I-Google ikhiphe ikhodi yomthombo ye-AI yayo "TAPAS"

 

I-Google imemezele ukukhishwa kwekhodi yomthombo ye- "TAPAS" (I-TAble PArSing), inethiwekhi ye-neural (ukuhlakanipha okungekhona okwangempela) ithuthukiswe ngaphakathi ukuze kuphendulwe umbuzo ngolimi lwemvelo bese uthola impendulo kusuka ku-database yobudlelwano noma ispredishithi.

Ukuze uthole imiphumela emihle ku-TAPAS, onjiniyela abaphethe iphrojekthi bazinikele ekuqeqesheni inethiwekhi ye-neural enamabhangqa ayisigidi esingu-6.2 table to text taken from Wikipedia. Ukuqinisekisa, inethiwekhi ye-neural kwakudingeka ibuyise amagama alahlekile kuwo womabili amatafula nasemibhalweni eyayingaqeqeshiwe kuyo. Ukunemba kokutakula kwakungu-71,4% njengokuhlolwa kwebenchmark kukhombisile ukuthi inethiwekhi ye-neural inikeza izimpendulo ezinembile noma eziqhathaniswayo kunama-algorithms ayizimbangi kuwo womathathu amasethi wedatha.

Mayelana neTAPAS

Ngokuyisisekelo okugxilwe kulo msebenzi ukukwazi ukubonisana, ukucubungula kanye nokuveza ulwazi ezihlobene nemibandela yombuzo eyenziwe ngumsebenzisi ngolimi lwemvelo, okwenza kube lula ukuthola ulwazi.

Isibonelo esiyisisekelo sokusetshenziswa kwe-TAPAS uma umsebenzisi efuna ukuhlola idatha yokuthengisa, imali engenayo, izicelo, phakathi kwezinye izinto. Ngaphandle kwalokho kufanele ukubheke lokho I-TAPAS ayigcini ngokuthola imininingwane ku-database kuphela, kepha futhi iyakwazi nokwenza izibalo, i-algorithm ibheka impendulo kumaseli wetafula, kokubili ngqo nangokusebenzisa ukuhlanganisa, isilinganiso kanye nabanye opharetha, ngaphandle kwalokho ingafuna impendulo phakathi kwamatafula amaningi ngasikhathi sinye.

IGoogle ithi iTapas Outperforms noma ifanisa ama-algorithms aphezulu amathathu womthombo ovulekile ukuhlaziya idatha ephathelene nobudlelwano. Amandla kaTapas okukhipha izinto ezithile kusuka ezinqolobaneni ezinkulu zedatha nazo zingaziboleka ekuthuthukiseni amandla okuphendula.

Ngaphansi, I-Tapas isebenzisa ukwehluka kwe-BERT inqubo yolimi yemvelo isetshenziswe ekusesheni okwenziwe yinjini yeGoogle.

I-BERT inikeza ukucacisa okukhulu kunezindlela zendabuko ngoba ivumela i-AI ukuthi ihlole ukulandelana kombhalo hhayi kuphela ukusuka kwesobunxele kuye kwesokudla noma kwesokudla kuye kwesobunxele njengokujwayelekile, kepha kwenza ngasikhathi sinye.

Uhlobo olwenziwe i-Google yi-TAPAS luvumela i-AI ukuthi ingabheki kuphela umbuzo obuzwe abasebenzisi nedatha abafuna ukuyibuza, kepha futhi nesakhiwo samatafula obudlelwano lapho idatha igcinwa khona.

Ungayifaka kanjani i-TAPAS ku-Linux?

Kusukela I-TAPAS empeleni iyimodeli ye-BERT ngakho-ke inezidingo ezifanayo. Lokhu kusho ukuthi imodeli enkulu ingaqeqeshwa ngobude bokulandelana obungu-512 obuzodinga i-TPU.

Ukuze ukwazi ukufaka i-TAPAS ku-Linux sidinga i-protocol compiler, engatholakala ekusatshalalisweni okuningi kweLinux.

Ku-Debian, Ubuntu nokutholakala kwalokhu, singafaka umhlanganisi ngomyalo olandelayo:

sudo apt-get install protobuf-compiler

Endabeni ye-Arch Linux, Manjaro, Arco Linux noma enye into etholakala ku-Arch Linux, sifaka ne:

sudo pacman -S protobuf

Manje ukuze sikwazi ukufaka i-TAPAS, kufanele nje sithole ikhodi yomthombo bese sihlanganisa nemiyalo elandelayo:

git clone https://github.com/google-research/tapas
cd tapas
pip install -e .

Futhi ukusebenzisa i-suite yokuhlola, sisebenzisa umtapo wezincwadi we-tox ongasetshenziswa ngokubiza:

pip install tox
tox

Ukusuka lapha i-AI kuzofanele iqeqeshwe endaweni yesithakazelo. Noma amamodeli athile aqeqeshiwe ahlinzekwa endaweni yokugcina izinto zeGitHub.

Ngaphezu kwalokho, ungasebenzisa izinketho ezahlukahlukene zokumisa, njengokukhetha ubude obungu-max ukudala ukulandelana okufushane. Lokhu kuzonciphisa ukunemba kepha futhi kuzokwenza imodeli ye-GPU iqeqeshwe. Enye indlela ukunciphisa usayizi we-batch (train_batch_size), kepha lokhu kungenzeka kuthinte nokunemba.

Ekugcineni uma ufuna ukwazi kabanzi ngayo Mayelana nale AI, ungabheka imininingwane yokusetshenziswa, ukwenziwa nolunye ulwazi Kulesi sixhumanisi esilandelayo.


Okuqukethwe yi-athikili kunamathela ezimisweni zethu ze izimiso zokuhlelela. Ukubika iphutha chofoza lapha.

Yiba ngowokuqala ukuphawula

Shiya umbono wakho

Ikheli lakho le ngeke ishicilelwe.

*

*

  1. Ubhekele imininingwane: Miguel Ángel Gatón
  2. Inhloso yedatha: Lawula Ugaxekile, ukuphathwa kwamazwana.
  3. Ukusemthethweni: Imvume yakho
  4. Ukuxhumana kwemininingwane: Imininingwane ngeke idluliselwe kubantu besithathu ngaphandle kwesibopho esisemthethweni.
  5. Isitoreji sedatha: Idatabase ebanjwe yi-Occentus Networks (EU)
  6. Amalungelo: Nganoma yisiphi isikhathi ungakhawulela, uthole futhi ususe imininingwane yakho.

i-bool(iqiniso)