Esinye sezona zixhobo zaziwayo zokuphuhlisa iimodeli zokufunda koomatshini yiTensorFlow. Sisebenzisa iTensorFlow kwizicelo ezininzi kumashishini ahlukeneyo.
Kule post, siza kujonga ezinye zeemodeli ze-TensorFlow AI. Ngenxa yoko, sinokwenza iinkqubo ezikrelekrele.
Siza kuphinda sihambe ngesakhelo esinikezelwa yiTensorFlow ekudaleni iimodeli ze-AI. Ngoko masiqalise!
Intshayelelo emfutshane yeTensorFlow
I-TensorFlow kaGoogle ngumthombo ovulekileyo yokufunda umatshini iphakheji yesoftware. Ibandakanya izixhobo zoqeqesho kunye nokuthunyelwa iimodeli zokufunda ngomatshini kumaqonga amaninzi. kunye nezixhobo, kunye nenkxaso yokufunda nzulu kunye amanethiwekhi.
I-TensorFlow yenza ukuba abaphuhlisi benze iimodeli zosetyenziso olwahlukeneyo. Oku kubandakanya ukuqondwa komfanekiso kunye nesandi, ukusetyenzwa kolwimi lwendalo, kunye umbono wekhompyutha. Sisixhobo esomeleleyo nesiguquguqukayo esinenkxaso yoluntu ngokubanzi.
Ukufakela iTensorFlow kwikhompyuter yakho ungachwetheza oku kwifestile yomyalelo wakho:
pip install tensorflow
Zisebenza Njani Iimodeli ze-AI?
Iimodeli ze-AI ziinkqubo zekhompyuter. Ke ngoko, zenzelwe ukwenza imisebenzi efuna ngokuqhelekileyo ingqondo yomntu. Ukuqaphela umfanekiso kunye nentetho kunye nokwenza izigqibo yimizekelo yemisebenzi enjalo. Iimodeli ze-AI ziphuhliswa kwiiseti zedatha ezinkulu.
Basebenzisa ubuchule bokufunda ngoomatshini ukwenza uqikelelo kunye nokwenza iintshukumo. Banosetyenziso oluninzi, kubandakanya iimoto eziziqhubayo, abancedisi bomntu, kunye noxilongo lwezonyango.
Ke, zeziphi iimodeli zeTensorFlow AI ezidumileyo?
ResNet
I-ResNet, okanye iResidual Network, luhlobo lwe-convolution inethiwekhi yomnatha. Siyisebenzisela ukuhlelwa kwemifanekiso kunye ukubona into. Yaphuhliswa ngabaphandi beMicrosoft ngo-2015. Kwakhona, ihluke kakhulu ngokusetyenziswa koqhagamshelwano oluseleyo.
Olu nxibelelwano luvumela uthungelwano ukuba lufunde ngempumelelo. Ngenxa yoko, kunokwenzeka ngokwenza ukuba ulwazi luhambe ngokukhululekileyo phakathi kweeleya.
I-ResNet inokuphunyezwa kwi-TensorFlow ngokusebenzisa i-Keras API. Ibonelela ngenqanaba eliphezulu, ujongano olusebenziseka lula lokudala nokuqeqesha uthungelwano lwe-neural.
Kuhlohlwa i-ResNet
Emva kokufaka iTensorFlow, ungasebenzisa iKeras API ukwenza imodeli yeResNet. I-TensorFlow ibandakanya i-Keras API, ngoko awudingi ukuyifaka ngabanye.
Ungangenisa imodeli ye-ResNet ukusuka kwi-tensorflow.keras.applications. Kwaye, unokukhetha inguqulelo yeResNet oza kuyisebenzisa, umzekelo:
from tensorflow.keras.applications import ResNet50
Unokusebenzisa le khowudi ilandelayo ukulayisha iintsimbi eziqeqeshwe kwangaphambili zeResNet:
model = ResNet50(weights='imagenet')
Ngokukhetha ipropathi include_top=False, ungasebenzisa imodeli yoqeqesho olongezelelweyo okanye ukulungisa kakuhle iseti yakho yedatha.
model = ResNet50(weights='imagenet', include_top=False, input_shape=(224, 224, 3))
IiNdawo zosetyenziso zeResNet
I-ResNet ingasetyenziswa kuhlelo lwemifanekiso. Ke, unokwahlulahlula iifoto ngokwamaqela amaninzi. Okokuqala, kufuneka uqeqeshe imodeli ye-ResNet kwiseti yedatha enkulu yeefoto ezibhaliweyo. Ke, iResNet inokuqikelela iklasi yemifanekiso engazange ibonwe ngaphambili.
