Hugging Face's logo

Buckets:
hf-doc-build
/
doc-dev

Files
xet
hf-doc-build/doc-dev / course /pr_1069 /fr /chapter3 /4.html
rtrm's picture
rtrm
download
raw
62.7 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Un entraînement complet&quot;,&quot;local&quot;:&quot;un-entraînement-complet&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Préparer l’entraînement&quot;,&quot;local&quot;:&quot;préparer-lentraînement&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;La boucle d’entraînement&quot;,&quot;local&quot;:&quot;la-boucle-dentraînement&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;La boucle d’évaluation&quot;,&quot;local&quot;:&quot;la-boucle-dévaluation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Optimisez votre boucle d’entraînement avec 🤗 <i> Accelerate </i>&quot;,&quot;local&quot;:&quot;optimisez-votre-boucle-dentraînement-avec--i-accelerate-i&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:1}">
 <link href="/docs/course/pr_1069/fr/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/entry/start.cea6db46.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/scheduler.37c15a92.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/singletons.2b29b91f.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/index.18351ede.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/paths.f6fdf97f.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/entry/app.3f6640b1.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/index.2bf4358c.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/nodes/0.b777de11.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/each.e59479a4.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/nodes/25.fcc9ae6b.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/Tip.363c041f.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/Youtube.1e50a667.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/CodeBlock.4e987730.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/CourseFloatingBanner.6add7356.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/fr/_app/immutable/chunks/getInferenceSnippets.24b50994.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Un entraînement complet&quot;,&quot;local&quot;:&quot;un-entraînement-complet&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Préparer l’entraînement&quot;,&quot;local&quot;:&quot;préparer-lentraînement&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;La boucle d’entraînement&quot;,&quot;local&quot;:&quot;la-boucle-dentraînement&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;La boucle d’évaluation&quot;,&quot;local&quot;:&quot;la-boucle-dévaluation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Optimisez votre boucle d’entraînement avec 🤗 <i> Accelerate </i>&quot;,&quot;local&quot;:&quot;optimisez-votre-boucle-dentraînement-avec--i-accelerate-i&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="un-entraînement-complet" 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="#un-entraînement-complet"><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>Un entraînement complet</span></h1> <div class="flex space-x-1 absolute z-10 right-0 top-0"><a href="https://discuss.huggingface.co/t/chapter-3-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> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/Dh9CL8fyG80" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-1g0xmca">Maintenant nous allons voir comment obtenir les mêmes résultats que dans la dernière section sans utiliser la classe <code>Trainer</code>. Encore une fois, nous supposons que vous avez fait le traitement des données dans la section 2. Voici un court résumé couvrant tout ce dont vous aurez besoin :</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> datasets <span class="hljs-keyword">import</span> load_dataset
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, DataCollatorWithPadding
raw_datasets = load_dataset(<span class="hljs-string">&quot;glue&quot;</span>, <span class="hljs-string">&quot;mrpc&quot;</span>)
checkpoint = <span class="hljs-string">&quot;bert-base-uncased&quot;</span>
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
<span class="hljs-keyword">def</span> <span class="hljs-title function_">tokenize_function</span>(<span class="hljs-params">example</span>):
