{"id":195285,"date":"2021-03-30T14:57:00","date_gmt":"2021-03-30T14:57:00","guid":{"rendered":"https:\/\/datascientest.com\/?p=195285"},"modified":"2023-11-09T14:09:39","modified_gmt":"2023-11-09T14:09:39","slug":"deep-convolutional-gan-la-creation-par-convolution","status":"publish","type":"post","link":"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution","title":{"rendered":"Deep Convolutional GAN : la cr\u00e9ation par convolution"},"content":{"rendered":"\t\t<div data-elementor-type=\"wp-post\" data-elementor-id=\"195285\" class=\"elementor elementor-195285\" data-elementor-post-type=\"post\">\n\t\t\t\t\t\t\t<div class=\"elementor-element elementor-element-b733dfe e-flex e-con-boxed e-con\" data-id=\"b733dfe\" data-element_type=\"container\" data-settings=\"{&quot;content_width&quot;:&quot;boxed&quot;}\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-e6541fa elementor-widget elementor-widget-text-editor\" data-id=\"e6541fa\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t<style>\/*! elementor - v3.15.0 - 02-08-2023 *\/\n.elementor-widget-text-editor.elementor-drop-cap-view-stacked .elementor-drop-cap{background-color:#69727d;color:#fff}.elementor-widget-text-editor.elementor-drop-cap-view-framed .elementor-drop-cap{color:#69727d;border:3px solid;background-color:transparent}.elementor-widget-text-editor:not(.elementor-drop-cap-view-default) .elementor-drop-cap{margin-top:8px}.elementor-widget-text-editor:not(.elementor-drop-cap-view-default) .elementor-drop-cap-letter{width:1em;height:1em}.elementor-widget-text-editor .elementor-drop-cap{float:left;text-align:center;line-height:1;font-size:50px}.elementor-widget-text-editor .elementor-drop-cap-letter{display:inline-block}<\/style>\t\t\t\t<p><b>Les r\u00e9seaux de neurones sont maintenant omnipr\u00e9sents dans le monde de la data, mais s\u2019il y a bien un concept novateur qui a \u00e9merg\u00e9 ces derni\u00e8res ann\u00e9es, ce sont les r\u00e9seaux GANs. Ces r\u00e9seaux qui apprennent en s&rsquo;affrontant sont particuli\u00e8rement efficaces, et si ce concept est combin\u00e9 avec l\u2019\u00e9tat de l\u2019art des techniques d\u2019image processing, vous obtenez ce que nous allons aborder dans cet article, les DCGANs ou Deep Convolutional Generative Adversarial Networks.<\/b><\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-f20cfee e-flex e-con-boxed e-con\" data-id=\"f20cfee\" data-element_type=\"container\" data-settings=\"{&quot;content_width&quot;:&quot;boxed&quot;}\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-5a8922d elementor-widget elementor-widget-heading\" data-id=\"5a8922d\" data-element_type=\"widget\" data-widget_type=\"heading.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t<style>\/*! elementor - v3.15.0 - 02-08-2023 *\/\n.elementor-heading-title{padding:0;margin:0;line-height:1}.elementor-widget-heading .elementor-heading-title[class*=elementor-size-]>a{color:inherit;font-size:inherit;line-height:inherit}.elementor-widget-heading .elementor-heading-title.elementor-size-small{font-size:15px}.elementor-widget-heading .elementor-heading-title.elementor-size-medium{font-size:19px}.elementor-widget-heading .elementor-heading-title.elementor-size-large{font-size:29px}.elementor-widget-heading .elementor-heading-title.elementor-size-xl{font-size:39px}.elementor-widget-heading .elementor-heading-title.elementor-size-xxl{font-size:59px}<\/style><h3 class=\"elementor-heading-title elementor-size-default\">Qu\u2019est ce que le concept des GANs ?<\/h3>\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-f3bae8f elementor-widget elementor-widget-text-editor\" data-id=\"f3bae8f\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<p><span style=\"font-weight: 400;\">Pour rappel, le concept des <\/span><b>r\u00e9seaux GANs<\/b><span style=\"font-weight: 400;\"> r\u00e9side dans l\u2019affrontement de deux r\u00e9seaux, le G\u00e9n\u00e9rateur (G) et le Discriminateur (D). Le G\u00e9n\u00e9rateur apprend \u00e0 g\u00e9n\u00e9rer de toute pi\u00e8ce une<\/span><b> image r\u00e9aliste<\/b><span style=\"font-weight: 400;\"> (similaire \u00e0 celles d\u2019un jeu de donn\u00e9es) qui va \u00eatre jug\u00e9e par le Discriminateur qui a d\u00e9j\u00e0 rencontr\u00e9e des images du jeu de donn\u00e9es, si le r\u00e9seau D est tromp\u00e9, alors il va devoir apprendre de son erreur, si le r\u00e9seau D a juste alors c\u2019est le r\u00e9seau G qui va devoir s\u2019am\u00e9liorer.