A7D6E2');
}
);
function showMenu() {
var topNav = document.getElementById("scikit-navbar");
if (topNav.className === "navbar") {
topNav.className += " responsive";
} else {
topNav.className = "navbar";
}
};
</script>
</head>
<body>
<div class="header-wrapper">
<div class="header">
<p class="logo"><a href="../index.html">
<img src="../_static/scikit-learn-logo-small.png" alt="Logo"/>
</a>
</p><div class="navbar" id="scikit-navbar">
<ul>
<li><a href="../index.html">Home</a></li>
<li><a href="../install.html">Installation</a></li>
<li class="btn-li"><div class="btn-group">
<a href="../documentation.html">Documentation</a>
<a class="btn dropdown-toggle" data-toggle="dropdown">
<span class="caret"></span>
</a>
<ul class="dropdown-menu">
<li class="link-title">Scikit-learn <script>document.write(DOCUMENTATION_OPTIONS.VERSION + (VERSION_SUBDIR ? " (" + VERSION_SUBDIR + ")" : ""));</script></li>
<li><a href="../tutorial/index.html">Tutorials</a></li>
<li><a href="../user_guide.html">User guide</a></li>
<li><a href="../modules/classes.html">API</a></li>
<li><a href="../glossary.html">Glossary</a></li>
<li><a href="../faq.html">FAQ</a></li>
<li><a href="index.html">Development</a></li>
<li><a href="../roadmap.html">Roadmap</a></li>
<li><a href="../about.html">About us</a></li>
<li class="divider"></li>
<script>if (VERSION_SUBDIR != "stable") document.write('<li><a href="http://scikit-learn.org/stable/documentation.html">Stable version</a></li>')</script>
<script>if (VERSION_SUBDIR != "dev") document.write('<li><a href="http://scikit-learn.org/dev/documentation.html">Development version</a></li>')</script>
<li><a href="http://scikit-learn.org/dev/versions.html">All available versions</a></li>
<li><a href="../_downloads/scikit-learn-docs.pdf">PDF documentation</a></li>
</ul>
</div>
</li>
<li><a href="../auto_examples/index.html">Examples</a></li>
</ul>
<a href="javascript:void(0);" onclick="showMenu()">
<div class="nav-icon">
<div class="hamburger-line"></div>
<div class="hamburger-line"></div>
<div class="hamburger-line"></div>
</div>
</a>
<div class="search_form">
<div class="gcse-search" id="cse" style="width: 100%;"></div>
</div>
</div> <!-- end navbar --></div>
</div>
<!-- GitHub "fork me" ribbon -->
<a href="https://github.com/scikit-learn/scikit-learn">
<img class="fork-me"
style="position: absolute; top: 0; right: 0; border: 0;"
src="../_static/img/forkme.png"
alt="Fork me on GitHub" />
</a>
<div class="content-wrapper">
<div class="sphinxsidebar">
<div class="sphinxsidebarwrapper">
<div class="rel">
<div class="rellink">
<a href="utilities.html"
accesskey="P">Previous
<br/>
<span class="smallrellink">
Utilities for...
</span>
<span class="hiddenrellink">
Utilities for Developers
</span>
</a>
</div>
<div class="spacer">
</div>
<div class="rellink">
<a href="advanced_installation.html"
accesskey="N">Next
<br/>
<span class="smallrellink">
Advanced inst...
</span>
<span class="hiddenrellink">
Advanced installation instructions
</span>
</a>
</div>
<!-- Ad a link to the 'up' page -->
<div class="spacer">
</div>
<div class="rellink">
<a href="index.html">
Up
<br/>
<span class="smallrellink">
Developer’s Guide
</span>
<span class="hiddenrellink">
Developer’s Guide
</span>
</a>
</div>
</div>
<p class="doc-version"><b>scikit-learn v0.21.3</b><br/>
<a href="http://scikit-learn.org/dev/versions.html">Other versions</a></p>
<p class="citing">Please <b><a href="../about.html#citing-scikit-learn" style="font-size: 110%;">cite us </a></b>if you use the software.</p>
<ul>
<li><a class="reference internal" href="#">How to optimize for speed</a><ul>
<li><a class="reference internal" href="#python-cython-or-c-c">Python, Cython or C/C++?</a></li>
<li><a class="reference internal" href="#profiling-python-code">Profiling Python code</a></li>
<li><a class="reference internal" href="#memory-usage-profiling">Memory usage profiling</a></li>
<li><a class="reference internal" href="#performance-tips-for-the-cython-developer">Performance tips for the Cython developer</a><ul>
<li><a class="reference internal" href="#using-openmp">Using OpenMP</a></li>
</ul>
</li>
<li><a class="reference internal" href="#profiling-compiled-extensions">Profiling compiled extensions</a><ul>
<li><a class="reference internal" href="#using-yep-and-gperftools">Using yep and gperftools</a></li>
<li><a class="reference internal" href="#using-gprof">Using gprof</a></li>
<li><a class="reference internal" href="#using-valgrind-callgrind-kcachegrind">Using valgrind / callgrind / kcachegrind</a><ul>
<li><a class="reference internal" href="#kcachegrind">kcachegrind</a></li>
</ul>
</li>
</ul>
</li>
<li><a class="reference internal" href="#multi-core-parallelism-using-joblib-parallel">Multi-core parallelism using <code class="docutils literal"><span class="pre">joblib.Parallel</span></code></a></li>
<li><a class="reference internal" href="#a-simple-algorithmic-trick-warm-restarts">A simple algorithmic trick: warm restarts</a></li>
</ul>
</li>
</ul>
</div>
</div>
<input type="checkbox" id="nav-trigger" class="nav-trigger" checked />
<label for="nav-trigger"></label>
<div class="content">
<div class="documentwrapper">
<div class="bodywrapper">
<div class="body" role="main">
<div class="section" id="how-to-optimize-for-speed">
<span id="performance-howto"></span><h1>How to optimize for speed<a class="headerlink" href="#how-to-optimize-for-speed" title="Permalink to this headline">¶</a></h1>
<p>The following gives some practical guidelines to help you write efficient
code for the scikit-learn project.</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p>While it is always useful to profile your code so as to <strong>check
performance assumptions</strong>, it is also highly recommended
to <strong>review the literature</strong> to ensure that the implemented algorithm
is the state of the art for the task before investing into costly
implementation optimization.</p>
<p>Times and times, hours of efforts invested in optimizing complicated
implementation details have been rendered irrelevant by the subsequent
discovery of simple <strong>algorithmic tricks</strong>, or by using another algorithm
altogether that is better suited to the problem.</p>
