session 5 - steven bamfordstevenbamford.com/python/2014/mpags_python_session5.pdf• for creating,...
TRANSCRIPT
Session 5: Extreme Python
An introduction to scientific programming with
• Managing your environment
• Efficiently handling large datasets
• Optimising your code
• Squeezing out extra speed
• Writing robust code
• Graphical interfaces
Managing your environment
• Some good things about Python
• lots of modules from many sources
• ongoing development of Python and modules
• Some bad things about Python
• lots of modules from many sources
• ongoing development of Python and modules
• A solution
• Maintain (or have option to create) separate environments (or manifests) for different projects
• virtualenv
• general Python solution – http://virtualenv.pypa.io
• modules are installed with pip – https://pip.pypa.io
$ pip install virtualenv # install virtualenv
$ virtualenv ENV1 # create a new environment ENV1
$ source ENV/bin/activate # set PATH to our environment
(ENV1)$ pip install emcee # install modules into ENV1
(ENV1)$ pip install numpy==1.8.2 # install specific version
(ENV1)$ python # use our custom environment
(ENV1)$ deactivate # return our PATH to normal
• virtualenv
• can record current state of modules to a 'requirements' file
• using that file can always recreate the same environment
(ENV1)$ pip freeze > requirements.txt
$ cat requirements.txt
emcee==2.1.0
numpy==1.8.2
$ deactivate
$ virtualenv ENV2
$ source ENV2/bin/activate (ENV2)$ pip install -r requirements.txt
• conda – http://conda.pydata.org
• specific to the Anaconda Python distribution
• similar to 'pip', but can install binaries (not just python)
• avoid using pip within conda environment (although possible)
$ conda create -n ENV1 # create a new environment ENV1
$ source activate ENV1 # set PATH to our environment
$ conda install numpy # install modules into ENV1
$ conda install -c thebamf emcee # install from binstar
$ source deactivate # return our PATH to normal
$ conda list –n ENV1 -e > requirements.txt # clone ENV1
$ conda create -n ENV2 --file requirements.txt # to ENV2
• Updating packages
$ conda update --all
$ conda update scipy emcee
OR
$ pip install --upgrade
$ pip install --upgrade scipy emcee
Efficiently handling large datasets
• Python has tools for accessing most (all?) databases
• e.g. MySQL, SQLite, MongoDB, Postgres, …
• Allow one to work with huge datasets
• Data can be at remote locations
• However, most databases are designed for webserver use
• Not optimised for data analysis
• http://pytables.github.io
• For creating, storing and analysing datasets
• from simple, small tables to complex, huge datasets
• standard HDF5 file format
• incredibly fast – even faster with indexing
• uses on the fly block compression
• designed for modern systems
• fast multi-code CPU; large, slow memory
• "in-kernel" – data and algorithm are sent to CPU in optimal way
• "out-of-core" – avoids loading whole dataset into memory
>>> from tables import *
>>> h5file = openFile("test.h5", mode = "w")
>>> x = h5file.createArray("/", "x", arange(1000)) >>> y = h5file.createArray("/", "y", sqrt(arange(1000)))
>>> h5file.close()
• Can store many things in one HDF5 file (like FITS)
• Tree structure
• Everything in a group (starting with root group, '/')
• Data stored in leaves
• Arrays (e.g. n-dimensional images)
>>> class MyTable(IsDescription):
z = Float32Col() >>> table = h5file.createTable("/", "mytable", MyTable)
>>> row = table.row >>> for i in xrange(1000):
row["z"] = i**(3.0/2.0)
row.append() >>> table.flush()
>>> z = table.cols.z
• Tables (columns with different formats)
• described by a class
• accessed by a row iterator
>>> r = h5file.createArray("/", "r", np.zeros(1000))
>>> xyz = Expr("x*y*z") >>> xyz.setOutput(r)
>>> xyz.eval() /r (Array(1000,)) ''
atom := Float64Atom(shape=(), dflt=0.0)
maindim := 0 flavor := 'numpy'
byteorder := 'little' chunkshape := None
>>> r.read(0, 10) array([ 0. , 1. , 7.99999986, 26.9999989 ,
64. , 124.99999917, 216.00000085, 343.00001259,
511.99999124, 729. ])
• Expr enables in-kernel & out-of-core operations
>>> r_bigish = [ row['z'] for row in
table.where('(z > 1000) & (z <= 2000)' ]
>>> for big in table.where('z > 10000;'): ... print('A big z is {}'.format(big['z'])
• where enables in-kernel selections
• There is also a where in Expr
• Python Data Analysis Library • http://pandas.pydata.org
• Easy-to-use data structures • DataFrame (more friendly recarray) • Handles missing data (more friendly masked array) • read and write various data formats • data-alignment
• tries to be helpful, though not always intuitive • Easy to combine data tables • Surprisingly fast!