\u00a0<\/span><\/p><p><span style=\"font-weight: 400;\">Cette m\u00e9thode permet de se retrouver dans une <\/span><a href=\"https:\/\/datascientest.com\/algorithme-de-classification-definition-et-principaux-modeles#:~:text=Une%20classification%20supervis%C3%A9e&amp;text=C'est%2D%C3%A0%2Ddire,pour%20ensuite%20r%C3%A9aliser%20la%20pr%C3%A9diction.\"><span style=\"font-weight: 400;\">situation de classification supervis\u00e9e<\/span><\/a><span style=\"font-weight: 400;\"> et l\u2019apprentissage se fait \u00e0 travers les <\/span><b>loss functions<\/b><span style=\"font-weight: 400;\"> (fonctions de pertes) des deux r\u00e9seaux, comme pour des r\u00e9seaux traditionnels.<\/span><\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-b0f87d5 elementor-widget__width-initial elementor-widget elementor-widget-image\" data-id=\"b0f87d5\" data-element_type=\"widget\" data-widget_type=\"image.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t<style>\/*! elementor - v3.15.0 - 02-08-2023 *\/\n.elementor-widget-image{text-align:center}.elementor-widget-image a{display:inline-block}.elementor-widget-image a img[src$=\".svg\"]{width:48px}.elementor-widget-image img{vertical-align:middle;display:inline-block}<\/style>\t\t\t\t\t\t\t\t\t\t\t\t<img decoding=\"async\" width=\"800\" height=\"439\" src=\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/Discriminateur.png\" class=\"attachment-large size-large wp-image-195293\" alt=\"\" srcset=\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/Discriminateur.png 1024w, https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/Discriminateur-300x165.png 300w, https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/Discriminateur-768x422.png 768w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-d0bee2e elementor-widget elementor-widget-text-editor\" data-id=\"d0bee2e\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<p><span style=\"font-weight: 400;\">Si vous souhaitez en savoir plus, nous avons un article sur le <\/span><a href=\"https:\/\/datascientest.com\/gan-machine-learning\"><span style=\"font-weight: 400;\">concept des GANs<\/span><\/a><span style=\"font-weight: 400;\"> et leurs utilisations.<\/span><\/p><p><span style=\"font-weight: 400;\">Rapidement ont \u00e9t\u00e9 pr\u00e9sent\u00e9es les limites de cette m\u00e9thode, les images g\u00e9n\u00e9r\u00e9es n\u2019\u00e9taient pas assez r\u00e9alistes, notamment quand on essayait d\u2019augmenter la qualit\u00e9 ou la finesse de celles-ci. Cette limite est due \u00e0 la structure elle-m\u00eame des<\/span><b> r\u00e9seaux G\u00e9n\u00e9rateur et Discriminateur<\/b><span style=\"font-weight: 400;\">, de simples r\u00e9seaux denses (feed-forward ou fully connected en anglais) ne permettent pas d\u2019exploiter au mieux toutes les sp\u00e9cificit\u00e9s d\u2019une image. Pour cela il faut utiliser des<\/span><b> r\u00e9seaux plus complexes<\/b><span style=\"font-weight: 400;\"> appel\u00e9es r\u00e9seaux de convolutions (CNNs ou ConvNets pour Convolutional Neural Networks).\u00a0<\/span><\/p><p><span style=\"font-weight: 400;\">Avant d\u2019entrer dans le d\u00e9tail de l\u2019impl\u00e9mentation du DCGAN, il est n\u00e9cessaire de <\/span><b>comprendre les concepts cl\u00e9s<\/b><span style=\"font-weight: 400;\"> qui composent les CNNs. En particulier les <\/span><a href=\"https:\/\/datascientest.com\/convolutional-neural-network\"><span style=\"font-weight: 400;\">convolution layers<\/span><\/a><span style=\"font-weight: 400;\">, si ceux-ci vous sont \u00e9trangers vous pourrez les retrouver dans l\u2019un de nos articles.<\/span><\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-32001e7 e-flex e-con-boxed e-con\" data-id=\"32001e7\" data-element_type=\"container\" data-settings=\"{&quot;content_width&quot;:&quot;boxed&quot;}\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-8e94ec2 elementor-widget elementor-widget-heading\" data-id=\"8e94ec2\" data-element_type=\"widget\" data-widget_type=\"heading.