<p class="last">The section <a class="reference internal" href="#warm-restarts"><span class="std std-ref">A simple algorithmic trick: warm restarts</span></a> gives an example of such a trick.</p>
</div>
<div class="section" id="python-cython-or-c-c">
<h2>Python, Cython or C/C++?<a class="headerlink" href="#python-cython-or-c-c" title="Permalink to this headline">¶</a></h2>
<p>In general, the scikit-learn project emphasizes the <strong>readability</strong> of
the source code to make it easy for the project users to dive into the
source code so as to understand how the algorithm behaves on their data
but also for ease of maintainability (by the developers).</p>
<p>When implementing a new algorithm is thus recommended to <strong>start
implementing it in Python using Numpy and Scipy</strong> by taking care of avoiding
looping code using the vectorized idioms of those libraries. In practice
this means trying to <strong>replace any nested for loops by calls to equivalent
Numpy array methods</strong>. The goal is to avoid the CPU wasting time in the
Python interpreter rather than crunching numbers to fit your statistical
model. It’s generally a good idea to consider NumPy and SciPy performance tips:
<a class="reference external" href="https://scipy.github.io/old-wiki/pages/PerformanceTips">https://scipy.github.io/old-wiki/pages/PerformanceTips</a></p>
<p>Sometimes however an algorithm cannot be expressed efficiently in simple
vectorized Numpy code. In this case, the recommended strategy is the
following:</p>
<blockquote>
<div><ol class="arabic simple">
<li><strong>Profile</strong> the Python implementation to find the main bottleneck and
isolate it in a <strong>dedicated module level function</strong>. This function
will be reimplemented as a compiled extension module.</li>
<li>If there exists a well maintained BSD or MIT <strong>C/C++</strong> implementation
of the same algorithm that is not too big, you can write a
<strong>Cython wrapper</strong> for it and include a copy of the source code
of the library in the scikit-learn source tree: this strategy is
used for the classes <a class="reference internal" href="../modules/generated/sklearn.svm.LinearSVC.html#sklearn.svm.LinearSVC" title="sklearn.svm.LinearSVC"><code class="xref py py-class docutils literal"><span class="pre">svm.LinearSVC</span></code></a>, <a class="reference internal" href="../modules/generated/sklearn.svm.SVC.html#sklearn.svm.SVC" title="sklearn.svm.SVC"><code class="xref py py-class docutils literal"><span class="pre">svm.SVC</span></code></a> and
<a class="reference internal" href="../modules/generated/sklearn.linear_model.LogisticRegression.html#sklearn.linear_model.LogisticRegression" title="sklearn.linear_model.LogisticRegression"><code class="xref py py-class docutils literal"><span class="pre">linear_model.LogisticRegression</span></code></a> (wrappers for liblinear
and libsvm).</li>
<li>Otherwise, write an optimized version of your Python function using
<strong>Cython</strong> directly. This strategy is used
for the <a class="reference internal" href="../modules/generated/sklearn.linear_model.ElasticNet.html#sklearn.linear_model.ElasticNet" title="sklearn.linear_model.ElasticNet"><code class="xref py py-class docutils literal"><span class="pre">linear_model.ElasticNet</span></code></a> and
<a class="reference internal" href="../modules/generated/sklearn.linear_model.SGDClassifier.html#sklearn.linear_model.SGDClassifier" title="sklearn.linear_model.SGDClassifier"><code class="xref py py-class docutils literal"><span class="pre">linear_model.SGDClassifier</span></code></a> classes for instance.</li>
<li><strong>Move the Python version of the function in the tests</strong> and use
it to check that the results of the compiled extension are consistent
with the gold standard, easy to debug Python version.</li>
<li>Once the code is optimized (not simple bottleneck spottable by
profiling), check whether it is possible to have <strong>coarse grained
parallelism</strong> that is amenable to <strong>multi-processing</strong> by using the
<code class="docutils literal"><span class="pre">joblib.Parallel</span></code> class.</li>
</ol>
</div></blockquote>
<p>When using Cython, use either</p>
<blockquote>
<div>$ python setup.py build_ext -i
$ python setup.py install</div></blockquote>
<p>to generate C files. You are responsible for adding .c/.cpp extensions along
with build parameters in each submodule <code class="docutils literal"><span class="pre">setup.py</span></code>.</p>
<p>C/C++ generated files are embedded in distributed stable packages. The goal is
to make it possible to install scikit-learn stable version
on any machine with Python, Numpy, Scipy and C/C++ compiler.</p>
</div>
<div class="section" id="profiling-python-code">
<span id="id1"></span><h2>Profiling Python code<a class="headerlink" href="#profiling-python-code" title="Permalink to this headline">¶</a></h2>
<p>In order to profile Python code we recommend to write a script that
loads and prepare you data and then use the IPython integrated profiler
for interactively exploring the relevant part for the code.</p>
<p>Suppose we want to profile the Non Negative Matrix Factorization module
of scikit-learn. Let us setup a new IPython session and load the digits
dataset and as in the <a class="reference internal" href="../auto_examples/classification/plot_digits_classification.html#sphx-glr-auto-examples-classification-plot-digits-classification-py"><span class="std std-ref">Recognizing hand-written digits</span></a> example:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">In</span> <span class="p">[</span><span class="mi">1</span><span class="p">]:</span> <span class="kn">from</span> <span class="nn">sklearn.decomposition</span> <span class="k">import</span> <span class="n">NMF</span>
<span class="n">In</span> <span class="p">[</span><span class="mi">2</span><span class="p">]:</span> <span class="kn">from</span> <span class="nn">sklearn.datasets</span> <span class="k">import</span> <span class="n">load_digits</span>
<span class="n">In</span> <span class="p">[</span><span class="mi">3</span><span class="p">]:</span> <span class="n">X</span> <span class="o">=</span> <span class="n">load_digits</span><span class="p">()</span><span class="o">.</span><span class="n">data</span>