>>> import pandas as pd
>>> t = np.arange(0.0, 1.0, 0.1)
>>> s = pd.Series(t)
>>> x = t + np.random.normal(scale=0.1, size=t.shape)
>>> y = x**2 + np.random.normal(scale=0.5, size=t.shape)
>>> df1 = pd.DataFrame({ 'x' : x, 'y' : y}, index = t)
>>> df1.plot()
>>> df1.plot('x', 'y', kind='scatter')
• 1D Series, 2D DataFrame, 3D Panel
• Various ways of creating these structures
>>> t2 = t[::2] >>> z = t2**3 + np.random.normal(scale=1.0, size=t2.shape)
>>> df2 = pd.DataFrame({ 'z' : z, 'z2' : z**2}, index = t2)
>>> df3 = df1.join(df2)
>>> df3.sort('y')
• Always indexed
• Indices used when joining tables
• Handling data fairly intuitive, similar to numpy slices
• Powerful and flexible
• Good documentation
data = pyfits.getdata('myfunkydata.fits')
df = pd.DataFrame(data)
# store DataFrame to HDF5
store = pd.HDFStore('myfunkydata.h5', mode='w', complib='blosc', complevel=5)
store['funkydata'] = df
# some time later…
store = pd.HDFStore('myfunkydata.h5', mode='r')
df = store['funkydata']
• DataFrame can be created directly from a recarray
• Pandas data structures can be efficiently stored on disk
• based on PyTables
Graphical interfaces
Several modules to construct GUIs in Python
wxpython is one of the most popular
http://www.wxpython.org
E.g., https://github.com/bamford/control/
Django
a high-level Python Web framework that encourages rapid development and clean, pragmatic design.
and many others, e.g. Zope (massive), web2py (light), …
• Give your scientific code a friendly face! • easy configuration • monitor progress • particularly for public code, cloud computing, HPC
• An (unsicentific) example, followed by another one
Optimising your code
timeit – use in interpreter, script or command line
Options:
-s S, --setup=S
statement to be executed once initially (default pass)
-n N, --number=N
how many times to execute 'statement' (default: take ~0.2 sec total)
-r N, --repeat=N
how many times to repeat the timer (default 3)
iPython magic version
python -m timeit [-n N] [-r N] [-s S] [statement ...]
%timeit # one line %%timeit # whole notebook cell
# fastest way to calculate x**5?
$ python -m timeit -s 'from math import pow; x = 1.23' 'x*x*x*x*x'
10000000 loops, best of 3: 0.161 usec per loop
$ python -m timeit -s 'from math import pow; x = 1.23' 'x**5'
10000000 loops, best of 3: 0.111 usec per loop
$ python -m timeit -s 'from math import pow; x = 1.23' 'pow(x, 5)'
1000000 loops, best of 3: 0.184 usec per loop
• Understand which parts of your code limit its execution time
• print summary to screen, or save file for detailed analysis
From shell
From iPython
Lots of functionality… see docs
$ python -m cProfile –o program.prof my_program.py
%prun -D program.prof my_function()
%%prun # profile an entire notebook cell
Nice visualisation with snakeviz – http://jiffyclub.github.io/snakeviz/
In iPython:
$ conda install –c thebamf snakeviz
OR
$ pip install snakeviz
%load_ext snakeviz
%snakeviz my_function()
%%snakeviz # profile entire cell
Writing robust code
• Several testing frameworks
• unittest is the main Python module
• nose is a third-party module that nicely automates testing
Squeezing out extra speed
• Python includes modules for writing "parallel" programs:
• threaded – limited by the Global Interpreter Lock
• multiprocessing – generally more useful
from multiprocessing import Pool
def f(x): return x*x
pool = Pool(processes=4) # start 4 worker processes
z = range(10) print pool.map(f, z) # apply f to each element of z in parallel
from multiprocessing import Process from time import sleep
def f(name): print('Hello {}, I am going to sleep now'.format(name)) sleep(3) print('OK, finished sleeping')
if __name__ == '__main__': p = Process(target=f, args=(lock, 'Steven')) p.start() # start additional process sleep(1) # carry on doing stuff print 'Wow, how lazy is that function!' p.join() # wait for process to complete
$ python thinking.py Hello Steven, I am going to sleep now Wow, how lazy is that function! OK, finished sleeping
(Really, should use a lock to avoid writing output to screen at same time)
Cython is used for compiling Python-like code to machine-code • supports a big subset of the Python language • conditions and loops run 2-8x faster, overall 30% faster for plain Python
code • add types for speedups (hundreds of times) • easily use native libraries (C/C++/Fortran) directly
• Cython code is turned into C code • uses the CPython API and runtime
• Coding in Cython is like coding in Python and C at the same time!