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t<h3 class=\"elementor-heading-title elementor-size-default\">Comment construire le discriminateur ?<\/h3>\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-43cb314 elementor-widget elementor-widget-text-editor\" data-id=\"43cb314\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<p><span style=\"font-weight: 400;\">Dans notre DCGAN, le r\u00e9seau Discriminateur doit, \u00e0 partir d\u2019une image r\u00e9elle ou fausse, <\/span><b>\u00e9tablir une classification<\/b><span style=\"font-weight: 400;\">. On pourrait donc se ramener \u00e0 un CNN classique type LeNet, mais il a \u00e9t\u00e9 \u00e9tabli par les cr\u00e9ateurs de l\u2019architecture DCGAN que certains param\u00e8tres sp\u00e9cifiques \u00e9taient plus int\u00e9ressants, passons les en revue.\u00a0<\/span><\/p><ul><li style=\"font-weight: 400;\" aria-level=\"1\"><b>Downsampling<\/b><span style=\"font-weight: 400;\"> : pour r\u00e9duire progressivement les dimensions de nos couches de convolutions et<\/span><b> \u00e9viter le sur-apprentissage<\/b><span style=\"font-weight: 400;\">, il est d\u2019usage d\u2019utiliser les couches de pooling. Ici se sont r\u00e9v\u00e9l\u00e9es plus efficaces des couches de convolution avec stride, c\u2019est-\u00e0-dire que le filtre de convolution va se d\u00e9placer de plus qu\u2019une 1 case \u00e0 chaque fois, voici une <\/span><b>repr\u00e9sentation<\/b><span style=\"font-weight: 400;\"> permettant de mieux comprendre le fonctionnement de celles-ci :<\/span><\/li><\/ul>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-9e9bad9 elementor-widget elementor-widget-image\" data-id=\"9e9bad9\" data-element_type=\"widget\" data-widget_type=\"image.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<img decoding=\"async\" width=\"244\" height=\"259\" src=\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/no_padding_no_strides.gif\" class=\"attachment-large size-large wp-image-195288\" alt=\"\" \/>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-d05791d elementor-widget__width-initial elementor-widget elementor-widget-image\" data-id=\"d05791d\" data-element_type=\"widget\" data-widget_type=\"image.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<img decoding=\"async\" width=\"294\" height=\"288\" src=\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/no_padding_strides2.gif\" class=\"attachment-large size-large wp-image-195295\" alt=\"\" \/>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-43112e6 elementor-widget elementor-widget-text-editor\" data-id=\"43112e6\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<ul><li style=\"font-weight: 400;\" aria-level=\"1\"><b>Batch Normalization<\/b><span style=\"font-weight: 400;\"> : cette normalisation des poids apr\u00e8s chacune des couches de convolution (par stride donc) permet un <\/span><b>apprentissage plus stable<\/b><span style=\"font-weight: 400;\"> et plus rapide du r\u00e9seau, par cons\u00e9quent il n\u2019est pas n\u00e9cessaire d\u2019utiliser de biais dans nos diff\u00e9rentes couches !<\/span><\/li><li style=\"font-weight: 400;\" aria-level=\"1\"><b style=\"background-color: transparent; color: var( --e-global-color-text ); font-family: var( --e-global-typography-text-font-family ), Sans-serif; text-align: var(--text-align);\">Activation Function<\/b><span style=\"background-color: transparent; color: var( --e-global-color-text ); font-family: var( --e-global-typography-text-font-family ), Sans-serif; text-align: var(--text-align);\"> : la fonction ReLU semblait, dans le cadre de<\/span><b style=\"background-color: transparent; color: var( --e-global-color-text ); font-family: var( --e-global-typography-text-font-family ), Sans-serif; text-align: var(--text-align);\"> l\u2019entra\u00eenement du Discriminateur<\/b><span style=\"background-color: transparent; color: var( --e-global-color-text ); font-family: var( --e-global-typography-text-font-family ), Sans-serif; text-align: var(--text-align);\">, converger vers un \u00e9tat o\u00f9 tous les poids du r\u00e9seau \u00e9taient inactifs, on utilise donc une variante de cette fonction appel\u00e9e LeakyReLU\u00a0<\/span><\/li><\/ul>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-bb9bb95 e-flex e-con-boxed e-con\" data-id=\"bb9bb95\" data-element_type=\"container\" data-settings=\"{&quot;content_width&quot;:&quot;boxed&quot;}\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-d9dc860 elementor-widget elementor-widget-heading\" data-id=\"d9dc860\" data-element_type=\"widget\" data-widget_type=\"heading.