{"id":705,"date":"2026-10-09T17:39:20","date_gmt":"2026-10-09T15:39:20","guid":{"rendered":"https:\/\/www.ranksignal.fr\/blog\/?p=705"},"modified":"2026-10-14T17:25:03","modified_gmt":"2026-10-14T15:25:03","slug":"qu-est-ce-qu-un-gru-gated-recurrent-unit","status":"publish","type":"post","link":"https:\/\/www.ranksignal.fr\/blog\/deep-learning\/qu-est-ce-qu-un-gru-gated-recurrent-unit\/","title":{"rendered":"Qu&rsquo;est-ce qu&rsquo;un GRU ? Deux portes pour une m\u00e9moire l\u00e9g\u00e8re"},"content":{"rendered":"<p><\/p>\n<p>Imagine que tu lises une critique de film avec pour seul support une page de brouillon. \u00c0 chaque phrase, tu choisis : qu&rsquo;est-ce que j&rsquo;efface, qu&rsquo;est-ce que je garde, qu&rsquo;est-ce que j&rsquo;ajoute ? Un GRU fait exactement ce choix, mot apr\u00e8s mot, sur une page unique. Tu vas voir comment il s&rsquo;y prend, ce qu&rsquo;il simplifie par rapport au LSTM et o\u00f9 sa l\u00e9g\u00e8ret\u00e9 se paie.<\/p>\n<p><strong>Sommaire<\/strong><\/p>\n<ul>\n<li><a href=\"#l-essentiel-en-30-secondes\">L&rsquo;essentiel en 30 secondes<\/a><\/li>\n<li><a href=\"#un-seul-carnet-au-lieu-de-deux-ce-qu-est-un-gru\">Un seul carnet au lieu de deux : ce qu&rsquo;est un GRU<\/a><\/li>\n<li><a href=\"#les-deux-portes-qui-decide-de-quoi\">Les deux portes : qui d\u00e9cide de quoi<\/a><\/li>\n<li><a href=\"#un-pas-de-calcul-du-brouillon-a-l-etat-suivant\">Un pas de calcul : du brouillon \u00e0 l&rsquo;\u00e9tat suivant<\/a><\/li>\n<li><a href=\"#gru-contre-lstm-ce-que-la-simplification-retire\">GRU contre LSTM : ce que la simplification retire<\/a><\/li>\n<li><a href=\"#un-gru-au-travail-lire-un-avis-de-film\">Un GRU au travail : lire un avis de film<\/a><\/li>\n<li><a href=\"#leger-ne-veut-pas-dire-gratuit\">L\u00e9ger ne veut pas dire gratuit<\/a><\/li>\n<\/ul>\n<h2 id=\"l-essentiel-en-30-secondes\">L&rsquo;essentiel en 30 secondes<\/h2>\n<ul>\n<li>Un GRU (<em>gated recurrent unit<\/em>) est une <strong>variante all\u00e9g\u00e9e du LSTM<\/strong> : il a lui aussi des portes internes, mais il est beaucoup plus simple.<\/li>\n<li>Il ne garde <strong>qu&rsquo;un seul \u00e9tat<\/strong>, l&rsquo;\u00e9tat cach\u00e9, l\u00e0 o\u00f9 le LSTM en entretient deux.<\/li>\n<li>Il a <strong>deux portes<\/strong> : la porte de mise \u00e0 jour <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/> et la porte de r\u00e9initialisation <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/>.<\/li>\n<li>Le nouvel \u00e9tat est <strong>un m\u00e9lange entre ancien \u00e9tat et candidat<\/strong>, dos\u00e9 par <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/>.<\/li>\n<li>Moins de param\u00e8tres et une convergence parfois plus rapide, mais <strong>un risque de surapprentissage plus \u00e9lev\u00e9<\/strong> et une m\u00e9moire moins fine \u00e0 tr\u00e8s long terme.<\/li>\n<\/ul>\n<h2 id=\"un-seul-carnet-au-lieu-de-deux-ce-qu-est-un-gru\">Un seul carnet au lieu de deux : ce qu&rsquo;est un GRU<\/h2>\n<p>Un <strong>GRU<\/strong> est une cellule de <a href=\"\/blog\/deep-learning\/que-sont-les-reseaux-recurrents\/\" title=\"r\u00e9seau r\u00e9current\">r\u00e9seau r\u00e9current<\/a> pens\u00e9e comme une version simplifi\u00e9e du LSTM. Il lutte comme lui contre le gradient \u00e9vanescent, et garde du LSTM quatre ingr\u00e9dients : l&rsquo;addition, le produit de Hadamard, un contexte port\u00e9 d&rsquo;un pas \u00e0 l&rsquo;autre, et des portes qui contr\u00f4lent l&rsquo;information dans la couche.