UGoogle ubhengeze ukukhutshwa kwekhowudi yomthombo we "TAPAS" (I-TABLE PARING), inethiwekhi ye-neural (kukubhadla okungeyonyani) iphuhliswe ngaphakathi ngenjongo yokuphendula umbuzo ngolwimi lwendalo kwaye ufumane impendulo kwi-database yobudlelwane okanye i-spreadsheet.
Ukuze ufumane iziphumo ezizezona zilungileyo kwi-TAPAS, Abaphuhlisi abaphethe iprojekthi bazinikele ekuqeqesheni inethiwekhi ye-neural kunye ne-6.2 yezigidi zezibini ukusuka kwitafile ukuya kwisicatshulwa esithathwe kwiWikipedia. Ukuqinisekisa, inethiwekhi ye-neural kwafuneka ibuyisele amagama angekhoyo zombini kwiitheyibhile nakwimibhalo engazange iqeqeshwe. Ukubuyiswa kokuchaneka kwaba yi-71,4% njengovavanyo lwebenchmark lubonise ukuba inethiwekhi ye-neural ibonelela ngeempendulo ezichanekileyo okanye ezithelekisekayo kune-algorithms ekhuphisanayo kuzo zontathu iiseti zedatha.
Malunga ne-TAPAS
Ngokusisiseko Ingqwalasela yale projekthi kukukwazi ukubonisana, inkqubo kunye nokubonisa ulwazi ezinxulumene nemigaqo yombuzo owenziwe ngumsebenzisi ngolwimi lwendalo, uququzelela ukufunyanwa kolwazi ngomlinganiselo omkhulu.
Umzekelo osisiseko wokusebenzisa i-TAPAS kukuba umsebenzisi ufuna ukuvavanya intengiso, ingeniso, izicelo, phakathi kwezinye izinto. Ukongeza, kufuneka kuthathelwe ingqalelo oko I-TAPAS ayiphelelanga nje ekufumaneni ulwazi kwiziko ledatha, kodwa ikwanakho ukwenza izibalo, i-algorithm ifuna impendulo kwiiseli zeetafile, zombini ngokuthe ngqo kunye nendlela yokongeza, i-avareji kunye nabanye abaqhubi, kwaye inokukhangela impendulo phakathi kweetafile ezininzi ngexesha elinye.
UGoogle uthi iTapas ibetha okanye idibanisa zontathu iialgorithms ezinkulu ezivulelekileyo ukuhlalutya idatha yobudlelwane. Isakhono seTapas sokukhupha izinto ezithile kwiindawo ezinkulu zokugcina idatha sinokuziboleka ekuphuculeni amandla okuphendula.
Ngaphantsi, I-Tapas isebenzisa ulwahlulo lwe-BERT yobuchule bokwenza ulwimi lwendalo isetyenziswe kukhangelo olwenziwe yinjini kaGoogle.
I-BERT ibonelela ngokuchaneka okuphezulu kuneendlela zemveli kuba ivumela i-AI ukuba ivavanye ukulandelelana kwesicatshulwa kungekhona nje ukusuka ekhohlo ukuya ekunene okanye ekunene ukuya ekhohlo njengesiqhelo, kodwa ukwenza zombini ngexesha elinye.
Inguqulelo ephunyezwe nguGoogle kwi-TAPAS ivumela i-AI ingathatheli ingqalelo kuphela umbuzo obuzwe ngabasebenzisi kunye nedatha abafuna ukuyibuza, kodwa kunye nesakhiwo seetafile zobudlelwane apho idatha igcinwa khona.
Uyifaka njani iTAPAS kwiLinux?
Ekubeni I-TAPAS ngokwesiseko yimodeli ye-BERT kwaye ke ineemfuno ezifanayo. Oku kuthetha ukuba imodeli enkulu inokuqeqeshwa ngokulandelelana ubude be-512 eya kufuna i-TPU.
Ukukwazi ukufaka i-TAPAS kwiLinux sifuna umqokeleli weprotoc, enokufumaneka kunikezelo oluninzi lweLinux.
KwiDebian, Ubuntu kunye nezinto eziphuma kuzo, sinokufaka umqokeleli ngalo myalelo ulandelayo:
sudo apt-get install protobuf-compiler
Kwimeko yeArch Linux, Manjaro, Arco Linux okanye nayiphi na enye into ephuma kwiArch Linux, siyifaka nge:
sudo pacman -S protobuf
Ngoku ukuze sikwazi ukufaka i-TAPAS, kufuneka sifumane ikhowudi yomthombo kwaye siyihlanganise nale miyalelo ilandelayo:
git clone https://github.com/google-research/tapas
cd tapas
pip install -e .
Kwaye ukuqhuba i-suite yovavanyo, sisebenzisa ilayibrari ye-tox enokuthi iqhutywe ngokufowuna:
pip install tox
tox
Ukusuka apha i-AI kuya kufuneka iqeqeshwe kwindawo yomdla. Nangona Ezinye iimodeli esele ziqeqeshiwe zinikezelwa kwindawo yokugcina iGitHub.
Ukongeza, ungasebenzisa iinketho zoqwalaselo ezahlukeneyo, ezifana nokhetho max_seq_ubude ukwenza ulandelelwano olufutshane. Oku kuya kunciphisa ukuchaneka kodwa kwakhona kwenza imodeli ye-GPU iqeqesheke. Enye inketho kukunciphisa ubungakanani bebhetshi (train_batch_size), kodwa oku kuya kuchaphazela ukuchaneka mhlawumbi.
Okokugqibela ukuba ufuna ukwazi ngakumbi ngayo Malunga nale AI, unokujongana neenkcukacha zokusetyenziswa, ukwenziwa kunye nolunye ulwazi Kule khonkco ilandelayo.