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t<h3 class=\"elementor-heading-title elementor-size-default\">Construction du G\u00e9n\u00e9rateur<\/h3>\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-c37a594 elementor-widget elementor-widget-text-editor\" data-id=\"c37a594\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<p><span style=\"font-weight: 400;\">Le r\u00f4le du r\u00e9seau G\u00e9n\u00e9rateur est de<\/span><b> passer d\u2019un vecteur al\u00e9atoire \u00e0 une image<\/b><span style=\"font-weight: 400;\"> capable de tromper le Discriminateur, il faut donc utiliser de nouvelles techniques qu\u2019on pourrait d\u00e9crire comme un r\u00e9seau convolutif invers\u00e9 :\u00a0<\/span><\/p><ul><li style=\"font-weight: 400;\" aria-level=\"1\"><b>Upsampling<\/b><span style=\"font-weight: 400;\"> : pour augmenter la dimension de notre vecteur vers une image, il faut utiliser ce qu\u2019on appelle des <\/span><b>convolutions transpos\u00e9es<\/b><span style=\"font-weight: 400;\"> comme ci-dessous :<\/span><\/li><\/ul>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-9d9f01c elementor-widget__width-initial elementor-widget elementor-widget-image\" data-id=\"9d9f01c\" data-element_type=\"widget\" data-widget_type=\"image.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<img decoding=\"async\" width=\"344\" height=\"387\" src=\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/no_padding_no_strides_transposed.gif\" class=\"attachment-large size-large wp-image-195296\" alt=\"no_padding_no_strides_transposed\" \/>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-24a626f elementor-widget elementor-widget-text-editor\" data-id=\"24a626f\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<ul><li style=\"font-weight: 400;\" aria-level=\"1\"><b>Batch Normalization<\/b><span style=\"font-weight: 400;\"> : la m\u00eame logique est appliqu\u00e9e que pour le Discriminateur<\/span><\/li><li style=\"font-weight: 400;\" aria-level=\"1\"><b>Activation Function<\/b><span style=\"font-weight: 400;\"> : ici on utilise la fonction ReLU dans toutes les couches sauf la derni\u00e8re o\u00f9 on utilise une <\/span><b>fonction tanh<\/b><span style=\"font-weight: 400;\">, parce qu\u2019il est plus efficace d\u2019utiliser des images normalis\u00e9es en input du Discriminateur !<\/span>\u00a0<\/li><\/ul>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-831fbfb elementor-align-center button-principal elementor-widget elementor-widget-button\" data-id=\"831fbfb\" data-element_type=\"widget\" data-widget_type=\"button.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t<div class=\"elementor-button-wrapper\">\n\t\t\t<a class=\"elementor-button elementor-button-link elementor-size-md\" href=\"https:\/\/datascientest.com\/formation-deep-learning\">\n\t\t\t\t\t\t<span class=\"elementor-button-content-wrapper\">\n\t\t\t\t\t\t<span class=\"elementor-button-text\">Comprendre les mod\u00e8les de Deep Learning<\/span>\n\t\t<\/span>\n\t\t\t\t\t<\/a>\n\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-a409d6a elementor-widget elementor-widget-text-editor\" data-id=\"a409d6a\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<p>Voici plusieurs illustrations de fonctions d\u2019activation :<\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-7bf2c49 elementor-widget__width-initial elementor-widget elementor-widget-image\" data-id=\"7bf2c49\" data-element_type=\"widget\" data-widget_type=\"image.