{"id":89110,"date":"2020-07-18T19:42:10","date_gmt":"2020-07-18T17:42:10","guid":{"rendered":"https:\/\/prohoster.info\/blog\/administrirovanie\/obshhij-obzor-arhitektury-servisa-dlya-oczenki-vneshnosti-na-osnove-nejronnyh-setej"},"modified":"2020-07-18T19:42:10","modified_gmt":"2020-07-18T17:42:10","slug":"obshhij-obzor-arhitektury-servisa-dlya-oczenki-vneshnosti-na-osnove-nejronnyh-setej","status":"publish","type":"post","link":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/obshhij-obzor-arhitektury-servisa-dlya-oczenki-vneshnosti-na-osnove-nejronnyh-setej","title":{"rendered":"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"<p><img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/f624e995fbee3496f1f7890b98106df4.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<h2>\u00davod<\/h2>\n<p><\/p>\n<p>Ahoj!<\/p>\n<p><\/p>\n<p>V tomto \u010dl\u00e1nku se pod\u011bl\u00edm o zku\u0161enosti s n\u00e1vrhem mikroslu\u017ebov\u00e9 architektury pro projekt vyu\u017e\u00edvaj\u00edc\u00ed neuronov\u00e9 s\u00edt\u011b.<\/p>\n<p><\/p>\n<p>Pov\u00edme si o po\u017eadavc\u00edch na architekturu, pod\u00edv\u00e1me se na r\u016fzn\u00e9 struktur\u00e1ln\u00ed diagramy, podrobn\u011b prozkoum\u00e1me jednotliv\u00e9 komponenty hotov\u00e9 architektury a posoud\u00edme technick\u00e9 metriky \u0159e\u0161en\u00ed.<\/p>\n<p><\/p>\n<p>P\u0159\u00edjemn\u00e9 \u010dten\u00ed!<\/p>\n<p><noindex><a rel=\"nofollow\" name=\"habracut\"><\/a><\/noindex><\/p>\n<h2>N\u011bkolik slov o \u00faloze a jej\u00edm \u0159e\u0161en\u00ed<\/h2>\n<p><\/p>\n<p>Hlavn\u00ed my\u0161lenka spo\u010d\u00edv\u00e1 v tom, \u017ee na z\u00e1klad\u011b fotografie poskytneme hodnocen\u00ed atraktivity \u010dlov\u011bka na desetibodov\u00e9 \u0161k\u00e1le.<\/p>\n<p><\/p>\n<p>V tomto \u010dl\u00e1nku se odch\u00fdl\u00edme od popisu jak pou\u017e\u00edvan\u00fdch neuronov\u00fdch s\u00edt\u00ed, tak procesu p\u0159\u00edpravy dat a u\u010den\u00ed. Nicm\u00e9n\u011b, v jedn\u00e9 z dal\u0161\u00edch publikac\u00ed se ur\u010dit\u011b vr\u00e1t\u00edme k podrobn\u011bj\u0161\u00edmu zkoum\u00e1n\u00ed pipeline hodnocen\u00ed.<\/p>\n<p><\/p>\n<p>Te\u010f se projdeme po pipeline hodnocen\u00ed z vy\u0161\u0161\u00edho pohledu, p\u0159i\u010dem\u017e se zam\u011b\u0159\u00edme na interakci mikroslu\u017eeb v kontextu celkov\u00e9 architektury projektu.\u00a0<\/p>\n<p><\/p>\n<p>P\u0159i pr\u00e1ci na pipeline hodnocen\u00ed atraktivity byla \u00faloha rozd\u011blena na n\u00e1sleduj\u00edc\u00ed slo\u017eky:<\/p>\n<p><\/p>\n<ol>\n<li>Detekce obli\u010dej\u016f na fotografi\u00edch<\/li>\n<li>Hodnocen\u00ed ka\u017ed\u00e9ho z obli\u010dej\u016f<\/li>\n<li>Renderov\u00e1n\u00ed v\u00fdsledku<\/li>\n<\/ol>\n<p><\/p>\n<p>Prvn\u00ed \u00fakol je \u0159e\u0161en pomoc\u00ed p\u0159edtr\u00e9novan\u00e9ho <noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/abs\/1604.02878\">MTCNN<\/a><\/noindex>. Pro druh\u00fd \u00fakol byla nau\u010dena konvolu\u010dn\u00ed neuronov\u00e1 s\u00ed\u0165 na PyTorch, kde byl pou\u017eit jako backbone <noindex><a rel=\"nofollow\" href=\"https:\/\/arxiv.org\/abs\/1512.03385\">ResNet34<\/a><\/noindex> \u2013 z pohledu \u201ekvalita \/ rychlost inferov\u00e1n\u00ed na CPU\u201c<\/p>\n<p>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/181cc89a2cf5295bdcd4b838216f96d6.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p><i>Funk\u010dn\u00ed diagram pipeline hodnocen\u00ed<\/i><\/p>\n<p><\/p>\n<h2>Anal\u00fdza po\u017eadavk\u016f na architekturu projektu<\/h2>\n<p><\/p>\n<p>V \u017eivotn\u00edm cyklu <noindex><a rel=\"nofollow\" href=\"https:\/\/en.wikipedia.org\/wiki\/Machine_learning\">ML<\/a><\/noindex> projektu jsou f\u00e1ze pr\u00e1ce na architektu\u0159e a automatizaci nasazen\u00ed modelu \u010dasto jedny z nejv\u00edce \u010dasov\u011b a zdrojov\u011b n\u00e1ro\u010dn\u00fdch.