How to optimize for speed¶
The following gives some practical guidelines to help you write efficient code for the scikit-learn project.
Note
While it is always useful to profile your code so as to check performance assumptions, it is also highly recommended to review the literature to ensure that the implemented algorithm is the state of the art for the task before investing into costly implementation optimization.
Times and times, hours of efforts invested in optimizing complicated implementation details have been rended irrelevant by the late discovery of simple algorithmic tricks, or by using another algorithm altogether that is better suited to the problem.
The section A sample algorithmic trick: warm restarts for cross validation gives an example of such a trick.
Python, Cython or C/C++?¶
In general, the scikit-learn project emphasizes the readability 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 maintanability (by the developers).
When implementing a new algorithm is thus recommended to start implementing it in Python using Numpy and Scipy by taking care of avoiding looping code using the vectorized idioms of those libraries. In practice this means trying to replace any nested for loops by calls to equivalent Numpy array methods. The goal is to avoid the CPU wasting time in the Python interpreter rather than crunching numbers to fit your statistical model.
Sometimes however an algorithm cannot be expressed efficiently in simple vectorized Numpy code. In this case, the recommended strategy is the following:
- Profile the Python implementation to find the main bottleneck and isolate it in a dedicated module level function. This function will be reimplemented as a compiled extension module.
- If there exists a well maintained BSD or MIT C/C++ implementation of the same algorithm that is not too big, you can write a Cython wrapper 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 svm.LinearSVC, svm.SVC and linear_model.LogisticRegression (wrappers for liblinear and libsvm).
- Otherwise, write an optimized version of your Python function using Cython directly. This strategy is used for the linear_model.ElasticNet and linear_model.SGDClassifier classes for instance.
- Move the Python version of the function in the tests and use it to check that the results of the compiled extension are consistent with the gold standard, easy to debug Python version.
- Once the code is optimized (not simple bottleneck spottable by profiling), check whether it is possible to have coarse grained parallelism that is amenable to multi-processing by using the joblib.Parallel class.
When using Cython, include the generated C source code alongside with the Cython source code. The goal is to make it possible to install the scikit on any machine with Python, Numpy, Scipy and C/C++ compiler.
Profiling Python code¶
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.
Suppose we want to profile the Non Negative Matrix Factorization module of the scikit. Let us setup a new IPython session and load the digits dataset and as in the NMF for digits feature extraction example:
In [1]: from scikits.learn.decomposition import NMF
In [2]: from scikits.learn.datasets import load_digits
In [3]: X = load_digits().data
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:
In [4]: %timeit NMF(n_components=16, tol=1e-2).fit(X)
1 loops, best of 3: 1.7 s per loop
To have have a look at the overall performance profile using the %prun magic command:
In [5]: %prun -l nmf.py NMF(n_components=16, tol=1e-2).fit(X)
14496 function calls in 1.682 CPU seconds
Ordered by: internal time
List reduced from 90 to 9 due to restriction <'nmf.py'>
ncalls tottime percall cumtime percall filename:lineno(function)
36 0.609 0.017 1.499 0.042 nmf.py:151(_nls_subproblem)
1263 0.157 0.000 0.157 0.000 nmf.py:18(_pos)
1 0.053 0.053 1.681 1.681 nmf.py:352(fit_transform)
673 0.008 0.000 0.057 0.000 nmf.py:28(norm)
1 0.006 0.006 0.047 0.047 nmf.py:42(_initialize_nmf)
36 0.001 0.000 0.010 0.000 nmf.py:36(_sparseness)
30 0.001 0.000 0.001 0.000 nmf.py:23(_neg)
1 0.000 0.000 0.000 0.000 nmf.py:337(__init__)
1 0.000 0.000 1.681 1.681 nmf.py:461(fit)
The totime columns 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 cumtime column.
Note the use of the -l nmf.py 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.
Here is the begining of the output of the same command without the -l nmf.py filter:
In [5] %prun NMF(n_components=16, tol=1e-2).fit(X)
16159 function calls in 1.840 CPU seconds
Ordered by: internal time
ncalls tottime percall cumtime percall filename:lineno(function)
2833 0.653 0.000 0.653 0.000 {numpy.core._dotblas.dot}
46 0.651 0.014 1.636 0.036 nmf.py:151(_nls_subproblem)
1397 0.171 0.000 0.171 0.000 nmf.py:18(_pos)
2780 0.167 0.000 0.167 0.000 {method 'sum' of 'numpy.ndarray' objects}
1 0.064 0.064 1.840 1.840 nmf.py:352(fit_transform)
1542 0.043 0.000 0.043 0.000 {method 'flatten' of 'numpy.ndarray' objects}
337 0.019 0.000 0.019 0.000 {method 'all' of 'numpy.ndarray' objects}
2734 0.011 0.000 0.181 0.000 fromnumeric.py:1185(sum)
2 0.010 0.005 0.010 0.005 {numpy.linalg.lapack_lite.dgesdd}
748 0.009 0.000 0.065 0.000 nmf.py:28(norm)
...