</pre></div>
</div>
<p>Before starting the profiling session and engaging in tentative
optimization iterations, it is important to measure the total execution
time of the function we want to optimize without any kind of profiler
overhead and save it somewhere for later reference:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">In</span> <span class="p">[</span><span class="mi">4</span><span class="p">]:</span> <span class="o">%</span><span class="n">timeit</span> <span class="n">NMF</span><span class="p">(</span><span class="n">n_components</span><span class="o">=</span><span class="mi">16</span><span class="p">,</span> <span class="n">tol</span><span class="o">=</span><span class="mf">1e-2</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X</span><span class="p">)</span>
<span class="mi">1</span> <span class="n">loops</span><span class="p">,</span> <span class="n">best</span> <span class="n">of</span> <span class="mi">3</span><span class="p">:</span> <span class="mf">1.7</span> <span class="n">s</span> <span class="n">per</span> <span class="n">loop</span>
</pre></div>
</div>
<p>To have a look at the overall performance profile using the <code class="docutils literal"><span class="pre">%prun</span></code>
magic command:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">In</span> <span class="p">[</span><span class="mi">5</span><span class="p">]:</span> <span class="o">%</span><span class="n">prun</span> <span class="o">-</span><span class="n">l</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span> <span class="n">NMF</span><span class="p">(</span><span class="n">n_components</span><span class="o">=</span><span class="mi">16</span><span class="p">,</span> <span class="n">tol</span><span class="o">=</span><span class="mf">1e-2</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X</span><span class="p">)</span>
<span class="mi">14496</span> <span class="n">function</span> <span class="n">calls</span> <span class="ow">in</span> <span class="mf">1.682</span> <span class="n">CPU</span> <span class="n">seconds</span>
<span class="n">Ordered</span> <span class="n">by</span><span class="p">:</span> <span class="n">internal</span> <span class="n">time</span>
<span class="n">List</span> <span class="n">reduced</span> <span class="kn">from</span> <span class="mi">90</span> <span class="n">to</span> <span class="mi">9</span> <span class="n">due</span> <span class="n">to</span> <span class="n">restriction</span> <span class="o"><</span><span class="s1">'nmf.py'</span><span class="o">></span>
<span class="n">ncalls</span> <span class="n">tottime</span> <span class="n">percall</span> <span class="n">cumtime</span> <span class="n">percall</span> <span class="n">filename</span><span class="p">:</span><span class="n">lineno</span><span class="p">(</span><span class="n">function</span><span class="p">)</span>
<span class="mi">36</span> <span class="mf">0.609</span> <span class="mf">0.017</span> <span class="mf">1.499</span> <span class="mf">0.042</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">151</span><span class="p">(</span><span class="n">_nls_subproblem</span><span class="p">)</span>
<span class="mi">1263</span> <span class="mf">0.157</span> <span class="mf">0.000</span> <span class="mf">0.157</span> <span class="mf">0.000</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">18</span><span class="p">(</span><span class="n">_pos</span><span class="p">)</span>
<span class="mi">1</span> <span class="mf">0.053</span> <span class="mf">0.053</span> <span class="mf">1.681</span> <span class="mf">1.681</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">352</span><span class="p">(</span><span class="n">fit_transform</span><span class="p">)</span>
<span class="mi">673</span> <span class="mf">0.008</span> <span class="mf">0.000</span> <span class="mf">0.057</span> <span class="mf">0.000</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">28</span><span class="p">(</span><span class="n">norm</span><span class="p">)</span>
<span class="mi">1</span> <span class="mf">0.006</span> <span class="mf">0.006</span> <span class="mf">0.047</span> <span class="mf">0.047</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">42</span><span class="p">(</span><span class="n">_initialize_nmf</span><span class="p">)</span>
<span class="mi">36</span> <span class="mf">0.001</span> <span class="mf">0.000</span> <span class="mf">0.010</span> <span class="mf">0.000</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">36</span><span class="p">(</span><span class="n">_sparseness</span><span class="p">)</span>
<span class="mi">30</span> <span class="mf">0.001</span> <span class="mf">0.000</span> <span class="mf">0.001</span> <span class="mf">0.000</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">23</span><span class="p">(</span><span class="n">_neg</span><span class="p">)</span>
<span class="mi">1</span> <span class="mf">0.000</span> <span class="mf">0.000</span> <span class="mf">0.000</span> <span class="mf">0.000</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">337</span><span class="p">(</span><span class="fm">__init__</span><span class="p">)</span>
<span class="mi">1</span> <span class="mf">0.000</span> <span class="mf">0.000</span> <span class="mf">1.681</span> <span class="mf">1.681</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">461</span><span class="p">(</span><span class="n">fit</span><span class="p">)</span>
</pre></div>
</div>
<p>The <code class="docutils literal"><span class="pre">tottime</span></code> column is the most interesting: it gives to total time spent
executing the code of a given function ignoring the time spent in executing the
sub-functions. The real total time (local code + sub-function calls) is given by
the <code class="docutils literal"><span class="pre">cumtime</span></code> column.</p>
<p>Note the use of the <code class="docutils literal"><span class="pre">-l</span> <span class="pre">nmf.py</span></code> that restricts the output to lines that
contains the “nmf.py” string. This is useful to have a quick look at the hotspot
of the nmf Python module it-self ignoring anything else.</p>