<\/p>\n<p>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/4949f217d31c863176122ed87b43a88c.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p><i>\u017divotn\u00ed cyklus ML projektu<\/i><\/p>\n<p><\/p>\n<p>Tento projekt nen\u00ed v\u00fdjimkou \u2013 bylo rozhodnuto zabalit pipeline hodnocen\u00ed do online slu\u017eby, k tomu bylo pot\u0159eba proniknout do architektury. Byly stanoveny n\u00e1sleduj\u00edc\u00ed z\u00e1kladn\u00ed po\u017eadavky:<\/p>\n<p><\/p>\n<ol>\n<li>Jednotn\u00e9 \u00falo\u017ei\u0161t\u011b log\u016f \u2013 v\u0161echny slu\u017eby by m\u011bly zapisovat logy na jedno m\u00edsto, aby bylo snadn\u00e9 je analyzovat<\/li>\n<li>Mo\u017enost horizont\u00e1ln\u00edho \u0161k\u00e1lov\u00e1n\u00ed slu\u017eby hodnocen\u00ed \u2013 jako\u017eto nejpravd\u011bpodobn\u011bj\u0161\u00edho bottlenecku<\/li>\n<li>Pro hodnocen\u00ed ka\u017ed\u00e9ho obr\u00e1zku by m\u011blo b\u00fdt vy\u010dlen\u011bno stejn\u00e9 mno\u017estv\u00ed procesorov\u00fdch zdroj\u016f \u2013 aby nedoch\u00e1zelo k v\u00fdkyv\u016fm v rozlo\u017een\u00ed \u010dasu na inferenci<\/li>\n<li>Rychl\u00e9 (p\u0159e)nasazen\u00ed jak konkr\u00e9tn\u00edch slu\u017eeb, tak stacku jako celku<\/li>\n<li>Mo\u017enost vyu\u017e\u00edvat spole\u010dn\u00e9 objekty v r\u016fzn\u00fdch slu\u017eb\u00e1ch, pokud je to pot\u0159eba.<\/li>\n<\/ol>\n<p><\/p>\n<h2>Architektura<\/h2>\n<p><\/p>\n<p>Po anal\u00fdze po\u017eadavk\u016f se uk\u00e1zalo, \u017ee mikroslu\u017ebov\u00e1 architektura se hod\u00ed prakticky ide\u00e1ln\u011b.<\/p>\n<p><\/p>\n<p>Abychom se zbavili zbyte\u010dn\u00fdch starost\u00ed, byl jako front-end zvolen Telegram API.<\/p>\n<p><\/p>\n<p>Nejprve se pod\u00edv\u00e1me na struktur\u00e1ln\u00ed diagram hotov\u00e9 architektury, pot\u00e9 p\u0159ejdeme k popisu jednotliv\u00fdch komponent\u016f a formalizujeme proces \u00fasp\u011b\u0161n\u00e9ho zpracov\u00e1n\u00ed obrazu.<\/p>\n<p>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/163ec019341e9d71a1e0854c6026bd01.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p><i>Struktur\u00e1ln\u00ed diagram hotov\u00e9 architektury<\/i><\/p>\n<p><\/p>\n<p>Pod\u00edvejme se podrobn\u011bji na jednotliv\u00e9 komponenty diagramu a ur\u010d\u00edme jejich Single Responsibility p\u0159i hodnocen\u00ed obrazu.<\/p>\n<p><\/p>\n<h3>Mikroslu\u017eba \u201eattrai-telegram-bot\u201c<\/h3>\n<p><\/p>\n<p>Tato mikroslu\u017eba encapsuluje ve\u0161ker\u00e9 interakce s Telegram API. M\u016f\u017eeme vymezit dva hlavn\u00ed sc\u00e9n\u00e1\u0159e \u2013 pr\u00e1ci s u\u017eivatelsk\u00fdm obrazem a pr\u00e1ci s v\u00fdsledkem pipeline hodnocen\u00ed. Prozkoum\u00e1me oba sc\u00e9n\u00e1\u0159e v obecn\u00e9m smyslu.<\/p>\n<p><\/p>\n<p>P\u0159i p\u0159ij\u00edm\u00e1n\u00ed u\u017eivatelsk\u00e9 zpr\u00e1vy s obrazem:<\/p>\n<p><\/p>\n<ol>\n<li>Prob\u00edh\u00e1 filtrov\u00e1n\u00ed, kter\u00e9 se skl\u00e1d\u00e1 z n\u00e1sleduj\u00edc\u00edch kontrol:\n<ul>\n<li>Existence optim\u00e1ln\u00ed velikosti obrazu<\/li>\n<li>Po\u010dtu obraz\u016f u\u017eivatele, kter\u00e9 ji\u017e \u010dekaj\u00ed ve front\u011b<\/li>\n<\/ul>\n<\/li>\n<li>Pokud obraz \u00fasp\u011b\u0161n\u011b projde p\u0159edb\u011b\u017en\u00fdm filtrem, je ulo\u017een do docker volume.