Keras, Tensorflow va Torch kutubxonalari.

Загружено в:

08.08.2023

Скачано:

0

Размер:

2081.892578125 KB

Скачать

Mavzu: Keras, Tensorflow va Torch kutubxonalari.
Reja:
1. Keras, Tensorflow va Torch kutubxonalari xaqida
2. K e r a sd a c h iz i q li r e gr e ss iy a m od eli n i qurish
3.
Python keras va tensorf low kutubx ona l aridan foyda l anish.
K eras – b u P y t hon-da y oz i lgan, T hea n o y oki Tensor f low-ning a sos i ga quril g an
O chiq K odli net ro n t a r m o q kut u b x o n a s i dir. U m odull i , te z k o r v a ish l at ish uc hu n
q ulay holatda G o o g l e k o m pa n i y a s i m uhandisi F ran s ua C hol e t is h l a b c hi q q a n.
K eras – b u hisob- ki toblar n i a m a lg a os hi r ish u c hun Tensor Fl ow, CNTK y ok i
Thea n o “ Backend” k ut u bxonalar i d a n f o y d al a n a dig a n , y u qo r i dara j ada i s hla y d i gan
A PI(a p p lic a ti o n p rogra mm in g interfa c e) di r.
K eras ni m a?
K era s ni n g i sh l a b c hiq ilishiga a s osiy sab a b shuki – u n ga c ha bos h qa ne y r o n
t ar m oqla r i ku t ub x o n a la r i d a n f o y dal a ni s h n o qula y roq bo ` lgan y a ’ n i sint a ksisi
qi y in ro q b o ` l g a n. K e r as y uqo r i - d ara j a li A P I m od ella r n i y aratish, qa t la m l ar n i

belgila s h y oki b i r n e c h t a k iri s h- c hiqi s h m o d e l la r i n i bo s h qar is h i m konini ber a di.
K eras m od e l n i y o`q ot ish (Loss f unction) v a opti m i zatsi y a ( O pti m izat i on function)
f u n k si y a s i bil a n ko m pil y atsi y a qil a d i , va m odelni o `q it i sh ni f i t f u nksi y asi o r qa l i
a m al g a oshiradi.
K era s d a ne y r o n t ar m o q qurish
K eras ne y ron t a r m o q m ode l ini j u da tez t a y y or l ash i m k o nini b e r ad i . B i r ne c h a
q ato r l i oddiy tar m o q m ode l ini y arati s h u ch un , K er a s bu b o rada y ordam b e r i shi
m u m kin . Q u y id a g i m i s o lga qar an g:
K era s d a ne y r o n t ar m o q qurish ( I n pu t )
m odel = Sequ e nti a l()
m odel.add(Dense ( 512, i np u t _ s hap e =(7 8 4 , )))
m od el.a d d (Ac ti v a ti o n('re l u'))
m odel.add(Dro p o ut (0.2))

K era s d a ne y r o n t ar m o q qurish ( H idd e n)
m odel = Sequ e nti a l()
m odel. ad d(Dense ( 51 2 , i npu t_ shape=(784 , )))
m odel.add(Act i va t i on('relu'))
m odel. ad d(Dropout ( 0. 2 ))
m odel.add(Dense ( 51 2 ))
m odel.add(Activation('re l u'))
m odel. ad d(Dropout ( 0. 2 ))
K era s d a ne y r o n t ar m o q qurish ( O ut p ut )
m odel = Sequ e nti a l()
m odel.add(Dense ( 512, i np u t _ s hap e =(7 8 4 , )))
m odel.add(Activation('re l u'))
m odel.add(Dro p o ut (0.2))

m odel.add(Dense ( 51 2 ))
m od el. ad d( Ac t iva tio n ( 'r e l u ' ))
m odel.add(Dro p o ut (0.2))
m odel. ad d(Dense ( 10 ))
m odel.add(Act i va t i on('sof t m ax'))
K era s ni n g a s o s i y t u s h un c h a la r i
K era s d a g i a so s iy s truk tura bu m od e l. M od ellar b ir qa n cha q a t l a m lar d an t as h k i l
t o p i s h i m u m kin . K er a s d a qat l a m larning bir ne c hta t uri m av j u d .
◦ Ket m a- k et tar ti bl i ( S e quent ia l ) m odel
◦ K o n v ol y u t s ion qa t l a m
◦ M a xP o oli n g q a t l am
◦ Zi c h ( De n s e ) qatl a m
◦ Dr o p o u t q a t l am

