Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Tokenisation <i> Unigram </i>","local":"tokenisation-i-unigram-i","sections":[{"title":"Algorithme d’entraînement","local":"algorithme-dentraînement","sections":[],"depth":2},{"title":"Algorithme de tokenisation","local":"algorithme-de-tokenisation","sections":[],"depth":2},{"title":"Retour à l’entraînement","local":"retour-à-lentraînement","sections":[],"depth":2},{"title":"Implémentation d’ <i> Unigram </i>","local":"implémentation-d-i-unigram-i","sections":[],"depth":2}],"depth":1}"> | |
| <link href="/docs/course/pr_1007/fr/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/entry/start.0dcef240.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/scheduler.37c15a92.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/singletons.f3bf6ff5.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/index.18351ede.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/paths.4129b87c.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/entry/app.e06c0eda.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/index.2bf4358c.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/nodes/0.d3d72fb7.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/each.e59479a4.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/nodes/50.1a777884.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/Tip.363c041f.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/Youtube.1e50a667.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/CodeBlock.4e987730.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/CourseFloatingBanner.6add7356.js"> | |
| <link rel="modulepreload" href="/docs/course/pr_1007/fr/_app/immutable/chunks/getInferenceSnippets.1837c472.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"Tokenisation <i> Unigram </i>","local":"tokenisation-i-unigram-i","sections":[{"title":"Algorithme d’entraînement","local":"algorithme-dentraînement","sections":[],"depth":2},{"title":"Algorithme de tokenisation","local":"algorithme-de-tokenisation","sections":[],"depth":2},{"title":"Retour à l’entraînement","local":"retour-à-lentraînement","sections":[],"depth":2},{"title":"Implémentation d’ <i> Unigram </i>","local":"implémentation-d-i-unigram-i","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="tokenisation-i-unigram-i" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#tokenisation-i-unigram-i"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Tokenisation <i> Unigram </i></span></h1> <div class="flex space-x-1 absolute z-10 right-0 top-0"><a href="https://discuss.huggingface.co/t/chapter-6-questions" target="_blank"><img alt="Ask a Question" class="!m-0" src="https://img.shields.io/badge/Ask%20a%20question-ffcb4c.svg?logo=data:image/svg+xml;base64,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"></a> <div class="relative colab-dropdown "> <button class=" " type="button"> <img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"> </button> </div> <div class="relative colab-dropdown "> <button class=" " type="button"> <img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"> </button> </div></div> <p data-svelte-h="svelte-ejzgmf">L’algorithme <em>Unigram</em> est souvent utilisé dans <em>SentencePiece</em>, qui est l’algorithme de tokenization utilisé par des modèles comme ALBERT, T5, mBART, Big Bird et XLNet.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/TGZfZVuF9Yc" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-e1td63">💡 Cette section couvre <em>Unigram</em> en profondeur, allant jusqu’à montrer une implémentation complète. Vous pouvez passer directement à la fin si vous souhaitez simplement avoir un aperçu général de l’algorithme de tokénisation.</p></div> <h2 class="relative group"><a id="algorithme-dentraînement" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#algorithme-dentraînement"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Algorithme d’entraînement</span></h2> <p data-svelte-h="svelte-e69bka">Comparé au BPE et <em>WordPiece</em>, <em>Unigram</em> fonctionne dans l’autre sens : il part d’un grand vocabulaire et enlève des <em>tokens</em> jusqu’à atteindre la taille de vocabulaire désirée. Il existe plusieurs options pour construire ce vocabulaire de base. Nous pouvons par exemple prendre les sous-chaînes les plus courantes dans les mots prétokénisés ou appliquer le BPE sur le corpus initial avec une grande taille de vocabulaire.</p> <p data-svelte-h="svelte-17f0cxc">À chaque étape de l’entraînement, l’algorithme <em>Unigram</em> calcule une perte sur le corpus compte tenu du vocabulaire actuel. Ensuite, pour chaque symbole du vocabulaire, l’algorithme calcule de combien la perte globale augmenterait si le symbole était supprimé et recherche les symboles qui l’augmenteraient le moins. Ces symboles ont un effet moindre sur la perte globale du corpus, ils sont donc en quelque sorte « moins nécessaires » et sont les meilleurs candidats à la suppression.</p> <p>Comme il s’agit d’une opération très coûteuse, nous ne nous contentons pas de supprimer le symbole unique associé à la plus faible augmentation de la perte mais le<!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi></mrow><annotation encoding="application/x-tex">p</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.625em;vertical-align:-0.1944em;"></span><span class="mord mathnormal">p</span></span></span></span><!-- HTML_TAG_END --> pourcent des symboles associés à la plus faible augmentation de la perte.<!