<\/li>\n<li>Do fronty \u201eto_estimate\u201c se produkuje \u00fakol, ve kter\u00e9m, mimo jin\u00e9, figuruj\u00ed cesta k obrazu ulo\u017een\u00e9mu v na\u0161em volume.<\/li>\n<li>Pokud v\u00fd\u0161e uveden\u00e9 kroky prob\u011bhnou \u00fasp\u011b\u0161n\u011b \u2013 u\u017eivateli bude zasl\u00e1na zpr\u00e1va s p\u0159ibli\u017enou dobou zpracov\u00e1n\u00ed obrazu, kter\u00e1 se vypo\u010d\u00edt\u00e1v\u00e1 na z\u00e1klad\u011b po\u010dtu \u00fakol\u016f ve front\u011b. V p\u0159\u00edpad\u011b chyby bude u\u017eivatel jasn\u011b informov\u00e1n \u2013 zasl\u00e1n\u00edm zpr\u00e1vy s informacemi o tom, co mohlo zklamat. <\/li>\n<\/ol>\n<p><\/p>\n<p>Tato mikroslu\u017eba funguje tak\u00e9 jako celery worker, kter\u00fd naslouch\u00e1 front\u011b \u201eafter_estimate\u201c, ur\u010den\u00e9 pro \u00fakoly, kter\u00e9 pro\u0161ly pipeline hodnocen\u00ed.<\/p>\n<p><\/p>\n<p>P\u0159i p\u0159ij\u00edm\u00e1n\u00ed nov\u00e9ho \u00fakolu z \u201eafter_estimate\u201c:<\/p>\n<p><\/p>\n<ol>\n<li>Pokud byl obraz \u00fasp\u011b\u0161n\u011b zpracov\u00e1n \u2013 ode\u0161leme v\u00fdsledek u\u017eivateli, pokud ne \u2013 informujeme o chyb\u011b.<\/li>\n<li>Odstran\u00edme obraz, kter\u00fd je v\u00fdsledkem pipeline hodnocen\u00ed.<\/li>\n<\/ol>\n<p><\/p>\n<h3>Mikroslu\u017eba hodnocen\u00ed \u201eattrai-estimator\u201c<\/h3>\n<p><\/p>\n<p>Tato mikroslu\u017eba je celery worker a encapsuluje v\u0161echno, co souvis\u00ed s pipeline hodnocen\u00ed obrazu. Algoritmus zde funguje pouze jeden \u2013 prozkoumejme jej.<\/p>\n<p><\/p>\n<p>P\u0159i p\u0159ij\u00edm\u00e1n\u00ed nov\u00e9ho \u00fakolu z \u201eto_estimate\u201c:<\/p>\n<p><\/p>\n<ol>\n<li>Proch\u00e1z\u00edme obr\u00e1zek hodnotic\u00edm pipeline:\n<ol>\n<li>Nahr\u00e1v\u00e1me obr\u00e1zek do pam\u011bti<\/li>\n<li>P\u0159izp\u016fsobujeme obr\u00e1zek na po\u017eadovanou velikost<\/li>\n<li>Vyhled\u00e1v\u00e1me v\u0161echny tv\u00e1\u0159e (MTCNN)<\/li>\n<li>Hodnot\u00edme v\u0161echny tv\u00e1\u0159e (zabal\u00edme tv\u00e1\u0159e nalezen\u00e9 v p\u0159edchoz\u00edm kroku do d\u00e1vky a provedeme inferenci pomoc\u00ed ResNet34)<\/li>\n<li>Renderujeme kone\u010dn\u00fd obr\u00e1zek\n<ol>\n<li>Kresl\u00edme bounding boxy<\/li>\n<li>Kresl\u00edme hodnocen\u00ed<\/li>\n<\/ol>\n<\/li>\n<\/ol>\n<\/li>\n<li>Odstra\u0148ujeme u\u017eivatelsk\u00fd (p\u016fvodn\u00ed) obr\u00e1zek<\/li>\n<li>Ukl\u00e1d\u00e1me v\u00fdstup z hodnotic\u00edho pipeline<\/li>\n<li>Vkl\u00e1d\u00e1me \u00falohu do fronty 'after_estimate', kterou sleduje d\u0159\u00edve zm\u00edn\u011bn\u00e1 mikroservis 'attrai-telegram-bot'<\/li>\n<\/ol>\n<p><\/p>\n<h3>Graylog (+ mongoDB + Elasticsearch)<\/h3>\n<p><\/p>\n<p><noindex><a rel=\"nofollow\" href=\"https:\/\/www.graylog.org\/\">Graylog<\/a><\/noindex> \u2014 je to \u0159e\u0161en\u00ed pro centralizovan\u00e9 \u0159\u00edzen\u00ed log\u016f. V tomto projektu byl pou\u017e\u00edv\u00e1n k jeho prim\u00e1rn\u00edmu \u00fa\u010delu.<\/p>\n<p><\/p>\n<p>Volba padla pr\u00e1v\u011b na n\u011bj, a ne na v\u0161em obvykl\u00fd <noindex><a rel=\"nofollow\" href=\"https:\/\/www.elastic.co\/what-is\/elk-stack\">ELK<\/a><\/noindex> stack, kv\u016fli pohodl\u00ed pr\u00e1ce s n\u00edm z Pythonu. V\u0161echno, co pot\u0159ebujete ud\u011blat pro logov\u00e1n\u00ed do Graylog, je p\u0159idat GELFTCPHandler z bal\u00ed\u010dku <noindex><a rel=\"nofollow\" href=\"https:\/\/pypi.org\/project\/graypy\/\">graypy<\/a><\/noindex> k ostatn\u00edm root logger handler\u016fm na\u0161\u00ed python mikroservisu.