ProHoster > Blog > internet nhau > Kuburitswa kwemuchina wekudzidza system TensorFlow 2.0

Kuburitswa kwemuchina wekudzidza system TensorFlow 2.0

Introduced kuburitswa kwakakosha kwechikuva chekudzidza muchina TensorFlow 2.0, iyo inopa yakagadzirira-yakagadzirwa mashandisirwo eakasiyana-siyana akadzika muchina kudzidza algorithms, yakapusa programming interface yekuvaka modhi muPython, uye yakaderera-level interface yeC ++ mutauro iyo inokutendera iwe kudzora kuvaka uye kuuraya magirafu emakomputa. Iyo system code yakanyorwa muC ++ uye Python uye inoparadzirwa ne pasi peApache rezinesi.

Iyi puratifomu yakatanga kugadzirwa neGoogle Brain timu uye inoshandiswa mumasevhisi eGoogle kucherechedzwa kwekutaura, kuona zviso mumifananidzo, kuona kufanana kwemifananidzo, kusefa spam muGmail, sarudzo nhau muGoogle News uye kuronga kushandura uchifunga nezvezvinoreva. Yakagoverwa muchina kudzidza masisitimu anogona kugadzirwa pane yakajairwa Hardware, nekuda kweTensorFlow's yakavakirwa-mukati tsigiro yekugovera macalculation mumaCPU akawanda kana maGPU.

TensorFlow inopa raibhurari yeakagadzirira-akagadzirwa manhamba ekuverenga algorithms anoitwa kuburikidza nedata kuyerera magirafu. Node mumagirafu akadaro dzinoshandisa masvomhu kana mapoinzi ekuisa/zvinobuda, nepo mipendero yegirafu inomiririra dhata dhata dzakawanda (tensors) dzinoyerera pakati penodhi.
Manodhi anogona kugoverwa kumidziyo yemakomputa uye kuitiswa asynchronously, panguva imwe chete kugadzirisa ese akakodzera ivo kamwechete, izvo zvinoita kuti zvikwanise kuronga panguva imwe chete yekushanda kwemanodhi muneural network nekufananidza pamwe chete nekuita activation yemaneuroni muuropi.

Chakanyanya kutariswa mukugadzirira shanduro itsva yaive pakurerutsa uye nyore kushandisa. Vamwe zvitsva:

  • Iyo itsva-yepamusoro-level API yakakurudzirwa yekuvaka uye kudzidzisa modhi Keras, iyo inopa akati wandei maficha sarudzo dzekuvaka modhi (Sequential, Functional, Subclassing) nekugona kushandiswa pakarepo (pasina pre-kuunganidza) uye ine nyore debugging mechanism;
  • Yakawedzerwa API tf.distribute.Strategy kusangano kudzidza kwakagoverwa mhando dzine shanduko shoma kune iripo kodhi. Mukuwedzera kune mukana wekuparadzira masvomhu mhiri maGPU akawanda, tsigiro yekuyedza iripo yekukamura nzira yekudzidza kuita mapurosesa akawanda akazvimirira uye kugona kushandisa gore TPU (Tensor processing unit);
  • Panzvimbo peiyo declarative modhi yekugadzira girafu nekuuraya kuburikidza netf.Session, zvinokwanisika kunyora zvakajairika mabasa muPython, iyo, uchishandisa call ku tf.function, inogona kushandurwa kuita magirafu uye yozoitwa kure, serialized, kana optimized. kuitira kuvandudza kushanda;
  • Akawedzera muturikiri AutoGraph, iyo inoshandura rukova rwemirairo yePython kuva TensorFlow matauriro, ichibvumira Python code kuti ishandiswe mukati tf.function-decorated, tf.data, tf.distribute, uye tf.keras mabasa;
  • SavedModel inobatanidza iyo modhi yekutsinhana fomati uye inowedzera rutsigiro rwekuchengetedza uye kudzoreredza modhi nyika. Models dzakagadzirirwa TensorFlow zvino dzinogona kushandiswa mukati TensorFlow Lite (panharembozha), TensorFlow JS (mubrowser kana Node.js), TensorFlow Kushumira ΠΈ TensorFlow Hub;
  • Iwo tf.train.Optimizers uye tf.keras.Optimizers APIs akabatanidzwa; pachinzvimbo checompute_gradients, kirasi itsva yakurudzirwa kuverengera magradients. Gradient Tape;
  • Yakawedzera kuita basa kana uchishandisa GPU.
    Iko kumhanya kwemuenzaniso wekudzidzira pane masisitimu ane NVIDIA Volta uye Turing GPUs yakawedzera kusvika katatu;

  • Kuitwa Yakakura API kucheneswa, akawanda mafoni akatumidzwa zita kana kubviswa, rutsigiro rwepasi rose shanduko munzira dzemubatsiri dzakamira. Panzvimbo pe tf.app, tf.flags, tf.logging, absl-py API itsva inokurudzirwa. Kuti uenderere mberi nekushandisa API yekare, compat.v1 module yakagadzirwa.

Source: opennet.ru

Voeg