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi></mrow><annotation encoding="application/x-tex">p</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.625em;vertical-align:-0.1944em;"></span><span class="mord mathnormal">p</span></span></span></span><!-- HTML_TAG_END --> est un hyperparamètre que vous pouvez contrôler, valant généralement 10 ou 20. Ce processus est ensuite répété jusqu’à ce que le vocabulaire ait atteint la taille souhaitée.</p> <p data-svelte-h="svelte-1h5svcj">Notez que nous ne supprimons jamais les caractères de base, afin de nous assurer que tout mot peut être tokenisé.</p> <p data-svelte-h="svelte-c0t4xq">Tout ceci peut paraître encore un peu vague. En effet, la partie principale de l’algorithme est de calculer une perte sur le corpus et de voir comment elle change lorsque nous supprimons certains <em>tokens</em> du vocabulaire mais nous n’avons pas encore expliqué comment le faire. Cette étape repose sur l’algorithme de tokénisation <em>Unigram</em>, nous allons donc l’aborder à présent.</p> <p data-svelte-h="svelte-13r8aw9">Nous allons réutiliser le corpus des exemples précédents :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">"hug"</span>, <span class="hljs-number">10</span>), (<span class="hljs-string">"pug"</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">"pun"</span>, <span class="hljs-number">12</span>), (<span class="hljs-string">"bun"</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">"hugs"</span>, <span class="hljs-number">5</span>) <span class="hljs-meta"># <span class="hljs-string">"câlin"</span>, <span class="hljs-string">"carlin"</span>, <span class="hljs-string">"jeu de mots"</span>, <span class="hljs-string">"brioche"</span>, <span class="hljs-string">"câlins"</span>...</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1f1ss07">et pour cet exemple, nous prendrons toutes les sous-chaînes strictes pour le vocabulaire initial :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-selector-attr">[<span class="hljs-string">"h"</span>, <span class="hljs-string">"u"</span>, <span class="hljs-string">"g"</span>, <span class="hljs-string">"hu"</span>, <span class="hljs-string">"ug"</span>, <span class="hljs-string">"p"</span>, <span class="hljs-string">"pu"</span>, <span class="hljs-string">"n"</span>, <span class="hljs-string">"un"</span>, <span class="hljs-string">"b"</span>, <span class="hljs-string">"bu"</span>, <span class="hljs-string">"s"</span>, <span class="hljs-string">"hug"</span>, <span class="hljs-string">"gs"</span>, <span class="hljs-string">"ugs"</span>]</span><!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="algorithme-de-tokenisation" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#algorithme-de-tokenisation"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Algorithme de tokenisation</span></h2> <p data-svelte-h="svelte-igrasz">Un modèle <em>Unigram</em> est un type de modèle de langage qui considère que chaque <em>token</em> est indépendant des <em>tokens</em> qui le précèdent. Il s’agit du modèle de langage le plus simple, dans le sens où la probabilité du <em>token</em> X compte tenu du contexte précédent est simplement la probabilité du <em>token</em> X. Ainsi, si nous utilisions un modèle de langage <em>Unigram</em> pour générer du texte, nous prédirions toujours le <em>token</em> le plus courant.</p> <p data-svelte-h="svelte-rrnk76">La probabilité d’un <em>token</em> donné est sa fréquence (le nombre de fois que nous le trouvons) dans le corpus original, divisée par la somme de toutes les fréquences de tous les <em>tokens</em> dans le vocabulaire (pour s’assurer que la somme des probabilités est égale à 1). Par exemple, <code>"ug"</code> est présent dans <code>"hug"</code>, <code>"pug"</code>, et <code>"hugs"</code>. Il a donc une fréquence de 20 dans notre corpus.</p> <p data-svelte-h="svelte-uw5nyj">Voici les fréquences de tous les sous-mots possibles dans le vocabulaire :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">"h"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">15</span>) (<span class="hljs-string">"u"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">36</span>) (<span class="hljs-string">"g"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">20</span>) (<span class="hljs-string">"hu"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">15</span>) (<span class="hljs-string">"ug"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">20</span>) (<span class="hljs-string">"p"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">17</span>) (<span class="hljs-string">"pu"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">17</span>) (<span class="hljs-string">"n"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">16</span>) | |
| (<span class="hljs-string">"un"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">16</span>) (<span class="hljs-string">"b"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>) (<span class="hljs-string">"bu"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>) (<span class="hljs-string">"s"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>) (<span class="hljs-string">"hug"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">15</span>) (<span class="hljs-string">"gs"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>) (<span class="hljs-string">"ugs"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1bp6f0h">Ainsi, la somme de toutes les fréquences est de 210 et la probabilité du sous-mot <code>"ug"</code> est donc de 20/210.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1qr0cxz">✏️ <strong>A votre tour !