ukukhutshwa okubalulekileyo kweqonga lokufunda lomatshini , ebonelela ngokuphunyezwa kokulungelelaniswa kweendlela ezahlukeneyo zokufunda koomatshini obunzulu, i-interface yeprogram elula yeemodeli zokwakha kwiPython, kunye ne-interface yezinga eliphantsi lolwimi lwe-C ++ oluvumela ukuba ulawule ukwakhiwa kunye nokwenziwa kweegrafu zokubala. Ikhowudi yenkqubo ibhalwe kwiC ++ kunye nePython kunye phantsi kwelayisensi ye-Apache.
Iqonga laphuhliswa liqela leGoogle Brain kwaye lisetyenziswa kwiinkonzo zikaGoogle zokuqondwa kwentetho, ukuchonga ubuso kwiifoto, ukugqiba ukufana kwemifanekiso, ukucoca ugaxekile kwiGmail, Iindaba kwiiNdaba zikaGoogle kunye nokulungelelanisa ukuguqulelwa kuthathelwa ingqalelo intsingiselo. Iinkqubo zokufunda koomatshini abasasazwayo zinokudalwa kwi-hardware eqhelekileyo, ngenxa yenkxaso eyakhelwe-ngaphakathi ye-TensorFlow yokusasaza izibalo kwii-CPU ezininzi okanye ii-GPU.
I-TensorFlow ibonelela ngethala leencwadi le-algorithms yokubala yamanani esele yenziwe ngokusetyenziswa kweegrafu zedatha. Iindawo ezikwiigrafu ezilolo hlobo ziphumeza imisebenzi yezibalo okanye amanqaku egalelo/imveliso, ngelixa imiphetho yegrafu imele uluhlu lwedatha olune-multidimensional (i-tensors) ehamba phakathi kweenodi.
I-Nodes inokwabelwa kwizixhobo ze-computing kwaye iqhutywe ngokulinganayo, ngaxeshanye ukucutshungulwa kwazo zonke ii-theors ezifanelekileyo kubo kanye, okwenza kube lula ukulungelelanisa ukusebenza kwangaxeshanye kwee-nodes kwinethiwekhi ye-neural ngokufanisa kunye nokusebenza kwangaxeshanye kwee-neurons kwingqondo.
Ingqwalasela ephambili ekulungiseleleni inguqulelo entsha yayikukwenza lula kunye nokulula ukuyisebenzisa. :
- I-API entsha yezinga eliphezulu iye yacetywayo yokwakha kunye noqeqesho lwemizekelo , enika iinketho ezininzi zojongano lweendlela zokwakha (Ulandelelwano, oluSebenzayo, uHlelo olungaphantsi) ngokukwazi uku (ngaphandle kokuhlanganiswa kwangaphambili) kunye nesixhobo esilula sokucoca;
- I-API eyongeziweyo ukwenzela umbutho iimodeli ezinotshintsho oluncinci kwikhowudi ekhoyo. Ukongeza kwithuba lokusasaza izibalo kulo lonke , inkxaso yovavanyo iyafumaneka ukwahlula inkqubo yokufunda ibe ziiprosesa ezininzi ezizimeleyo kunye nokukwazi ukusebenzisa ilifu (Iyunithi yokulungisa i-Tensor);
- Endaweni yemodeli ebhengezayo yokwakha igrafu ngokubulawa ngokusebenzisa i-tf.Session, kunokwenzeka ukuba ubhale imisebenzi eqhelekileyo kwiPython, ethi, usebenzisa umnxeba kwi-tf.function, inokuguqulwa ibe yigrafu kwaye emva koko iqhutywe, i-serialized, okanye iphuculwe. kuphuculo lokusebenza;
- Kongezwe umguquli , eguqula umlambo wemiyalelo yePython ibe yintetho yeTensorFlow, evumela ikhowudi yePython ukuba isetyenziswe ngaphakathi kwetf.function-decorated, tf.data, tf.distribute, kunye nemisebenzi ye-tf.keras;
- ISavedModel idibanisa imodeli yotshintshiselwano ifomathi kwaye yongeza inkxaso yokugcina kunye nokubuyisela imodeli yeemeko. Iimodeli ezidityaniselwe iTensorFlow ngoku zingasetyenziswa kuyo (kwizixhobo eziphathwayo), (kwibhrawuza okanye kwiNode.js), ΠΈ ;
- I-tf.train.Optimizers kunye ne-tf.keras.Optimizers APIs zidityanisiwe; endaweni ye-compute_gradients, udidi olutsha lucetyiwe ukuba kubalwe ukuthambeka. ;
- Ukonyusa kakhulu ukusebenza xa usebenzisa i-GPU.
Isantya soqeqesho lwemodeli kwiinkqubo kunye ne-NVIDIA Volta kunye ne-Turing GPUs iye yanda ukuya kumaxesha amathathu; - Ukucocwa kwe-API enkulu, iifowuni ezininzi eziqanjwe ngokutsha okanye zisusiwe, inkxaso yeenguqu zehlabathi jikelele kwiindlela zomncedisi zimisiwe. Endaweni ye-tf.app, tf.flags, tf.logging, i-absl-py API entsha iyacetywa. Ukuqhubeka nokusebenzisa i-API endala, imodyuli ye-comat.v1 ilungisiwe.
umthombo: opennet.ru