<p>Here is the beginning of the output of the same command without the <code class="docutils literal"><span class="pre">-l</span> <span class="pre">nmf.py</span></code>
filter:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">In</span> <span class="p">[</span><span class="mi">5</span><span class="p">]</span> <span class="o">%</span><span class="n">prun</span> <span class="n">NMF</span><span class="p">(</span><span class="n">n_components</span><span class="o">=</span><span class="mi">16</span><span class="p">,</span> <span class="n">tol</span><span class="o">=</span><span class="mf">1e-2</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X</span><span class="p">)</span>
<span class="mi">16159</span> <span class="n">function</span> <span class="n">calls</span> <span class="ow">in</span> <span class="mf">1.840</span> <span class="n">CPU</span> <span class="n">seconds</span>
<span class="n">Ordered</span> <span class="n">by</span><span class="p">:</span> <span class="n">internal</span> <span class="n">time</span>
<span class="n">ncalls</span> <span class="n">tottime</span> <span class="n">percall</span> <span class="n">cumtime</span> <span class="n">percall</span> <span class="n">filename</span><span class="p">:</span><span class="n">lineno</span><span class="p">(</span><span class="n">function</span><span class="p">)</span>
<span class="mi">2833</span> <span class="mf">0.653</span> <span class="mf">0.000</span> <span class="mf">0.653</span> <span class="mf">0.000</span> <span class="p">{</span><span class="n">numpy</span><span class="o">.</span><span class="n">core</span><span class="o">.</span><span class="n">_dotblas</span><span class="o">.</span><span class="n">dot</span><span class="p">}</span>
<span class="mi">46</span> <span class="mf">0.651</span> <span class="mf">0.014</span> <span class="mf">1.636</span> <span class="mf">0.036</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">151</span><span class="p">(</span><span class="n">_nls_subproblem</span><span class="p">)</span>
<span class="mi">1397</span> <span class="mf">0.171</span> <span class="mf">0.000</span> <span class="mf">0.171</span> <span class="mf">0.000</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">18</span><span class="p">(</span><span class="n">_pos</span><span class="p">)</span>
<span class="mi">2780</span> <span class="mf">0.167</span> <span class="mf">0.000</span> <span class="mf">0.167</span> <span class="mf">0.000</span> <span class="p">{</span><span class="n">method</span> <span class="s1">'sum'</span> <span class="n">of</span> <span class="s1">'numpy.ndarray'</span> <span class="n">objects</span><span class="p">}</span>
<span class="mi">1</span> <span class="mf">0.064</span> <span class="mf">0.064</span> <span class="mf">1.840</span> <span class="mf">1.840</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">352</span><span class="p">(</span><span class="n">fit_transform</span><span class="p">)</span>
<span class="mi">1542</span> <span class="mf">0.043</span> <span class="mf">0.000</span> <span class="mf">0.043</span> <span class="mf">0.000</span> <span class="p">{</span><span class="n">method</span> <span class="s1">'flatten'</span> <span class="n">of</span> <span class="s1">'numpy.ndarray'</span> <span class="n">objects</span><span class="p">}</span>
<span class="mi">337</span> <span class="mf">0.019</span> <span class="mf">0.000</span> <span class="mf">0.019</span> <span class="mf">0.000</span> <span class="p">{</span><span class="n">method</span> <span class="s1">'all'</span> <span class="n">of</span> <span class="s1">'numpy.ndarray'</span> <span class="n">objects</span><span class="p">}</span>
<span class="mi">2734</span> <span class="mf">0.011</span> <span class="mf">0.000</span> <span class="mf">0.181</span> <span class="mf">0.000</span> <span class="n">fromnumeric</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">1185</span><span class="p">(</span><span class="nb">sum</span><span class="p">)</span>
<span class="mi">2</span> <span class="mf">0.010</span> <span class="mf">0.005</span> <span class="mf">0.010</span> <span class="mf">0.005</span> <span class="p">{</span><span class="n">numpy</span><span class="o">.</span><span class="n">linalg</span><span class="o">.</span><span class="n">lapack_lite</span><span class="o">.</span><span class="n">dgesdd</span><span class="p">}</span>
<span class="mi">748</span> <span class="mf">0.009</span> <span class="mf">0.000</span> <span class="mf">0.065</span> <span class="mf">0.000</span> <span class="n">nmf</span><span class="o">.</span><span class="n">py</span><span class="p">:</span><span class="mi">28</span><span class="p">(</span><span class="n">norm</span><span class="p">)</span>
<span class="o">...</span>
</pre></div>
</div>
<p>The above results show that the execution is largely dominated by
dot products operations (delegated to blas). Hence there is probably
no huge gain to expect by rewriting this code in Cython or C/C++: in
this case out of the 1.7s total execution time, almost 0.7s are spent
in compiled code we can consider optimal. By rewriting the rest of the
Python code and assuming we could achieve a 1000% boost on this portion
(which is highly unlikely given the shallowness of the Python loops),
we would not gain more than a 2.4x speed-up globally.</p>
<p>Hence major improvements can only be achieved by <strong>algorithmic
improvements</strong> in this particular example (e.g. trying to find operation
that are both costly and useless to avoid computing then rather than
trying to optimize their implementation).</p>
<p>It is however still interesting to check what’s happening inside the
<code class="docutils literal"><span class="pre">_nls_subproblem</span></code> function which is the hotspot if we only consider
Python code: it takes around 100% of the accumulated time of the module. In
order to better understand the profile of this specific function, let
us install <code class="docutils literal"><span class="pre">line_profiler</span></code> and wire it to IPython:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span>$ pip install line_profiler
</pre></div>
</div>
<ul>