<\/p>\n<p>Reprenons le brouillon. Le LSTM tient deux pages : une m\u00e9moire courte <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-14b463d0ecd5b350ced6cf1d6a12eef3_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#104;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"10\" style=\"vertical-align: 0px;\"\/>, utile \u00e0 l&rsquo;instant <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-b4e3cbf5d4c5c6d9b702dd139f14c147_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#116;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"6\" style=\"vertical-align: 0px;\"\/>, et une m\u00e9moire longue <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-41a04eeea923a1a0c28094a8a4680525_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#99;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/>. Le GRU n&rsquo;en tient <strong>qu&rsquo;une seule<\/strong>, son \u00e9tat cach\u00e9 <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-14b463d0ecd5b350ced6cf1d6a12eef3_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#104;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"10\" style=\"vertical-align: 0px;\"\/>, qui porte \u00e0 la fois ce qu&rsquo;il faut sortir maintenant et ce qu&rsquo;il faut retenir pour la suite.<\/p>\n<blockquote>\n<p><strong>\ud83d\udca1 Intuition<\/strong> Pense \u00e0 une page de brouillon que tu r\u00e9\u00e9cris \u00e0 chaque nouvelle phrase. Deux curseurs r\u00e8glent la r\u00e9\u00e9criture : l&rsquo;un dit combien de l&rsquo;ancien texte tu conserves, l&rsquo;autre combien tu consultes de l&rsquo;ancien texte avant de r\u00e9diger le nouveau.<\/p>\n<\/blockquote>\n<p>Malgr\u00e9 un nombre de param\u00e8tres r\u00e9duit, un GRU <strong>peut converger plus vite<\/strong> et atteindre des performances comparables \u00e0 celles du LSTM.<\/p>\n<h2 id=\"les-deux-portes-qui-decide-de-quoi\">Les deux portes : qui d\u00e9cide de quoi<\/h2>\n<p>Les deux portes sont calcul\u00e9es de la m\u00eame fa\u00e7on, \u00e0 partir de l&rsquo;entr\u00e9e courante <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-e95410b7a7a6a79cbcbdc56ca038a68e_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#120;&#95;&#116;\" title=\"Rendered by QuickLaTeX.com\" height=\"11\" width=\"15\" style=\"vertical-align: -3px;\"\/> et de l&rsquo;\u00e9tat pr\u00e9c\u00e9dent <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-b42e7d0229e1695947d3b9bbecaae71d_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#104;&#94;&#123;&#40;&#116;&#45;&#49;&#41;&#125;\" title=\"Rendered by QuickLaTeX.com\" height=\"16\" width=\"42\" style=\"vertical-align: 0px;\"\/> :<\/p>\n<p><p class=\"ql-center-displayed-equation\" style=\"line-height: 24px;\"><span class=\"ql-right-eqno\"> &nbsp; <\/span><span class=\"ql-left-eqno\"> &nbsp; <\/span><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-d7f7bbabbbd9d55a8253374e865f271d_l3.png\" height=\"24\" width=\"525\" class=\"ql-img-displayed-equation quicklatex-auto-format\" alt=\"&#92;&#91; &#122;&#94;&#123;&#40;&#116;&#41;&#125;&#61;&#92;&#115;&#105;&#103;&#109;&#97;&#92;&#98;&#105;&#103;&#108;&#40;&#98;&#95;&#122;&#43;&#85;&#95;&#122;&#104;&#94;&#123;&#40;&#116;&#45;&#49;&#41;&#125;&#43;&#87;&#95;&#122;&#120;&#95;&#116;&#92;&#98;&#105;&#103;&#114;&#41;&#44;&#92;&#113;&#113;&#117;&#97;&#100;&#32;&#114;&#94;&#123;&#40;&#116;&#41;&#125;&#61;&#92;&#115;&#105;&#103;&#109;&#97;&#92;&#98;&#105;&#103;&#108;&#40;&#98;&#95;&#114;&#43;&#85;&#95;&#114;&#104;&#94;&#123;&#40;&#116;&#45;&#49;&#41;&#125;&#43;&#87;&#95;&#114;&#120;&#95;&#116;&#92;&#98;&#105;&#103;&#114;&#41;&#46;&#92;&#93;\" title=\"Rendered by QuickLaTeX.com\"\/><\/p><\/p>\n<p>Chaque porte a son biais (<img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-751f1fe2814f49852094cc2c217075b9_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#98;&#95;&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"15\" width=\"15\" style=\"vertical-align: -3px;\"\/>, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-bf1457977f0d30415a1af36540dda25e_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#98;&#95;&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"15\" width=\"14\" style=\"vertical-align: -3px;\"\/>) et ses matrices de poids : <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4caed22919a1780df1b6310b338b904e_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#87;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"19\" style=\"vertical-align: 0px;\"\/> pond\u00e8re l&rsquo;entr\u00e9e, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-2b60fc262803f27ba3717d8ec4eb656d_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#85;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"13\" style=\"vertical-align: 0px;\"\/> pond\u00e8re le pass\u00e9. La <a href=\"\/blog\/maths\/qu-est-ce-qu-une-sigmoide\/\" title=\"sigmo\u00efde\">sigmo\u00efde<\/a> <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-1c9cc40f96a1492e298e7da85a2c1692_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#92;&#115;&#105;&#103;&#109;&#97;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"11\" style=\"vertical-align: 0px;\"\/> ram\u00e8ne chaque composante entre 0 et 1, ce qui donne \u00e0 la porte son r\u00f4le d&rsquo;interrupteur progressif.<\/p>\n<p>La <strong>porte de mise \u00e0 jour <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/><\/strong> descend de la porte d&rsquo;oubli du LSTM : quand ses valeurs sont proches de 1, l&rsquo;information importante est conserv\u00e9e ; proches de 0, l&rsquo;information sans importance est r\u00e9\u00e9crite.<\/p>\n<p>La <strong>porte de r\u00e9initialisation <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/><\/strong> descend de la porte d&rsquo;entr\u00e9e, avec une diff\u00e9rence de taille : elle n&rsquo;est plus ind\u00e9pendante, elle est reli\u00e9e \u00e0 la porte de mise \u00e0 jour. Elle r\u00e8gle la part de l&rsquo;\u00e9tat pr\u00e9c\u00e9dent ignor\u00e9e quand on calcule le nouvel \u00e9tat.<\/p>\n<h2 id=\"un-pas-de-calcul-du-brouillon-a-l-etat-suivant\">Un pas de calcul : du brouillon \u00e0 l&rsquo;\u00e9tat suivant<\/h2>\n<p>Une fois <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/> et <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/> connues, le GRU fabrique un <strong>\u00e9tat candidat<\/strong> <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-177b472bf86e5f546878eecd3036ffda_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#92;&#116;&#105;&#108;&#100;&#101;&#32;&#104;&#94;&#123;&#40;&#116;&#41;&#125;\" title=\"Rendered by QuickLaTeX.com\" height=\"17\" width=\"24\" style=\"vertical-align: 0px;\"\/>, le brouillon du nouvel \u00e9tat, puis le m\u00e9lange avec l&rsquo;ancien :<\/p>\n<p><p class=\"ql-center-displayed-equation\" style=\"line-height: 24px;\"><span class=\"ql-right-eqno\"> &nbsp; <\/span><span class=\"ql-left-eqno\"> &nbsp; <\/span><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c34892536a1f3411a84ac02065f97f54_l3.png\" height=\"24\" width=\"582\" class=\"ql-img-displayed-equation