Tafole ea likateng[Pata][Bontša]
Lilemong tsa morao tjena, marang-rang a neural a ntse a tumme kaha a bonts'itse hore a sebetsa hantle haholo mesebetsing e mengata e fapaneng.
Li bonts'itsoe e le khetho e ntle bakeng sa ho amohela litšoantšo le molumo, ho sebetsana le puo ea tlhaho, esita le ho bapala lipapali tse thata tse kang Go le chess.
Ka poso ena, ke tla u tsamaisa ts'ebetsong eohle ea ho koetlisa neural network. Ke tla bolela le ho hlalosa mehato eohle ea ho koetlisa neural network.
Ha ke ntse ke tla feta mehato eo nka ratang ho eketsa mohlala o bonolo ho etsa bonnete ba hore ho na le mohlala o sebetsang hape.
Kahoo, tloo, 'me re ithute ho sebetsana le marang-rang a neural
Ha re qaleng ka bonolo 'me re botse hore na ke eng marangrang a neural sebakeng sa pele.
Hantle-ntle Neural Networks ke Eng?
Marang-rang a Neural ke komporo ea komporo e etsisang tšebetso ea boko ba motho. Ba ka ithuta ho tsoa mefuteng e mengata ea data le mekhoa ea bona eo batho ba ka fumanang ho le thata ho e lemoha.
Marang-rang a Neural a se a tumme lilemong tsa morao tjena ka lebaka la ho feto-fetoha ha 'ona mesebetsing e kang ho lemoha litšoantšo le molumo, ho sebetsana le puo ea tlhaho, le ho etsa mohlala oa pele.
Ka kakaretso, marang-rang a neural ke sesebelisoa se matla bakeng sa mefuta e mengata ea lits'ebetso mme re na le monyetla oa ho fetola tsela eo re atamelang mefuta e mengata ea mesebetsi.
Ke Hobane’ng ha re Lokela ho Tseba ka Tsona?
Ho bohlokoa ho utloisisa marang-rang a neural hobane a lebisitse ho sibollotsoe mafapheng a fapaneng, ho kenyeletsoa pono ea komporo, temoho ea puo, le ts'ebetso ea puo ea tlhaho.
Ka mohlala, marang-rang a marang-rang a bohareng ba tsoelo-pele ea morao tjena likoloing tse itsamaisang, litšebeletso tsa phetolelo ea othomathike, esita le tlhahlobo ea bongaka.
Ho utloisisa hore na marang-rang a neural a sebetsa joang le hore na a ka a qapa joang ho re thusa ho theha lits'ebetso tse ncha tse ncha. 'Me, mohlomong, e ka lebisa ho litšibollo tse kholoanyane nakong e tlang.
Tlhokomeliso Mabapi le Thupelo
Joalokaha ke boletse ka holimo, ke rata ho hlalosa mehato ea ho koetlisa neural network ka ho fana ka mohlala. Ho etsa sena, re lokela ho bua ka dataset ea MNIST. Ke khetho e tsebahalang bakeng sa ba qalang ba batlang ho qala ka marang-rang a neural.
MNIST ke khutsufatso e emelang Modified National Institute of Standards and Technology. Ke pokello ea lintlha tse ngotsoeng ka letsoho tse atisang ho sebelisoa ho koetlisa le ho leka mekhoa ea ho ithuta ka mochini, haholo li-neural network.
Pokello e na le linepe tse 70,000 tsa grayscale tsa lipalo tse ngotsoeng ka letsoho ho tloha ho 0 ho isa ho 9.
Lethathamo la lintlha tsa MNIST ke letšoao le tsebahalang la tlhophiso ea litšoantšo mesebetsi. E sebelisoa khafetsa bakeng sa ho ruta le ho ithuta kaha e kopane 'me ho bonolo ho sebetsana le eona empa e ntse e hlahisa phephetso e boima bakeng sa li-algorithms tsa ho ithuta ka mochini ho araba.
Lenane la boitsebiso la MNIST le tšehetsoa ke meralo e mengata ea ho ithuta ka mochini le lilaebrari, ho kenyeletsoa TensorFlow, Keras, le PyTorch.
Joale re tseba ka MNIST dataset, ha re qaleng ka mehato ea rona ea ho koetlisa neural network.