Ket m a-ket t a r t i b l i ( S e que n tial) m od e l
Ket m a-ket t a r t ibli m ode ln i n g a so s iy g' o y asi shun c h a ki Ker a s q a t la m l ar i ni k e t m a -
k et t a r t ib da jo y lash t i r is h dir. Ya ` n i q a tla m lar tu r g a n t ar ti bi b o ` y i c ha is h g a
tush i r ila d i
Kon v o l y utsion q a tl a m
Kon v o l y utsion q a tl a m – b u m as s ivlar u st i da b ir ne c h t a f i l t r l a r d an f o y dal a n g a n
h o l da ula r d a n tu r li xil xu s u s i y a t la r n i a j r a tib olish uc hu n qo`ll an ilad i g a n qa t la m di r
Kon v o l y utsion q a tl a m
ReLU-ni f a o ll as h ti r i s h fu n k t si y as i d a n f o y dal a n a dig a n va kirish shak l i(i n put s h a pe)
3 2 0 x 3 2 0 x3 b o ' lgan 4 8 t a 3x3 o'lcha m dagi f i lt r i bo`lg a n konvo l yutsi o n q at la m .
i nput_sh a pe= ( 32 0 , 3 2 0, 3 ) #this is the i np u t sh a p e of an i m a ge 320 x 320 x 3
m odel.add( C o nv2D(48, (3, 3 ), ac t i va ti o n = ' r e l u', i np u t _ s h a pe= i n put_ s hape))
Kon v o l y utsion q a tl a m ning b oshqacha k o `rinishi:
m odel. ad d(Conv2 D ( 4 8 , ( 3 , 3 ), a c t ivati o n = 're l u'))
Ma x Po oling qatlam

Maxp o oli n g - b u har bi r fi lte r dan o`tka z il a y otkan q iy m atlar d a n e n g katta s i n i
o l i sh di r. O datda m axpoo l ing qan a da y dir m a ss i v da q o`llan ilga nda n a tij a viy
m assivning o ` l c ha m i ki chk i na r o q b o ` la d i .
( 2 , 2) fi lt e rga e ga b o `lg a n m axpooling qa t l a m i
m odel.add(M a x P ool i ng2D ( po o l_size= ( 2, 2)))
Z i ch (D e nse) q at lam
B o shqa c ha qi l i b a y tga n da, zi c h q a t lam b i r-b i r i ga t o'l i q bog'l an g a n q at la m di r, y a ' n i
q a t la m d a g i b a r c h a n e y r onl ar ke y in gi q a t l a m b i lan b o g'langan va bu q a t l a m d an j u d a
ko ` p f o y dal a nil a d i .
2 56 ta c h iqish q i y m atla r i bo`lg a n z ich q a tl a m
m odel. ad d(Dense ( 25 6 , ac t i vat i o n = ' re l u'))
Dro p out qa t la m i

