Ezinsukwini ezithile ezedlule kwethulwe inguqulo entsha kubalulekile epulatifomu yokufunda ngomshini I-TensorFlow 2.0, lokho inikeza ukusetshenziswa okungaphandle kwebhokisi kwama-algorithms ahlukahlukene wokufunda komshini ojulile, isikhombimsebenzisi esibonakalayo sokuhlela samamodeli wokwakha ePython kanye ne-interface esezingeni eliphansi le-C ++ ekuvumela ukuthi ulawule ukwakhiwa nokwenziwa kwemidwebo yekhompyutha.
Isiteji yasungulwa ekuqaleni yithimba le-Google Brain futhi isetshenziswa amasevisi akwa-Google ngokubona izwi, ukubonwa kobuso ezithombeni, kunquma ukufana kwezithombe, ukuhlunga ogaxekile ku-Gmail, khetha izindaba ku-Google News bese uhlela ukuhumusha ngokusho kwencazelo.
I-TensorFlow inikeza umtapo wolwazi lwe-computer algorithms Izinombolo ezingaphandle kwebhokisi zenziwe ngamashadi wokugeleza kwedatha. Ama-node kumagrafu anjalo asebenzisa ukusebenza kwezibalo noma amaphuzu wokungena / okuphuma, kuyilapho imiphetho yegrafu imelela amasethi wedatha ahlukahlukene (ama-tensors) ageleza phakathi kwama-node.
Ama-node angabelwa amadivayisi wekhompiyutha futhi asebenze ngokuvumelanayo, ngasikhathi sinye ukucubungula wonke ama-tensors afanele ngasikhathi sinye, kukuvumela ukuthi uhlele ukusebenza ngasikhathi sinye kwamaqhuqhuva kunethiwekhi ye-neural ngokufanisa nokudubula ngasikhathi sinye kwama-neuron ebuchosheni.
Izinhlelo ezisatshalaliswa zokufunda ngomshini zingakhiwa kwimishini ejwayelekile, sibonga ukusekelwa okwakhelwe ngaphakathi kuTensorFlow yokwelula ukusebenzisa ikhompyutha kuma-CPU amaningi noma ama-GPU. I-TensorFlow ingasebenza kuma-CPU amaningi nama-GPU (ngezandiso zokuzikhethela ze-CUDA zenjongo ejwayelekile yekhompyutha kumayunithi wokucubungula ihluzo)
I-TensorFlow iyatholakala kuma-Linux angama-64-bit, ama-macOS, nezinkundla zeselula ezifaka i-Android ne-iOS. Ikhodi yesistimu ibhalwe nge-C ++ nePython futhi isatshalaliswa ngaphansi kwelayisense le-Apache.
Izici ezintsha eziyinhloko zeTensorFlow 2.0
Ngokukhishwa kwale nguqulo entsha ukunakwa okuyinhloko lizibophezele ekwenzeni lula futhi kube lula ukusetshenziswa, kunjalo ukuthi ukwakha nokuqeqesha amamodeli, i-Keras API entsha esezingeni eliphakeme iphakanyisiwe enikeza izinketho eziningana zezindawo zokuhlangana zokwakha amamodeli (okulandelanayo, okusebenzayo, okungaphansi) okungenzeka kwenziwe kwawo ngokushesha (ngaphandle kokuhlanganiswa kokuqala) nangendlela elula yokulungisa iphutha.
Kungezwe i-tf.distribute.Str Strategy API ukuhlela ukuqeqeshwa kwemodeli okusatshalalisiwes ngokuguqulwa okuncane kwikhodi ekhona. Ngaphezu kwekhono lokusabalalisa izibalo kuma-GPU amaningi, kukhona ukwesekwa kokuhlola okutholakalayo kokuhlukanisa inqubo yokufunda kuma-processor amaningi azimele kanye nekhono lokusebenzisa ifu TPU (Tensor Processing Unit).
Esikhundleni semodeli yokwakhiwa kwegrafu esetshenziswayo eyenziwe nge-tf.Session, kungenzeka ukuthi ubhale imisebenzi ejwayelekile yePython engaguqulwa ibe ngamagrafu ngokubiza i-tf. Ukusebenza bese kwenziwa kude, kwenziwe i-serial noma kwenziwa ngcono ukuthuthukisa ukusebenza.
Kungeziwe umhumushi we-AutoGraph oguqula ukugeleza komyalo wePython kube izinkulumo zeTensorFlow, ekuvumela ukuthi usebenzise ikhodi ye-Python ngaphakathi kwe-tf.function, tf.data, tf.distribute, kanye nemisebenzi tf.keras.
I-SavedModel ihlanganise ifomethi yokushintshaniswa kwemodeli futhi yengeza ukwesekwa kokonga nokubuyisa isimo samamodeli. Izinhlobo ezihlanganisiwe zeTensorFlow manje sezingasetshenziswa kuTensorFlow Lite (kumadivayisi eselula), iTensorFlow JS (kusiphequluli noma iNode.js), iTensorFlow Serving, neTensorFlow Hub.
I-tf.train.Optimizers kanye ne-tf.keras.Optimizers APIs ahlanganisiwe, Esikhundleni sama-compute_gradients, kuhlongozwa isigaba esisha se-GradientTape ukwenza ama-gradients.
Futhi ukusebenza kwale nguqulo entsha bekuphakeme kakhulu lapho usebenzisa i-GPU. Isivinini sokuqeqeshwa kwemodeli ezinhlelweni ezine-NVIDIA Volta nama-Turing GPUs sikhuphuke safinyelela kathathu.
Ama-API amaningi wokuhlanza, izingcingo eziningi ziqanjwa kabusha noma ziyasuswa, ukusekelwa kokuguquguqukayo komhlaba wonke ezindleleni zokusiza kwephuliwe. Esikhundleni se-tf.app, tf.flags, tf.logging, kuhlongozwa i-API entsha ye-absl-py. Ukuqhubeka nokusebenzisa i-API yakudala, imodyuli ye-Compat.v1 isilungisiwe.
Uma ufuna ukwazi kabanzi ngalo ungabonisana isixhumanisi esilandelayo.