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.
Hence major improvements can only be achieved by algorithmic improvements 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).
It is however still interesting to check what’s happening inside the _nls_subproblem function which is the hotspot if we only consider Python code: it takes around 100% of the cumulated time of the module. In order to better understand the profile of this specific function, let us install line-prof and wire it to IPython:
$ pip install line-profiler
$ vim ~/.ipython/ipy_user_conf.py
Ensure the following lines are present:
import IPython.ipapi
ip = IPython.ipapi.get()
Towards the end of the file, define the %lprun magic:
import line_profiler
ip.expose_magic('lprun', line_profiler.magic_lprun)
Now restart IPython and let us use this new toy:
In [1]: from scikits.learn.datasets import load_digits
In [2]: from scikits.learn.decomposition.nmf import _nls_subproblem, NMF
In [3]: X = load_digits().data
In [4]: %lprun -f _nls_subproblem NMF(n_components=16, tol=1e-2).fit(X)
Timer unit: 1e-06 s
File: scikits/learn/decomposition/nmf.py
Function: _nls_subproblem at line 137
Total time: 1.73153 s
Line # Hits Time Per Hit % Time Line Contents
==============================================================
137 def _nls_subproblem(V, W, H_init, tol, max_iter):
138 """Non-negative least square solver
...
170 """
171 48 5863 122.1 0.3 if (H_init < 0).any():
172 raise ValueError("Negative values in H_init passed to NLS solver.")
173
174 48 139 2.9 0.0 H = H_init
175 48 112141 2336.3 5.8 WtV = np.dot(W.T, V)
176 48 16144 336.3 0.8 WtW = np.dot(W.T, W)
177
178 # values justified in the paper
179 48 144 3.0 0.0 alpha = 1
180 48 113 2.4 0.0 beta = 0.1
181 638 1880 2.9 0.1 for n_iter in xrange(1, max_iter + 1):
182 638 195133 305.9 10.2 grad = np.dot(WtW, H) - WtV
183 638 495761 777.1 25.9 proj_gradient = norm(grad[np.logical_or(grad < 0, H > 0)])
184 638 2449 3.8 0.1 if proj_gradient < tol:
185 48 130 2.7 0.0 break
186
187 1474 4474 3.0 0.2 for inner_iter in xrange(1, 20):
188 1474 83833 56.9 4.4 Hn = H - alpha * grad
189 # Hn = np.where(Hn > 0, Hn, 0)
190 1474 194239 131.8 10.1 Hn = _pos(Hn)
191 1474 48858 33.1 2.5 d = Hn - H
192 1474 150407 102.0 7.8 gradd = np.sum(grad * d)
193 1474 515390 349.7 26.9 dQd = np.sum(np.dot(WtW, d) * d)
...
By looking at the top values of the % Time column it is really easy to pin-point the most expensive expressions that would deserve additional care.
Performance tips for the Cython developer¶
If the 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. for loops over vector components, nested evaluation of conditional expression, scalar arithmetics...), it is probably adequate to extract the hotspot portion of the code as a standalone function in a .pyx file and add static type declarations and then use cython_ to generate a C program suitable to be compiled as a Python extension module.
The official documentation available http://docs.cython.org/ contains 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.
TODO: html report, type declarations, bound checks, division by zero checks, memory alignement, direct blas calls...
Profiling compiled extensions¶
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 instrospect what’s happening inside the compiled extension it-self.
In order to profile compiled Python extensions one could use gprof after having recompiled the project with gcc -pg and using the python-dbg variant of the interpreter on debian / ubuntu: however this approach requires to also have numpy and scipy recompiled with -pg which is rather complicated to get working.
Fortunately there exist two alternative profilers that don’t require you to recompile everything.
Using google-perftools¶
TODO
Note
google-perftools provides a nice ‘line by line’ report mode that can be triggered with the --lines option. However this does not seem to work correctly at the time of writing. This issue can be tracked on the project issue tracker.
Using valgrind / callgrind / kcachegrind¶
TODO
Multi-core parallelism using joblib.Parallel¶
TODO: give a simple teaser example here.
Checkout the official joblib documentation:
A sample algorithmic trick: warm restarts for cross validation¶
TODO: demonstrate the warm restart tricks for cross validation of linear regression with Coordinate Descent.