NumPy Tutorial presented at SciPy2010

Description

Kittens&DragonsNumPyTutorialpresentedatSciPy2010Stéfan van der WaltStellenbosch University, South Africa28 June 2010 Setup• Tutorial layout• Setup import numpy as npThe NumPy ndarrayprint np.__version__ # version 1.3 or greaterBroadcastingIndexingStructured arrays Point your browser to the problem set atUniversal functionshttp://mentat.za.net/numpy/kittensThe __array_interface__OptimisationUpdate, wrap-up & questions9th Python in Science Conference (SciPy) 2010 3 / 48• Tutorial layout• SetupThe NumPy ndarray• ndarray• Rating: Kitten• Data buffers• Dimensions• Data-type• Strides• Flags• Base Pointer• TheNumPyndarrayBroadcastingIndexingStructured arraysUniversal functionsThe __array_interface__OptimisationUpdate, wrap-up & questions9th Python in Science Conference (SciPy) 2010 4 / 48Revision: Structureofanndarray• Tutorial layout Taking a look atnumpy/core/include/numpy/ndarraytypes.h:• SetupThe NumPy ndarray• ndarray• Rating: Kitten typedef struct PyArrayObject {• Data buffersPyObject_HEAD• Dimensions• Data-type char *data; /* pointer to data buffer */• Stridesint nd; /* number of dimensions */• Flags• Base Pointer npy_intp *dimensions; /* size in each dimension */•npy_intp *strides; /* bytes to jump to getBroadcasting* to the next element inIndexing* each ...

Sujets

• Python
• Calcul (mathématiques)
• Tutoriel

Informations

• Publié par: mtoledan
• Publié le: 27 juin 2011
• Nombre de lectures: 221
• Langue: English

Extrait

NumPy

n n Kitte s & Drago s

Tutorialpresetedat n

SciPy2010

StéfanvanderWalt StellenboschUniversity,South

28 June 2010

Africa

² Tutorial layout ² etSup The NumPy ndarray Broadcasting

Indexing Structured arrays Universalfunctions The_fretecaraariny__

Optimisation

Update,wrap-up&questions

Setup

p

i m p o r tnspmaynu r i n ts i o n _ _np . _ _ v e r

p #

version1.3or

Point your browser to the problem set at

9th Python in Science Conference (SciPy) 2010

htt:p//mentat.za.net/numpy/kittens

g r e a t e r

3 / 48

² ² ² ² ² ²

Tutorial layout Setup

The NumPy ndarray ndarray Rating: Kitten Data buffers Dimensions Data-type Strides Flags Base Pointer

² ² ² ² ²

Broadcasting

Indexing

Structured arrays

Universalfunctions

The_yarraafretni_ce_

Optimisation

Update,wrap-up&questions

9th Python in Science Conference (SciPy) 2010

TheNumPyndarray

4 / 48

otadtnrep*io/ 485/ /idhcaeni*noisnem *tot oget xtelhene ybet*/mutptsjo nu/*ermbdiofnsme ubatreff/* /*size oisn/* */n. oP*/etni hisobjec*Decreft dnleteoi *t / / u*op toerntoi*P/ / /*yarralanigiro dimeeach tin * emen * n P / snoi

Revision:Structureofadarray n n

² Tutorial layout ² Setup The NumPy ndarray ² arndyra ² Rating: Kitten ² Data buffers ² ionsmensDi ² Data-type ² sertdiS ² Flags ² Base Pointer ² Broadcasting Indexing Structured arrays Universalfunctions The_arrainy_rfteeac_ Optimisation Update,wrap-up&questions

Taking a look at munc/yptypyseh./ydnraarude/numpore/incl

fereneec*s } / P

P y A r r a y D e s c r * d e s c r ; _ intf l a g s ; P y O b e c t * w e a k r e f l i s t ; j y A r r a y O b e c t ; j

t y p e d e f s tj r u c t {P y A r r a y O b c t e j P y O b e c t _ H E A D c h a r* d a t a ; intnd ; n y i n t * d i m e n s i o n s ; npy_i n tp* s t r i d e s ;

9th Python in Science Conference (SciPy) 2010

y O b j e c t * b a s e ;

:

/ /*krearwFo pysetrto*t/rtcu ags*/*Fl

A homoge eous container n

c h a r* d a t a ;

/*pointertodatabuffer*/

Data is just a pointer to bytes in memory:

I I I I

n[16]:x=np.array([1,2,3])

n [ 2 2 ] : x . d t y p e Out t y p e ( d[ 2 2 ] :’ i n t 3 2 ’) b y t e s# 4

n [ 1 8 ] : x . _ _ a r r a y _ i n t e r f a c e _ _ [’ d a t a ’] Out , 2 6 3 1 6 6 2 4 a l s e ) F ([ 1 8 ] :

n [ 2 1 ] : str ( x . d a t a ) Out[ 2 1 ] : ’’ \ x01 \ x00 \ x00 \ x00 \ x02 \ x00 \ x00 \ x00 \ x03 \ x00 \ x00 \ x00