quicklatex-auto-format\" alt=\"&#92;&#91; &#92;&#116;&#105;&#108;&#100;&#101;&#32;&#104;&#94;&#123;&#40;&#116;&#41;&#125;&#61;&#92;&#116;&#97;&#110;&#104;&#92;&#98;&#105;&#103;&#108;&#40;&#87;&#95;&#104;&#120;&#95;&#116;&#43;&#114;&#94;&#123;&#40;&#116;&#41;&#125;&#92;&#111;&#100;&#111;&#116;&#32;&#85;&#95;&#104;&#104;&#94;&#123;&#40;&#116;&#45;&#49;&#41;&#125;&#92;&#98;&#105;&#103;&#114;&#41;&#44;&#92;&#113;&#113;&#117;&#97;&#100;&#32;&#104;&#94;&#123;&#40;&#116;&#41;&#125;&#61;&#122;&#94;&#123;&#40;&#116;&#41;&#125;&#92;&#111;&#100;&#111;&#116;&#32;&#104;&#94;&#123;&#40;&#116;&#45;&#49;&#41;&#125;&#43;&#92;&#98;&#105;&#103;&#108;&#40;&#49;&#45;&#122;&#94;&#123;&#40;&#116;&#41;&#125;&#92;&#98;&#105;&#103;&#114;&#41;&#92;&#111;&#100;&#111;&#116;&#92;&#116;&#105;&#108;&#100;&#101;&#32;&#104;&#94;&#123;&#40;&#116;&#41;&#125;&#46;&#92;&#93;\" title=\"Rendered by QuickLaTeX.com\"\/><\/p><\/p>\n<p>Lis la premi\u00e8re \u00e9galit\u00e9 : le pass\u00e9 <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-b42e7d0229e1695947d3b9bbecaae71d_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#104;&#94;&#123;&#40;&#116;&#45;&#49;&#41;&#125;\" title=\"Rendered by QuickLaTeX.com\" height=\"16\" width=\"42\" style=\"vertical-align: 0px;\"\/> n&rsquo;entre dans le brouillon qu&rsquo;apr\u00e8s avoir travers\u00e9 le filtre <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/>. Si <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/> est proche de 0, le brouillon ne repose presque plus que sur l&rsquo;entr\u00e9e <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-e95410b7a7a6a79cbcbdc56ca038a68e_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#120;&#95;&#116;\" title=\"Rendered by QuickLaTeX.com\" height=\"11\" width=\"15\" style=\"vertical-align: -3px;\"\/> : c&rsquo;est la part \u00ab ignor\u00e9e \u00bb de l&rsquo;\u00e9tat pr\u00e9c\u00e9dent. Cette porte d\u00e9cide donc quelle nouvelle information entre et quelle part utile du pass\u00e9 reste.<\/p>\n<p>La seconde \u00e9galit\u00e9 est le c\u0153ur du dispositif. Les poids <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/> et <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-0d370947d5cda05da7ccba83a4e9107a_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#49;&#45;&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"39\" style=\"vertical-align: 0px;\"\/> s&rsquo;additionnent \u00e0 1 : le nouvel \u00e9tat est <strong>une moyenne pond\u00e9r\u00e9e<\/strong>. Avec <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/> proche de 1, l&rsquo;\u00e9tat traverse le pas presque inchang\u00e9 ; avec <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/> proche de 0, il est remplac\u00e9 par le candidat.<\/p>\n<p>Le LSTM, lui, r\u00e8gle l&rsquo;oubli et l&rsquo;ajout avec deux portes ind\u00e9pendantes, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-9c09a708375fde2676da319bcdfe8b24_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#102;\" title=\"Rendered by QuickLaTeX.com\" height=\"16\" width=\"10\" style=\"vertical-align: -4px;\"\/> et <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-695d9d59bd04859c6c99e7feb11daab6_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#105;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"6\" style=\"vertical-align: 0px;\"\/>. Le GRU n&rsquo;en utilise qu&rsquo;une : tout ce qu&rsquo;il garde de l&rsquo;ancien, il le retire au nouveau.<\/p>\n<h2 id=\"gru-contre-lstm-ce-que-la-simplification-retire\">GRU contre LSTM : ce que la simplification retire<\/h2>\n<p>Le tableau met face \u00e0 face les deux cellules, telles que leurs \u00e9quations les d\u00e9crivent :<\/p>\n<table>\n<thead>\n<tr>\n<th><\/th>\n<th>LSTM<\/th>\n<th>GRU<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u00c9tats entretenus<\/td>\n<td>deux : <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-14b463d0ecd5b350ced6cf1d6a12eef3_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#104;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"10\" style=\"vertical-align: 0px;\"\/> et <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-41a04eeea923a1a0c28094a8a4680525_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#99;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/><\/td>\n<td>un seul : <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-14b463d0ecd5b350ced6cf1d6a12eef3_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#104;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"10\" style=\"vertical-align: 0px;\"\/><\/td>\n<\/tr>\n<tr>\n<td>Portes<\/td>\n<td>trois (oubli, entr\u00e9e, sortie), ind\u00e9pendantes<\/td>\n<td>deux (mise \u00e0 jour, r\u00e9initialisation), reli\u00e9es<\/td>\n<\/tr>\n<tr>\n<td>Blocs de poids dans les \u00e9quations<\/td>\n<td>quatre (<img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-9c09a708375fde2676da319bcdfe8b24_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#102;\" title=\"Rendered by QuickLaTeX.com\" height=\"16\" width=\"10\" style=\"vertical-align: -4px;\"\/>, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-695d9d59bd04859c6c99e7feb11daab6_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#105;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"6\" style=\"vertical-align: 0px;\"\/>, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-d208fd391fa57c168dc0f151de829fee_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#103;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"9\" style=\"vertical-align: -4px;\"\/>, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-673447f599d9874842c6998af0efc0ed_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#111;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/>)<\/td>\n<td>trois (<img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/>, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/>, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-76ec5ff1ccb97f752aba6f42c9c25c3c_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#92;&#116;&#105;&#108;&#100;&#101;&#32;&#104;\" title=\"Rendered by QuickLaTeX.com\" height=\"17\" width=\"10\" style=\"vertical-align: 0px;\"\/>)<\/td>\n<\/tr>\n<tr>\n<td>Valeurs rendues par la couche PyTorch<\/td>\n<td>sortie, \u00e9tat cach\u00e9 et \u00e9tat de cellule<\/td>\n<td>sortie et \u00e9tat cach\u00e9<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n<figure data-wp-context=\"{&quot;imageId&quot;:&quot;6adef46947a41&quot;}\" data-wp-interactive=\"core\/image\" data-wp-key=\"6adef46947a41\" class=\"wp-block-image size-large wp-lightbox-container\"><img loading=\"lazy\" decoding=\"async\" width=\"2400\" height=\"1500\" data-wp-class--hide=\"state.isContentHidden\" data-wp-class--show=\"state.isContentVisible\" data-wp-init=\"callbacks.setButtonStyles\" data-wp-on--click=\"actions.showLightbox\" data-wp-on--load=\"callbacks.setButtonStyles\" data-wp-on--pointerdown=\"actions.preloadImage\" data-wp-on--pointerenter=\"actions.preloadImageWithDelay\" data-wp-on--pointerleave=\"actions.cancelPreload\" data-wp-on-window--resize=\"callbacks.setButtonStyles\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_cellule_lstm.png\" alt=\"Cellule LSTM : en haut, la m\u00e9moire longue c passe par un bloc \u00ab oubli \u00bb puis un bloc \u00ab ajout \u00bb ; en dessous, trois portes (oubli f, entr\u00e9e i avec candidat g, sortie o avec tanh de c) produisent la m\u00e9moire courte h(t)\" class=\"wp-image-630\" srcset=\"https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_cellule_lstm.png 2400w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_cellule_lstm-300x188.png 300w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_cellule_lstm-1024x640.png 