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<img decoding=\"async\" width=\"800\" height=\"439\" src=\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/GANS.png\" class=\"attachment-large size-large wp-image-195297\" alt=\"GANS\" srcset=\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/GANS.png 1024w, https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/GANS-300x165.png 300w, https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/GANS-768x422.png 768w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-3ae7b36 elementor-widget elementor-widget-text-editor\" data-id=\"3ae7b36\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<p><span style=\"font-weight: 400;\">Par ailleurs, les constructions particuli\u00e8res des Discriminateurs et G\u00e9n\u00e9rateurs rendent inutiles les couches denses qu\u2019on retrouve habituellement dans les <\/span><a href=\"https:\/\/datascientest.com\/convolutional-neural-network\"><span style=\"font-weight: 400;\">CNNs<\/span><\/a><span style=\"font-weight: 400;\">.<\/span><\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-d9a7735 e-flex e-con-boxed e-con\" data-id=\"d9a7735\" data-element_type=\"container\" data-settings=\"{&quot;content_width&quot;:&quot;boxed&quot;}\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-7937dc4 elementor-widget elementor-widget-heading\" data-id=\"7937dc4\" data-element_type=\"widget\" data-widget_type=\"heading.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t<h3 class=\"elementor-heading-title elementor-size-default\">Entra\u00eenement ou training loop<\/h3>\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-7567b5a elementor-widget elementor-widget-text-editor\" data-id=\"7567b5a\" data-element_type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t<p><span style=\"font-weight: 400;\">Si vous avez bien suivi jusqu\u2019ici vous avez remarqu\u00e9 qu\u2019il y a une \u00e9tape essentielle dont nous n\u2019avons pas encore parl\u00e9, c\u2019est la training loop, ou phase d\u2019entra\u00eenement. Si vous \u00eates d\u00e9j\u00e0 familier avec le <\/span><b>training des r\u00e9seaux de neurones<\/b><span style=\"font-weight: 400;\"> vous savez qu\u2019il existe un <\/span><b>nombre quasi-infini d\u2019hyperparam\u00e8tres<\/b><span style=\"font-weight: 400;\"> \u00e0 tester. Cette probl\u00e9matique est logiquement 2 fois plus pr\u00e9sente avec l\u2019entra\u00eenement des DCGANs, c\u2019est pourquoi il est d\u2019usage d\u2019utiliser des <\/span><b>param\u00e8tres empiriquement prouv\u00e9s<\/b><span style=\"font-weight: 400;\"> comme efficaces, notamment par les auteurs du papier.<\/span><\/p><p><span style=\"font-weight: 400;\">Nous pr\u00e9senterons ici simplement les grandes lignes :\u00a0<\/span><\/p><ul><li style=\"font-weight: 400;\" aria-level=\"1\"><b>Initialisation <\/b><span style=\"font-weight: 400;\">: tous les poids sont initialis\u00e9s suivant une distribution Gaussienne sp\u00e9cifique permettant d\u2019\u00e9viter des probl\u00e8mes de convergences durant l\u2019apprentissage.<\/span><\/li><li style=\"font-weight: 400;\" aria-level=\"1\"><b>Descente de gradient <\/b><span style=\"font-weight: 400;\">: celle-ci se fait de mani\u00e8re stochastique sur des mini-batch, pour la m\u00eame raison que l\u2019initialisation : cela stabilise l\u2019entra\u00eenement.\u00a0<\/span><\/li><\/ul><ul><li style=\"font-weight: 400;\" aria-level=\"1\"><b>Optimizer <\/b><span style=\"font-weight: 400;\">: on utilise l\u2019optimizer Adam qui permet une mise \u00e0 jour en continu du taux d\u2019apprentissage. Des param\u00e8tres sp\u00e9cifiques pour le<\/span><b> momentum \u03b21<\/b><span style=\"font-weight: 400;\"> et <\/span><b>l\u2019initialisation de la learning rate<\/b><span style=\"font-weight: 400;\"> sont utilis\u00e9s pour stabiliser l\u2019entra\u00eenement.<\/span><\/li><\/ul><p>\u00a0<\/p><p><span style=\"font-weight: 400;\">Beaucoup de d\u00e9tails techniques ont \u00e9t\u00e9 \u00e9voqu\u00e9s, cependant ce sont des param\u00e8tres de plus en plus utilis\u00e9s en <\/span><a href=\"https:\/\/datascientest.com\/deep-learning-definition\"><span style=\"font-weight: 400;\">Deep Learning<\/span><\/a><span style=\"font-weight: 400;\">. Par exemple, les convolutions transpos\u00e9es ou \u00e0 stride sont utilis\u00e9es dans de nombreuses applications d\u2019apprentissage profond telles que l\u2019image inpainting, la <\/span><b>segmentation s\u00e9mantique<\/b><span style=\"font-weight: 400;\"> ou la <\/span><b>super-r\u00e9solution d\u2019images<\/b><span style=\"font-weight: 400;\">. De m\u00eame, l\u2019optimizer Adam est un optimizer tr\u00e8s robuste et omnipr\u00e9sent en Deep Learning.