</strong> Ecrivez le code permettant de calculer les fréquences ci-dessus et vérifiez que les résultats affichés sont corrects, de même que la somme totale.</p></div> <p>Maintenant, pour tokeniser un mot donné, nous examinons toutes les segmentations possibles en <em data-svelte-h="svelte-rqf02o">tokens</em> et calculons la probabilité de chacune d’entre elles selon le modèle <em data-svelte-h="svelte-1qp6mex">Unigram</em>. Puisque tous les <em data-svelte-h="svelte-rqf02o">tokens</em> sont considérés comme indépendants, cette probabilité est juste le produit de la probabilité de chaque <em data-svelte-h="svelte-1ayo9ht">token</em>. Par exemple, la tokenisation <code data-svelte-h="svelte-1n2m4po">["p", "u", "g"]</code> de <code data-svelte-h="svelte-1gjdq76">"pug"</code> a la probabilité : | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><mi>P</mi><mo stretchy="false">(</mo><mo stretchy="false">[</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>p</mi><mi mathvariant="normal">"</mi><mo separator="true">,</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>u</mi><mi mathvariant="normal">"</mi><mo separator="true">,</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>g</mi><mi mathvariant="normal">"</mi><mo stretchy="false">]</mo><mo stretchy="false">)</mo><mo>=</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>p</mi><mi mathvariant="normal">"</mi><mo stretchy="false">)</mo><mo>×</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>u</mi><mi mathvariant="normal">"</mi><mo stretchy="false">)</mo><mo>×</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>g</mi><mi mathvariant="normal">"</mi><mo stretchy="false">)</mo><mo>=</mo><mfrac><mn>5</mn><mn>210</mn></mfrac><mo>×</mo><mfrac><mn>36</mn><mn>210</mn></mfrac><mo>×</mo><mfrac><mn>20</mn><mn>210</mn></mfrac><mo>=</mo><mn>0.000389</mn></mrow><annotation encoding="application/x-tex">P([``p", ``u", ``g"]) = P(``p") \times P(``u") \times P(``g") = \frac{5}{210} \times \frac{36}{210} \times \frac{20}{210} = 0.000389</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">([</span><span class="mord">‘‘</span><span class="mord mathnormal">p</span><span class="mord">"</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">‘‘</span><span class="mord mathnormal">u</span><span class="mord">"</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">‘‘</span><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="mord">"</span><span class="mclose">])</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal">p</span><span class="mord">"</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal">u</span><span class="mord">"</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="mord">"</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">5</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">36</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">20</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">0.000389</span></span></span></span></span><!-- HTML_TAG_END --></p> <p>Comparativement, la tokenization <code data-svelte-h="svelte-42m5r0">["pu", "g"]</code> a la probabilité : | |
| <!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><mi>P</mi><mo stretchy="false">(</mo><mo stretchy="false">[</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>p</mi><mi>u</mi><mi mathvariant="normal">"</mi><mo separator="true">,</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>g</mi><mi mathvariant="normal">"</mi><mo stretchy="false">]</mo><mo stretchy="false">)</mo><mo>=</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>p</mi><mi>u</mi><mi mathvariant="normal">"</mi><mo stretchy="false">)</mo><mo>×</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>g</mi><mi mathvariant="normal">"</mi><mo stretchy="false">)</mo><mo>=</mo><mfrac><mn>5</mn><mn>210</mn></mfrac><mo>×</mo><mfrac><mn>20</mn><mn>210</mn></mfrac><mo>=</mo><mn>0.0022676</mn></mrow><annotation encoding="application/x-tex">P([``pu", ``g"]) = P(``pu") \times P(``g") = \frac{5}{210} \times \frac{20}{210} = 0.0022676</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">([</span><span class="mord">‘‘</span><span class="mord mathnormal">p</span><span class="mord mathnormal">u</span><span class="mord">"</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">‘‘</span><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="mord">"</span><span class="mclose">])</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal">p</span><span class="mord mathnormal">u</span><span class="mord">"</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="mord">"</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">5</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">20</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">0.0022676</span></span></span></span></span><!-- HTML_TAG_END --></p> <p data-svelte-h="svelte-d6sbl2">donc celle-là est beaucoup plus probable. En général, les tokénisations comportant le moins de <em>tokens</em> possible auront la probabilité la plus élevée (en raison de la division par 210 répétée pour chaque <em>token</em>), ce qui correspond à ce que nous voulons intuitivement : diviser un mot en un nombre de <em>tokens</em> le plus faible possible.</p> <p data-svelte-h="svelte-l2t88i">La tokenisation d’un mot avec le modèle <em>Unigram</em> est donc la tokenisation avec la plus haute probabilité. Dans l’exemple de <code>"pug"</code>, voici les probabilités que nous obtiendrions pour chaque segmentation possible :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[<span class="hljs-string">"p"</span>, <span class="hljs-string">"u"</span>, <span class="hljs-string">"g"</span>] : 0.000389 | |