9th Python in Science Conference (SciPy) 2010

7 / 48

² Tutorial layout ² ptuSe The NumPy ndarray ² naydarr ² Rating: Kitten ² ata buffers D ² neisDminos ² eyp-taDat ² Stsderi ² Flags ² Base Pointer ²

Broadcasting Indexing Structured arrays Universalfunctions The_ayarrecreafi_tn_ Optimisation Update,wrap-up&questions

Dime sio s n n

I I

intnd ; n p _ n t p * d i m e n s i o n s ; y i

n [ 3 ] : x = np . a r r a y ( [ ] ) n [ 4 ] : x . s h a p e Out[4(0]:,)

n [ 5 ] : np . a r r a y ( 0 ) . s h a p e Out[5]:()

/*numberofdimensions*/ /*sizeineachdimension*/

n[8]:x=np.random.random((3,2,3,3)) In [ 9 ] : x . s h a p e Out)3,3,3(2,9[:] In 1 0 ] : x . n d i m [ Out[ 1 0 ] : 4

9th Python in Science Conference (SciPy) 2010

8 / 48

Datatypedescriptors

P y A r r a y _ D e s c r * d e s c r ;/*Poi*/eptsurtctnreotyt

Common types in include int, oat, bool:

I I I

n [ 1 9 ] : np . a r r a y ([ -1 , 0 , 1] , d t y p e = int ) Out1[a:]9ayrr-1([,1,0]) n [ 2 0 ] : np . a r r a y ([ -1 , 0 , 1] , d t y p e = f l o a t ) Out , 0. . ] ) 1 r r a y ([ -1. ,[ 2 0 ] : a n [ 2 1 ] : np . a r r a y ([ -1 , 0 , 1] , d t y p e = b o o l ) Oute = b o o l ) t y p  d[ 2 1 ] : a False , T r u e ] , r r a y ([ True ,

Eachiteminthearrayhastohavethesametype(occupyaxednrofbytesin memory), but that does not mean a type has to consist of a single item:

I

. n[2]:dt=np.dtype([(’ v a l u e ’, np int ) , (’ s t a t u s ’, n[3]:np.array([(0,True),(1,False)],dtype=dt) Out[ 3 ] : array([(0,True),(1,False)], dtype=[(’ v a l u e ’,’ < i4 ’) , (’ s t a t u s ’,’ | b1 ’)])

This is called astructured array.

9th Python in Science Conference (SciPy) 2010

np.bool)])

9 / 48

² Tutorial layout ² upetS The NumPy ndarray ² dnraary ² Rating: Kitten ² Data buffers ² snoiiDsnem ² taDpea-ty ² estridS ² Flags ² Base Pointer ² Broadcasting Indexing Structured arrays Universalfunctions The_raar_ynietfrcae_ Optimisation Update,wrap-up&questions

Strides

I I I I I

n p y _ i n t p * s t r i d e s ;

/*bytestojumptoget*/ /*tothenextelement*/

n [ 3 7 ] : x = np . a r a n g e ( 1 2 ) . r e s h a p e ( ( 3 , 4 ) )

n [ 3 8 ] : x Out[ 3 8 ] : a r r a y ([[ 0 , 1 , 2 , 3] , [ 4 , 5 , 6 , 7] , [ 8 , 9 , 10 , 1 1 ] ] )

n [ 3 9 ] : x . d t y p e Out t y p e ( d[ 3 9 ] :’ i n t 3 2 ’)

n [ 4 0 ] : x . d t y p e . i t e m s i z e Out[ 4 0 ] : 4

n [ 4 1 ] : x . s t r i d e s Out,416)4[(:]1 t e m s i z e ) , i t e m s i z e(4* i # k i p _ b y t e s _ c o l s ,_ b y t e s _ r o w( s k i p

9th Python in Science Conference (SciPy) 2010

)

10 / 48

Flags

intf l a g s ;

In [ 6 6 ] : x = np . a r r a y ([1 , 2 , 3

In [ 6 7 ] : x . f l a g s Out[ 6 7 ] : C C O N T I O S T r u e F_O N T IG UOUS:T r u e C O N D A T A T r u e W : W R I T E A B L E : T r u e A L I G N E D : T r u e U P D A T E I F C O P Y : F a l s e

In [ 6 8 ] : z . f l a g s Out[ 6 8 ] : C C O N T I O S F a l s e F_C O N T IG UOUS:F a l s e O W N D A T A : F a l s e W R I T E A B L E : T r u e A L I G N E : T r u e D U P D A T E I F C O P Y : F a l s e

9th Python in Science Conference (SciPy) 2010

# ])

/*Flags*/

C - c o n t i g u o u s F o r t r a n - c o n t i g u o u s are we r e s p o n s i b l e for m e m o r y h a n d l i n g ? ay we c h a n g e the d a t a ? m a p p r o p r i a t e h a r d w a r e a l i g n m e n t u p d a t e b a s e on d e a l l o c a t i o n ?

11 / 48