<\/span><span style=\"font-weight: 400;\"><br \/><\/span><span style=\"font-weight: 400;\"><br \/><\/span><span style=\"font-weight: 400;\">La \u201csimplicit\u00e9\u201d du DCGAN contribue \u00e0 son succ\u00e8s. Cependant, un certain goulot d\u2019\u00e9tranglement a \u00e9t\u00e9 atteint : l\u2019augmentation de la complexit\u00e9 du g\u00e9n\u00e9rateur n\u2019am\u00e9liore pas n\u00e9cessairement la qualit\u00e9 de l\u2019image. C\u2019est pourquoi il y a eu de nombreuses am\u00e9liorations r\u00e9centes : les cGANs, les styleGANs, cycleGANs etc..<\/span><span style=\"font-weight: 400;\"><br \/><\/span><span style=\"font-weight: 400;\"><br \/><\/span><span style=\"font-weight: 400;\">Vous pouvez retrouver les GANs, les DCGANs et les am\u00e9liorations qui ont depuis \u00e9t\u00e9 apport\u00e9es \u00e0 ces r\u00e9seaux dans le module \u201c<\/span><b>Generative Adversarial Network\u201d<\/b><span style=\"font-weight: 400;\"> de notre formation Deep Learning for Computer Vision.<\/span><\/p>\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<div class=\"elementor-element elementor-element-79e102c elementor-align-center button-principal elementor-widget elementor-widget-button\" data-id=\"79e102c\" data-element_type=\"widget\" data-widget_type=\"button.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t<div class=\"elementor-button-wrapper\">\n\t\t\t<a class=\"elementor-button elementor-button-link elementor-size-md\" href=\"https:\/\/datascientest.com\/formation-deep-learning\">\n\t\t\t\t\t\t<span class=\"elementor-button-content-wrapper\">\n\t\t\t\t\t\t<span class=\"elementor-button-text\">D\u00e9couvrir la formation Deep Learning<\/span>\n\t\t<\/span>\n\t\t\t\t\t<\/a>\n\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-b31976b e-flex e-con-boxed e-con\" data-id=\"b31976b\" data-element_type=\"container\" data-settings=\"{&quot;content_width&quot;:&quot;boxed&quot;}\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-9423a3a elementor-widget elementor-widget-html\" data-id=\"9423a3a\" data-element_type=\"widget\" data-widget_type=\"html.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t<script type=\"application\/ld+json\">\r\n{\r\n  \"@context\": \"https:\/\/schema.org\",\r\n  \"@type\": \"FAQPage\",\r\n  \"mainEntity\": [{\r\n    \"@type\": \"Question\",\r\n    \"name\": \"Qu\u2019est ce que le concept des GANs ?\",\r\n    \"acceptedAnswer\": {\r\n      \"@type\": \"Answer\",\r\n      \"text\": \"Pour rappel, le concept des r\u00e9seaux GANs r\u00e9side dans l\u2019affrontement de deux r\u00e9seaux, le G\u00e9n\u00e9rateur (G) et le Discriminateur (D). Le G\u00e9n\u00e9rateur apprend \u00e0 g\u00e9n\u00e9rer de toute pi\u00e8ce une image r\u00e9aliste (similaire \u00e0 celles d\u2019un jeu de donn\u00e9es) qui va \u00eatre jug\u00e9e par le Discriminateur qui a d\u00e9j\u00e0 rencontr\u00e9e des images du jeu de donn\u00e9es, si le r\u00e9seau D est tromp\u00e9, alors il va devoir apprendre de son erreur, si le r\u00e9seau D a juste alors c\u2019est le r\u00e9seau G qui va devoir s\u2019am\u00e9liorer.\"\r\n    }\r\n  },{\r\n    \"@type\": \"Question\",\r\n    \"name\": \"Comment construire le discriminateur ?\",\r\n    \"acceptedAnswer\": {\r\n      \"@type\": \"Answer\",\r\n      \"text\": \"Dans notre DCGAN, le r\u00e9seau Discriminateur doit, \u00e0 partir d\u2019une image r\u00e9elle ou fausse, \u00e9tablir une classification. On pourrait donc se ramener \u00e0 un CNN classique type LeNet, mais il a \u00e9t\u00e9 \u00e9tabli par les cr\u00e9ateurs de l\u2019architecture DCGAN que certains param\u00e8tres sp\u00e9cifiques \u00e9taient plus int\u00e9ressants, passons les en revue.