Dro p out qatla m i o `q it ish jara y o n ida ne y ron tar m oqd a gi ah a m i y ati y o ’ q b o ’lga n
n e y r o n l a r n i ta s hl a b yubo r ish a m a li n i ba j a r i s h uc hu n ishl a til a di
K eras m odelni ishga t u s huri s h , o'q i tish va b a holash
Model t u z i l g a n d a n so'n g , o'qitish bo s hlan a di. M o de l n i birin c hi b o ' lib lo s s
f u n k t s i y asi va o pti m i za t s i y a funktsi y asi b il an ko m pil y atsi y a q i lish talab qilin ad i.
Bu t a r m o q o g ' i r l i k k o effitsi e ntla r i ni (weig h ts) o' z ga r ti ri shga i m ko n ber ad i va
x atoli k ni(loss) m ini m alla s htiradi. L o ss f unksi y asi – b u q u r il g a n m odelning o ’ q u v
t anl a n m aga nisba t an q a ncha l ik t o ’ g’ri s hak l l a nti r ilg a nl i gini b a h olash usu l i
hisoblana d i.
O p ti m i zat s i y a f un ksi y asi – bu t ar m o qning o g ' i r l i k koe f fitsi e ntla r ini (wi eg ht s ) va
o z od ( bias) la r i n i o` z ga r ti r gan h o ld a x a tolikni ka m a y tir u vc h i fu n ksi y adir.
m odel.co m pile( l oss = 'm ean_squar ed _error', opti m izer='ada m ')
K eras m odelni ishga tus h uri s h , o' qi ti s h va b a hol a sh
Modelni o `q iti sh n i b o s h l a s h u c h u n u n g a v a lid a tsi y a va o ` qitish(tra i ni n g)
m a` lu m otl a r i n i be r ish ker a k shuningdek p a ket o ` l c ha m i va si k l l a r so n i ham
belgilan ad i
m od el.fi t ( X _tr a i n , X_ t r a i n , ba t c h _siz e =32, e p ochs=10,
validation_d ata=( x _ val, y _va l ) )
Ox i r gi b o s q ich e sa m ode l ni t e st m a`lu m o t lari b i lan b a h olash
s c o re = m odel. e v a lu a te ( x_t e st, y _t e st , bat c h_size = 32)

Ke r a sd a c h iz i q li r e gr e ss iy a m od eli n i qu r ish
Modelni o`qit g andan so `ng n a t i ja qu y ida g i ch a bo ` l a d i
Bosh l a n g ` ich o` g irli k k oeffitsie n tla r i(ini t ial w e i g h t s ): w : 0.37, b : 0.00
O p t i m a l la s h t i ril g an o` g i rli k k oeffitsientl a ri: w: 3.70, b : 0.61

$Py t hon Vir t ual envi r on m ent (venv) c:\>p y t hon -m v env c:\path\ t o\m y env M isol u c hun “ M a chine L e a rnin g ” no m li p ap k a m a v j ud. Shu p a p k ani i c h i g a k irib c m d o y nasi o r q ali y uq o rida g i k od i sh g a t us huril a d i N atij a da M a ch i neLe a r n i ng pap k as i n i ng ic h i da “ v en v ” no m l i p a p k a v a un i ng i c h i d a j o ri y k o m p y u t er g a o ’r n a t il g an python n i ng v irtual m uhiti y ar a t i la d i Py t hon Vir t ual envi r on m ent (venv)$ $P y thon Vi r t ual e n v i ronm e n t ( ve n v ) ni ish g a t ushi r ish Endi yaratilgan virtual muh i tni ishga tushirib ke r akli k u tu b x o nalarni o’rnatish loz i m b o ’ladi, bu a malni c m d o rqali yoki “ Machin eL ear n ing” loyihasini PyCha r m das t uri orqali ochib, shu muhitda h a m a malga o shirish mumkin. venv\Script s \activ a te. b at N aitjada v i r t u al muhit ish g a tushadi V e n v da ker a k li pa k et la r ni o ’ r na t ish Bu yer d an quyidagi b u yr u q orqali ma v jud o’rnatilgan p aketlar n i ko’r i sh m u m kin pip list$

pip install pa c kage- n a me
pip install t en s o rfl o w
pip install numpy
pip install k eras
Endi shu yerda ba r c h a k er a kli bo’lgan pa k etlar n i
m axsus b uyruq o rqali o’rn a tish m u m kinV
e n v da ker a k li pa k et la r ni o ’ r na t ish Us h bu m o dullar o’rn a tilgan d an
k eyin pip li s t buyrugi o r q ali tekshirib ko’r a miz
P yt h on virt u al muhitda ( v e nv) j o riy m uhitda o’rnatilg a n m o dullarni saqlash v a
k e yin g i l o y i h a uc h un shu mo dullarni i s hl a tish imk o n i ya t i m a vjud. Bu n ing u ch u n
quyidagilar a m alga oshirili s hi l o zim
pip free z e
V
e n v da ker a k li pa k et la r ni o ’ r na t ish

