Isikhokelo esipheleleyo seKeras: yintoni kwaye isebenza njani

Informatec Digital » Intelligence Intelligence » Isikhokelo esipheleleyo seKeras: yintoni kwaye isebenza njani

I-Keras ibe lilitye lembombo lehlabathi lokufunda nzulu. Eli thala leencwadi, lenzelwe ukuququzelela ukuyilwa kunye noqeqesho lweemodeli zothungelwano lwe-neural, libalasele ngokulula kwalo, ukuguquguquka, kunye nojongano olubeka phambili amava omphuhlisi. Ngamanye amazwi, ivumela abasebenzisi bawo onke amanqanaba ezakhono ukuba bafunise ubuchule bokufunda obunzulu obunamandla ngaphandle kokujongana nobunzima obunxulunyaniswa nezikhokelo zomgangatho ophantsi.

Ukuba usaqala kwihlabathi lobukrelekrele bokwenziwa okanye sele uphonononga izicelo zalo kangangexesha elithile, mhlawumbi ukhe weva ngeKeras. Kweli nqaku, siza kukunika ukujonga nzulu kuyo. Yintoni iKeras, isebenza njani, ziziphi iingenelo ezibonelela ngazo, kweziphi iimeko ezisetyenziswayo kwaye kutheni kufuneka ufunde ukuyiqonda?.

Yintoni iKeras?

I-Keras lithala leencwadi elivulekileyo le ukwakhiwa kunye noqeqesho lwenethiwekhi ye-neural, ebhalwe kwiPython. Yazalwa ngo-2015 ngu UFrançois Chollet njengenxalenye yeprojekthi yophando I-ONEIROS (I-Open-end Neuro-Electronic Intelligent Robot Operating System) kwaye yenzelwe ukwenza lula ukusetyenziswa kwe-algorithms yokufunda nzulu.

Ifilosofi yayo isekelwe kwimodyuli, ukwandiswa kunye nokulula kokusetyenziswa.. Ngokungafaniyo nezinye iindawo zobugcisa, iKeras ibonelela nge ujongano olukwinqanaba eliphezulu abaphuhlisi abanokwakha ngayo imifuziselo entsonkothileyo ngaphandle kokukhathazeka malunga neenkcukacha zomgangatho ophantsi. Nangona yayiqale ngokuhambelana nezinto ezahlukeneyo ezingasemva ezifana neTheano okanye iMicrosoft Cognitive Toolkit (CNTK), namhlanje isebenza ngokuyintloko TensorFlow, ibe API yayo esemthethweni.

Isetyenziselwa ntoni iKeras?

Injongo ephambili yeKeras kukuququzelela uphuhliso lweemodeli ze Ukufunda nzulu, ukuvumela umsebenzisi ukuba agxininise kwingqiqo yemodeli endaweni yokusebenza kwangaphakathi kwesakhelo. Phakathi usetyenziso oluphambili lweKeras zi:

• Ukwahlula-hlula komfanekiso usebenzisa i-convolutional neural networks (CNN).
• Ukusetyenzwa kolwimi lwendalo kunye nohlalutyo lombhalo kunye ne-LSTM okanye iinethiwekhi zeTransformer.
• Uqikelelo lothotho lwexesha njengamaxabiso okanye idatha yemozulu.
• Ukudala iinkqubo zeengcebiso ukwenza amava omsebenzisi alungele wena.
• Ukunikezelwa koqeqesho ukuthatha ithuba leemo ezingqongileyo ngeGPU okanye iTPU.

Ndiyabulela kwinkxaso yeplatform eninzi, iimodeli zeKeras zinokuqhutywa kwizixhobo eziphathwayo (iOS, Android), kwiziphequluli zewebhu, nakwiindawo zamafu njengeGoogle Cloud ngeTensorFlow Serving.

Iimpawu eziphambili zeKeras

I-Keras igqame ngaphakathi kwi-artificial intelligence ecosystem enkosi kuthotho lwe iimpawu ezizodwa kunye nokwahlula:

1. Ujongano olulula nolucacileyo: Uyilo lwayo lwenzelwe ukuvumela abasebenzisi ukuba bakhe iimodeli ngokukhawuleza ngaphandle kokubhala inani elikhulu lekhowudi.

2. Iindlela zokwakha eziguquguqukayo: Ibonelela ngeendlela ezimbini zokuvelisa iimodeli: Sequential API, ilungele ukwakhiwa kwemigca, kunye I-API esebenzayo, evumela ukwakha iimodeli ezinzima ngakumbi njengalezo ezinamagalelo amaninzi okanye iziphumo.

3. Ukuhambelana kweqonga: Inokuqhutywa TensorFlow, CNTK, Theano kunye nezinye iimeko-bume, kuquka iinkqubo ezizinzisiweyo okanye iimeko zewebhu, usebenzisa amathala eencwadi afana neKeras.js okanye iWebDNN.

4. Inkxaso kwiindidi ezahlukeneyo zothungelwano: Ibandakanya izixhobo zokuyila Convolutional Neural Networks (CNN), Uthungelwano lweNeural oluQhelekileyo (RNN) okanye iimodeli ezivelisayo. Ikwaququzelela ukuguqulwa kwevidiyo, umbhalo, okanye idatha yemifanekiso.

5. Uqeqesho novavanyo olwenziwe lula: I-Keras ikuvumela ukuba ulawule ngokulula imisebenzi yokulahleka, izilungisi, kunye neemetrics. Ukusebenzisa imiyalelo efana fit, evaluate o predict, umsebenzisi unokuqeqesha kwaye avavanye imifuziselo ngamanyathelo nje ambalwa.