I-ResNet isenokusetyenziselwa imisebenzi yokubona izinto ezinjengokubona izinto kwiifoto. Singakwenza oku ngokuqeqesha kuqala imodeli ye-ResNet kwingqokelela yeefoto ezibhalwe ngeebhokisi zokubopha izinto. Emva koko, sinokusebenzisa imodeli efundiweyo ukubona izinto kwimifanekiso emitsha.
Sinokusebenzisa i-ResNet kwimisebenzi yolwahlulo lwesemantic. Ke, sinokwabela ileyibhile yesemantic kwipixel nganye emfanekisweni.
liqala
Ukuqaliswa yimodeli yokufunda enzulu ekwaziyo ukuqaphela izinto kwimifanekiso. UGoogle wayibhengeza ngo-2014, kwaye ihlalutya imifanekiso yeesayizi ezahlukeneyo isebenzisa iileya ezininzi. Nge-Inception, imodeli yakho inokuqonda umfanekiso ngokuchanekileyo.
I-TensorFlow sisixhobo esomeleleyo sokwenza kunye nokusebenzisa iimodeli zokuQala. Ibonelela ngenqanaba eliphezulu kunye nojongano olusebenzisekayo loqeqesho kuthungelwano lwe-neural. Ke ngoko, i-Inception yimodeli ethe ngqo yokufaka isicelo kubaphuhlisi.
Ukufakela Ukuqaliswa
Ungayifaka i-Inception ngokuchwetheza lo mgca wekhowudi.
from tensorflow.keras.applications import InceptionV3
IiNdawo zokuQala zokuSebenzisa
Imodeli yokuQala inokusetyenziselwa ukutsala amanqaku kuyo ukufunda okunzulu iimodeli ezifana ne-Generative Adversarial Networks (GANs) kunye ne-Autoencoders.
Imodeli yokuQala inokulungiswa kakuhle ukuchonga iimpawu ezithile. Kwakhona, sinokukwazi ukuxilonga ukuphazamiseka okuthile kwizicelo zokucinga zonyango ezifana ne-X-reyi, i-CT, okanye i-MRI.
Imodeli yokuQala inokulungiswa kakuhle ukujonga umgangatho womfanekiso. Singavavanya ukuba ingaba umfanekiso ufuze okanye ukrwada.
Ukuqala kungasetyenziselwa imisebenzi yohlalutyo lwevidiyo njengokulandela umkhondo kunye nokubonwa kwesenzo.
BHALA
I-BERT (i-Bidirectional Encoder Representations ezivela kwiTransformers) yimodeli yenethiwekhi ye-neural ephuhliswe kwangaphambili eqeqeshwe nguGoogle. Sinokuyisebenzisa kwimisebenzi eyahlukeneyo yolwimi lwendalo. Le misebenzi inokwahluka ukusuka kuhlelo lweetekisi ukuya ekuphenduleni imibuzo.
I-BERT yakhiwe kwi-architecture ye-transformer. Ke ngoko, unokuphatha umthamo omkhulu wegalelo lombhalo ngelixa uqonda udibaniso lwamagama.
I-BERT ngumzekelo oqeqeshwe kwangaphambili onokuthi udibanise kwizicelo ze-TensorFlow.
I-TensorFlow ibandakanya imodeli ye-BERT eqeqeshwe kwangaphambili kunye nengqokelela yezixhobo zokulungisa kakuhle kunye nokusebenzisa i-BERT kwimisebenzi eyahlukeneyo. Ke ngoko, unokudibanisa ngokulula ubuchule be-BERT bokusetyenzwa kolwimi lwendalo.