$Kompyu t er x ot irasida ( mi s ol uchun C:\U s er \ De s c t op) “requir e m e n t s.tx t ” f a yl ya r atiladi v a ushbu f a ylga ras m da ko’ r satilgan m o dul l ar v a ularning vers i yalari h a qi d agi ma’lum o t k o’chirib o lina d iV e n v da ker a k li pa k et la r ni o ’ r na t ish Us h bu y aratil g an “requir e ments . tx t ” f a yli, k eyinc h al i k y ar a tilishi m u mkin b o ’lg a n v irtual m uhit uc h un ta y y or m o dullar r o ’yhati hisoblanadi v a j o r i y l o y iha u c h un “v e n v” yar a til g a n dan k eyin q u yidagi b u yruq o rqali fa y l ic h ida k eltiril ga n m o dullarni bir ur u nish d a o ’ r n atish m u mkin D emak P ython virtual m uh i ti v a “ pip” orqali l o yiha uc h un kerakli ba r c h a m o dullar o’rnatilga n dan k e yin virtual muhitni “ d eact i va t e ” qilish k er a k b o ’ladi$

Foydala n ilgan a d a bi y o tl ar
Aureli a n Geron, Hands on M a c h in e Lear n i n g w i t h S c ikit- L ea r n
Keras&Tensorfl o w // Sec on d edition Con c ept s , T o o l s, a n d Te c hn i q u e s to Bu i l d
Int e lli g e n t S y st e m s , 20 1 9 , 5 10 P a g e s .
Pr i m oz P o to c nik, Neur a l Netw o rks: MA T LAB e x a m p l es // Neur a l Netwo rk s
c o u rs e (pra ct ic a l exa m pl e s ) © 2012
h t t ps://towa r dsdat a s c i ence . c o m /f i rst-n e ur al -network- f or - beg in n e rs- e xpl a i n ed-
w ith-code - 4c f d37 e 06eaf
h t t ps://w w w. m athw o rks. c o m /help/ t h in gsp e ak / cre a te- an d-tra i n- a - feed f orwa r d -
neural- n e t w ork.ht m l
ht t ps://w w w. m athw o rks. c o m /help/d e e p lea r ni n g/ e xa m ples/c r ea te -si m ple -
de e p-lea r ning-ne t w o r k-for-cl a ssi f ic a tion .h t m l
https: / /w w w.t u torial s po i n t . c o m /
art i fi c i a l_ n eur a l _ network/a r ti f i c i al_neur al _n e t w or k _b asic_conc e pt s . htm

Mavzu: Keras, Tensorflow va Torch kutubxonalari. Reja: 1. Keras, Tensorflow va Torch kutubxonalari xaqida 2. K e r a sd a c h iz i q li r e gr e ss iy a m od eli n i qurish 3. Python keras va tensorf low kutubx ona l aridan foyda l anish. K eras – b u P y t hon-da y oz i lgan, T hea n o y oki Tensor f low-ning a sos i ga quril g an O chiq K odli net ro n t a r m o q kut u b x o n a s i dir. U m odull i , te z k o r v a ish l at ish uc hu n q ulay holatda G o o g l e k o m pa n i y a s i m uhandisi F ran s ua C hol e t is h l a b c hi q q a n. K eras – b u hisob- ki toblar n i a m a lg a os hi r ish u c hun Tensor Fl ow, CNTK y ok i Thea n o “ Backend” k ut u bxonalar i d a n f o y d al a n a dig a n , y u qo r i dara j ada i s hla y d i gan A PI(a p p lic a ti o n p rogra mm in g interfa c e) di r. K eras ni m a? K era s ni n g i sh l a b c hiq ilishiga a s osiy sab a b shuki – u n ga c ha bos h qa ne y r o n t ar m oqla r i ku t ub x o n a la r i d a n f o y dal a ni s h n o qula y roq bo ` lgan y a ’ n i sint a ksisi qi y in ro q b o ` l g a n. K e r as y uqo r i - d ara j a li A P I m od ella r n i y aratish, qa t la m l ar n i