 <span class="hljs-keyword">return</span> tokenizer(example[<span class="hljs-string">&quot;sentence1&quot;</span>], example[<span class="hljs-string">&quot;sentence2&quot;</span>], truncation=<span class="hljs-literal">True</span>)
tokenized_datasets = raw_datasets.<span class="hljs-built_in">map</span>(tokenize_function, batched=<span class="hljs-literal">True</span>)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="préparer-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="#préparer-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>Préparer l’entraînement</span></h3> <p data-svelte-h="svelte-17hvgx4">Avant d’écrire réellement notre boucle d’entraînement, nous devons définir quelques objets. Les premiers sont les <em>dataloaders</em> que nous utiliserons pour itérer sur les batchs. Mais avant de pouvoir définir ces chargeurs de données, nous devons appliquer un peu de post-traitement à nos <code>tokenized_datasets</code>, pour prendre soin de certaines choses que le <code>Trainer</code> fait pour nous automatiquement. Spécifiquement, nous devons :</p> <ul data-svelte-h="svelte-s8mdl4"><li>supprimer les colonnes correspondant aux valeurs que le modèle n’attend pas (comme les colonnes <code>sentence1</code> et <code>sentence2</code>),</li> <li>renommer la colonne <code>label</code> en <code>labels</code> (parce que le modèle s’attend à ce que l’argument soit nommé <code>labels</code>),</li> <li>définir le format des jeux de données pour qu’ils retournent des tenseurs PyTorch au lieu de listes.</li></ul> <p data-svelte-h="svelte-1vhf3m8">Notre <code>tokenized_datasets</code> a une méthode pour chacune de ces étapes :</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 -->tokenized_datasets = tokenized_datasets.remove_columns([<span class="hljs-string">&quot;sentence1&quot;</span>, <span class="hljs-string">&quot;sentence2&quot;</span>, <span class="hljs-string">&quot;idx&quot;</span>])
tokenized_datasets = tokenized_datasets.rename_column(<span class="hljs-string">&quot;label&quot;</span>, <span class="hljs-string">&quot;labels&quot;</span>)
tokenized_datasets.set_format(<span class="hljs-string">&quot;torch&quot;</span>)
tokenized_datasets[<span class="hljs-string">&quot;train&quot;</span>].column_names<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1jligxe">Nous pouvons alors vérifier que le résultat ne comporte que des colonnes que notre modèle acceptera :</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">&quot;attention_mask&quot;</span>, <span class="hljs-string">&quot;input_ids&quot;</span>, <span class="hljs-string">&quot;labels&quot;</span>, <span class="hljs-string">&quot;token_type_ids&quot;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-15ft7dc">Maintenant que cela est fait, nous pouvons facilement définir nos <em>dataloaders</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">from</span> torch.utils.data <span class="hljs-keyword">import</span> DataLoader
train_dataloader = DataLoader(
 tokenized_datasets[<span class="hljs-string">&quot;train&quot;</span>], shuffle=<span class="hljs-literal">True</span>, batch_size=<span class="hljs-number">8</span>, collate_fn=data_collator
)
eval_dataloader = DataLoader(
 tokenized_datasets[<span class="hljs-string">&quot;validation&quot;</span>], batch_size=<span class="hljs-number">8</span>, collate_fn=data_collator
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1rk3y8n">Pour vérifier rapidement qu’il n’y a pas d’erreur dans le traitement des données, nous pouvons inspecter un batch comme celui-ci :</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">for</span> batch <span class="hljs-keyword">in</span> train_dataloader:
 <span class="hljs-keyword">break</span>
{k: v.shape <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> batch.items()}<!-- 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">&#x27;attention_mask&#x27;</span>: torch.Size([<span class="hljs-number">8</span>, <span class="hljs-number">65</span>]),
 <span class="hljs-string">&#x27;input_ids&#x27;</span>: torch.Size([<span class="hljs-number">8</span>, <span class="hljs-number">65</span>]),
 <span class="hljs-string">&#x27;labels&#x27;</span>: torch.Size([<span class="hljs-number">8</span>]),