Mehato ea Motheo ea ho Koetlisa Neural Network
Kenya Lilaebrari tse Hlokehang
Ha u qala ho koetlisa marang-rang a neural, ho bohlokoa ho ba le lisebelisoa tse hlokahalang ho rala le ho koetlisa mohlala. Mohato oa pele oa ho theha marang-rang a neural ke ho kenya lilaebrari tse hlokahalang tse kang TensorFlow, Keras, le NumPy.
Lilaebrari tsena li sebetsa e le litšiea tsa kaho bakeng sa nts'etsopele ea neural network le ho fana ka bokhoni ba bohlokoa. Motsoako oa lilaebrari tsena o lumella ho theha meralo e tsoetseng pele ea marang-rang a methapo le koetliso e potlakileng.
Ho qala mohlala oa rona; re tla tlisa lilaebrari tse hlokahalang, tse kenyelletsang TensorFlow, Keras, le NumPy. TensorFlow ke moralo o bulehileng oa ho ithuta oa mochini, Keras ke API ea neural network ea boemo bo holimo, 'me NumPy ke laeborari ea lipalo ea Python ea komporo.
import tensorflow as tf
from tensorflow import keras
import numpy as np
Kenya Dataset
Hona joale dataset e tlameha ho kenngoa. Dataset ke sete ea data eo marang-rang a neural a tla koetlisoa ho eona. Ena e ka ba mofuta ofe kapa ofe oa data, ho kenyeletsoa linepe, molumo, le mongolo.
Ho bohlokoa ho arola dataset ka likarolo tse peli: e 'ngoe bakeng sa koetliso ea neural network le e' ngoe bakeng sa ho lekola ho nepahala ha mohlala o koetlisitsoeng. Lilaebrari tse 'maloa, ho kenyeletsoa TensorFlow, Keras, le PyTorch, li ka sebelisoa ho kenya dataset.
Mohlala oa rona, re sebelisa le Keras ho kenya dataset ea MNIST. Ho na le linepe tsa koetliso tse 60,000 le litšoantšo tsa liteko tse 10,000 lethathamong la data.
mnist = keras.datasets.mnist
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()
Fetola Lintlha
Ts'ebetso ea data preprocessing ke mohato oa bohlokoa oa ho koetlisa neural network. E kenyelletsa ho lokisa le ho hloekisa data pele e kenngoa ho neural network.
Ho eketsa boleng ba pixel, ho tloaeleha ha data, le ho fetolela lileibole ho khouto e le 'ngoe ke mehlala ea lits'ebetso tsa ho lokisa esale pele. Mekhoa ena e thusa neural network ho ithuta ka katleho le ka nepo.
Ho lokisa data esale pele ho ka thusa ho fokotsa ho fetella le ho ntlafatsa ts'ebetso ea neural network.
U tlameha ho lokisa lintlha pele u koetlisa neural network. Sena se kenyelletsa ho fetola lileibole ho khouto e le 'ngoe e chesang le ho lekanya boleng ba pixel hore e be lipakeng tsa 0 le 1.
train_images = train_images / 255.0
test_images = test_images / 255.0
train_labels = keras.utils.to_categorical(train_labels, 10)
test_labels = keras.utils.to_categorical(test_labels, 10)
Hlalosa Mohlala
Mokhoa oa ho hlalosa mokhoa oa marang-rang oa neural o kenyelletsa ho theha mohaho oa eona, joalo ka palo ea lihlopha, palo ea li-neurons ka lera, mesebetsi ea ts'ebetso, le mofuta oa marang-rang (feedfor, recurrent, or convolutional).
Moralo oa neural network oo u o sebelisang o khethoa ke mofuta oa bothata boo u lekang ho bo rarolla. Moralo o hlophisitsoeng hantle oa marang-rang oa neural o ka thusa ho ithuta ka marang-rang ka ho etsa hore e sebetse hantle le ho nepahala.
Ke nako ea ho hlalosa mofuta oa neural network ntlheng ena. Sebelisa mohlala o bonolo o nang le lihlopha tse peli tse patiloeng, e 'ngoe le e' ngoe e na le li-neuron tsa 128, le sekhahla sa softmax se hlahisoang, se nang le li-neurons tsa 10, mohlala ona.
model = keras.Sequential([
keras.layers.Flatten(input_shape=(28, 28)),
keras.layers.Dense(128, activation='relu'),
keras.layers.Dense(128, activation='relu'),
keras.layers.Dense(10, activation='softmax')
])
Kopanya Mohlala
Ts'ebetso ea tahlehelo, optimizer, le metrics li tlameha ho hlalosoa nakong ea ho bokella mohlala oa neural network. Bokhoni ba neural network ba ho lepa sephetho ka nepo bo lekantsoe ke ts'ebetso ea tahlehelo.
