Neural network. Zvose izvi zviri kuenda kupi?

Chinyorwa chine zvikamu zviviri:

1. Tsanangudzo pfupi yemamwe magadzirirwo etiweki ekuona chinhu mumifananidzo uye chikamu chemifananidzo ine zvinonzwisisika zvinongedzo kune zviwanikwa kwandiri. Ndakaedza kusarudza tsananguro dzevhidhiyo uye zviri nani muRussia.
2. Chikamu chechipiri ndechekuedza kunzwisisa kutungamira kwekuvandudzwa kweneural network architectures. Uye matekinoroji anobva pavari.

Mufananidzo 1 - Kunzwisisa neural network architectures hazvisi nyore

Izvo zvese zvakatanga nekugadzira maviri edemo maapplication echinhu kupatsanura uye kuonekwa pane Android foni:

• Kumashure-kumagumo demo, kana data ichigadziriswa pane sevha uye ichiendeswa kune foni. Mufananidzo wemhando yemhando mitatu yemapere: brown, dema uye teddy.
• Kumberi-kumagumo demokana data ichigadziriswa pafoni pachayo. Kuonekwa kwezvinhu (chinhu chekuonekwa) chemhando nhatu: hazelnuts, maonde nemazuva.

Pane musiyano pakati pemabasa emhando yemifananidzo, kuona chinhu mumufananidzo uye mufananidzo segmentation. Naizvozvo, paive nechido chekutsvaga kuti ndeapi neural network architecture anoona zvinhu mumifananidzo uye ndezvipi zvinogona kupatsanura. Ndakawana iyo inotevera mienzaniso yezvivakwa zvine zvinonzwisisika zvinongedzo kune zviwanikwa kwandiri:

• Mutsara wezvivakwa zvakavakirwa paR-CNN (Rmatunhu ane Convolution Neural Networks maficha): R-CNN, Fast R-CNN, Inokurumidza R-CNN, Mask R-CNN. Kuti uone chinhu chiri mumufananidzo, mabhokisi ekusungirira anopihwa uchishandisa iyo Region Proposal Network (RPN) michina. Pakutanga, iyo inononoka Kutsvaga yekutsvaga nzira yakashandiswa pachinzvimbo cheRPN. Ipapo matunhu akasarudzwa akaganhurirwa anopihwa kune yekupinza yeyakajairwa neural network yekuiswa. Iyo R-CNN dhizaini ine yakajeka "ye" zvishwe pamusoro penzvimbo shoma, inokwana kusvika 2000 inomhanya kuburikidza neAlexNet yemukati network. Akajeka "zve" zvishwe zvinononotsa kukurumidza kugadzirisa mufananidzo. Huwandu hwezvishwe zvakajeka zvinomhanya nemukati neural network inodzikira neimwe vhezheni itsva yezvivakwa, uye akawanda edzimwe shanduko anogadzirwa kuti awedzere kumhanya uye kutsiva basa rekuona chinhu nechikamu chechinhu muMask R-CNN.
• YOLO (You Only Look Once) ndiyo yekutanga neural network yakaziva zvinhu munguva chaiyo pane nharembozha. Chimiro chakasiyana: kusiyanisa zvinhu mukumhanya kumwe (ingotarisa kamwe chete). Ndiko kuti, muYOLO architecture hapana akajeka "ye" zvishwe, ndicho chikonzero network inoshanda nekukurumidza. Semuenzaniso, kuenzanisa uku: muNumPy, paunenge uchiita mabasa nematrices, hapanawo akajeka "nokuda" zvishwe, izvo muNumPy zvinoshandiswa pamatunhu akaderera ezvivakwa kuburikidza nemutauro wepurogiramu yeC. YOLO inoshandisa grid yemahwindo akafanotsanangurwa. Kudzivirira chinhu chimwe kubva pakutsanangurwa kakawanda, hwindo rinopindirana coefficient (IoU) rinoshandiswa. Imharadzano oVer Union). Iyi dhizaini inoshanda pamusoro pehupamhi uye ine yakakwirira kusimba: Modhi inogona kudzidziswa pamifananidzo asi ichiri kuita zvakanaka pamifananidzo inodhonzwa nemaoko.
• SSD (Sginin Sinopisa MultiBox Detector) - iyo yakabudirira zvikuru "hacks" yeYOLO architecture inoshandiswa (somuenzaniso, kwete-yakanyanya kudzvinyirira) uye itsva inowedzerwa kuita kuti neural network ishande nekukurumidza uye nenzira yakarurama. Chimiro chakasiyana: kusiyanisa zvinhu mune imwe kumhanya uchishandisa yakapihwa gidhi remahwindo (default bhokisi) pane mufananidzo piramidhi. Iyo piramidhi yemufananidzo yakavharidzirwa mune convolution tensor kuburikidza inotevedzana convolution uye pooling mashandiro (nemax-pooling oparesheni, chiyero chepakati chinodzikira). Neiyi nzira, zvese zvakakura uye zvidiki zvinhu zvinotemerwa mune imwechete network inomhanya.
• MobileSSD (fambaNetV2+ SSD) inobatanidza maviri neural network architectures. First network MobileNetV2 inoshanda nekukurumidza uye inowedzera kuzivikanwa kwechokwadi. MobileNetV2 inoshandiswa pachinzvimbo cheVGG-16, iyo yakatanga kushandiswa mukati chinyorwa chekutanga. Yechipiri SSD network inosarudza nzvimbo yezvinhu zviri mumufananidzo.
• SqueezeNet - idiki kwazvo asi yakarurama neural network. Nayo pachayo, haigadzirise dambudziko rekuonekwa kwechinhu. Nekudaro, inogona kushandiswa mukusanganiswa kwezvivakwa zvakasiyana. Uye inoshandiswa mune nharembozha. Chinhu chinosiyanisa ndechekuti iyo data inotanga kudzvanywa kuita ina 1 × 1 convolutional mafirita uye yozowedzerwa kuita ina 1 × 1 uye ina 3 × 3 convolutional mafirita. Imwe iteration yakadaro yedata compression-kuwedzera inonzi "Fire Module".
• DeepLab (Semantic Image Segmentation ine Deep Convolutional Nets) - kupatsanurwa kwezvinhu zviri mumufananidzo. Chinhu chakasiyana chezvivakwa ndeye dilated convolution, inochengetedza kugadziriswa kwenzvimbo. Izvi zvinoteverwa ne-post-processing stage of the results using graphical probabilistic model (conditional random field), iyo inokubvumira kubvisa ruzha ruduku muchikamu uye kuvandudza kunaka kwemufananidzo wakakamurwa. Kuseri kwezita rinotyisa "graphical probabilistic modhi" inovanza yakajairwa Gaussian sefa, iyo inofananidzwa nemapoinzi mashanu.
• Ndakaedza kufunga mudziyo RefineDet (Kamwe-pfuti Tsanangurirament Neural Network yeChinhu Theection), asi handina kunzwisisa zvakawanda.
• Ndakatarisawo kuti tekinoroji ye "kutarisisa" inoshanda sei: vhidhiyo1, vhidhiyo2, vhidhiyo3. Chinhu chakasarudzika che "kutarisisa" kwekuvaka ndiko kusarudzwa otomatiki kwematunhu ekuwedzera kutarisisa mumufananidzo (RoI, Reions of Iinterest) uchishandisa neural network inonzi Attention Unit. Matunhu ekuwedzera kutarisa akafanana nemabhokisi anosungira, asi kusiyana nawo, haana kugadzika mumufananidzo uye anogona kunge ane miganhu yakasviba. Zvadaro, kubva kumatunhu ekuwedzera kutarisisa, zviratidzo (zvimiro) zvakaparadzaniswa, izvo "zvinodyiswa" kune anodzokororwa neural network ane architecture. LSDM, GRU kana Vanilla RNN. Recurrent neural networks vanokwanisa kuongorora hukama hwezvinhu munhevedzano. Recurrent neural network yakatanga kushandiswa kuturikira zvinyorwa kune mimwe mitauro, uye iko zvino kushandura mifananidzo kune zvinyorwa и text to image.