1024w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_cellule_lstm-768x480.png 768w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_cellule_lstm-1536x960.png 1536w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_cellule_lstm-2048x1280.png 2048w\" sizes=\"auto, (max-width: 2400px) 100vw, 2400px\" \/><button\n\t\t\tclass=\"lightbox-trigger\"\n\t\t\ttype=\"button\"\n\t\t\taria-haspopup=\"dialog\"\n\t\t\tdata-wp-bind--aria-label=\"state.thisImage.triggerButtonAriaLabel\"\n\t\t\tdata-wp-init=\"callbacks.initTriggerButton\"\n\t\t\tdata-wp-on--click=\"actions.showLightbox\"\n\t\t\tdata-wp-style--right=\"state.thisImage.buttonRight\"\n\t\t\tdata-wp-style--top=\"state.thisImage.buttonTop\"\n\t\t>\n\t\t\t<svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"12\" height=\"12\" fill=\"none\" viewBox=\"0 0 12 12\">\n\t\t\t\t<path fill=\"#fff\" d=\"M2 0a2 2 0 0 0-2 2v2h1.5V2a.5.5 0 0 1 .5-.5h2V0H2Zm2 10.5H2a.5.5 0 0 1-.5-.5V8H0v2a2 2 0 0 0 2 2h2v-1.5ZM8 12v-1.5h2a.5.5 0 0 0 .5-.5V8H12v2a2 2 0 0 1-2 2H8Zm2-12a2 2 0 0 1 2 2v2h-1.5V2a.5.5 0 0 0-.5-.5H8V0h2Z\" \/>\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n\n\n<p style=\"text-align: center;\"><em>Rep\u00e8re la ligne du haut, celle de la m\u00e9moire longue <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-41a04eeea923a1a0c28094a8a4680525_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#99;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/> : c&rsquo;est pr\u00e9cis\u00e9ment ce que le GRU supprime. Dans le sch\u00e9ma d&rsquo;un GRU, il ne resterait que la m\u00e9moire courte <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-14b463d0ecd5b350ced6cf1d6a12eef3_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#104;\" title=\"Rendered by QuickLaTeX.com\" height=\"12\" width=\"10\" style=\"vertical-align: 0px;\"\/>, pilot\u00e9e par <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/> et <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/>.<\/em><\/p>\n<p>Cette ligne en moins se traduit par <strong>moins de param\u00e8tres et moins de calcul<\/strong>. Le GRU sait quand m\u00eame traiter de longues s\u00e9quences sans oublier les entr\u00e9es pass\u00e9es, et il se comporte bien sur de nombreuses t\u00e2ches textuelles, en reconnaissance vocale ou en g\u00e9n\u00e9ration musicale.<\/p>\n<h2 id=\"un-gru-au-travail-lire-un-avis-de-film\">Un GRU au travail : lire un avis de film<\/h2>\n<p>Question concr\u00e8te : comment faire lire un avis \u00e0 un r\u00e9seau pour qu&rsquo;il rende un verdict, positif ou n\u00e9gatif ? On empile des couches GRU qui avalent le texte mot apr\u00e8s mot. Avant de l&rsquo;assembler, regarde la brique seule. Les extraits supposent que <code>numpy<\/code> et <code>torch<\/code> sont d\u00e9j\u00e0 import\u00e9s ; on fabrique un tenseur al\u00e9atoire de dimensions 10, 5, 3 et on le passe dans une couche GRU :<\/p>\n<pre><code>data_tensor = torch.tensor(np.random.random((10, 5, 3)),\n                           dtype=torch.float32)\nGRU =nn.GRU(input_size=3, hidden_size=10,\n            num_layers=1, batch_first=True)\noutput, hidden = GRU(data_tensor)\n<\/code><\/pre>\n<p>Chaque \u00e9l\u00e9ment de la s\u00e9quence est d\u00e9crit par 3 nombres (<code>input_size=3<\/code>) et l&rsquo;\u00e9tat cach\u00e9 en compte 10 (<code>hidden_size=10<\/code>). Remarque la derni\u00e8re ligne : la couche rend <code>output<\/code> et <code>hidden<\/code>, rien d&rsquo;autre. <strong>Pas d&rsquo;\u00e9tat de cellule<\/strong>, comme le tableau de la section pr\u00e9c\u00e9dente l&rsquo;annon\u00e7ait.