Ho eketsa ho nepahala ha netweke ea neural nakong ea boikoetliso, optimizer e fetola boima ba eona. Ho sebetsa ha marang-rang a methapo nakong ea koetliso ho lekoa ho sebelisoa metrics. Mohlala o tlameha ho etsoa pele marang-rang a neural a ka koetlisoa.
Mohlala oa rona, re tlameha hona joale ho aha mohlala.
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
Koetlisa Mohlala
Ho fetisa dataset e lokisitsoeng ka marang-rang a marang-rang ha u ntse u fetola litekanyo tsa marang-rang ho fokotsa ts'ebetso ea tahlehelo ho tsejoa e le ho koetlisa marang-rang a marang-rang.
Lintlha tsa netefatso li sebelisetsoa ho lekola marang-rang a neural nakong ea koetliso ho latela katleho ea ona le ho thibela ho feta. Ts'ebetso ea koetliso e ka nka nako, ka hona ho bohlokoa ho etsa bonnete ba hore marang-rang a neural a koetlisoa ka nepo ho thibela ho se sebetse hantle.
Re sebelisa lintlha tsa koetliso, joale re ka koetlisa mohlala. Ho etsa sena, re tlameha ho hlalosa boholo ba batch (palo ea lisampole tse sebetsitsoeng pele mohlala o nchafatsoa) le palo ea linako (palo ea ho pheta-pheta ho dataset e felletseng).
model.fit(train_images, train_labels, epochs=10, batch_size=32)
Ho Lekola Mohlala
Ho lekola ts'ebetso ea netweke ea neural ho dataset ea liteko ke mokhoa oa ho e lekola. Mothating ona, marang-rang a koetlisitsoeng a neural a sebelisoa ho sebetsana le dataset ea liteko, 'me ho nepahala hoa hlahlojoa.
Hore na marang-rang a methapo a ka bolela esale pele sephetho se nepahetseng ho tsoa ho data e ncha, e sa lekoang ke tekanyo ea ho nepahala ha eona. Ho sekaseka mohlala ho ka thusa ho tseba hore na marang-rang a neural a sebetsa hantle hakae le ho fana ka maikutlo a mekhoa ea ho e ntlafatsa le ho feta.
Qetellong re ka hlahloba ts'ebetso ea mohlala re sebelisa lintlha tsa tlhahlobo ka mor'a koetliso.
test_loss, test_acc = model.evaluate(test_images, test_labels)
print('Test accuracy:', test_acc)
Ke phetho! Re koetlisitse marang-rang a marang-rang ho bona linomoro ho dataset ea MNIST.
Ho tloha ho lokisetsa lintlha ho ea ho hlahloba katleho ea mohlala o koetlisitsoeng, koetliso ea marang-rang ea methapo e kenyelletsa mekhoa e mengata. Litaelo tsena li thusa ba qalang ho aha le ho koetlisa marang-rang a neural.
Ba qalang ba batlang ho sebelisa marang-rang a neural ho sebetsana le litaba tse fapaneng ba ka etsa joalo ka ho latela litaelo tsena.
Ho Bona Mohlala ka Monahano
A re lekeng ho bona ka mahlo a kelello seo re se entseng ka mohlala ona ho utloisisa hamolemo.
Sephutheloana sa Matplotlib se sebelisoa snippet ea khoutu ena ho rala khetho e sa reroang ea linepe ho tsoa pokellong ea data ea koetliso. Taba ea pele, re kenya module ea "pyplot" ea Matplotlib mme re e bitsa "plt". Joale, ka boholo ba lisenthimithara tse 10 ho isa ho tse 10, re etsa setšoantšo se nang le mela e 5 le litšiea tse 5 tsa subplots.