Izinto eziluncedo zokusebenzisa iKeras

Zininzi izibonelelo ezinikezelwa leli thala leencwadi, nokuba uyafunda okanye sele uligcisa kweli candelo:

• Igophe lokufunda eliphantsi: Kulula ukuyiqonda kunye nokusetyenziswa, nakubantu abanamava amancinane kwiSifundo soomatshini.
• Amaxwebhu acacileyo nabanzi: Inemizekelo emininzi, izifundo kunye neeforam zoncedo.
• Ilungele umboniso okhawulezayo: ikuvumela ukuba uvavanye izimvo ezintsha ngokukhawuleza.
• Amandla aqinisekisiweyo: Ngokudityaniswa ne-TensorFlow, uzuza njengelifa ubungakanani bayo kunye nokusebenza kwayo.
• Uluntu olusebenzayo: Inamawaka abasebenzisi kunye nabaphuhlisi kwihlabathi liphela abahlala benegalelo kuphuculo, iimodeli kunye nemizekelo.

Iimodeli ezongezelelweyo kunye nezixhobo

I-Keras ihamba ngaphaya kwe-API elula. Ikwabonisa izixhobo ezininzi ezakhelwe ngaphakathi ezikhawulezisa uphuhliso lwesisombululo.

Keras izicelo: Ingqokelela ye iimodeli eziqeqeshwe kwangaphambili kwi-ImageNet, kuquka iVGG16, VGG19, ResNet50, Xception, InceptionV3, DenseNet, kunye neNASNet. Ezi zinokusetyenziswa ngokuthe ngqo kuhlelo, ukutsalwa kweempawu, okanye njengesiseko sokufunda ngokugqithiselwayo.

KerasTuner: Ilungele i fumana ngokuzenzekelayo ezona zicwangciso zehyperparameterEsi sixhobo sikongela ixesha elininzi ngexesha lokulungisa imodeli.

KerasNLP: Ulwandiso olukhethekileyo lwe ukwenziwa kolwimi kwendalo. Ikuvumela ukuba usebenze ngeemodeli eziqeqeshwe kwangaphambili ezifana ne-BERT, i-RoBERTa, okanye i-GPT2, ukuququzelela ukuhlelwa, ukuguqulela, kunye nemisebenzi yokuhlalutya imvakalelo.

KerasCV: Yenzelwe iiprojekthi ze umbono ongeyonyani. Yenza lula umsebenzi kunye nokuhlelwa kwezinto, ulwahlulo lwemifanekiso, ukufunyanwa kwezinto, kunye nokuveliswa kwemifanekiso.

I-Keras isebenza njani inyathelo ngenyathelo?

Ubume bemodyuli ngundoqo ekusebenzeni kwayo. Nantsi indlela esebenza ngayo ngaphakathi:

1. Ukwakha imodeli

Umsebenzisi uchaza i-architecture usebenzisa i-APIs ezifana Ngokulandelelana o ezisebenzayoIileya zongezwa, amagalelo kunye neziphumo zichaziwe, kwaye idatha iya kuhamba njani ngemodeli isekiwe.

2. Ukuqulunqa imodeli

I-algorithm yokuphucula iqwalaselwe (isilungisi), umsebenzi welahleko (ukulahlekelwa) kunye neemetrics zovavanyo (iimetwork).

3. Uqeqesho

Indlela isetyenziswa fit() ukondla imodeli ngedatha yoqeqesho. Imijikelo yokufunda yenzeke apho imodeli ilungisa ubunzima bayo.

4. Uvandlakanyo

Con evaluate(), ukusebenza komzekelo kulinganiswa kwidatha engazange ibonwe ngaphambili, ukuchonga umthamo wayo jikelele.

5. Ukuqikelela

Enkosi kwindlela predict(), unokusebenzisa imodeli kwidatha entsha kwaye ufumane iziphumo eziphathekayo.

Imizekelo yesicelo

Ukwahlula-hlula komfanekiso

Ukusebenzisa imifuziselo efana neVGG16 okanye iResNet, unokwahlula imifanekiso ngokwamaqela ahlukeneyo (umzekelo, izilwanyana, iimpawu zendlela, okanye izimbo zobugcisa). Oku kungundoqo kumacandelo afana nokhuseleko, ukuthengisa, kunye namayeza.

Uqikelelo lwamaxabiso exesha elizayo

Ngamanethiwekhi afana ne-LSTM, unokuqikelela ixabiso lesitokhwe okanye ubushushu bexesha elizayo. Idla ngokusetyenziswa kwezemali, kwiklimatology, nakuhlalutyo lwemarike.

Ukuveliswa kombhalo

Yenza izivakalisi okanye imihlathi efana naleyo yombhali usebenzisa iimodeli eziqeqeshwe ezincwadini, amanqaku, okanye iincoko. Iluncedo kakhulu kwimidlalo, abancedisi benyani, okanye ukwenziwa komxholo ozenzekelayo.

Singatsho ukuba I-Keras inedemokhrasi yokusetyenziswa kweenethiwekhi ze-neuralEnkosi kwindlela esebenziseka lula, amawakawaka abantu akwazile ukungena kwihlabathi lokufunda ngoomatshini ngaphandle kokufuna iPhD kwimathematika. Ukudityaniswa kwayo kunye ne-TensorFlow, indlela yayo yeemodyuli, kunye nokukhula kwayo rhoqo kuyenza ibe sisixhobo sanamhlanje, esinamandla esinekamva.

Inqaku elidibeneyo:

Yintoni iTensorFlow kwaye ibuguqula njani ubukrelekrele bokwenziwa