<li><p class="first"><strong>Under IPython 0.13+</strong>, first create a configuration profile:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span>$ ipython profile create
</pre></div>
</div>
<p>Then register the line_profiler extension in
<code class="docutils literal"><span class="pre">~/.ipython/profile_default/ipython_config.py</span></code>:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">c</span><span class="o">.</span><span class="n">TerminalIPythonApp</span><span class="o">.</span><span class="n">extensions</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="s1">'line_profiler'</span><span class="p">)</span>
<span class="n">c</span><span class="o">.</span><span class="n">InteractiveShellApp</span><span class="o">.</span><span class="n">extensions</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="s1">'line_profiler'</span><span class="p">)</span>
</pre></div>
</div>
<p>This will register the <code class="docutils literal"><span class="pre">%lprun</span></code> magic command in the IPython terminal
application and the other frontends such as qtconsole and notebook.</p>
</li>
</ul>
<p>Now restart IPython and let us use this new toy:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">In</span> <span class="p">[</span><span class="mi">1</span><span class="p">]:</span> <span class="kn">from</span> <span class="nn">sklearn.datasets</span> <span class="k">import</span> <span class="n">load_digits</span>
<span class="n">In</span> <span class="p">[</span><span class="mi">2</span><span class="p">]:</span> <span class="kn">from</span> <span class="nn">sklearn.decomposition.nmf</span> <span class="k">import</span> <span class="n">_nls_subproblem</span><span class="p">,</span> <span class="n">NMF</span>
<span class="n">In</span> <span class="p">[</span><span class="mi">3</span><span class="p">]:</span> <span class="n">X</span> <span class="o">=</span> <span class="n">load_digits</span><span class="p">()</span><span class="o">.</span><span class="n">data</span>
<span class="n">In</span> <span class="p">[</span><span class="mi">4</span><span class="p">]:</span> <span class="o">%</span><span class="n">lprun</span> <span class="o">-</span><span class="n">f</span> <span class="n">_nls_subproblem</span> <span class="n">NMF</span><span class="p">(</span><span class="n">n_components</span><span class="o">=</span><span class="mi">16</span><span class="p">,</span> <span class="n">tol</span><span class="o">=</span><span class="mf">1e-2</span><span class="p">)</span><span class="o">.</span><span class="n">fit</span><span class="p">(</span><span class="n">X</span><span class="p">)</span>
<span class="n">Timer</span> <span class="n">unit</span><span class="p">:</span> <span class="mf">1e-06</span> <span class="n">s</span>
<span class="n">File</span><span class="p">:</span> <span class="n">sklearn</span><span class="o">/</span><span class="n">decomposition</span><span class="o">/</span><span class="n">nmf</span><span class="o">.</span><span class="n">py</span>
<span class="n">Function</span><span class="p">:</span> <span class="n">_nls_subproblem</span> <span class="n">at</span> <span class="n">line</span> <span class="mi">137</span>
<span class="n">Total</span> <span class="n">time</span><span class="p">:</span> <span class="mf">1.73153</span> <span class="n">s</span>
<span class="n">Line</span> <span class="c1"># Hits Time Per Hit % Time Line Contents</span>
<span class="o">==============================================================</span>
<span class="mi">137</span> <span class="k">def</span> <span class="nf">_nls_subproblem</span><span class="p">(</span><span class="n">V</span><span class="p">,</span> <span class="n">W</span><span class="p">,</span> <span class="n">H_init</span><span class="p">,</span> <span class="n">tol</span><span class="p">,</span> <span class="n">max_iter</span><span class="p">):</span>
<span class="mi">138</span> <span class="s2">"""Non-negative least square solver</span>
<span class="s2"> ...</span>
<span class="s2"> 170 """</span>
<span class="mi">171</span> <span class="mi">48</span> <span class="mi">5863</span> <span class="mf">122.1</span> <span class="mf">0.3</span> <span class="k">if</span> <span class="p">(</span><span class="n">H_init</span> <span class="o"><</span> <span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">any</span><span class="p">():</span>
<span class="mi">172</span> <span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span><span class="s2">"Negative values in H_init passed to NLS solver."</span><span class="p">)</span>
<span class="mi">173</span>
<span class="mi">174</span> <span class="mi">48</span> <span class="mi">139</span> <span class="mf">2.9</span> <span class="mf">0.0</span> <span class="n">H</span> <span class="o">=</span> <span class="n">H_init</span>
<span class="mi">175</span> <span class="mi">48</span> <span class="mi">112141</span> <span class="mf">2336.3</span> <span class="mf">5.8</span> <span class="n">WtV</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">W</span><span class="o">.</span><span class="n">T</span><span class="p">,</span> <span class="n">V</span><span class="p">)</span>
<span class="mi">176</span> <span class="mi">48</span> <span class="mi">16144</span> <span class="mf">336.3</span> <span class="mf">0.8</span> <span class="n">WtW</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">W</span><span class="o">.</span><span class="n">T</span><span class="p">,</span> <span class="n">W</span><span class="p">)</span>
<span class="mi">177</span>
<span class="mi">178</span> <span class="c1"># values justified in the paper</span>
<span class="mi">179</span> <span class="mi">48</span> <span class="mi">144</span> <span class="mf">3.0</span> <span class="mf">0.0</span> <span class="n">alpha</span> <span class="o">=</span> <span class="mi">1</span>
<span class="mi">180</span> <span class="mi">48</span> <span class="mi">113</span> <span class="mf">2.4</span> <span class="mf">0.0</span> <span class="n">beta</span> <span class="o">=</span> <span class="mf">0.1</span>
<span class="mi">181</span> <span class="mi">638</span> <span class="mi">1880</span> <span class="mf">2.9</span> <span class="mf">0.1</span> <span class="k">for</span> <span class="n">n_iter</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">max_iter</span> <span class="o">+</span> <span class="mi">1</span><span class="p">):</span>
