tseem ceeb tshaj tawm ntawm tshuab kev kawm platform , uas muab kev npaj ua tiav ntawm ntau yam sib sib zog nqus kev kawm algorithms, qhov kev sib txuas lus yooj yim rau kev tsim qauv hauv Python, thiab qib qis rau cov lus C ++ uas tso cai rau koj los tswj kev tsim kho thiab ua tiav cov lej suav. Lub system code sau hauv C ++ thiab Python thiab nyob rau hauv Apache daim ntawv tso cai.
Lub platform yog Ameslikas tsim los ntawm pab pawg Google Brain thiab siv hauv Google cov kev pabcuam rau kev paub txog kev hais lus, txheeb xyuas lub ntsej muag hauv cov duab, txiav txim siab qhov zoo sib xws ntawm cov duab, lim tawm spam hauv Gmail, xov xwm hauv Google Xov Xwm thiab teeb tsa kev txhais lus suav nrog lub ntsiab lus. Kev faib tshuab kev kawm tuaj yeem tsim los ntawm cov khoom siv tus qauv, ua tsaug rau TensorFlow's built-in kev txhawb nqa rau kev faib cov lej hla ntau CPUs lossis GPUs.
TensorFlow muab lub tsev qiv ntawv ntawm kev npaj ua lej suav algorithms siv los ntawm cov ntaub ntawv ntws graphs. Cov nodes hauv cov duab no siv cov lej ua haujlwm lossis cov ntsiab lus nkag / tso tawm, thaum cov npoo ntawm daim duab sawv cev rau cov ntaub ntawv sib txawv (tensors) uas ntws ntawm cov nodes.
Cov nodes tuaj yeem raug xa mus rau cov cuab yeej suav thiab ua tiav asynchronously, ib txhij ua tag nrho cov thesors tsim nyog rau lawv ib zaug, uas ua rau nws muaj peev xwm los teeb tsa lub sijhawm ua haujlwm ntawm cov nodes hauv lub neural network los ntawm kev sib piv nrog kev ua haujlwm ib txhij ntawm cov neurons hauv lub hlwb.
Lub hom phiaj tseem ceeb hauv kev npaj cov ntawv tshiab yog ntawm kev yooj yim thiab yooj yim ntawm kev siv. :
- API tshiab qib siab tau raug npaj rau kev tsim thiab kev cob qhia qauv , uas muab ntau yam kev xaiv interface rau cov qauv tsim (Sequential, Functional, Subclassing) nrog lub peev xwm (tsis muaj pre-compilation) thiab nrog ib tug yooj yim debugging mechanism;
- Ntxiv API rau lub koom haum cov qauv nrog cov kev hloov me me rau cov cai uas twb muaj lawm. Ntxiv nrog rau qhov muaj peev xwm nthuav tawm cov lej suav thoob plaws , kev sim kev txhawb nqa muaj rau kev faib cov txheej txheem kev kawm mus rau ntau tus txheej txheem ywj pheej thiab muaj peev xwm siv huab (Tensor processing unit);
- Hloov ntawm cov qauv tshaj tawm ntawm kev tsim cov duab nrog kev ua tiav los ntawm tf.Session, nws muaj peev xwm sau cov haujlwm zoo tib yam hauv Python, uas, siv hu rau tf.function, tuaj yeem hloov mus rau hauv cov duab thiab tom qab ntawd remotely tua, serialized, lossis optimized txhawm rau txhim kho kev ua haujlwm;
- Ntxiv tus txhais lus , uas hloov cov kwj ntawm Python cov lus txib rau hauv TensorFlow kab lus, tso cai rau Python code siv hauv tf.function-decorated, tf.data, tf.distribute, thiab tf.keras functions;
- SavedModel koom ua ke cov qauv pauv hloov pauv thiab ntxiv kev txhawb nqa rau kev txuag thiab kho cov qauv qauv. Cov qauv muab tso ua ke rau TensorFlow tam sim no tuaj yeem siv rau hauv (ntawm mobile devices), (hauv browser lossis Node.js), ΠΈ ;
- Cov tf.train.Optimizers thiab tf.keras.Optimizers APIs tau koom ua ke; es tsis txhob ntawm compute_gradients, chav kawm tshiab tau raug npaj rau kev suav cov gradients ;
- Kev ua tau zoo nce ntxiv thaum siv GPU.
Kev ceev ntawm cov qauv kev cob qhia ntawm cov tshuab nrog NVIDIA Volta thiab Turing GPUs tau nce mus txog peb zaug; - Kev ntxuav API Loj, ntau lub npe hu ua lossis tshem tawm, kev txhawb nqa rau kev hloov pauv thoob ntiaj teb hauv cov txheej txheem pab nres. Es tsis txhob tf.app, tf.flags, tf.logging, ib tug tshiab absl-py API yog thov. Txhawm rau txuas ntxiv siv API qub, cov qauv compat.v1 tau npaj.
Tau qhov twg los: opennet.ru