<\/p>\n<p><\/p>\n<p>J\u00e1, jako \u010dlov\u011bk, kter\u00fd d\u0159\u00edve pracoval pouze s ELK stackem, jsem celkov\u011b m\u011bl pozitivn\u00ed zku\u0161enost b\u011bhem pr\u00e1ce s Graylog. Jedin\u00e1 v\u011bc, kter\u00e1 m\u011b zklamala \u2013 p\u0159evaha funkc\u00ed Kibany nad webov\u00fdm rozhran\u00edm Graylog.<\/p>\n<p><\/p>\n<h3>RabbitMQ<\/h3>\n<p><\/p>\n<p><noindex><a rel=\"nofollow\" href=\"https:\/\/www.rabbitmq.com\/\">RabbitMQ<\/a><\/noindex> \u2014 je to zprost\u0159edkovatel zpr\u00e1v zalo\u017een\u00fd na protokolu AMQP.<\/p>\n<p><\/p>\n<p>V tomto projektu byl pou\u017e\u00edv\u00e1n jako <noindex><a rel=\"nofollow\" href=\"https:\/\/docs.celeryproject.org\/en\/latest\/getting-started\/brokers\/#broker-overview\">nejstabiln\u011bj\u0161\u00ed a \u010dasem prov\u011b\u0159en\u00fd<\/a><\/noindex> zprost\u0159edkovatel pro Celery a fungoval v durable re\u017eimu.<\/p>\n<p><\/p>\n<h3>Redis<\/h3>\n<p><\/p>\n<p><noindex><a rel=\"nofollow\" href=\"https:\/\/redis.io\/\">Redis<\/a><\/noindex> \u2014 je to NoSQL datab\u00e1ze, kter\u00e1 pracuje se strukturami dat typu 'kl\u00ed\u010d \u2014 hodnota'<\/p>\n<p><\/p>\n<p>N\u011bkdy je pot\u0159eba pou\u017e\u00edvat v r\u016fzn\u00fdch python mikroservisech spole\u010dn\u00e9 objekty, kter\u00e9 implementuj\u00ed n\u011bjak\u00e9 datov\u00e9 struktury.<\/p>\n<p><\/p>\n<p>Nap\u0159\u00edklad v Redis se uchov\u00e1v\u00e1 hashmapa typu 'telegram_user_id =&gt; po\u010det aktivn\u00edch \u00faloh ve front\u011b', co\u017e umo\u017e\u0148uje omezit po\u010det po\u017eadavk\u016f od jednoho u\u017eivatele na ur\u010ditou hodnotu, \u010d\u00edm\u017e se zabr\u00e1n\u00ed DoS \u00fatok\u016fm.<\/p>\n<p><\/p>\n<h3>Formalizujeme proces \u00fasp\u011b\u0161n\u00e9ho zpracov\u00e1n\u00ed obr\u00e1zku<\/h3>\n<p><\/p>\n<ol>\n<li>U\u017eivatel ode\u0161le obr\u00e1zek do Telegram bota<\/li>\n<li>'attrai-telegram-bot' p\u0159ij\u00edm\u00e1 zpr\u00e1vu od Telegram API a analyzuje ji<\/li>\n<li>\u00daloha s obr\u00e1zkem se p\u0159id\u00e1v\u00e1 do asynchronn\u00ed fronty 'to_estimate'<\/li>\n<li>U\u017eivatel dost\u00e1v\u00e1 zpr\u00e1vu s pl\u00e1novan\u00fdm \u010dasem hodnocen\u00ed<\/li>\n<li>'attrai-estimator' bere \u00falohu z fronty 'to_estimate', proch\u00e1z\u00ed ji hodnotic\u00edm pipeline a produkuje \u00falohu do fronty 'after_estimate'<\/li>\n<li>'attrai-telegram-bot', kter\u00fd sleduje frontu 'after_estimate', pos\u00edl\u00e1 v\u00fdsledek u\u017eivateli<\/li>\n<\/ol>\n<p><\/p>\n<h2>DevOps<\/h2>\n<p><\/p>\n<p>A kone\u010dn\u011b, po p\u0159ehledu architektury, se m\u016f\u017eeme p\u0159esunout k nem\u00e9n\u011b zaj\u00edmav\u00e9 \u010d\u00e1sti \u2014 DevOps<\/p>\n<p><\/p>\n<h3>Docker Swarm<\/h3>\n<p><\/p>\n<p>\u00a0<\/p>\n<p>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/ba539d1c8197c7961d6e5e6a5c7e83c6.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p><noindex><a rel=\"nofollow\" href=\"https:\/\/docs.docker.com\/engine\/swarm\/\">Docker Swarm <\/a><\/noindex>\u00a0\u2014 syst\u00e9m clusterov\u00e1n\u00ed, jeho\u017e funk\u010dnost je realizov\u00e1na uvnit\u0159 Docker Engine a je dostupn\u00e1 p\u0159\u00edmo z krabice.<\/p>\n<p><\/p>\n<p>Pomoc\u00ed \u201eroje\u201c m\u016f\u017eeme rozd\u011blit v\u0161echny uzly na\u0161eho clusteru na 2 typy \u2013 worker a manager. Na stroj\u00edch prvn\u00edho typu se nasazuj\u00ed skupiny kontejner\u016f (stacky), stroje druh\u00e9ho typu zaji\u0161\u0165uj\u00ed \u0161k\u00e1lov\u00e1n\u00ed, vyva\u017eov\u00e1n\u00ed a <noindex><a rel=\"nofollow\" href=\"https:\/\/docs.docker.com\/engine\/swarm\/#feature-highlights\">dal\u0161\u00ed skv\u011bl\u00e9 funkce<\/a><\/noindex>. Managery jsou obvykle tak\u00e9 workery.