\"\r\n    }\r\n  }]\r\n}\r\n<\/script>\r\n\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"<p>Les r\u00e9seaux de neurones sont maintenant omnipr\u00e9sents dans le monde de la data, mais s\u2019il y a bien un concept novateur qui a \u00e9merg\u00e9 ces derni\u00e8res ann\u00e9es, ce sont les r\u00e9seaux GANs. Ces r\u00e9seaux qui apprennent en s&rsquo;affrontant sont particuli\u00e8rement efficaces, et si ce concept est combin\u00e9 avec l\u2019\u00e9tat de l\u2019art des techniques d\u2019image processing, [&hellip;]<\/p>\n","protected":false},"author":77,"featured_media":195287,"comment_status":"open","ping_status":"open","sticky":false,"template":"elementor_theme","format":"standard","meta":{"content-type":"","inline_featured_image":false},"categories":[2497],"tags":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO Premium plugin v21.7 (Yoast SEO v21.7) - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>Deep Convolutional GAN : la cr\u00e9ation par convolution<\/title>\n<meta name=\"description\" content=\"Les r\u00e9seaux de neurones sont maintenant omnipr\u00e9sents dans le monde de la data, mais s\u2019il y a bien un concept novateur qui a \u00e9merg\u00e9 ces derni\u00e8res ann\u00e9es, ce sont les r\u00e9seaux GANs.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution\" \/>\n<meta property=\"og:locale\" content=\"fr_FR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Deep Convolutional GAN : la cr\u00e9ation par convolution\" \/>\n<meta property=\"og:url\" content=\"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution\" \/>\n<meta property=\"og:site_name\" content=\"Formation Data Science | DataScientest.com\" \/>\n<meta property=\"article:published_time\" content=\"2021-03-30T14:57:00+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2023-11-09T14:09:39+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/Deep-conventional.png\" \/>\n\t<meta property=\"og:image:width\" content=\"1024\" \/>\n\t<meta property=\"og:image:height\" content=\"562\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Raphael Kassel\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u00c9crit par\" \/>\n\t<meta name=\"twitter:data1\" content=\"Raphael Kassel\" \/>\n\t<meta name=\"twitter:label2\" content=\"Dur\u00e9e de lecture estim\u00e9e\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution\",\"url\":\"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution\",\"name\":\"Deep Convolutional GAN : la cr\u00e9ation par convolution\",\"isPartOf\":{\"@id\":\"https:\/\/datascientest.com\/#website\"},\"datePublished\":\"2021-03-30T14:57:00+00:00\",\"dateModified\":\"2023-11-09T14:09:39+00:00\",\"author\":{\"@id\":\"https:\/\/datascientest.com\/#\/schema\/person\/6eca16f63daaafcbc96e15549b9d4ee1\"},\"description\":\"Les r\u00e9seaux de neurones sont maintenant omnipr\u00e9sents dans le monde de la data, mais s\u2019il y a bien un concept novateur qui a \u00e9merg\u00e9 ces derni\u00e8res ann\u00e9es, ce sont les r\u00e9seaux GANs.\",\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution\"]}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/datascientest.com\/#website\",\"url\":\"https:\/\/datascientest.com\/\",\"name\":\"Formation Data Science | DataScientest.com\",\"description\":\"Leader de la formation en Data Science\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/datascientest.com\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"fr-FR\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/datascientest.com\/#\/schema\/person\/6eca16f63daaafcbc96e15549b9d4ee1\",\"name\":\"Raphael Kassel\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/datascientest.com\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/11\/raph.jpeg\",\"contentUrl\":\"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/11\/raph.jpeg\",\"caption\":\"Raphael Kassel\"},\"description\":\"Raphael Kassel est le Head of Data Programs chez DataScientest. Il supervise une \u00e9quipe Data, impl\u00e9mente de nouveaux parcours comme \\\"Chef de projet IA\\\" et \\\"Data Product Manager\\\", et contribue \u00e0 la modernisation des op\u00e9rations du groupe. En collaboration \u00e9troite avec l'\u00e9quipe de d\u00e9veloppement, il adapte l'accompagnement acad\u00e9mique aux \u00e9volutions de l'\u00e9quipe data et \u00e0 l'internationalisation.