<\/p>\n<p>Pour le vrai classifieur, l&rsquo;avis brut est converti en indices de mots, compl\u00e9t\u00e9 par du remplissage, puis transform\u00e9 en vecteurs de 300 dimensions. Ces vecteurs entrent dans un GRU de 3 couches de 256 unit\u00e9s avec un dropout de 0,5, et une couche lin\u00e9aire suivie d&rsquo;une sigmo\u00efde fournit la pr\u00e9diction.<\/p>\n\n\n<figure data-wp-context=\"{&quot;imageId&quot;:&quot;6adef4694807b&quot;}\" data-wp-interactive=\"core\/image\" data-wp-key=\"6adef4694807b\" class=\"wp-block-image size-large wp-lightbox-container\"><img loading=\"lazy\" decoding=\"async\" width=\"2250\" height=\"1260\" data-wp-class--hide=\"state.isContentHidden\" data-wp-class--show=\"state.isContentVisible\" data-wp-init=\"callbacks.setButtonStyles\" data-wp-on--click=\"actions.showLightbox\" data-wp-on--load=\"callbacks.setButtonStyles\" data-wp-on--pointerdown=\"actions.preloadImage\" data-wp-on--pointerenter=\"actions.preloadImageWithDelay\" data-wp-on--pointerleave=\"actions.cancelPreload\" data-wp-on-window--resize=\"callbacks.setButtonStyles\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_architecture_sentiments.png\" alt=\"Pipeline de classification d'avis : texte brut, indices de mots avec remplissage, embedding de 300 dimensions, GRU \u00e0 3 couches de 256 unit\u00e9s avec dropout de 0,5, couche lin\u00e9aire et sigmo\u00efde, pr\u00e9diction positive ou n\u00e9gative\" class=\"wp-image-703\" srcset=\"https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_architecture_sentiments.png 2250w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_architecture_sentiments-300x168.png 300w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_architecture_sentiments-1024x573.png 1024w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_architecture_sentiments-768x430.png 768w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_architecture_sentiments-1536x860.png 1536w, https:\/\/www.ranksignal.fr\/wp-content\/uploads\/2026\/10\/b1_ch01_architecture_sentiments-2048x1147.png 2048w\" sizes=\"auto, (max-width: 2250px) 100vw, 2250px\" \/><button\n\t\t\tclass=\"lightbox-trigger\"\n\t\t\ttype=\"button\"\n\t\t\taria-haspopup=\"dialog\"\n\t\t\tdata-wp-bind--aria-label=\"state.thisImage.triggerButtonAriaLabel\"\n\t\t\tdata-wp-init=\"callbacks.initTriggerButton\"\n\t\t\tdata-wp-on--click=\"actions.showLightbox\"\n\t\t\tdata-wp-style--right=\"state.thisImage.buttonRight\"\n\t\t\tdata-wp-style--top=\"state.thisImage.buttonTop\"\n\t\t>\n\t\t\t<svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"12\" height=\"12\" fill=\"none\" viewBox=\"0 0 12 12\">\n\t\t\t\t<path fill=\"#fff\" d=\"M2 0a2 2 0 0 0-2 2v2h1.5V2a.5.5 0 0 1 .5-.5h2V0H2Zm2 10.5H2a.5.5 0 0 1-.5-.5V8H0v2a2 2 0 0 0 2 2h2v-1.5ZM8 12v-1.5h2a.5.5 0 0 0 .5-.5V8H12v2a2 2 0 0 1-2 2H8Zm2-12a2 2 0 0 1 2 2v2h-1.5V2a.5.5 0 0 0-.5-.5H8V0h2Z\" \/>\n\t\t\t<\/svg>\n\t\t<\/button><figcaption class=\"wp-element-caption\">Suis le sch\u00e9ma de gauche \u00e0 droite : le GRU occupe l&rsquo;\u00e9tape centrale et transforme une suite de vecteurs en une repr\u00e9sentation que la couche lin\u00e9aire r\u00e9sume en une seule probabilit\u00e9.<\/figcaption><\/figure>\n\n\n<p>Le code qui fixe ces dimensions tient en quelques lignes :<\/p>\n<pre><code># Hyperparameters\nno_layers = 3\nvocab_size = len(vocab) + 1  # extra 1 for padding\nembedding_dim = 300\noutput_dim = 1\nhidden_dim = 256\n# Initialize the model\nmodel = SentimentRNN(no_layers, vocab_size, hidden_dim,\n                     embedding_dim, drop_prob=0.5)\n<\/code><\/pre>\n<p><code>vocab<\/code> est le vocabulaire des avis, auquel on ajoute une place pour le remplissage ; <code>SentimentRNN<\/code> est le mod\u00e8le qui assemble embedding, couches GRU et couche lin\u00e9aire. <code>output_dim = 1<\/code> correspond \u00e0 la probabilit\u00e9 unique que l&rsquo;avis soit positif, et le dropout stabilise l&rsquo;entra\u00eenement. Le verdict n&rsquo;ayant que deux issues, l&rsquo;entra\u00eenement s&rsquo;appuie sur l&rsquo;<a href=\"\/blog\/deep-learning\/l-entropie-croisee-binaire-bceloss\/\" title=\"entropie crois\u00e9e binaire\">entropie crois\u00e9e binaire<\/a> et l&rsquo;optimiseur Adam.