Ebe, re sebelisa for loop ho pheta-pheta li-subplots, re bonts'a setšoantšo se tsoang ho dataset ea koetliso ho e 'ngoe le e 'ngoe. Ho hlahisa setšoantšo, ho sebelisoa mosebetsi oa "imshow", 'me khetho ea "cmap" e behiloe ho 'grey' ho hlahisa linepe ka mokhoa o moputsoa. Sehlooho sa karolo ka 'ngoe se boetse se behiloe ho leibole ea setšoantšo se amanang le pokello.
Qetellong, re sebelisa mosebetsi oa "show" ho bontša litšoantšo tse reriloeng setšoantšong. Ts'ebetso ena e re lumella ho lekola ka mahlo sampole ea linepe ho tsoa pokellong ea data, e ka re thusang ho utloisisa datha le ho tsebahatsa mathata afe kapa afe a ka bang teng.
import matplotlib.pyplot as plt
# Plot a random sample of images
fig, axes = plt.subplots(nrows=5, ncols=5, figsize=(10,10))
for i, ax in enumerate(axes.flat):
ax.imshow(train_images[i], cmap='gray')
ax.set_title(f"Label: {train_labels[i].argmax()}")
ax.axis('off')
plt.show()
Mehlala ea Bohlokoa ea Neural Network
- Feedforward Neural Networks (FFNN): Mofuta o bonolo oa marang-rang a marang-rang ao boitsebiso bo tsamaeang feela ka tsela e le 'ngoe, ho tloha ho lera la ho kenya letsoho ho ea ho lera la tlhahiso ka lera le le leng kapa tse ngata tse patiloeng.
- Convolutional Neural Networks (CNN): Neural network e sebelisoang hangata ho lemoha le ho lokisa litšoantšo. Li-CNN li reretsoe ho lemoha le ho ntša likarolo litšoantšong ka bo eona.
- Marang-rang a Neural Networks (RNN): Neural network e sebelisoang hangata ho lemoha le ho lokisa litšoantšo. Li-CNN li reretsoe ho lemoha le ho ntša likarolo litšoantšong ka bo eona.
- Marang-rang a Memori ea Nako e Khutšoane (LSTM): Mofuta oa RNN o etselitsoe ho hlola taba ea ho nyamela li-gradient ho li-RNN tse tloaelehileng. Litšepeho tsa nako e telele tsa data tse latellanang li ka nkuoa hamolemo ka li-LSTM.
- Li-autoencoder: Ho ithuta ho sa laoleheng neural network eo marang-rang a rutoang ho hlahisa lintlha tsa eona tse kentsoeng sebakeng sa eona sa tlhahiso. Khatello ea data, ho lemoha ka mokhoa o sa hlakang, le ho hlakola litšoantšo kaofela li ka etsoa ka li-autoencoder.
- Maqhubu a Hlahisang Adversarial (GAN): A generative neural network ke mofuta oa neural network o rutoang ho hlahisa data e ncha e ka bapisoang le dataset ea koetliso. Li-GAN li entsoe ka marang-rang a mabeli: marang-rang a jenereithara a etsang data e ncha le marang-rang a khethollo a hlahlobang boleng ba data e entsoeng.
Qetella, Mehato ea Hao e Latelang e Lokela ho ba Efe?
Lekola lisebelisoa le lithuto tse 'maloa tsa inthaneteng ho ithuta haholoanyane ka koetliso ea neural network. Ho sebetsa mererong kapa mehlala ke mokhoa o mong oa ho utloisisa hantle marang-rang a neural.
Qala ka mehlala e bonolo joalo ka mathata a likarolo tsa binary kapa mesebetsi ea ho arola litšoantšo, ebe u ea mesebetsing e thata joalo ka puo ea tlhaho kapa ho matlafatsa thuto.
Ho sebetsa mererong ho u thusa ho fumana boiphihlelo ba 'nete le ho ntlafatsa tsebo ea hau ea koetliso ea neural network.
U kanna ua kenela ho ithuta ka mochini oa inthaneteng le lihlopha tsa marang-rang a marang-rang le liforamu ho sebelisana le baithuti le litsebi tse ling, ho arolelana mosebetsi oa hau, le ho amohela maikutlo le thuso.
LSRS MONRAD-KROHN
⁶ĵKe ka be ke ratile ho bona lenaneo la python bakeng sa phokotso ea liphoso. Li-node tse khethehileng tsa khetho bakeng sa liphetoho tsa boima ho lera le latelang