| [<span class="hljs-string">"p"</span>, <span class="hljs-string">"ug"</span>] : 0.0022676 | |
| [<span class="hljs-string">"pu"</span>, <span class="hljs-string">"g"</span>] : 0.0022676<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1wahuxe">Ainsi, <code>"pug"</code> sera tokenisé comme <code>["p", "ug"]</code> ou <code>["pu", "g"]</code>, selon la segmentation rencontrée en premier (notez que dans un corpus plus large, les cas d’égalité comme celui-ci seront rares).</p> <p data-svelte-h="svelte-l2f5bk">Dans ce cas-ci, cela a été facile de trouver toutes les segmentations possibles et de calculer leurs probabilités, mais en général ce sera un peu plus difficile. Il existe un algorithme classique utilisé pour cela, appelé <em>algorithme de Viterbi</em>. Essentiellement, on peut construire un graphe pour détecter les segmentations possibles d’un mot donné en disant qu’il existe une branche du caractère <em>a</em> au caractère <em>b</em> si le sous-mot de <em>a</em> à <em>b</em> est dans le vocabulaire, et attribuer à cette branche la probabilité du sous-mot.</p> <p data-svelte-h="svelte-151naiu">Pour trouver le chemin qui va avoir le meilleur score dans ce graphe, l’algorithme de Viterbi détermine, pour chaque position dans le mot, la segmentation avec le meilleur score qui se termine à cette position. Puisque nous allons du début à la fin, ce meilleur score peut être trouvé en parcourant en boucle tous les sous-mots se terminant à la position actuelle, puis en utilisant le meilleur score de tokenization de la position à laquelle ce sous-mot commence. Ensuite, il suffit de dérouler le chemin emprunté pour arriver à la fin.</p> <p data-svelte-h="svelte-fvrpie">Prenons un exemple en utilisant notre vocabulaire et le mot <code>"unhug"</code>. Pour chaque position, les sous-mots avec les meilleurs scores se terminant là sont les suivants :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-attribute">Character</span> <span class="hljs-number">0</span> (u): <span class="hljs-string">"u"</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">171429</span>) | |
| <span class="hljs-attribute">Character</span> <span class="hljs-number">1</span> (n): <span class="hljs-string">"un"</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">076191</span>) | |
| <span class="hljs-attribute">Character</span> <span class="hljs-number">2</span> (h): <span class="hljs-string">"un"</span> <span class="hljs-string">"h"</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">005442</span>) | |
| <span class="hljs-attribute">Character</span> <span class="hljs-number">3</span> (u): <span class="hljs-string">"un"</span> <span class="hljs-string">"hu"</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">005442</span>) | |
| <span class="hljs-attribute">Character</span> <span class="hljs-number">4</span> (g): <span class="hljs-string">"un"</span> <span class="hljs-string">"hug"</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">005442</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1tix5c5">Ainsi, <code>"unhug"</code> serait tokenisé comme <code>["un", "hug"]</code>.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-6lgye7">✏️ <strong>A votre tour !</strong> Déterminer la tokenization du mot <code>"huggun"</code> et son score.</p></div> <h2 class="relative group"><a id="retour-à-lentraînement" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#retour-à-lentraînement"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Retour à l’entraînement</span></h2> <p data-svelte-h="svelte-1bpewg3">Maintenant que nous avons vu comment fonctionne la tokenisation, nous pouvons nous plonger un peu plus profondément dans la perte utilisée pendant l’entraînement. À n’importe quelle étape, cette perte est calculée en tokenisant chaque mot du corpus, en utilisant le vocabulaire courant et le modèle <em>Unigram</em> déterminé par les fréquences de chaque <em>token</em> dans le corpus (comme vu précédemment).</p> <p data-svelte-h="svelte-1lp44gc">Chaque mot du corpus a un score, et la perte est le négatif du logarithme de ces scores, c’est-à-dire la somme pour tous les mots du corpus de tous les <code>-log(P(word))</code>.</p> <p data-svelte-h="svelte-2uprq6">Revenons à notre exemple avec le corpus suivant :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">"hug"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">10</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"pug"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"pun"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">12</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"bun"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"hugs"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-inowz0">La tokenisation de chaque mot avec leurs scores respectifs est :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-string">"hug"</span>: [<span class="hljs-string">"hug"</span>] <span class="hljs-comment">(score 0.071428)</span> | |
| <span class="hljs-string">"pug"</span>: [<span class="hljs-string">"pu"</span>, <span class="hljs-string">"g"</span>] <span class="hljs-comment">(score 0.007710)</span> | |
| <span class="hljs-string">"pun"</span>: [<span class="hljs-string">"pu"</span>, <span class="hljs-string">"n"</span>] <span class="hljs-comment">(score 0.006168)</span> | |
| <span class="hljs-string">"bun"</span>: [<span class="hljs-string">"bu"</span>, <span class="hljs-string">"n"</span>] <span class="hljs-comment">(score 0.001451)</span> | |
