gagarumin sakin dandali na koyon inji , wanda ke ba da shirye-shiryen aiwatar da shirye-shiryen daban-daban na ilmantarwa mai zurfi na na'ura, mai sauƙin shirye-shirye na shirye-shirye don gina samfurori a cikin Python, da ƙananan ƙananan ƙananan harshe don harshen C ++ wanda ke ba ku damar sarrafa gine-gine da aiwatar da zane-zane. An rubuta lambar tsarin a cikin C++ da Python kuma ƙarƙashin lasisin Apache.
Ƙungiyar Brain ta Google ta samo asali ne ta dandalin kuma ana amfani da ita a cikin ayyukan Google don gane magana, gano fuskoki a cikin hotuna, ƙayyade kamancen hotuna, tace spam a Gmail, labarai a cikin Google News da tsara fassarar la'akari da ma'anar. Ana iya ƙirƙira tsarin koyo na inji da aka rarraba akan daidaitaccen kayan aiki, godiya ga ginanniyar tallafi na TensorFlow don rarraba ƙididdiga a cikin CPUs ko GPUs da yawa.
TensorFlow yana ba da ɗakin karatu na shirye-shiryen ƙididdige ƙididdiga na ƙididdiga waɗanda aka aiwatar ta hanyar zane-zanen bayanai. Nodes a cikin irin waɗannan jadawali suna aiwatar da ayyukan lissafi ko abubuwan shigarwa/fitarwa, yayin da gefuna na jadawali ke wakiltar tsararrun bayanai masu girma dabam (tensors) waɗanda ke gudana tsakanin nodes.
Ana iya sanya nodes zuwa na'urorin kwamfuta kuma a aiwatar da su ba tare da izini ba, tare da sarrafa duk abubuwan da suka dace da su lokaci guda, wanda ke ba da damar tsara ayyukan nodes na lokaci guda a cikin hanyar sadarwa ta jijiyoyi ta hanyar kwatankwacin kunnawa lokaci guda na neurons a cikin kwakwalwa.
Babban abin da aka fi mayar da hankali wajen shirya sabon sigar shine akan sauƙaƙe da sauƙin amfani. :
- An gabatar da sabon babban matakin API don ƙirar gini da horo , wanda ke ba da zaɓuɓɓukan dubawa da yawa don ƙirar gini (Mai-da-biyu, Aiki, Subclassing) tare da damar (ba tare da riga-kafi ba) kuma tare da tsarin lalata mai sauƙi;
- API ɗin da aka ƙara domin kungiya samfura tare da ƙaramin canje-canje zuwa lambar data kasance. Baya ga yuwuwar yada lissafi a fadin , tallafin gwaji yana samuwa don rarraba tsarin ilmantarwa zuwa na'urori masu zaman kansu da yawa da kuma ikon yin amfani da girgije (Naúrar sarrafa Tensor);
- Maimakon ƙirar ƙira na gina jadawali tare da kisa ta hanyar tf.Session, yana yiwuwa a rubuta ayyuka na yau da kullun a Python, wanda, ta amfani da kira zuwa tf.function, za'a iya canza shi zuwa jadawali sannan a aiwatar da shi daga nesa, serialized, ko inganta shi. don ingantaccen aiki;
- Ƙara mai fassara , wanda ke canza rafi na umarnin Python zuwa maganganun TensorFlow, yana ba da damar yin amfani da lambar Python a cikin tf.function-decorated, tf.data, tf.distribute, da ayyukan tf.keras;
- SavedModel yana haɓaka tsarin musayar ƙirar ƙira kuma yana ƙara goyan baya don adanawa da dawo da jahohin ƙira. Ana iya amfani da samfuran da aka haɗa don TensorFlow yanzu a ciki (akan wayoyin hannu), (a cikin browser ko Node.js), и ;
- An haɗa tf.train.Optimizers da tf.keras.Optimizers APIs; maimakon compute_gradients, an gabatar da sabon aji don kirga gradients ;
- Ƙarfafa aiki sosai lokacin amfani da GPU.
Gudun horon samfurin akan tsarin tare da NVIDIA Volta da Turing GPUs ya karu har sau uku; - Babban tsaftacewar API, yawancin kira da aka sake suna ko cirewa, goyan bayan masu canjin duniya a hanyoyin taimako sun tsaya. Maimakon tf.app, tf.flags, tf.logging, sabon API na absl-py ana samarwa. Don ci gaba da amfani da tsohuwar API, an shirya tsarin compat.v1.
source: budenet.ru