 <span class="hljs-string">&#x27;token_type_ids&#x27;</span>: torch.Size([<span class="hljs-number">8</span>, <span class="hljs-number">65</span>])}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-hfva83">Notez que les formes réelles seront probablement légèrement différentes pour vous puisque nous avons défini <code>shuffle=True</code> pour le chargeur de données d’entraînement et que nous <em>paddons</em> à la longueur maximale dans le batch.</p> <p data-svelte-h="svelte-hgf9jd">Maintenant que nous en avons terminé avec le prétraitement des données (un objectif satisfaisant mais difficile à atteindre pour tout praticien d’apprentissage automatique), passons au modèle. Nous l’instancions exactement comme nous l’avons fait dans la section précédente :</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> AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=<span class="hljs-number">2</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-mg010b">Pour s’assurer que tout se passera bien pendant l’entraînement, nous transmettons notre batch à ce 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 -->outputs = model(**batch)
<span class="hljs-built_in">print</span>(outputs.loss, outputs.logits.shape)<!-- 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 -->tensor(<span class="hljs-number">0.5441</span>, grad_fn=&lt;NllLossBackward&gt;) torch.Size([<span class="hljs-number">8</span>, <span class="hljs-number">2</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-iwaa2x">Tous les modèles 🤗 <em>Transformers</em> renvoient la perte lorsque les <code>labels</code> sont fournis. Nous obtenons également les logits (deux pour chaque entrée de notre batch, donc un tenseur de taille 8 x 2).</p> <p data-svelte-h="svelte-ahwb4">Nous sommes presque prêts à écrire notre boucle d’entraînement ! Il nous manque juste deux choses : un optimiseur et un planificateur de taux d’apprentissage. Puisque nous essayons de reproduire à la main ce que fait la fonction <code>Trainer</code>, utilisons les mêmes paramètres par défaut. L’optimiseur utilisé par <code>Trainer</code> est <code>AdamW</code>, qui est le même qu’Adam, mais avec une torsion pour la régularisation par décroissance de poids (voir <a href="https://arxiv.org/abs/1711.05101" rel="nofollow"><em>Decoupled Weight Decay Regularization</em></a> par Ilya Loshchilov et Frank Hutter) :</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> torch.optim <span class="hljs-keyword">import</span> AdamW
optimizer = AdamW(model.parameters(), lr=<span class="hljs-number">5e-5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-wgcpmp">Enfin, le planificateur du taux d’apprentissage utilisé par défaut est juste une décroissance linéaire de la valeur maximale (5e-5) à 0. Pour le définir correctement, nous devons connaître le nombre d’étapes d’entraînement que nous prendrons, qui est le nombre d’époques que nous voulons exécuter multiplié par le nombre de batch d’entraînement (qui est la longueur de notre <em>dataloader</em> d’entraînement). Le <code>Trainer</code> utilise trois époques par défaut, nous allons donc suivre ça :</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> get_scheduler
num_epochs = <span class="hljs-number">3</span>
num_training_steps = num_epochs * <span class="hljs-built_in">len</span>(train_dataloader)
lr_scheduler = get_scheduler(
 <span class="hljs-string">&quot;linear&quot;</span>,
 optimizer=optimizer,
 num_warmup_steps=<span class="hljs-number">0</span>,
 num_training_steps=num_training_steps,
)
<span class="hljs-built_in">print</span>(num_training_steps)<!-- 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">1377</span><!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="la-boucle-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="#la-boucle-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>La boucle d’entraînement</span></h3> <p data-svelte-h="svelte-rbvecl">Une dernière chose : nous voulons utiliser le GPU si nous en avons un (sur un CPU, l’entraînement peut prendre plusieurs heures au lieu de quelques minutes). Pour ce faire, nous définissons un <code>device</code> sur lequel nous allons placer notre modèle et nos batchs :</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> torch
device = torch.device(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-keyword">if</span> torch.cuda.is_available() <span class="hljs-keyword">else</span> torch.device(<span class="hljs-string">&quot;cpu&quot;</span>)
model.to(device)
