ukukhishwa okubalulekile kweplathifomu yokufunda komshini , okunikeza ukusetshenziswa osekwenziwe kakade kwama-algorithms okufunda omshini ajulile, isixhumi esibonakalayo esilula sokuhlela samamodeli wokwakha e-Python, kanye nesixhumi esibonakalayo esisezingeni eliphansi solimi lwe-C++ olukuvumela ukuthi ulawule ukwakhiwa nokwenza amagrafu okubalayo. Ikhodi yesistimu ibhalwe ku-C ++ naku-Python kanye ngaphansi kwelayisensi ye-Apache.
Inkundla yasungulwa ithimba le-Google Brain futhi isetshenziswa ezinsizeni ze-Google ukuze kuqashelwe inkulumo, ukuhlonza ubuso ezithombeni, ukucacisa ukufana kwezithombe, ukuhlunga ugaxekile ku-Gmail, Izindaba Ezindabeni ze-Google nokuhlela ukuhumusha kucatshangelwa incazelo. Amasistimu okufunda emishini esabalalisiwe angadalwa ngehadiwe evamile, ngenxa yosekelo olwakhelwe ngaphakathi lwe-TensorFlow lokusabalalisa izibalo kuma-CPU amaningi noma ama-GPU.
I-TensorFlow inikeza umtapo wolwazi wokubala amanani enziwe ngomumo osetshenziswa ngamagrafu okugeleza kwedatha. Amanodi akumagrafu anjalo asebenzisa imisebenzi yezibalo noma amaphuzu okufaka/okukhiphayo, kuyilapho amaphethelo egrafu amelela ama-multidimensional data array (ama-tensor) ageleza phakathi kwamanodi.
Ama-Node angabelwa kumadivayisi e-computing futhi enziwe ngokuhambisanayo, kanyekanye acubungule wonke ama-thesors afanele wona ngesikhathi esisodwa, okwenza kube nokwenzeka ukuhlela ukusebenza ngesikhathi esisodwa kwama-node kunethiwekhi ye-neural ngokufanisa nokusebenza kanyekanye kwama-neurons ebuchosheni.
Okugxilwe kakhulu ekulungiseleleni inguqulo entsha kwakuwukwenza lula kanye nokusetshenziswa kalula. :
- Kuphakanyiswe i-API entsha yezinga eliphezulu yokwakha nokuqeqeshwa , ehlinzeka ngezinketho zokusebenzelana eziningana zamamodeli wokwakha (Ukulandelana, Okusebenzayo, I-Subclassing) enekhono loku (ngaphandle kokuhlanganiswa kwangaphambili) kanye nendlela elula yokulungisa iphutha;
- I-API eyengeziwe okwenhlangano amamodeli anezinguquko ezincane kumakhodi akhona. Ngaphezu kwamathuba okusabalalisa izibalo yonkana , usekelo lokuhlola luyatholakala ekuhlukaniseni inqubo yokufunda ibe amaphrosesa ambalwa azimele kanye nekhono lokusebenzisa ifu (Iyunithi yokucubungula i-Tensor);
- Esikhundleni semodeli ememezelayo yokwakha igrafu ngokusebenza ngokusebenzisa i-tf.Session, kuyenzeka ukuthi ubhale imisebenzi evamile ku-Python, okuthi, kusetshenziswa ikholi eya ku-tf.function, iguqulelwe ibe amagrafu bese isenziwa ukude, i-serialized, noma ilungiselelwe. ukusebenza okuthuthukisiwe;
- Kwengezwe umhumushi , eguqula ukusakazwa kwemiyalo ye-Python ibe izinkulumo ze-TensorFlow, okuvumela ikhodi ye-Python ukuthi isetshenziswe ngaphakathi kwe-tf.function-decorated, tf.data, tf.distribute, kanye nemisebenzi ye-tf.keras;
- I-SavedModel ihlanganisa ifomethi yokushintshanisa imodeli futhi yengeza usekelo lokulondoloza nokubuyisela imodeli yesimo. Amamodeli ahlanganiselwe i-TensorFlow manje angasetshenziswa ku (kumadivayisi eselula), (kusiphequluli noma ku-Node.js), ΠΈ ;
- I-tf.train.Optimizers kanye ne-tf.keras.Optimizers APIs ahlanganisiwe; esikhundleni se-compute_gradients, kuphakanyiswe ikilasi elisha lokubala ama-gradient. ;
- Ukusebenza okukhuphuke kakhulu uma usebenzisa i-GPU.
Ijubane lokuqeqeshwa okuyimodeli kumasistimu ane-NVIDIA Volta kanye ne-Turing GPUs likhuphuke lafika izikhathi ezintathu; - Ukuhlanzwa okukhulu kwe-API, izingcingo eziningi eziqanjwe kabusha noma zisusiwe, ukusekelwa kokuguquguquka komhlaba wonke ezindleleni zomsizi kumisiwe. Esikhundleni se-tf.app, tf.flags, tf.logging, i-absl-py API entsha iyaphakanyiswa. Ukuze uqhubeke usebenzisa i-API endala, imojuli ye-comat.v1 isilungisiwe.
Source: opennet.ru