<span class="mi">182</span> <span class="mi">638</span> <span class="mi">195133</span> <span class="mf">305.9</span> <span class="mf">10.2</span> <span class="n">grad</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">WtW</span><span class="p">,</span> <span class="n">H</span><span class="p">)</span> <span class="o">-</span> <span class="n">WtV</span>
<span class="mi">183</span> <span class="mi">638</span> <span class="mi">495761</span> <span class="mf">777.1</span> <span class="mf">25.9</span> <span class="n">proj_gradient</span> <span class="o">=</span> <span class="n">norm</span><span class="p">(</span><span class="n">grad</span><span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">logical_or</span><span class="p">(</span><span class="n">grad</span> <span class="o"><</span> <span class="mi">0</span><span class="p">,</span> <span class="n">H</span> <span class="o">></span> <span class="mi">0</span><span class="p">)])</span>
<span class="mi">184</span> <span class="mi">638</span> <span class="mi">2449</span> <span class="mf">3.8</span> <span class="mf">0.1</span> <span class="k">if</span> <span class="n">proj_gradient</span> <span class="o"><</span> <span class="n">tol</span><span class="p">:</span>
<span class="mi">185</span> <span class="mi">48</span> <span class="mi">130</span> <span class="mf">2.7</span> <span class="mf">0.0</span> <span class="k">break</span>
<span class="mi">186</span>
<span class="mi">187</span> <span class="mi">1474</span> <span class="mi">4474</span> <span class="mf">3.0</span> <span class="mf">0.2</span> <span class="k">for</span> <span class="n">inner_iter</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">20</span><span class="p">):</span>
<span class="mi">188</span> <span class="mi">1474</span> <span class="mi">83833</span> <span class="mf">56.9</span> <span class="mf">4.4</span> <span class="n">Hn</span> <span class="o">=</span> <span class="n">H</span> <span class="o">-</span> <span class="n">alpha</span> <span class="o">*</span> <span class="n">grad</span>
<span class="mi">189</span> <span class="c1"># Hn = np.where(Hn > 0, Hn, 0)</span>
<span class="mi">190</span> <span class="mi">1474</span> <span class="mi">194239</span> <span class="mf">131.8</span> <span class="mf">10.1</span> <span class="n">Hn</span> <span class="o">=</span> <span class="n">_pos</span><span class="p">(</span><span class="n">Hn</span><span class="p">)</span>
<span class="mi">191</span> <span class="mi">1474</span> <span class="mi">48858</span> <span class="mf">33.1</span> <span class="mf">2.5</span> <span class="n">d</span> <span class="o">=</span> <span class="n">Hn</span> <span class="o">-</span> <span class="n">H</span>
<span class="mi">192</span> <span class="mi">1474</span> <span class="mi">150407</span> <span class="mf">102.0</span> <span class="mf">7.8</span> <span class="n">gradd</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="n">grad</span> <span class="o">*</span> <span class="n">d</span><span class="p">)</span>
<span class="mi">193</span> <span class="mi">1474</span> <span class="mi">515390</span> <span class="mf">349.7</span> <span class="mf">26.9</span> <span class="n">dQd</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">WtW</span><span class="p">,</span> <span class="n">d</span><span class="p">)</span> <span class="o">*</span> <span class="n">d</span><span class="p">)</span>
<span class="o">...</span>
</pre></div>
</div>
<p>By looking at the top values of the <code class="docutils literal"><span class="pre">%</span> <span class="pre">Time</span></code> column it is really easy to
pin-point the most expensive expressions that would deserve additional care.</p>
</div>
<div class="section" id="memory-usage-profiling">
<h2>Memory usage profiling<a class="headerlink" href="#memory-usage-profiling" title="Permalink to this headline">¶</a></h2>
<p>You can analyze in detail the memory usage of any Python code with the help of
<a class="reference external" href="https://pypi.org/project/memory_profiler/">memory_profiler</a>. First,
install the latest version:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span>$ pip install -U memory_profiler
</pre></div>
</div>
<p>Then, setup the magics in a manner similar to <code class="docutils literal"><span class="pre">line_profiler</span></code>.</p>
<ul>
<li><p class="first"><strong>Under IPython 0.11+</strong>, first create a configuration profile:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span>$ ipython profile create
</pre></div>
</div>
<p>Then register the extension in
<code class="docutils literal"><span class="pre">~/.ipython/profile_default/ipython_config.py</span></code>
alongside the line profiler:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">c</span><span class="o">.</span><span class="n">TerminalIPythonApp</span><span class="o">.</span><span class="n">extensions</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="s1">'memory_profiler'</span><span class="p">)</span>
<span class="n">c</span><span class="o">.</span><span class="n">InteractiveShellApp</span><span class="o">.</span><span class="n">extensions</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="s1">'memory_profiler'</span><span class="p">)</span>
</pre></div>
</div>
<p>This will register the <code class="docutils literal"><span class="pre">%memit</span></code> and <code class="docutils literal"><span class="pre">%mprun</span></code> magic commands in the
IPython terminal application and the other frontends such as qtconsole and
notebook.</p>
</li>
</ul>
<p><code class="docutils literal"><span class="pre">%mprun</span></code> is useful to examine, line-by-line, the memory usage of key
functions in your program. It is very similar to <code class="docutils literal"><span class="pre">%lprun</span></code>, discussed in the
previous section. For example, from the <code class="docutils literal"><span class="pre">memory_profiler</span></code> <code class="docutils literal"><span class="pre">examples</span></code>