Some material borrowed from Dag Sverre Seljebotn (University of Oslo) EuroSciPy 2010 presentation
Use cases:
• Performance-critical code • which does not translate to array-based approach (numpy / pytables) • existing Python code ! optimise critical parts
• Wrapping existing C/C++ libraries • particularly higher-level Pythonised wrapper • for one-to-one wrapping other tools might be better suited
Cython code must be compiled.
Two stages:
• A .pyx file is compiled by Cython to a .c file, containing the code of a Python extension module
• The .c file is compiled by a C compiler • Generated C code can be built without Cython installed • Cython is a developer dependency, not a build-time dependency • Generated C code works with Python 2.3+ • The result is a .so file (or .pyd on Windows) which can be imported
directly into a Python session
Ways of building Cython code:
• Run cython command-line utility and compile the resulting C file • use favourite build tool • for cross-system operation you need to query Python for the C build
options to use
• Use pyximport to importing Cython .pyx files as if they were .py files; building on the fly (recommended to start). • things get complicated if you must link to native libraries • larger projects tend to need a build phase anyway
• Write a distutils setup.py • standard way of distributing, building and installing Python modules
• Cython supports most of normal Python
• Most standard Python code can be used directly with Cython • typical speedups of (very roughly) a factor of two • should not ever slow down code – safe to try • name file .pyx or use pyimport = True
>>> import pyximport
>>> pyximport.install() >>> import mypyxmodule # converts and compiles on the fly
>>> pyximport.install(pyimport = True) >>> import mypymodule # converts and compiles on the fly # should fall back to Python if fails
• Big speedup from defining types of key variables
• Use native C-types (int, double, char *, etc.)
• Use Python C-types (Py_int_t, Py_float_t, etc.)
• Use cdef to declare variable types
• Also use cdef to declare C-only functions (with return type) • can also use cpdef to declare functions which are automatically treated
as C or Python depending on usage
• Don't forget function arguments (but note cdef not used here)
• Efficient algorithm to find first N prime numbers
def primes(kmax): p = [] k = 0 n = 2 while k < kmax: i = 0 while i < k and n % p[i] != 0: i = i + 1 if i == k: k = k + 1 p.append(n) n = n + 1 return p
$ python -m timeit -s 'import primes as p' 'p.primes(100)' 1000 loops, best of 3: 1.35 msec per loop
primes.py
$ python -m timeit -s 'import pyximport; pyximport.install(); import cprimes as p' 'p.primes(100)'
1000 loops, best of 3: 731 usec per loop 1.8x speedup
def primes(kmax): p = [] k = 0 n = 2 while k < kmax: i = 0 while i < k and n % p[i] != 0: i = i + 1 if i == k: k = k + 1 p.append(n) n = n + 1 return p
cprimes.pyx
def primes(int kmax): # declare types of parameters cdef int n, k, i # declare types of variables cdef int p[1000] # including arrays result = [] # can still use normal Python types if kmax > 1000: # in this case need to hardcode limit kmax = 1000 k = 0 n = 2 while k < kmax: i = 0 while i < k and n % p[i] != 0: i = i + 1 if i == k: p[k] = n k = k + 1 result.append(n) n = n + 1 return result # return Python object
xprimes.pyx
40.8 usec per loop
33x speedup
contains only C-code
• Cython provides a way to quickly access Numpy arrays with specified types and dimensionality
! for implementing fast specific algorithms
• Also provides way to create generalized Ufuncs
• Can be useful, but often using functions provided by numpy, scipy, numexpr or pytables will be easier and faster
• JIT: just in time compilation of Python functions
• Compilation for both CPU and GPU hardware
from numba import jit
@jit def sum2d(arr): M, N = arr.shape result = 0.0 for i in range(M): for j in range(N): result += arr[i,j] return result
a = np.arange(10000).reshape(1000,10) %timeit sum2d(a) 1000 loops, best of 3: 334 µs per loop %timeit sum2d(a) # without @jit 1 loops, best of 3: 2.15 µs per loop
The End
An introduction to scientific programming with