| <span class="hljs-string">"hugs"</span>: [<span class="hljs-string">"hug"</span>, <span class="hljs-string">"s"</span>] <span class="hljs-comment">(score 0.001701)</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-16ti1wj">Donc la perte est :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-attribute">10</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">071428</span>)) + <span class="hljs-number">5</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">007710</span>)) + <span class="hljs-number">12</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">006168</span>)) + <span class="hljs-number">4</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">001451</span>)) + <span class="hljs-number">5</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">001701</span>)) = <span class="hljs-number">169</span>.<span class="hljs-number">8</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-8p0mxb">Maintenant, nous devons calculer comment la suppression de chaque token affecte la perte. C’est plutôt fastidieux, donc nous allons le faire pour deux <em>tokens</em> ici et garder tout le processus pour quand nous aurons du code pour nous aider. Dans ce cas (très) particulier, nous avions deux tokenizations équivalentes de tous les mots. Par exmeple, comme nous l’avons vu précédemment, <code>"pug"</code> pourrait être tokenisé en <code>["p", "ug"]</code> avec le même score. Ainsi, enlever le token <code>"pu"</code> du vocabulaire donnera exactement la même perte.</p> <p data-svelte-h="svelte-igrdwl">D’un autre côté, supprimer le mot <code>"hug"</code> aggravera la perte, car la tokenisation de <code>"hug"</code> et <code>"hugs"</code> deviendra :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-string">"hug"</span>: [<span class="hljs-string">"hu"</span>, <span class="hljs-string">"g"</span>] <span class="hljs-comment">(score 0.006802)</span> | |
| <span class="hljs-string">"hugs"</span>: [<span class="hljs-string">"hu"</span>, <span class="hljs-string">"gs"</span>] <span class="hljs-comment">(score 0.001701)</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-dln45d">Ces changements entraîneront une augmentation de la perte de :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->- <span class="hljs-number">10</span> * (<span class="hljs-name">-log</span>(<span class="hljs-number">0.071428</span>)) + <span class="hljs-number">10</span> * (<span class="hljs-name">-log</span>(<span class="hljs-number">0.006802</span>)) = <span class="hljs-number">23.5</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1j4g1ck">Par conséquent, le token <code>"pu"</code> sera probablement retiré du vocabulaire, mais pas <code>"hug"</code>.</p> <h2 class="relative group"><a id="implémentation-d-i-unigram-i" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#implémentation-d-i-unigram-i"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Implémentation d’ <i> Unigram </i></span></h2> <p data-svelte-h="svelte-wg8gbk">Maintenant, implémentons tout ce que nous avons vu jusqu’à présent dans le code. Comme pour le BPE et <em>WordPiece</em>, ce n’est pas une implémentation efficace de l’algorithme <em>Unigram</em> (bien au contraire), mais elle devrait vous aider à le comprendre un peu mieux.</p> <p data-svelte-h="svelte-jpbogw">Nous allons utiliser le même corpus que précédemment comme exemple :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->corpus = [ | |
| <span class="hljs-string">"This is the Hugging Face Course."</span>, | |
| <span class="hljs-comment"># C'est le cours d'Hugging Face.</span> | |
| <span class="hljs-string">"This chapter is about tokenization."</span>, | |
| <span class="hljs-comment"># Ce chapitre traite de la tokenisation.</span> | |
| <span class="hljs-string">"This section shows several tokenizer algorithms."</span>, | |
| <span class="hljs-comment"># Cette section présente plusieurs algorithmes de *tokenizer*.</span> | |
| <span class="hljs-string">"Hopefully, you will be able to understand how they are trained and generate tokens."</span>, | |
| <span class="hljs-comment"># Avec un peu de chance, vous serez en mesure de comprendre comment ils sont entraînés et génèrent des *tokens*.</span> | |
| ]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1fbz4i8">Cette fois, nous allons utiliser <code>xlnet-base-cased</code> comme modèle :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer | |
| tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">"xlnet-base-cased"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-4qxqbb">Comme pour le BPE et <em>WordPiece</em>, nous commençons par compter le nombre d’occurrences de chaque mot dans le corpus :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> collections <span class="hljs-keyword">import</span> defaultdict | |
| word_freqs = defaultdict(<span class="hljs-built_in">int</span>) | |
| <span class="hljs-keyword">for</span> text <span class="hljs-keyword">in</span> corpus: | |
| words_with_offsets = tokenizer.backend_tokenizer.pre_tokenizer.pre_tokenize_str(text) | |
| new_words = [word <span class="hljs-keyword">for</span> word, offset <span class="hljs-keyword">in</span> words_with_offsets] | |
| <span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> new_words: | |
| word_freqs[word] += <span class="hljs-number">1</span> | |
