Wasu kwanaki da suka gabata an gabatar da sabon salo mahimmanci na dandalin ilmantarwa na na'ura TensorFlow 2.0, que yana samar da aiwatarwa daga cikin-akwatin na algorithms mai zurfin ilmantarwa, karamin tsarin shirye-shirye don ƙirar gine-gine a cikin Python da ƙaramin matakin keɓaɓɓu don C ++ wanda zai baka damar sarrafa gini da aiwatar da zane-zanen lissafi.
Dandalin kungiyar Google Brain ce ta kirkireshi kuma Google ke amfani dashi don ganewar murya, fitowar fuska a cikin hotuna, ƙayyade kamanceceniya da hotuna, bincika spam a cikin Gmel, zaɓi labarai a cikin Labaran Google kuma tsara fassarar gwargwadon ma'anar.
TensorFlow yana ba da ɗakin karatu na algorithms na kwamfuta Lambobin daga-akwatin da aka aiwatar ta hanyar sigogin bayanan bayanai. Nodes a cikin irin waɗannan zane-zane suna aiwatar da ayyukan lissafi ko wuraren shigarwa / fitarwa, yayin da gefunan jadawalin ke wakiltar bayanan bayanai masu yawa (tensors) waɗanda ke gudana tsakanin nodes.
Za'a iya sanya nodes zuwa na'urori masu sarrafa kwamfuta da kuma gudanar da aiki ba tare da bata lokaci ba, a lokaci guda ana aiki da dukkan masu auna su a lokaci guda, hakan zai baka damar tsara aiki tare a kowane lokaci a cikin hanyar sadarwar ta hanyar kwatancen aiki tare da harba kwayoyi a kwakwalwa.
Za'a iya gina tsarin ilmantarwa na na'ura akan ingantattun kayan aiki, godiya ga ginanniyar tallafi a cikin TensorFlow don faɗaɗa lissafi zuwa CPUs da yawa ko GPUs. TensorFlow na iya gudana akan CPU da yawa da GPUs (tare da zaɓin kari na CUDA don ƙididdigar babban manufar ƙididdigar sassan sarrafa hoto)
TensorFlow yana samuwa akan 64-bit Linux, macOS, da kuma dandamali ta hannu ciki har da Android da iOS. An rubuta lambar tsarin a cikin C ++ da Python kuma an rarraba ta ƙarƙashin lasisin Apache.
Babban sabon fasali na TensorFlow 2.0
Tare da fitowar wannan sabon sigar babban hankalin ara kanta don sauƙaƙawa da sauƙi na amfani, irin wannan shine batun ginawa da horar dashi, an gabatar da sabon sabon matakin Keras API wanda ke ba da zaɓuɓɓuka da yawa don musaya don gina samfuran (bi da bi, aiki, ƙarami) tare da yiwuwar aiwatar da su kai tsaye (ba tare da haɗawa na farko ba) kuma tare da sauƙaƙen tsarin lalatawa.
Tara tf.distribute.Strategy API don tsara rarraba samfurin horos tare da gyare-gyare kaɗan zuwa lambar data kasance. Baya ga damar rarraba lissafi ga GPU masu yawa, akwai tallafin gwaji don rarraba tsarin ilmantarwa a tsakanin masu sarrafawa masu zaman kansu da yawa da ikon amfani da girgije TPU (Tensor Processing Unit).
Maimakon samfurin zane-zanen hoto wanda aka bayyana tare da aiwatarwa ta hanyar tf.Session, yana yiwuwa a rubuta ayyukan Python na yau da kullun waɗanda za a iya canza su zuwa jadawalai ta hanyar kiran tf. aiki sannan a kashe nesa, a sanya su cikin tsari ko inganta don inganta aikin. yi.
An kara mai fassarar AutoGraph wanda ke juyar da umarnin Python zuwa maganganun TensorFlow, wanda zai baka damar amfani da lambar Python a cikin aikin tf.function, tf.data, tf.distribute, da tf.keras.
SavedModel ya haɗaka tsarin musayar samfurin kuma ya ƙara tallafi don adanawa da dawo da yanayin samfuran. Za a iya amfani da samfuran da aka harhada don TensorFlow a cikin TensorFlow Lite (a kan wayoyin hannu), TensorFlow JS (a cikin mai bincike ko Node.js), TensorFlow Serving, da TensorFlow Hub.
Tf.train.Optimizers da tf.keras.Optimizers APIs an haɗaka, Maimakon lissafin_ digiri, an samar da sabon tsarin karatun GradientTape don kirga gradients.
Hakanan aikin da aka yi a cikin wannan sabon sigar ya kasance mafi girma yayin amfani da GPU. Gudun horar da samfura akan tsarin tare da NVIDIA Volta da Turing GPUs ya karu har sau uku.
APIs masu yawa na tsaftacewa, kiraye-kiraye da yawa an sake suna ko cire su, tallafi don masu canjin yanayi a cikin hanyoyin mataimaki ya karye. Madadin tf.app, tf.flags, tf.logging, sabon samfil-py API aka gabatar. Don ci gaba da amfani da tsohuwar API, an shirya tsarin komputa nav.v1.
Idan kana son karin bayani game da shi zaka iya tuntuba mahada mai zuwa.