<\/p>\n<p>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/d195e2aab6d3552652102c2b5246abfc.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p><i>Cluster s jedn\u00edm leader managerem a t\u0159emi workery<\/i><\/p>\n<p><\/p>\n<p>Minim\u00e1ln\u00ed mo\u017en\u00e1 velikost clusteru je 1 uzel, kter\u00fd bude slou\u017eit z\u00e1rove\u0148 jako leader manager a worker. Vzhledem k velikosti projektu a minim\u00e1ln\u00edm po\u017eadavk\u016fm na odolnost v\u016f\u010di chyb\u00e1m byla p\u0159ijata rozhodnut\u00ed pou\u017e\u00edt pr\u00e1v\u011b tento p\u0159\u00edstup.<\/p>\n<p><\/p>\n<p>P\u0159edb\u00edhaje ud\u00e1losti, mohu \u0159\u00edct, \u017ee od prvn\u00ed produk\u010dn\u00ed dod\u00e1vky, kter\u00e1 prob\u011bhla v polovin\u011b \u010dervna, nebyly \u017e\u00e1dn\u00e9 probl\u00e9my spojen\u00e9 s touto organizac\u00ed clusteru (ale to neznamen\u00e1, \u017ee takov\u00e1 organizace je jak\u00fdmkoli zp\u016fsobem p\u0159ijateln\u00e1 pro st\u0159edn\u011b velk\u00e9 a velk\u00e9 projekty, na kter\u00e9 se vztahuj\u00ed po\u017eadavky na odolnost v\u016f\u010di chyb\u00e1m).<\/p>\n<p><\/p>\n<h3>Docker Stack<\/h3>\n<p><\/p>\n<p>V re\u017eimu \u201eroje\u201c za nasazen\u00ed stack\u016f (sady docker slu\u017eeb) zodpov\u00edd\u00e1 <noindex><a rel=\"nofollow\" href=\"https:\/\/docs.docker.com\/engine\/reference\/commandline\/stack\/\">docker stack<\/a><\/noindex><\/p>\n<p><\/p>\n<p>Podporuje docker-compose konfigurace a umo\u017e\u0148uje pou\u017e\u00edt dal\u0161\u00ed parametry pro nasazen\u00ed.\u00a0\u00a0<\/p>\n<p><\/p>\n<p>Nap\u0159\u00edklad pomoc\u00ed t\u011bchto parametr\u016f byly omezeny zdroje na ka\u017edou instanci mikroslu\u017eby hodnocen\u00ed (vy\u010dlenili jsme N jader na N instanc\u00ed, v samotn\u00e9 mikroslu\u017eb\u011b omezujeme po\u010det jader, kter\u00e1 pou\u017e\u00edv\u00e1 PyTorch, na jedno)<\/p>\n<p><\/p>\n<pre><code class=\"plaintext\">attrai_estimator:\n  image: 'erqups\/attrai_estimator:1.2'\n  deploy:\n    replicas: 4\n    resources:\n      limits:\n        cpus: '4'\n    restart_policy:\n      condition: on-failure\n      \u2026<\/code><\/pre>\n<p><\/p>\n<p>Je d\u016fle\u017eit\u00e9 poznamenat, \u017ee Redis, RabbitMQ a Graylog jsou stateful slu\u017eby a jejich \u0161k\u00e1lov\u00e1n\u00ed nen\u00ed tak jednoduch\u00e9 jako u \u201eattrai-estimator\u201c<\/p>\n<p><\/p>\n<h3>Odpov\u00edd\u00e1m p\u0159edem na ot\u00e1zku \u2014 pro\u010d ne Kubernetes?<\/h3>\n<p><\/p>\n<p>Zd\u00e1 se, \u017ee pou\u017eit\u00ed Kubernetes v projektech mal\u00e9ho a st\u0159edn\u00edho rozsahu je p\u0159\u00edt\u011b\u017e\u00ed, v\u0161echny pot\u0159ebn\u00e9 funkce lze z\u00edskat od Docker Swarm, kter\u00fd je pom\u011brn\u011b u\u017eivatelsky p\u0159\u00edv\u011btiv\u00fd jako orchestr\u00e1tor kontejner\u016f a m\u00e1 tak\u00e9 n\u00edzk\u00fd pr\u00e1h vstupu.<\/p>\n<p><\/p>\n<h3>Infrastruktura<\/h3>\n<p><\/p>\n<p>V\u0161echno to bylo nasazeno na VDS s n\u00e1sleduj\u00edc\u00edmi specifikacemi:<\/p>\n<p><\/p>\n<ul>\n<li>CPU: 4 j\u00e1dra Intel&reg; Xeon&reg; Gold 5120 CPU @ 2.20GHz<\/li>\n<li>RAM: 8 GB<\/li>\n<li>SSD: 160 GB<\/li>\n<\/ul>\n<p><\/p>\n<p>Po lok\u00e1ln\u00edm z\u00e1t\u011b\u017eov\u00e9m testov\u00e1n\u00ed se zd\u00e1lo, \u017ee p\u0159i v\u00fdrazn\u00e9m n\u00e1r\u016fstu u\u017eivatel\u016f bude t\u00e9to ma\u0161inky akor\u00e1t.<\/p>\n<p><\/p>\n<p>Ale hned po nasazen\u00ed jsem sd\u00edlel odkaz na jednu z nejpopul\u00e1rn\u011bj\u0161\u00edch v SNG imageboard\u016f (ano, tu zn\u00e1mou), po kter\u00e9m se lid\u00e9 za\u010dali zaj\u00edmat a b\u011bhem n\u011bkolika hodin slu\u017eba \u00fasp\u011b\u0161n\u011b zpracovala des\u00edtky tis\u00edc obr\u00e1zk\u016f. P\u0159itom v kritick\u00fdch momentech nebyly zdroje CPU a RAM ani z poloviny vyu\u017eity.