Kuhlohlwa i-BERT
Usebenzisa umphathi wepakethe yepip, ungafaka i-BERT kwiTensorFlow:
pip install tensorflow-gpu==2.2.0 # This installs TensorFlow with GPU support
pip install transformers==3.0.0 # This installs the transformers library, which includes BERT
Inguqulelo ye-CPU ye-TensorFlow inokufakwa ngokulula ngokufaka endaweni ye-tensorflow-gpu nge-tensorflow.
Emva kokufaka ithala leencwadi, unokungenisa imodeli ye-BERT kwaye uyisebenzisele imisebenzi eyahlukeneyo ye-NLP. Nantsi ikhowudi yesampulu yokulungisa kakuhle imodeli ye-BERT kwingxaki yohlelo lokubhaliweyo, umzekelo:
from transformers import BertForSequenceClassification
# Load the pre-trained BERT model
model = BertForSequenceClassification.from_pretrained("bert-base-uncased")
# Fine-tune the model on your text classification task
model.fit(training_data, labels)
# Make predictions on new data
predictions = model.predict(test_data)
IiNdawo zokusetyenziswa kwe-BERT
Unokwenza imisebenzi yokuhlelwa kwesicatshulwa. Ngokomzekelo, kunokwenzeka ukufezekisa Uhlalutyo lweemvakalelo, ukuhlelwa kwesihloko, kunye nokubhaqwa kwespam.
I-BERT ine- Ukwamkelwa kweQumrhu elinamagama (NER) uphawu. Ke, unokwazi kwaye uleyibhelishe amaqumrhu kwisicatshulwa esinje ngabantu kunye nemibutho.
Ingasetyenziselwa ukuphendula imibuzo ngokuxhomekeke kumxholo othile, njenge-injini yokukhangela okanye i-chatbot application.
I-BERT inokuba luncedo kwiNguqulelo yoLwimi ukunyusa ukuchaneka kwenguqulelo koomatshini.
I-BERT ingasetyenziselwa ukushwankathela okubhaliweyo. Ngenxa yoko, inokunika isishwankathelo esifutshane, esiluncedo samaxwebhu amade esicatshulwa.
Ilizwi elinzulu
Uphando lwe-Baidu lwenze i-DeepVoice, a Isicatshulwa esiya kwincoko imodeli yokudibanisa.
Yenziwe ngesakhelo seTensorFlow kwaye yaqeqeshwa kwingqokelela enkulu yedatha yelizwi.
I-DeepVoice ivelisa ilizwi kwigalelo lombhalo. I-DeepVoice yenza kube nokwenzeka ngokusebenzisa ubuchule bokufunda obunzulu. Yimodeli esekwe kwinethiwekhi ye-neural.
Ke ngoko, ihlalutya idatha yegalelo kwaye ivelise intetho isebenzisa inani elikhulu lamaleko eendawo eziqhagamshelweyo.
Kufakelwa i-DeepVoice
!pip install deepvoice
Kungenjalo;
# Clone the DeepVoice repository
!git clone https://github.com/r9y9/DeepVoice3_pytorch.git
%cd DeepVoice3_pytorch
!pip install -r requirements.txt
IiNdawo zokuSebenza ze-DeepVoice
Ungasebenzisa i-DeepVoice ukuvelisa intetho kubancedisi bomntu njengeAmazon Alexa kunye noMncedisi kaGoogle.
Kwakhona, i-DeepVoice ingasetyenziselwa ukuvelisa intetho kwizixhobo ezisebenza ngelizwi njengezithethi ezihlakaniphileyo kunye neenkqubo ezizisebenzelayo zasekhaya.
I-DeepVoice inokudala ilizwi kwizicelo zonyango lwentetho. Inokunceda abaguli abanengxaki yokuthetha baphucule indlela abathetha ngayo.
I-DeepVoice isenokusetyenziswa ukwenza intetho yemathiriyeli yokufundisa njengee-audiobooks kunye nee-apps zokufunda ulwimi.
Shiya iMpendulo