<\/p>\n<h2 id=\"leger-ne-veut-pas-dire-gratuit\">L\u00e9ger ne veut pas dire gratuit<\/h2>\n<p>Deux id\u00e9es re\u00e7ues \u00e0 ranger.<\/p>\n<p><strong>\u00ab Un GRU est un LSTM au rabais. \u00bb<\/strong> Non : \u00e0 performances comparables, il demande moins de param\u00e8tres et peut converger plus vite. Mais il est <strong>moins efficace que le LSTM<\/strong> pour mod\u00e9liser des d\u00e9pendances \u00e0 tr\u00e8s long terme ou des motifs complexes.<\/p>\n<blockquote>\n<p><strong>\u26a0\ufe0f Pi\u00e8ge<\/strong> Moins de param\u00e8tres, c&rsquo;est aussi <strong>plus de risque de surapprentissage<\/strong>. Un GRU l\u00e9ger n&rsquo;est pas un GRU sans r\u00e9glage de r\u00e9gularisation, comme le dropout de l&rsquo;exemple pr\u00e9c\u00e9dent.<\/p>\n<\/blockquote>\n<p><strong>\u00ab Les portes sont cod\u00e9es \u00e0 la main. \u00bb<\/strong> Non plus : <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/> et <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/> sont des calculs, avec des poids propres \u00e0 chaque porte, appliqu\u00e9s \u00e0 l&rsquo;entr\u00e9e et \u00e0 l&rsquo;\u00e9tat du moment. Rien n&rsquo;est fig\u00e9 d&rsquo;avance.<\/p>\n<blockquote>\n<p><strong>\ud83d\udccc \u00c0 retenir<\/strong> Un GRU est un r\u00e9seau r\u00e9current \u00e0 une seule m\u00e9moire et deux portes : <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-c409433a9e2dfcdb83360a974d243f18_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#114;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"8\" style=\"vertical-align: 0px;\"\/> filtre le pass\u00e9 pour le brouillon, <img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.ranksignal.fr\/wp-content\/ql-cache\/quicklatex.com-4586e340cb83d5b642972e97a288fec2_l3.png\" class=\"ql-img-inline-formula quicklatex-auto-format\" alt=\"&#122;\" title=\"Rendered by QuickLaTeX.com\" height=\"8\" width=\"9\" style=\"vertical-align: 0px;\"\/> dose le m\u00e9lange entre ancien \u00e9tat et brouillon.<\/p>\n<\/blockquote>","protected":false},"excerpt":{"rendered":"<p>Un GRU (Gated Recurrent Unit) est un LSTM all\u00e9g\u00e9 : un seul \u00e9tat, deux portes. D\u00e9couvre ce que font z et r, comment l&rsquo;\u00e9tat se met \u00e0 jour et ses limites.<\/p>\n","protected":false},"author":1,"featured_media":704,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-705","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-deep-learning"],"_links":{"self":[{"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/posts\/705","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/comments?post=705"}],"version-history":[{"count":3,"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/posts\/705\/revisions"}],"predecessor-version":[{"id":792,"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/posts\/705\/revisions\/792"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/media\/704"}],"wp:attachment":[{"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/media?parent=705"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/categories?post=705"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ranksignal.fr\/blog\/wp-json\/wp\/v2\/tags?post=705"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}