<\/p>\n<p>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/458d74dae00d455fbce1ca6a197325d7.png\" style=\"display:block;margin: 0 auto;\" \/><br \/>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/e9383c47369c606d26e89a69f4d22b5a.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<h3>Je\u0161t\u011b trochu grafiky<\/h3>\n<p><\/p>\n<p>Po\u010det unik\u00e1tn\u00edch u\u017eivatel\u016f a \u017e\u00e1dost\u00ed o hodnocen\u00ed od okam\u017eiku nasazen\u00ed, v z\u00e1vislosti na dni<\/p>\n<p>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/7112c78c8400d2ac73562243b16a015b.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<p>Rozd\u011blen\u00ed \u010dasu inferen\u010dn\u00edho pipeline hodnocen\u00ed<\/p>\n<p>\n<img decoding=\"async\" alt=\"Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed p\u0159ita\u017elivosti na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed\" src=\"\/wp-content\/uploads\/2020\/07\/4c892ce607c3ff3db466459729dd0876.png\" style=\"display:block;margin: 0 auto;\" \/><\/p>\n<h2>Z\u00e1v\u011bry<\/h2>\n<p><\/p>\n<p>Ve zkratce mohu \u0159\u00edci, \u017ee architektura a p\u0159\u00edstup k orchestrace kontejner\u016f se pln\u011b osv\u011bd\u010dily \u2013 i v nejv\u00edce zat\u00ed\u017een\u00fdch momentech nedo\u0161lo k pokles\u016fm a zpo\u017ed\u011bn\u00edm v dob\u011b zpracov\u00e1n\u00ed.\u00a0<\/p>\n<p><\/p>\n<p>Mysl\u00edm, \u017ee projekty mal\u00e9ho a st\u0159edn\u00edho rozsahu, kter\u00e9 ve sv\u00e9m procesu pou\u017e\u00edvaj\u00ed re\u00e1ln\u00fd \u010das inferenci neuronov\u00fdch s\u00edt\u00ed na CPU, mohou \u00fasp\u011b\u0161n\u011b p\u0159evz\u00edt praktikovan\u00e9 popsan\u00e9 v tomto \u010dl\u00e1nku.<\/p>\n<p><\/p>\n<p>Dod\u00e1m, \u017ee p\u016fvodn\u011b byl \u010dl\u00e1nek del\u0161\u00ed, ale aby se nep\u0159id\u00e1valo longread, rozhodl jsem se n\u011bkter\u00e9 body v tomto \u010dl\u00e1nku vynechat \u2013 k nim se vr\u00e1t\u00edme v n\u00e1sleduj\u00edc\u00edch publikac\u00edch.<\/p>\n<p><\/p>\n<p>Na bota si m\u016f\u017eete zkusit kliknout na Telegramu \u2013 @AttraiBot, fungovat bude minim\u00e1ln\u011b do konce podzimu 2020. P\u0159ipom\u00edn\u00e1m \u2013 \u017e\u00e1dn\u00e1 u\u017eivatelsk\u00e1 data nejsou ulo\u017eena \u2013 ani p\u016fvodn\u00ed obr\u00e1zky, ani v\u00fdsledky inferen\u010dn\u00edho pipeline \u2013 v\u0161e se po zpracov\u00e1n\u00ed vyma\u017ee.<\/p>\n<p>Zdroj: <a content=\"nofollow\" rel=\"nofollow\" href=\"https:\/\/habr.com\/ru\/post\/511332\/\">habr.com<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"<p>\u0412\u0441\u0442\u0443\u043f\u043b\u0435\u043d\u0438\u0435 \u041f\u0440\u0438\u0432\u0435\u0442! \u0412 \u0434\u0430\u043d\u043d\u043e\u0439 \u0441\u0442\u0430\u0442\u044c\u0435 \u044f \u043f\u043e\u0434\u0435\u043b\u044e\u0441\u044c \u043e\u043f\u044b\u0442\u043e\u043c \u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u0438\u044f \u043c\u0438\u043a\u0440\u043e\u0441\u0435\u0440\u0432\u0438\u0441\u043d\u043e\u0439 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b \u0434\u043b\u044f \u043f\u0440\u043e\u0435\u043a\u0442\u0430, \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u044e\u0449\u0435\u0433\u043e \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0435 \u0441\u0435\u0442\u0438. \u041f\u043e\u0433\u043e\u0432\u043e\u0440\u0438\u043c \u043e \u0442\u0440\u0435\u0431\u043e\u0432\u0430\u043d\u0438\u044f\u0445 \u043a \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0435, \u043f\u043e\u0441\u043c\u043e\u0442\u0440\u0438\u043c \u043d\u0430 \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u0435 \u0441\u0442\u0440\u0443\u043a\u0442\u0443\u0440\u043d\u044b\u0435 \u0434\u0438\u0430\u0433\u0440\u0430\u043c\u043c\u044b, \u0440\u0430\u0437\u0431\u0435\u0440\u0435\u043c \u043a\u0430\u0436\u0434\u044b\u0439 \u0438\u0437 \u043a\u043e\u043c\u043f\u043e\u043d\u0435\u043d\u0442\u043e\u0432 \u0433\u043e\u0442\u043e\u0432\u043e\u0439 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b, \u0430 \u0442\u0430\u043a\u0436\u0435 \u043e\u0446\u0435\u043d\u0438\u043c \u0442\u0435\u0445\u043d\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u043c\u0435\u0442\u0440\u0438\u043a\u0438 \u0440\u0435\u0448\u0435\u043d\u0438\u044f. \u041f\u0440\u0438\u044f\u0442\u043d\u043e\u0433\u043e \u0447\u0442\u0435\u043d\u0438\u044f! \u041f\u0430\u0440\u0443 \u0441\u043b\u043e\u0432 \u043e \u0437\u0430\u0434\u0430\u0447\u0435 \u0438 \u0435\u0435 \u0440\u0435\u0448\u0435\u043d\u0438\u0438 \u041e\u0441\u043d\u043e\u0432\u043d\u0430\u044f \u0438\u0434\u0435\u044f \u2013 \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u0444\u043e\u0442\u043e \u0434\u0430\u0442\u044c \u043e\u0446\u0435\u043d\u043a\u0443 [&hellip;]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":1,"featured_media":89111,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[688],"tags":[],"class_list":["post-89110","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-administrirovanie"],"aioseo_notices":[],"aioseo_head":"\n\t\t<!-- All in One SEO 5.0.3 - aioseo.com -->\n\t<meta name=\"description\" content=\".\" \/>\n\t<meta name=\"robots\" content=\"max-image-preview:large\" \/>\n\t<meta name=\"author\" content=\"Yuri Gagarin\"\/>\n\t<link rel=\"canonical\" href=\"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/obshhij-obzor-arhitektury-servisa-dlya-oczenki-vneshnosti-na-osnove-nejronnyh-setej\" \/>\n\t\t<meta name=\"generator\" content=\"All in One SEO (AIOSEO) 5.0.3\" \/>\n\t\t<meta property=\"og:locale\" content=\"cs_CZ\" \/>\n\t\t<meta property=\"og:site_name\" content=\"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b\" \/>\n\t\t<meta property=\"og:type\" content=\"article\" \/>\n\t\t<meta property=\"og:title\" content=\"\ud83e\udd47\u041e\u0431\u0449\u0438\u0439 \u043e\u0431\u0437\u043e\u0440 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b \u0441\u0435\u0440\u0432\u0438\u0441\u0430 \u0434\u043b\u044f \u043e\u0446\u0435\u043d\u043a\u0438 \u0432\u043d\u0435\u0448\u043d\u043e\u0441\u0442\u0438 \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0445 \u0441\u0435\u0442\u0435\u0439 | ProHoster\" \/>\n\t\t<meta property=\"og:description\" content=\".\" \/>\n\t\t<meta property=\"og:url\" content=\"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/obshhij-obzor-arhitektury-servisa-dlya-oczenki-vneshnosti-na-osnove-nejronnyh-setej\" \/>\n\t\t<meta property=\"og:image\" content=\"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg\" \/>\n\t\t<meta property=\"og:image:secure_url\" content=\"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg\" \/>\n\t\t<meta property=\"og:image:width\" content=\"350\" \/>\n\t\t<meta property=\"og:image:height\" content=\"350\" \/>\n\t\t<meta property=\"article:published_time\" content=\"2020-07-18T17:42:10+00:00\" \/>\n\t\t<meta property=\"article:modified_time\" content=\"2020-07-18T17:42:10+00:00\" \/>\n\t\t<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/prohoster\" \/>\n\t\t<meta property=\"article:author\" content=\"https:\/\/www.facebook.com\/prohoster\" \/>\n\t\t<!