directory:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">In</span> <span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="kn">from</span> <span class="nn">example</span> <span class="k">import</span> <span class="n">my_func</span>
<span class="n">In</span> <span class="p">[</span><span class="mi">2</span><span class="p">]</span> <span class="o">%</span><span class="n">mprun</span> <span class="o">-</span><span class="n">f</span> <span class="n">my_func</span> <span class="n">my_func</span><span class="p">()</span>
<span class="n">Filename</span><span class="p">:</span> <span class="n">example</span><span class="o">.</span><span class="n">py</span>
<span class="n">Line</span> <span class="c1"># Mem usage Increment Line Contents</span>
<span class="o">==============================================</span>
<span class="mi">3</span> <span class="nd">@profile</span>
<span class="mi">4</span> <span class="mf">5.97</span> <span class="n">MB</span> <span class="mf">0.00</span> <span class="n">MB</span> <span class="k">def</span> <span class="nf">my_func</span><span class="p">():</span>
<span class="mi">5</span> <span class="mf">13.61</span> <span class="n">MB</span> <span class="mf">7.64</span> <span class="n">MB</span> <span class="n">a</span> <span class="o">=</span> <span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="o">*</span> <span class="p">(</span><span class="mi">10</span> <span class="o">**</span> <span class="mi">6</span><span class="p">)</span>
<span class="mi">6</span> <span class="mf">166.20</span> <span class="n">MB</span> <span class="mf">152.59</span> <span class="n">MB</span> <span class="n">b</span> <span class="o">=</span> <span class="p">[</span><span class="mi">2</span><span class="p">]</span> <span class="o">*</span> <span class="p">(</span><span class="mi">2</span> <span class="o">*</span> <span class="mi">10</span> <span class="o">**</span> <span class="mi">7</span><span class="p">)</span>
<span class="mi">7</span> <span class="mf">13.61</span> <span class="n">MB</span> <span class="o">-</span><span class="mf">152.59</span> <span class="n">MB</span> <span class="k">del</span> <span class="n">b</span>
<span class="mi">8</span> <span class="mf">13.61</span> <span class="n">MB</span> <span class="mf">0.00</span> <span class="n">MB</span> <span class="k">return</span> <span class="n">a</span>
</pre></div>
</div>
<p>Another useful magic that <code class="docutils literal"><span class="pre">memory_profiler</span></code> defines is <code class="docutils literal"><span class="pre">%memit</span></code>, which is
analogous to <code class="docutils literal"><span class="pre">%timeit</span></code>. It can be used as follows:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">In</span> <span class="p">[</span><span class="mi">1</span><span class="p">]:</span> <span class="kn">import</span> <span class="nn">numpy</span> <span class="k">as</span> <span class="nn">np</span>
<span class="n">In</span> <span class="p">[</span><span class="mi">2</span><span class="p">]:</span> <span class="o">%</span><span class="n">memit</span> <span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="mf">1e7</span><span class="p">)</span>
<span class="n">maximum</span> <span class="n">of</span> <span class="mi">3</span><span class="p">:</span> <span class="mf">76.402344</span> <span class="n">MB</span> <span class="n">per</span> <span class="n">loop</span>
</pre></div>
</div>
<p>For more details, see the docstrings of the magics, using <code class="docutils literal"><span class="pre">%memit?</span></code> and
<code class="docutils literal"><span class="pre">%mprun?</span></code>.</p>
</div>
<div class="section" id="performance-tips-for-the-cython-developer">
<h2>Performance tips for the Cython developer<a class="headerlink" href="#performance-tips-for-the-cython-developer" title="Permalink to this headline">¶</a></h2>
<p>If profiling of the Python code reveals that the Python interpreter
overhead is larger by one order of magnitude or more than the cost of the
actual numerical computation (e.g. <code class="docutils literal"><span class="pre">for</span></code> loops over vector components,
nested evaluation of conditional expression, scalar arithmetic…), it
is probably adequate to extract the hotspot portion of the code as a
standalone function in a <code class="docutils literal"><span class="pre">.pyx</span></code> file, add static type declarations and
then use Cython to generate a C program suitable to be compiled as a
Python extension module.</p>
<p>The official documentation available at <a class="reference external" href="http://docs.cython.org/">http://docs.cython.org/</a> contains
a tutorial and reference guide for developing such a module. In the
following we will just highlight a couple of tricks that we found
important in practice on the existing cython codebase in the scikit-learn
project.</p>
<p>TODO: html report, type declarations, bound checks, division by zero checks,
memory alignment, direct blas calls…</p>
<ul class="simple">
<li><a class="reference external" href="https://www.youtube.com/watch?v=gMvkiQ-gOW8">https://www.youtube.com/watch?v=gMvkiQ-gOW8</a></li>
<li><a class="reference external" href="http://conference.scipy.org/proceedings/SciPy2009/paper_1/">http://conference.scipy.org/proceedings/SciPy2009/paper_1/</a></li>
<li><a class="reference external" href="http://conference.scipy.org/proceedings/SciPy2009/paper_2/">http://conference.scipy.org/proceedings/SciPy2009/paper_2/</a></li>
</ul>
<div class="section" id="using-openmp">
<h3>Using OpenMP<a class="headerlink" href="#using-openmp" title="Permalink to this headline">¶</a></h3>
<p>Since scikit-learn can be built without OpenMP support, it’s necessary to
protect each direct call to OpenMP. This can be done using the following
syntax:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="c1"># importing OpenMP</span>
<span class="n">IF</span> <span class="n">SKLEARN_OPENMP_SUPPORTED</span><span class="p">:</span>
<span class="n">cimport</span> <span class="n">openmp</span>
<span class="c1"># calling OpenMP</span>
<span class="n">IF</span> <span class="n">SKLEARN_OPENMP_SUPPORTED</span><span class="p">:</span>