Sezvatinoongorora zvivakwa izvi Ndakaona kuti hapana chandainzwisisa. Uye hazvisi izvo kuti yangu neural network ine matambudziko neyekutarisisa meshini. Kusikwa kwezvivakwa zvese izvi kwakafanana neimwe mhando yehackathon hombe, uko vanyori vanokwikwidza muma hacks. Hack inokurumidza mhinduro kune yakaoma software dambudziko. Ndiko kuti, hapana inooneka uye inonzwisisika inonzwisisika kubatana pakati pezvivakwa izvi zvese. Zvese zvinovabatanidza seti yeakabudirira hacks yavanokwereta kubva kune mumwe nemumwe, pamwe neyakajairwa kune vese. kuvhara-loop convolution operation (kukanganisa backpropagation, backpropagation). Aihwa maitiro ekufunga! Hazvisi pachena kuti chii chekuchinja uye maitiro ekugadzirisa zviripo zviripo.

Nekuda kwekushaikwa kwekubatana kunonzwisisika pakati pema hacks, iwo akaomarara zvakanyanya kurangarira uye kushandisa mukuita. Iyi izivo yakapatsanurwa. Pakanakisisa, nguva shomanana dzinonakidza uye dzisingatarisirwi dzinoyeukwa, asi zvizhinji zvezvinonzwisiswa uye zvisinganzwisisiki zvinopera kubva mumusoro mukati memazuva mashomanana. Zvichava zvakanaka kana muvhiki iwe uchiyeuka zvishoma zita rezvivakwa. Asi maawa anoverengeka uye kunyange mazuva enguva yokushanda akapedzerwa kurava zvinyorwa nokuona mavhidhiyo okudzokorora!

Mufananidzo 2 - Zoo yeNeural Networks

Vazhinji vanyori vezvinyorwa zvesainzi, mumaonero angu pachangu, vanoita zvese zvinogoneka kuti vaone kuti kunyangwe ruzivo urwu rwakapatsanuka harunzwisisike nemuverengi. Asi mitsara inobatanidzwa mumitsara gumi ine mafomula anotorwa "kunze kwemweya mutete" inyaya yechinyorwa chakasiyana (dambudziko. zivisa kana kuparara).

Nechikonzero ichi, pane kudikanwa kwekugadzirisa ruzivo uchishandisa neural network uye, nekudaro, kuwedzera kunaka kwekunzwisisa uye nemusoro. Naizvozvo, musoro mukuru wekuongororwa kwetekinoroji yega yega uye zvivakwa zveartificial neural network raive iro rinotevera basa: tsvagai kuti zvese zviri kuenda kupi, uye kwete mudziyo wechero chaiyo neural network zvakasiyana.

Zvose izvi zviri kuenda kupi? Mibairo mikuru:

• Huwandu hwemakina ekudzidza ekutanga mumakore maviri apfuura akadonha zvakanyanya. Chikonzero chinogona: "neural network haisisiri chinhu chitsva."
• Chero ani zvake anogona kugadzira inoshanda neural network kugadzirisa iri nyore dambudziko. Kuti uite izvi, tora yakagadzirira-yakagadzirwa modhi kubva ku "modhi zoo" uye dzidzisa yekupedzisira layer yeneural network (chinja pakudzidza) pane yakagadzirira-yakagadzirwa data kubva Google Dataset Search kana kubva 25 zviuru zveKaggle datasets mumahara Cloud Jupyter Notebook.
• Vagadziri vakakura veneural network vakatanga kugadzira "mienzaniso zoo" (muenzaniso zoo). Uchishandisa iwo unogona kukurumidza kugadzira application yekutengesa: TF Hub yeTensorFlow, MMDetection yePyTorch, Detectron yeCaffe2, chainer-modelzoo yeChainer uye другие.
• Neural network inoshanda mukati nguva chaiyo (chaiyo-nguva) pane nharembozha. Kubva pa10 kusvika ku50 mafuremu pasekondi.
• Kushandiswa kweneural network mumafoni (TF Lite), mumabrowser (TF.js) uye mukati zvinhu zvemumba (IoT, Iinternet of Things). Kunyanya mumafoni anototsigira neural network padanho rehardware (neural accelerators).
• “Mudziyo wose, mbatya, uye zvichida kunyange zvokudya zvichava nazvo IP-v6 kero uye kutaurirana" - Sebastian Thrun.
• Huwandu hwezvinyorwa pakudzidza kwemichina hwatanga kukura kudarika mutemo waMoore (kaviri makore maviri ega ega) kubvira 2015. Zviripachena, isu tinoda neural network yekuongorora zvinyorwa.
• Matekinoroji anotevera ari kuwedzera mukurumbira:
  • PyTorch - kufarirwa kuri kukura nekukurumidza uye kunoratidzika kunge kuri kupfuura TensorFlow.
  • Otomatiki sarudzo ye hyperparameters AutoML - mukurumbira uri kukura zvakanaka.
  • Kuderera zvishoma nezvishoma mukurongeka uye kuwedzera kwekumhanya kwekuverenga: fuzzy logic, algorithms kukurudzira, zvisizvo (zvinenge) kuverenga, quantization (apo zviyero zveneural network zvinoshandurwa kuva integers uye quantized), neural accelerators.
  • Chinjana mifananidzo kune zvinyorwa и text to image.
  • zvisikwa XNUMXD zvinhu kubva muvhidhiyo, iye zvino munguva chaiyo.
  • Chinhu chikuru pamusoro peDL ndechokuti kune data yakawanda, asi kuunganidza uye kuisa mazita hazvisi nyore. Naizvozvo, markup automation iri kusimukira (otomatiki annotation) yeNeural network uchishandisa neural network.
• Neneural network, Computer Science yakangoerekana yave sainzi yekuedza akasimuka dambudziko reproducibility.
• IT mari uye kufarirwa kweneural network kwakabuda panguva imwe chete apo komputa yakava kukosha kwemusika. Hupfumi huri kuchinja kubva pahupfumi hwegoridhe uye hwemari kuenda gold-currency-computing. Ona chinyorwa changu pa econophysics uye chikonzero chekuonekwa kweIT mari.

Zvishoma nezvishoma imwe itsva inooneka ML/DL programming methodology (Muchina Kudzidza & Kudzidza Kwakadzika), iyo yakavakirwa pakumiririra chirongwa seti yeakadzidziswa neural network modhi.

Mufananidzo 3 - ML / DL senzira itsva yekuronga

Zvisinei, hazvina kumboonekwa "neural network theory", mukati maunogona kufunga uye kushanda zvakarongeka. Izvo zvave kunzi "dzidziso" ndeyekuyedza, heuristic algorithms.

Manongedzo kune angu uye zvimwe zviwanikwa:

• Data Science Newsletter. Kunyanya kugadzirisa mufananidzo. Ani nani anoda kuigamuchira anofanira kutumira e-mail (foobar167<gaf-gaf>gmail<dot>com). Ini ndinotumira zvinongedzo kune zvinyorwa uye mavhidhiyo sezvo zvinhu zvinounganidza.
• General information runyoro rwezvidzidzo uye zvinyorwayandakapfuura uye yandinoda kupfuura.
• Makosi uye mavhidhiyo evanotanga, kwaunofanirwa kutanga kudzidza neural network. Plus bhurocha "Sumo kune Machine Kudzidza uye Artificial Neural Networks".
• Zvishandiso Zvinobatsira, apo munhu wose achawana chimwe chinhu chinonakidza ivo pachavo.
• Takaona zvichibatsira zvikuru. mavhidhiyo chiteshi chekuongorora zvinyorwa zvesainzi by Data Science. Tsvaga, nyorera kwavari uye upfuudze zvinongedzo kune vaunoshanda navo neniwo. Mienzaniso:
  • Mapepa maviri eMiniti
  • Henry AI Labs
  • Yannic Kilcher
  • CodeEmporium
  • Blog Chengwei Zhang aka Tony607 ine nhanho-ne-nhanho mirayiridzo uye yakavhurika sosi.

Спасибо за внимание!

Source: www.habr.com