device<!-- 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 -->device(<span class="hljs-built_in">type</span>=<span class="hljs-string">&#x27;cuda&#x27;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-unj8ze">Nous sommes maintenant prêts à entraîner ! Pour avoir une idée du moment où l’entraînement sera terminé, nous ajoutons une barre de progression sur le nombre d’étapes d’entraînement, en utilisant la bibliothèque <code>tqdm</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">from</span> tqdm.auto <span class="hljs-keyword">import</span> tqdm
progress_bar = tqdm(<span class="hljs-built_in">range</span>(num_training_steps))
model.train()
<span class="hljs-keyword">for</span> epoch <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(num_epochs):
 <span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> train_dataloader:
 batch = {k: v.to(device) <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> batch.items()}
 outputs = model(**batch)
 loss = outputs.loss
 loss.backward()
optimizer.step()
 lr_scheduler.step()
 optimizer.zero_grad()
 progress_bar.update(<span class="hljs-number">1</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-w5chrh">Vous pouvez voir que le cœur de la boucle d’entraînement ressemble beaucoup à celui de l’introduction. Nous n’avons pas demandé de rapport, donc cette boucle d’entraînement ne nous dira rien sur les résultats du modèle. Pour cela, nous devons ajouter une boucle d’évaluation.</p> <h3 class="relative group"><a id="la-boucle-dévaluation" 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="#la-boucle-dévaluation"><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>La boucle d’évaluation</span></h3> <p data-svelte-h="svelte-as7rjs">Comme nous l’avons fait précédemment, nous allons utiliser une métrique fournie par la bibliothèque 🤗 <em>Evaluate</em>. Nous avons déjà vu la méthode <code>metric.compute()</code>, mais les métriques peuvent en fait accumuler des batchs pour nous au fur et à mesure que nous parcourons la boucle de prédiction avec la méthode <code>add_batch()</code>. Une fois que nous avons accumulé tous les batchs, nous pouvons obtenir le résultat final avec <code>metric.compute()</code>. Voici comment implémenter tout cela dans une boucle d’évaluation :</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> evaluate
metric = evaluate.load(<span class="hljs-string">&quot;glue&quot;</span>, <span class="hljs-string">&quot;mrpc&quot;</span>)
model.<span class="hljs-built_in">eval</span>()
<span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> eval_dataloader:
 batch = {k: v.to(device) <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> batch.items()}
 <span class="hljs-keyword">with</span> torch.no_grad():
 outputs = model(**batch)
logits = outputs.logits
 predictions = torch.argmax(logits, dim=-<span class="hljs-number">1</span>)
 metric.add_batch(predictions=predictions, references=batch[<span class="hljs-string">&quot;labels&quot;</span>])
metric.compute()<!-- 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">&#x27;accuracy&#x27;</span>: <span class="hljs-number">0.8431372549019608</span>, <span class="hljs-string">&#x27;f1&#x27;</span>: <span class="hljs-number">0.8907849829351535</span>}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-h5j4wt">Une fois encore, vos résultats seront légèrement différents en raison du caractère aléatoire de l’initialisation de la tête du modèle et du mélange des données, mais ils devraient se situer dans la même fourchette.</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-1dfxuon">✏️ <strong>Essayez</strong> Modifiez la boucle d’entraînement précédente pour <em>finetuner</em> votre modèle sur le jeu de données SST-2.</p></div> <h3 class="relative group"><a id="optimisez-votre-boucle-dentraînement-avec--i-accelerate-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="#optimisez-votre-boucle-dentraînement-avec--i-accelerate-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>Optimisez votre boucle d’entraînement avec 🤗 &lt;i> Accelerate &lt;/i></span></h3> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/s7dy8QRgjJ0" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-159rzx0">La boucle d’entraînement que nous avons définie précédemment fonctionne bien sur un seul CPU ou GPU. Mais en utilisant la bibliothèque <a href="https://github.com/huggingface/accelerate" rel="nofollow">🤗 <em>Accelerate</em></a>, il suffit de quelques ajustements pour permettre un entraînement distribué sur plusieurs GPUs ou TPUs. En partant de la création des <em>dataloaders</em> d’entraînement et de validation, voici à quoi ressemble notre boucle d’entraînement manuel :</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> torch.optim <span class="hljs-keyword">import</span> AdamW