-- All in One SEO -->\n\n","aioseo_head_json":{"title":"\ud83e\udd47Obecn\u00fd p\u0159ehled architektury slu\u017eby pro hodnocen\u00ed vzhledu na z\u00e1klad\u011b neuronov\u00fdch s\u00edt\u00ed | ProHoster","description":".","canonical_url":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/obshhij-obzor-arhitektury-servisa-dlya-oczenki-vneshnosti-na-osnove-nejronnyh-setej","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"cs_CZ","og:site_name":"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b","og:type":"article","og:title":"\ud83e\udd47\u041e\u0431\u0449\u0438\u0439 \u043e\u0431\u0437\u043e\u0440 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b \u0441\u0435\u0440\u0432\u0438\u0441\u0430 \u0434\u043b\u044f \u043e\u0446\u0435\u043d\u043a\u0438 \u0432\u043d\u0435\u0448\u043d\u043e\u0441\u0442\u0438 \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0445 \u0441\u0435\u0442\u0435\u0439 | ProHoster","og:description":".","og:url":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/obshhij-obzor-arhitektury-servisa-dlya-oczenki-vneshnosti-na-osnove-nejronnyh-setej","og:image":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:secure_url":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:width":350,"og:image:height":350,"article:published_time":"2020-07-18T17:42:10+00:00","article:modified_time":"2020-07-18T17:42:10+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster","article:author":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":{"post_id":"89110","title":null,"description":null,"keywords":null,"keyphrases":null,"primary_term":null,"canonical_url":null,"og_title":null,"og_description":null,"og_object_type":"default","og_image_type":"default","og_image_url":null,"og_image_width":null,"og_image_height":null,"og_image_custom_url":null,"og_image_custom_fields":null,"og_video":null,"og_custom_url":null,"og_article_section":null,"og_article_tags":null,"twitter_use_og":false,"twitter_card":"default","twitter_image_type":"default","twitter_image_url":null,"twitter_image_custom_url":null,"twitter_image_custom_fields":null,"twitter_title":null,"twitter_description":null,"schema":{"blockGraphs":[],"customGraphs":[],"default":{"data":{"Article":[],"Course":[],"Dataset":[],"FAQPage":[],"Movie":[],"Person":[],"Product":[],"ProductReview":[],"Car":[],"Recipe":[],"Service":[],"SoftwareApplication":[],"WebPage":[]},"graphName":"","isEnabled":true},"graphs":[]},"schema_type":null,"schema_type_options":null,"pillar_content":false,"robots_default":true,"robots_noindex":false,"robots_noarchive":false,"robots_nosnippet":false,"robots_nofollow":false,"robots_noimageindex":false,"robots_noodp":false,"robots_notranslate":false,"robots_max_snippet":null,"robots_max_videopreview":null,"robots_max_imagepreview":"large","priority":null,"frequency":null,"local_seo":null,"seo_analyzer_scan_date":null,"breadcrumb_settings":null,"limit_modified_date":false,"reviewed_by":null,"ai":null,"created":"2021-02-28 13:16:27","updated":"2022-09-29 06:10:39","focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/89110","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/comments?post=89110"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/89110\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/media\/89111"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/media?parent=89110"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/categories?post=89110"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/tags?post=89110"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}