| word_freqs<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1q7qer4">Ensuite, nous devons initialiser notre vocabulaire à une taille plus grande que celle du vocabulaire que nous voudrons à la fin. Nous devons inclure tous les caractères de base (sinon nous ne serons pas en mesure de tokeniser chaque mot), mais pour les sous-chaînes plus grandes, nous ne garderons que les plus communs. AInsi nous les trions par fréquence :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->char_freqs = defaultdict(<span class="hljs-built_in">int</span>) | |
| subwords_freqs = defaultdict(<span class="hljs-built_in">int</span>) | |
| <span class="hljs-keyword">for</span> word, freq <span class="hljs-keyword">in</span> word_freqs.items(): | |
| <span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">len</span>(word)): | |
| char_freqs[word[i]] += freq | |
| <span class="hljs-comment"># Boucle à travers les sous-mots de longueur au moins égale à 2</span> | |
| <span class="hljs-keyword">for</span> j <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(i + <span class="hljs-number">2</span>, <span class="hljs-built_in">len</span>(word) + <span class="hljs-number">1</span>): | |
| subwords_freqs[word[i:j]] += freq | |
| <span class="hljs-comment"># Trier les sous-mots par fréquence</span> | |
| sorted_subwords = <span class="hljs-built_in">sorted</span>(subwords_freqs.items(), key=<span class="hljs-keyword">lambda</span> x: x[<span class="hljs-number">1</span>], reverse=<span class="hljs-literal">True</span>) | |
| sorted_subwords[:<span class="hljs-number">10</span>]<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[(<span class="hljs-string">'▁t'</span>, <span class="hljs-number">7</span>), (<span class="hljs-string">'is'</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">'er'</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">'▁a'</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">'▁to'</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">'to'</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">'en'</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">'▁T'</span>, <span class="hljs-number">3</span>), (<span class="hljs-string">'▁Th'</span>, <span class="hljs-number">3</span>), (<span class="hljs-string">'▁Thi'</span>, <span class="hljs-number">3</span>)]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-sue0nu">Nous regroupons les caractères avec les meilleurs sous-mots pour arriver à un vocabulaire initial de taille 300 :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->token_freqs = <span class="hljs-built_in">list</span>(char_freqs.items()) + sorted_subwords[: <span class="hljs-number">300</span> - <span class="hljs-built_in">len</span>(char_freqs)] | |
| token_freqs = {token: freq <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs}<!-- HTML_TAG_END --></pre></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1od7h8p">💡 <em>SentencePiece</em> utilise un algorithme plus efficace appelé <em>Enhanced Suffix Array</em> (ESA) pour créer le vocabulaire initial.</p></div> <p data-svelte-h="svelte-1bmy4ss">Ensuite, nous calculons la somme de toutes les fréquences, pour convertir les fréquences en probabilités. Pour notre modèle, nous allons stocker les logarithmes des probabilités, car c’est plus stable numériquement d’additionner des logarithmes que de multiplier des petits nombres. Cela simplifiera aussi le calcul de la perte du modèle :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> math <span class="hljs-keyword">import</span> log | |
| total_sum = <span class="hljs-built_in">sum</span>([freq <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs.items()]) | |
| model = {token: -log(freq / total_sum) <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs.items()}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-a28omg">Maintenant la fonction principale est celle qui tokenise les mots en utilisant l’algorithme de Viterbi. Comme nous l’avons vu précédemment, cet algorithme calcule la meilleure segmentation de chaque sous-chaîne du mot que nous allons stocker dans une variable nommée <code>best_segmentations</code>. Nous allons stocker un dictionnaire par position dans le mot (de 0 à sa longueur totale), avec deux clés : l’index du début du dernier <em>token</em> dans la meilleure segmentation et le score de la meilleure segmentation. Avec l’index du début du dernier <em>token</em>, nous serons en mesure de récupérer la segmentation complète une fois que la liste est complètement remplie.</p> <p data-svelte-h="svelte-1510vjt">Le remplissage de la liste se fait à l’aide de deux boucles seulement : la boucle principale passe en revue chaque position de départ et la seconde boucle essaie toutes les sous-chaînes commençant à cette position de départ. Si la sous-chaîne est dans le vocabulaire, nous avons une nouvelle segmentation du mot jusqu’à cette position finale que nous comparons à ce qui est dans <code>best_segmentations</code>.</p> <p data-svelte-h="svelte-y6xm80">Une fois que la boucle principale est terminée, nous commençons juste à la fin et sautons d’une position de départ à une autre, en enregistrant les <em>tokens</em> au fur et à mesure, jusqu’à ce que nous atteignions le début du mot :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">encode_word</span>(<span class="hljs-params">word, model</span>): | |
| best_segmentations = [{<span class="hljs-string">"start"</span>: <span class="hljs-number">0</span>, <span class="hljs-string">"score"</span>: <span class="hljs-number">1</span>}] + [ | |
| {<span class="hljs-string">"start"</span>: <span class="hljs-literal">None</span>, <span class="hljs-string">"score"</span>: <span class="hljs-literal">None</span>} <span class="hljs-keyword">for</span> _ <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">len</span>(word)) | |