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForSequenceClassification, get_scheduler
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=<span class="hljs-number">2</span>)
optimizer = AdamW(model.parameters(), lr=<span class="hljs-number">3e-5</span>)
device = torch.device(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-keyword">if</span> torch.cuda.is_available() <span class="hljs-keyword">else</span> torch.device(<span class="hljs-string">&quot;cpu&quot;</span>)
model.to(device)
num_epochs = <span class="hljs-number">3</span>
num_training_steps = num_epochs * <span class="hljs-built_in">len</span>(train_dataloader)
lr_scheduler = get_scheduler(
 <span class="hljs-string">&quot;linear&quot;</span>,
 optimizer=optimizer,
 num_warmup_steps=<span class="hljs-number">0</span>,
 num_training_steps=num_training_steps,
)
progress_bar = tqdm(<span class="hljs-built_in">range</span>(num_training_steps))
model.train()
<span class="hljs-keyword">for</span> epoch <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(num_epochs):
 <span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> train_dataloader:
 batch = {k: v.to(device) <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> batch.items()}
 outputs = model(**batch)
 loss = outputs.loss
 loss.backward()
optimizer.step()
 lr_scheduler.step()
 optimizer.zero_grad()
 progress_bar.update(<span class="hljs-number">1</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-3jm87s">Et voici les changements :</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-addition">+ from accelerate import Accelerator</span>
 from torch.optim import AdamW
 from transformers import AutoModelForSequenceClassification, get_scheduler
<span class="hljs-addition">+ accelerator = Accelerator()</span>
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)
 optimizer = AdamW(model.parameters(), lr=3e-5)
<span class="hljs-deletion">- device = torch.device(&quot;cuda&quot;) if torch.cuda.is_available() else torch.device(&quot;cpu&quot;)</span>
<span class="hljs-deletion">- model.to(device)</span>
<span class="hljs-addition">+ train_dataloader, eval_dataloader, model, optimizer = accelerator.prepare(</span>
<span class="hljs-addition">+ train_dataloader, eval_dataloader, model, optimizer</span>
<span class="hljs-addition">+ )</span>
num_epochs = 3
 num_training_steps = num_epochs * len(train_dataloader)
 lr_scheduler = get_scheduler(
 &quot;linear&quot;,
 optimizer=optimizer,
 num_warmup_steps=0,
 num_training_steps=num_training_steps
 )
progress_bar = tqdm(range(num_training_steps))
model.train()
 for epoch in range(num_epochs):
 for batch in train_dataloader:
<span class="hljs-deletion">- batch = {k: v.to(device) for k, v in batch.items()}</span>
 outputs = model(**batch)
 loss = outputs.loss
<span class="hljs-deletion">- loss.backward()</span>
<span class="hljs-addition">+ accelerator.backward(loss)</span>
optimizer.step()
 lr_scheduler.step()
 optimizer.zero_grad()
 progress_bar.update(1)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1uypok9">La première ligne à ajouter est la ligne d’importation. La deuxième ligne instancie un objet <code>Accelerator</code> qui va regarder l’environnement et initialiser la bonne configuration distribuée. 🤗 <em>Accelerate</em> gère le placement des périphériques pour vous, donc vous pouvez enlever les lignes qui placent le modèle sur le périphérique (ou, si vous préférez, les changer pour utiliser <code>accelerator.device</code> au lieu de <code>device</code>).