<span class="n">max_threads</span> <span class="o">=</span> <span class="n">openmp</span><span class="o">.</span><span class="n">omp_get_max_threads</span><span class="p">()</span>
<span class="n">ELSE</span><span class="p">:</span>
<span class="n">max_threads</span> <span class="o">=</span> <span class="mi">1</span>
</pre></div>
</div>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">Protecting the parallel loop, <code class="docutils literal"><span class="pre">prange</span></code>, is already done by cython.</p>
</div>
</div>
</div>
<div class="section" id="profiling-compiled-extensions">
<span id="profiling-compiled-extension"></span><h2>Profiling compiled extensions<a class="headerlink" href="#profiling-compiled-extensions" title="Permalink to this headline">¶</a></h2>
<p>When working with compiled extensions (written in C/C++ with a wrapper or
directly as Cython extension), the default Python profiler is useless:
we need a dedicated tool to introspect what’s happening inside the
compiled extension it-self.</p>
<div class="section" id="using-yep-and-gperftools">
<h3>Using yep and gperftools<a class="headerlink" href="#using-yep-and-gperftools" title="Permalink to this headline">¶</a></h3>
<p>Easy profiling without special compilation options use yep:</p>
<ul class="simple">
<li><a class="reference external" href="https://pypi.org/project/yep/">https://pypi.org/project/yep/</a></li>
<li><a class="reference external" href="http://fa.bianp.net/blog/2011/a-profiler-for-python-extensions">http://fa.bianp.net/blog/2011/a-profiler-for-python-extensions</a></li>
</ul>
</div>
<div class="section" id="using-gprof">
<h3>Using gprof<a class="headerlink" href="#using-gprof" title="Permalink to this headline">¶</a></h3>
<p>In order to profile compiled Python extensions one could use <code class="docutils literal"><span class="pre">gprof</span></code>
after having recompiled the project with <code class="docutils literal"><span class="pre">gcc</span> <span class="pre">-pg</span></code> and using the
<code class="docutils literal"><span class="pre">python-dbg</span></code> variant of the interpreter on debian / ubuntu: however
this approach requires to also have <code class="docutils literal"><span class="pre">numpy</span></code> and <code class="docutils literal"><span class="pre">scipy</span></code> recompiled
with <code class="docutils literal"><span class="pre">-pg</span></code> which is rather complicated to get working.</p>
<p>Fortunately there exist two alternative profilers that don’t require you to
recompile everything.</p>
</div>
<div class="section" id="using-valgrind-callgrind-kcachegrind">
<h3>Using valgrind / callgrind / kcachegrind<a class="headerlink" href="#using-valgrind-callgrind-kcachegrind" title="Permalink to this headline">¶</a></h3>
<div class="section" id="kcachegrind">
<h4>kcachegrind<a class="headerlink" href="#kcachegrind" title="Permalink to this headline">¶</a></h4>
<p><code class="docutils literal"><span class="pre">yep</span></code> can be used to create a profiling report.
<code class="docutils literal"><span class="pre">kcachegrind</span></code> provides a graphical environment to visualize this report:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="c1"># Run yep to profile some python script</span>
<span class="n">python</span> <span class="o">-</span><span class="n">m</span> <span class="n">yep</span> <span class="o">-</span><span class="n">c</span> <span class="n">my_file</span><span class="o">.</span><span class="n">py</span>
<span class="c1"># open my_file.py.callgrin with kcachegrind</span>
<span class="n">kcachegrind</span> <span class="n">my_file</span><span class="o">.</span><span class="n">py</span><span class="o">.</span><span class="n">prof</span>
</pre></div>
</div>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last"><code class="docutils literal"><span class="pre">yep</span></code> can be executed with the argument <code class="docutils literal"><span class="pre">--lines</span></code> or <code class="docutils literal"><span class="pre">-l</span></code> to compile
a profiling report ‘line by line’.</p>
</div>
</div>
</div>
</div>
<div class="section" id="multi-core-parallelism-using-joblib-parallel">
<h2>Multi-core parallelism using <code class="docutils literal"><span class="pre">joblib.Parallel</span></code><a class="headerlink" href="#multi-core-parallelism-using-joblib-parallel" title="Permalink to this headline">¶</a></h2>
<p>TODO: give a simple teaser example here.</p>
<p>Checkout the official joblib documentation:</p>
<ul class="simple">
<li><a class="reference external" href="https://joblib.readthedocs.io">https://joblib.readthedocs.io</a></li>
</ul>
</div>
<div class="section" id="a-simple-algorithmic-trick-warm-restarts">
<span id="warm-restarts"></span><h2>A simple algorithmic trick: warm restarts<a class="headerlink" href="#a-simple-algorithmic-trick-warm-restarts" title="Permalink to this headline">¶</a></h2>
<p>See the glossary entry for <a class="reference external" href="http://scikit-learn.org/dev/glossary.html#term-warm-start">warm_start</a></p>
</div>
</div>
</div>
</div>
</div>
<div class="clearer"></div>
</div>
</div>
<div class="footer">
© 2007 - 2019, scikit-learn developers (BSD License).
<a href="../_sources/developers/performance.rst.txt" rel="nofollow">Show this page source</a>
</div>
<div class="rel">
<div class="buttonPrevious">
<a href="utilities.html">Previous
</a>
</div>
<div class="buttonNext">
<a href="advanced_installation.html">Next
</a>
</div>
</div>
<script>
window.ga=window.ga||function(){(ga.q=ga.q||[]).push(arguments)};ga.l=+new Date;
ga('create', 'UA-22606712-2', 'auto');
ga('set', 'anonymizeIp', true);
ga('send', 'pageview');
</script>
<script async src='https://www.google-analytics.com/analytics.js'></script>
<script>
(function() {
var cx = '016639176250731907682:tjtqbvtvij0';
var gcse = document.createElement('script'); gcse.type = 'text/javascript'; gcse.async = true;
gcse.src = 'https://cse.google.com/cse.js?cx=' + cx;
var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(gcse, s);
})();
</script>
<script src="https://scikit-learn.org/versionwarning.js"></script>
<script src="https://scikit-learn.org/versionwarning.js"></script>
</body>
</html>