| ] | |
| <span class="hljs-keyword">for</span> start_idx <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">len</span>(word)): | |
| <span class="hljs-comment"># Doit être correctement rempli par les étapes précédentes de la boucle</span> | |
| best_score_at_start = best_segmentations[start_idx][<span class="hljs-string">"score"</span>] | |
| <span class="hljs-keyword">for</span> end_idx <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(start_idx + <span class="hljs-number">1</span>, <span class="hljs-built_in">len</span>(word) + <span class="hljs-number">1</span>): | |
| token = word[start_idx:end_idx] | |
| <span class="hljs-keyword">if</span> token <span class="hljs-keyword">in</span> model <span class="hljs-keyword">and</span> best_score_at_start <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>: | |
| score = model[token] + best_score_at_start | |
| <span class="hljs-comment"># Si nous avons trouvé une meilleure segmentation se terminant à end_idx, nous mettons à jour</span> | |
| <span class="hljs-keyword">if</span> ( | |
| best_segmentations[end_idx][<span class="hljs-string">"score"</span>] <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span> | |
| <span class="hljs-keyword">or</span> best_segmentations[end_idx][<span class="hljs-string">"score"</span>] > score | |
| ): | |
| best_segmentations[end_idx] = {<span class="hljs-string">"start"</span>: start_idx, <span class="hljs-string">"score"</span>: score} | |
| segmentation = best_segmentations[-<span class="hljs-number">1</span>] | |
| <span class="hljs-keyword">if</span> segmentation[<span class="hljs-string">"score"</span>] <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span>: | |
| <span class="hljs-comment"># Nous n'avons pas trouvé de tokenization du mot -> inconnu (<unk>)</span> | |
| <span class="hljs-keyword">return</span> [<span class="hljs-string">"<unk>"</span>], <span class="hljs-literal">None</span> | |
| score = segmentation[<span class="hljs-string">"score"</span>] | |
| start = segmentation[<span class="hljs-string">"start"</span>] | |
| end = <span class="hljs-built_in">len</span>(word) | |
| tokens = [] | |
| <span class="hljs-keyword">while</span> start != <span class="hljs-number">0</span>: | |
| tokens.insert(<span class="hljs-number">0</span>, word[start:end]) | |
| next_start = best_segmentations[start][<span class="hljs-string">"start"</span>] | |
| end = start | |
| start = next_start | |
| tokens.insert(<span class="hljs-number">0</span>, word[start:end]) | |
| <span class="hljs-keyword">return</span> tokens, score<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-dlk0sv">Nous pouvons déjà essayer notre modèle initial sur quelques mots :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-built_in">print</span>(encode_word(<span class="hljs-string">"Hopefully"</span>, model)) | |
| <span class="hljs-built_in">print</span>(encode_word(<span class="hljs-string">"This"</span>, model))<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->([<span class="hljs-string">'H'</span>, <span class="hljs-string">'o'</span>, <span class="hljs-string">'p'</span>, <span class="hljs-string">'e'</span>, <span class="hljs-string">'f'</span>, <span class="hljs-string">'u'</span>, <span class="hljs-string">'ll'</span>, <span class="hljs-string">'y'</span>], <span class="hljs-number">41.5157494601402</span>) | |
| ([<span class="hljs-string">'This'</span>], <span class="hljs-number">6.288267030694535</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1omgojr">Il est maintenant facile de calculer la perte du modèle sur le corpus !</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">compute_loss</span>(<span class="hljs-params">model</span>): | |
| loss = <span class="hljs-number">0</span> | |
| <span class="hljs-keyword">for</span> word, freq <span class="hljs-keyword">in</span> word_freqs.items(): | |
| _, word_loss = encode_word(word, model) | |
| loss += freq * word_loss | |
| <span class="hljs-keyword">return</span> loss<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-wxlxrf">Nous pouvons vérifier que cela fonctionne sur le modèle que nous avons :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->compute_loss(model)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-number">413.10377642940875</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-appzjd">Le calcul des scores pour chaque <em>token</em> n’est pas très difficile non plus. Il suffit de calculer la perte pour les modèles obtenus en supprimant chaque <em>token</em> :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> copy | |
| <span class="hljs-keyword">def</span> <span class="hljs-title function_">compute_scores</span>(<span class="hljs-params">model</span>): | |
| scores = {} | |
| model_loss = compute_loss(model) | |
| <span class="hljs-keyword">for</span> token, score <span class="hljs-keyword">in</span> model.items(): | |
| <span class="hljs-comment"># Nous gardons toujours les tokens de longueur 1.</span> | |
| <span class="hljs-keyword">if</span> <span class="hljs-built_in">len</span>(token) == <span class="hljs-number">1</span>: | |
| <span class="hljs-keyword">continue</span> | |
| model_without_token = copy.deepcopy(model) | |
| _ = model_without_token.pop(token) | |
| scores[token] = compute_loss(model_without_token) - model_loss | |