\",\"sameAs\":[\"https:\/\/www.linkedin.com\/in\/raphaelkassel\/\",\"Data Science\"],\"url\":\"https:\/\/datascientest.com\/author\/raphaelkassel\"}]}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"Deep Convolutional GAN : la cr\u00e9ation par convolution","description":"Les r\u00e9seaux de neurones sont maintenant omnipr\u00e9sents dans le monde de la data, mais s\u2019il y a bien un concept novateur qui a \u00e9merg\u00e9 ces derni\u00e8res ann\u00e9es, ce sont les r\u00e9seaux GANs.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution","og_locale":"fr_FR","og_type":"article","og_title":"Deep Convolutional GAN : la cr\u00e9ation par convolution","og_url":"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution","og_site_name":"Formation Data Science | DataScientest.com","article_published_time":"2021-03-30T14:57:00+00:00","article_modified_time":"2023-11-09T14:09:39+00:00","og_image":[{"width":1024,"height":562,"url":"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/03\/Deep-conventional.png","type":"image\/png"}],"author":"Raphael Kassel","twitter_card":"summary_large_image","twitter_misc":{"\u00c9crit par":"Raphael Kassel","Dur\u00e9e de lecture estim\u00e9e":"6 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution","url":"https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution","name":"Deep Convolutional GAN : la cr\u00e9ation par convolution","isPartOf":{"@id":"https:\/\/datascientest.com\/#website"},"datePublished":"2021-03-30T14:57:00+00:00","dateModified":"2023-11-09T14:09:39+00:00","author":{"@id":"https:\/\/datascientest.com\/#\/schema\/person\/6eca16f63daaafcbc96e15549b9d4ee1"},"description":"Les r\u00e9seaux de neurones sont maintenant omnipr\u00e9sents dans le monde de la data, mais s\u2019il y a bien un concept novateur qui a \u00e9merg\u00e9 ces derni\u00e8res ann\u00e9es, ce sont les r\u00e9seaux GANs.","inLanguage":"fr-FR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/datascientest.com\/deep-convolutional-gan-la-creation-par-convolution"]}]},{"@type":"WebSite","@id":"https:\/\/datascientest.com\/#website","url":"https:\/\/datascientest.com\/","name":"Formation Data Science | DataScientest.com","description":"Leader de la formation en Data Science","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/datascientest.com\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"fr-FR"},{"@type":"Person","@id":"https:\/\/datascientest.com\/#\/schema\/person\/6eca16f63daaafcbc96e15549b9d4ee1","name":"Raphael Kassel","image":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/datascientest.com\/#\/schema\/person\/image\/","url":"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/11\/raph.jpeg","contentUrl":"https:\/\/datascientest.com\/wp-content\/uploads\/2023\/11\/raph.jpeg","caption":"Raphael Kassel"},"description":"Raphael Kassel est le Head of Data Programs chez DataScientest. Il supervise une \u00e9quipe Data, impl\u00e9mente de nouveaux parcours comme \"Chef de projet IA\" et \"Data Product Manager\", et contribue \u00e0 la modernisation des op\u00e9rations du groupe. En collaboration \u00e9troite avec l'\u00e9quipe de d\u00e9veloppement, il adapte l'accompagnement acad\u00e9mique aux \u00e9volutions de l'\u00e9quipe data et \u00e0 l'internationalisation.","sameAs":["https:\/\/www.linkedin.com\/in\/raphaelkassel\/","Data Science"],"url":"https:\/\/datascientest.com\/author\/raphaelkassel"}]}},"_links":{"self":[{"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/posts\/195285"}],"collection":[{"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/users\/77"}],"replies":[{"embeddable":true,"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/comments?post=195285"}],"version-history":[{"count":0,"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/posts\/195285\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/media\/195287"}],"wp:attachment":[{"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/media?parent=195285"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/categories?post=195285"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/datascientest.com\/wp-json\/wp\/v2\/tags?post=195285"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}