</p> <p data-svelte-h="svelte-13ypbak">Ensuite, le gros du travail est fait dans la ligne qui envoie les <em>dataloaders</em>, le modèle, et l’optimiseur à <code>accelerator.prepare()</code>. Cela va envelopper ces objets dans le conteneur approprié pour s’assurer que votre entraînement distribué fonctionne comme prévu. Les changements restants à faire sont la suppression de la ligne qui met le batch sur le <code>device</code> (encore une fois, si vous voulez le garder, vous pouvez juste le changer pour utiliser <code>accelerator.device</code>) et le remplacement de <code>loss.backward()</code> par <code>accelerator.backward(loss)</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">⚠️ Afin de bénéficier de la rapidité offerte par les TPUs du Cloud, nous vous recommandons de rembourrer vos échantillons à une longueur fixe avec les arguments `padding=&quot;max_length&quot;` et `max_length` du <i data-svelte-h="svelte-ymquwj">tokenizer</i>.</div> <p data-svelte-h="svelte-1t4du1b">Si vous souhaitez faire un copier-coller pour jouer, voici à quoi ressemble la boucle d’entraînement complète avec 🤗 <i>Accelerate</i> :</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> accelerate <span class="hljs-keyword">import</span> Accelerator
<span class="hljs-keyword">from</span> torch.optim <span class="hljs-keyword">import</span> AdamW
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForSequenceClassification, get_scheduler
accelerator = Accelerator()
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=<span class="hljs-number">2</span>)
optimizer = AdamW(model.parameters(), lr=<span class="hljs-number">3e-5</span>)
train_dl, eval_dl, model, optimizer = accelerator.prepare(
 train_dataloader, eval_dataloader, model, optimizer
)
num_epochs = <span class="hljs-number">3</span>
num_training_steps = num_epochs * <span class="hljs-built_in">len</span>(train_dl)
lr_scheduler = get_scheduler(
 <span class="hljs-string">&quot;linear&quot;</span>,
 optimizer=optimizer,
 num_warmup_steps=<span class="hljs-number">0</span>,
 num_training_steps=num_training_steps,
)
progress_bar = tqdm(<span class="hljs-built_in">range</span>(num_training_steps))
model.train()
<span class="hljs-keyword">for</span> epoch <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(num_epochs):
 <span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> train_dl:
 outputs = model(**batch)
 loss = outputs.loss
 accelerator.backward(loss)
optimizer.step()
 lr_scheduler.step()
 optimizer.zero_grad()
 progress_bar.update(<span class="hljs-number">1</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-dkpmec">En plaçant ceci dans un script <code>train.py</code>, cela sera exécutable sur n’importe quel type d’installation distribuée. Pour l’essayer dans votre installation distribuée, exécutez la commande :</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 -->accelerate config<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1qilg93">qui vous demandera de répondre à quelques questions et enregistrera vos réponses dans un fichier de configuration utilisé par cette commande :</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 -->accelerate <span class="hljs-built_in">launch</span> train.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-16tpdes">qui lancera l’entraînement distribué.</p> <p data-svelte-h="svelte-1whbz23">Si vous voulez essayer ceci dans un <em>notebook</em> (par exemple, pour le tester avec des TPUs sur Colab), collez simplement le code dans une <code>training_function()</code> et lancez une dernière cellule avec :</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> accelerate <span class="hljs-keyword">import</span> notebook_launcher
notebook_launcher(training_function)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1qnllaj">Vous trouverez d’autres exemples dans le dépôt d’<a href="https://github.com/huggingface/accelerate/tree/main/examples" rel="nofollow">🤗 <em>Accelerate</em></a>.</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/chapter3/4.mdx" target="_blank"><span data-svelte-h="svelte-1kd6by1">&lt;</span> <span data-svelte-h="svelte-x0xyl0">&gt;</span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>
<script>
 {
 __sveltekit_1sfisyd = {
 assets: "/docs/course/pr_1069/fr",
 base: "/docs/course/pr_1069/fr",
 env: {}
 };
 const element = document.currentScript.parentElement;
 const data = [null,null];
 Promise.all([
 import("/docs/course/pr_1069/fr/_app/immutable/entry/start.cea6db46.js"),
 import("/docs/course/pr_1069/fr/_app/immutable/entry/app.3f6640b1.js")
 ]).then(([kit, app]) => {
 kit.start(app, element, {
 node_ids: [0, 25],
 data,
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
62.7 kB
·
Xet hash:
aa40f03c1ad91301a6f59611329e245b356062a5b6b1accc06d1dd7f608155ba

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.