| <span class="hljs-keyword">return</span> scores<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1eldf9t">Nous pouvons l’essayer sur un <em>token</em> donné :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->scores = compute_scores(model) | |
| <span class="hljs-built_in">print</span>(scores[<span class="hljs-string">"ll"</span>]) | |
| <span class="hljs-built_in">print</span>(scores[<span class="hljs-string">"his"</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1fskz9w">Puisque <code>"ll"</code> est utilisé dans la tokenisation de <code>"Hopefully"</code>, et que le supprimer nous fera probablement utiliser le token <code>"l"</code> deux fois à la place, nous nous attendons à ce qu’il ait une perte positive. <code>"his"</code> n’est utilisé qu’à l’intérieur du mot <code>"This"</code>, qui est tokenisé comme lui-même, donc nous nous attendons à ce qu’il ait une perte nulle. Voici les résultats :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-number">6.376412403623874</span> | |
| <span class="hljs-number">0.0</span><!-- HTML_TAG_END --></pre></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1c7o6zx">💡 Cette approche est très inefficace, c’est pourquoi <em>SentencePiece</em> utilise une approximation de la perte du modèle sans le <em>token</em> X. Au lieu de partir de zéro, il remplace simplement le <em>token</em> X par sa segmentation dans le vocabulaire restant. De cette façon, tous les scores peuvent être calculés en une seule fois, en même temps que la perte du modèle.</p></div> <p data-svelte-h="svelte-1l6kcco">Une fois tout cela en place, la dernière chose à faire est d’ajouter les <em>tokens</em> spéciaux utilisés par le modèle au vocabulaire, puis de boucler jusqu’à ce que nous ayons élagué suffisamment de <em>tokens</em> du vocabulaire pour atteindre la taille souhaitée :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->percent_to_remove = <span class="hljs-number">0.1</span> | |
| <span class="hljs-keyword">while</span> <span class="hljs-built_in">len</span>(model) > <span class="hljs-number">100</span>: | |
| scores = compute_scores(model) | |
| sorted_scores = <span class="hljs-built_in">sorted</span>(scores.items(), key=<span class="hljs-keyword">lambda</span> x: x[<span class="hljs-number">1</span>]) | |
| <span class="hljs-comment"># Supprime les tokens percent_to_remove ayant les scores les plus bas</span> | |
| <span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">int</span>(<span class="hljs-built_in">len</span>(model) * percent_to_remove)): | |
| _ = token_freqs.pop(sorted_scores[i][<span class="hljs-number">0</span>]) | |
| total_sum = <span class="hljs-built_in">sum</span>([freq <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs.items()]) | |
| model = {token: -log(freq / total_sum) <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs.items()}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-jzcl41">Ensuite, pour tokeniser un texte, il suffit d’appliquer la prétokénisation et d’utiliser la fonction <code>encode_word()</code> :</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">tokenize</span>(<span class="hljs-params">text, model</span>): | |
| words_with_offsets = tokenizer.backend_tokenizer.pre_tokenizer.pre_tokenize_str(text) | |
| pre_tokenized_text = [word <span class="hljs-keyword">for</span> word, offset <span class="hljs-keyword">in</span> words_with_offsets] | |
| encoded_words = [encode_word(word, model)[<span class="hljs-number">0</span>] <span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> pre_tokenized_text] | |
| <span class="hljs-keyword">return</span> <span class="hljs-built_in">sum</span>(encoded_words, []) | |
| tokenize(<span class="hljs-string">"This is the Hugging Face course."</span>, model)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[<span class="hljs-string">'▁This'</span>, <span class="hljs-string">'▁is'</span>, <span class="hljs-string">'▁the'</span>, <span class="hljs-string">'▁Hugging'</span>, <span class="hljs-string">'▁Face'</span>, <span class="hljs-string">'▁'</span>, <span class="hljs-string">'c'</span>, <span class="hljs-string">'ou'</span>, <span class="hljs-string">'r'</span>, <span class="hljs-string">'s'</span>, <span class="hljs-string">'e'</span>, <span class="hljs-string">'.'</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1y5vnmz">C’est tout pour <em>Unigram</em> ! Avec un peu de chance, vous vous sentez à présent être un expert des <em>tokenizers</em>. Dans la prochaine section, nous allons nous plonger dans les blocs de construction de la bibliothèque 🤗 <em>Tokenizers</em> et allons vous montrer comment vous pouvez les utiliser pour construire votre propre <em>tokenizer</em>.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/course/blob/main/chapters/fr/chapter6/7.mdx" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span data-svelte-h="svelte-x0xyl0">></span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p> | |
| <script> | |
| { | |
| __sveltekit_4ugnx6 = { | |
| assets: "/docs/course/pr_1007/fr", | |
| base: "/docs/course/pr_1007/fr", | |
| env: {} | |
| }; | |
| const element = document.currentScript.parentElement; | |
| const data = [null,null]; | |
| Promise.all([ | |
| import("/docs/course/pr_1007/fr/_app/immutable/entry/start.0dcef240.js"), | |
| import("/docs/course/pr_1007/fr/_app/immutable/entry/app.e06c0eda.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 50], | |
| data, | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 100 kB
- Xet hash:
- 1802571618a6f7578e7ab417d12b48ec87988b6ff3a40376be55cbf506575384
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.