Kutulutsidwa kwa makina ophunzirira makina TensorFlow 2.0

Yovomerezedwa ndi kumasulidwa kwakukulu kwa nsanja yophunzirira makina TensorFlow 2.0, yomwe imapereka kukhazikitsidwa kokonzeka kwa makina osiyanasiyana ophunzirira makina akuya, mawonekedwe osavuta a pulogalamu yamitundu yomanga ku Python, ndi mawonekedwe otsika a chilankhulo cha C ++ chomwe chimakulolani kuti muzitha kuyang'anira ntchito yomanga ndikuchita ma graph owerengera. Khodi yamakina imalembedwa mu C ++ ndi Python ndi wogawidwa ndi pansi pa chilolezo cha Apache.

Pulatifomu idapangidwa koyambirira ndi gulu la Google Brain ndipo imagwiritsidwa ntchito mu mautumiki a Google kuti azindikire mawu, kuzindikira nkhope muzithunzi, kudziwa kufanana kwa zithunzi, kusefa sipamu mu Gmail, kusankha nkhani mu Google News ndikukonzekera zomasulira poganizira tanthauzo lake. Makina ophunzirira makina ogawidwa amatha kupangidwa pazida zokhazikika, chifukwa cha thandizo la TensorFlow lothandizira kugawa mawerengedwe pama CPU angapo kapena ma GPU.

TensorFlow imapereka laibulale ya ma algorithms opangidwa okonzeka owerengera manambala omwe akhazikitsidwa kudzera pa ma graph a data. Manode mu ma graph oterowo amagwiritsa ntchito masamu kapena zolowetsa/zotulutsa, pomwe m'mphepete mwa graph imayimira mitundu ingapo ya data (ma tensor) omwe amayenda pakati pa node.
Nodes akhoza kuperekedwa kwa zipangizo kompyuta ndi kuphedwa asynchronously, imodzi processing onse thesors oyenera iwo mwakamodzi, zimene zimathandiza kuti bungwe ntchito imodzi ya mfundo mu neural network ndi fanizo ndi kutsegula munthawi yomweyo ma neuron mu ubongo.

Cholinga chachikulu pokonzekera Baibulo latsopanoli chinali chosavuta komanso chosavuta kugwiritsa ntchito. Ena zatsopano:

• API yatsopano yapamwamba yaperekedwa kuti ikhale yomanga ndi yophunzitsira Keras, yomwe imapereka zosankha zingapo zamawonekedwe amitundu yomanga (Sequential, Functional, Subclassing) ndi kuthekera kukhazikitsa mwamsanga (popanda kusonkhanitsa) komanso ndi njira yosavuta yochepetsera;
• API Yowonjezera tf.distribute.Strategy za bungwe maphunziro ogawa zitsanzo zokhala ndi zosintha zochepa pama code omwe alipo. Kuwonjezera kuthekera kufalitsa mawerengedwe kudutsa ma GPU ambiri, chithandizo choyesera chilipo pogawa njira yophunzirira kukhala mapurosesa angapo odziyimira pawokha komanso kuthekera kogwiritsa ntchito mtambo TPU (Tensor processing unit);
• M'malo mwa chitsanzo chofotokozera chopanga graph ndi kuphedwa kupyolera mu tf.Session, n'zotheka kulemba ntchito wamba mu Python, zomwe, pogwiritsa ntchito kuyitana kwa tf.function, zikhoza kusinthidwa kukhala ma graph ndiyeno kuchitidwa kutali, kusinthidwa, kapena kukonzedwa bwino. kwa ntchito yabwino;
• Wowonjezera womasulira AutoGraph, yomwe imasintha mtsinje wa malamulo a Python kukhala mawu a TensorFlow, kulola kuti code ya Python igwiritsidwe ntchito mkati mwa tf.function-decorated, tf.data, tf.distribute, ndi tf.keras ntchito;
• SavedModel imagwirizanitsa mtundu wa kusinthana kwachitsanzo ndikuwonjezera kuthandizira kupulumutsa ndi kubwezeretsa zigawo zachitsanzo. Mitundu yopangidwa ndi TensorFlow tsopano itha kugwiritsidwa ntchito TensorFlow Lite (pazida zam'manja), TensorFlow JS (mu msakatuli kapena Node.js), Kutumikira kwa TensorFlow и TensorFlow Hub;
• Ma tf.train.Optimizers ndi tf.keras.Optimizers API agwirizana; m'malo mwa compute_gradients, kalasi yatsopano yaperekedwa kuti iwerengere ma gradients. Tepi ya Gradient;
• Kuwonjezeka kwakukulu mukugwiritsa ntchito GPU.
  Kuthamanga kwa maphunziro a chitsanzo pa machitidwe omwe ali ndi NVIDIA Volta ndi Turing GPUs awonjezeka mpaka katatu;

• Zidachitidwa Kuyeretsa kwakukulu kwa API, mafoni ambiri otchulidwanso kapena kuchotsedwa, kuthandizira kwamitundu yosiyanasiyana yapadziko lonse lapansi munjira zothandizira kuyimitsidwa. M'malo mwa tf.app, tf.flags, tf.logging, absl-py API yatsopano ikuperekedwa. Kuti mupitirize kugwiritsa ntchito API yakale, gawo la compat.v1 lakonzedwa.

Source: opennet.ru