belgila s h y oki b i r n e c h t a k iri s h- c hiqi s h m o d e l la r i n i bo s h qar is h i m konini ber a di. K eras m od e l n i y o`q ot ish (Loss f unction) v a opti m i zatsi y a ( O pti m izat i on function) f u n k si y a s i bil a n ko m pil y atsi y a qil a d i , va m odelni o `q it i sh ni f i t f u nksi y asi o r qa l i a m al g a oshiradi. K era s d a ne y r o n t ar m o q qurish K eras ne y ron t a r m o q m ode l ini j u da tez t a y y or l ash i m k o nini b e r ad i . B i r ne c h a q ato r l i oddiy tar m o q m ode l ini y arati s h u ch un , K er a s bu b o rada y ordam b e r i shi m u m kin . Q u y id a g i m i s o lga qar an g: K era s d a ne y r o n t ar m o q qurish ( I n pu t ) m odel = Sequ e nti a l() m odel.add(Dense ( 512, i np u t _ s hap e =(7 8 4 , ))) m od el.a d d (Ac ti v a ti o n('re l u')) m odel.add(Dro p o ut (0.2))

K era s d a ne y r o n t ar m o q qurish ( H idd e n) m odel = Sequ e nti a l() m odel. ad d(Dense ( 51 2 , i npu t_ shape=(784 , ))) m odel.add(Act i va t i on('relu')) m odel. ad d(Dropout ( 0. 2 )) m odel.add(Dense ( 51 2 )) m odel.add(Activation('re l u')) m odel. ad d(Dropout ( 0. 2 )) K era s d a ne y r o n t ar m o q qurish ( O ut p ut ) m odel = Sequ e nti a l() m odel.add(Dense ( 512, i np u t _ s hap e =(7 8 4 , ))) m odel.add(Activation('re l u')) m odel.add(Dro p o ut (0.2))

m odel.add(Dense ( 51 2 )) m od el. ad d( Ac t iva tio n ( 'r e l u ' )) m odel.add(Dro p o ut (0.2)) m odel. ad d(Dense ( 10 )) m odel.add(Act i va t i on('sof t m ax')) K era s ni n g a s o s i y t u s h un c h a la r i K era s d a g i a so s iy s truk tura bu m od e l. M od ellar b ir qa n cha q a t l a m lar d an t as h k i l t o p i s h i m u m kin . K er a s d a qat l a m larning bir ne c hta t uri m av j u d . ◦ Ket m a- k et tar ti bl i ( S e quent ia l ) m odel ◦ K o n v ol y u t s ion qa t l a m ◦ M a xP o oli n g q a t l am ◦ Zi c h ( De n s e ) qatl a m ◦ Dr o p o u t q a t l am

Ket m a-ket t a r t i b l i ( S e que n tial) m od e l Ket m a-ket t a r t ibli m ode ln i n g a so s iy g' o y asi shun c h a ki Ker a s q a t la m l ar i ni k e t m a - k et t a r t ib da jo y lash t i r is h dir. Ya ` n i q a tla m lar tu r g a n t ar ti bi b o ` y i c ha is h g a tush i r ila d i Kon v o l y utsion q a tl a m Kon v o l y utsion q a tl a m – b u m as s ivlar u st i da b ir ne c h t a f i l t r l a r d an f o y dal a n g a n h o l da ula r d a n tu r li xil xu s u s i y a t la r n i a j r a tib olish uc hu n qo`ll an ilad i g a n qa t la m di r Kon v o l y utsion q a tl a m ReLU-ni f a o ll as h ti r i s h fu n k t si y as i d a n f o y dal a n a dig a n va kirish shak l i(i n put s h a pe) 3 2 0 x 3 2 0 x3 b o ' lgan 4 8 t a 3x3 o'lcha m dagi f i lt r i bo`lg a n konvo l yutsi o n q at la m . i nput_sh a pe= ( 32 0 , 3 2 0, 3 ) #this is the i np u t sh a p e of an i m a ge 320 x 320 x 3 m odel.add( C o nv2D(48, (3, 3 ), ac t i va ti o n = ' r e l u', i np u t _ s h a pe= i n put_ s hape)) Kon v o l y utsion q a tl a m ning b oshqacha k o `rinishi: m odel. ad d(Conv2 D ( 4 8 , ( 3 , 3 ), a c t ivati o n = 're l u')) Ma x Po oling qatlam

Keras, Tensorflow va Torch kutubxonalari.

08.08.2023

0

2081.892578125 KB

Похожие