Buckets:

rtrm's picture
download
raw
149 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Создание токенизатора, блок за блоком&quot;,&quot;local&quot;:&quot;building-a-tokenizer-block-by-block&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Получение корпуса текста&quot;,&quot;local&quot;:&quot;acquiring-a-corpus&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Создание токенизатора WordPiece с нуля&quot;,&quot;local&quot;:&quot;building-a-wordpiece-tokenizer-from-scratch&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Создание токенизатора BPE с нуля&quot;,&quot;local&quot;:&quot;building-a-bpe-tokenizer-from-scratch&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Создание токенизатора Unigram с нуля&quot;,&quot;local&quot;:&quot;building-a-unigram-tokenizer-from-scratch&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
<link href="/docs/course/pr_1069/ru/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/entry/start.48687cc8.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/scheduler.37c15a92.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/singletons.6f259016.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/index.18351ede.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/paths.930ed261.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/entry/app.b79a803d.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/index.2bf4358c.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/nodes/0.e11366e4.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/each.e59479a4.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/nodes/50.0007e926.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/Tip.363c041f.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/Youtube.1e50a667.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/CodeBlock.4e987730.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/CourseFloatingBanner.9ff4c771.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/DocNotebookDropdown.efc1fb7c.js">
<link rel="modulepreload" href="/docs/course/pr_1069/ru/_app/immutable/chunks/getInferenceSnippets.24b50994.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Создание токенизатора, блок за блоком&quot;,&quot;local&quot;:&quot;building-a-tokenizer-block-by-block&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Получение корпуса текста&quot;,&quot;local&quot;:&quot;acquiring-a-corpus&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Создание токенизатора WordPiece с нуля&quot;,&quot;local&quot;:&quot;building-a-wordpiece-tokenizer-from-scratch&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Создание токенизатора BPE с нуля&quot;,&quot;local&quot;:&quot;building-a-bpe-tokenizer-from-scratch&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Создание токенизатора Unigram с нуля&quot;,&quot;local&quot;:&quot;building-a-unigram-tokenizer-from-scratch&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="building-a-tokenizer-block-by-block" 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="#building-a-tokenizer-block-by-block"><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>Создание токенизатора, блок за блоком</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> <a href="https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/en/chapter6/section8.ipynb" target="_blank"><img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"></a> <a href="https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/master/course/en/chapter6/section8.ipynb" target="_blank"><img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"></a></div> <p data-svelte-h="svelte-b5kklz">Как мы уже видели в предыдущих разделах, токенизация состоит из нескольких этапов:</p> <ul data-svelte-h="svelte-1sajjkb"><li>Нормализация (любая необходимая очистка текста, например, удаление пробелов или подчеркиваний, нормализация Unicode и т. д.)</li> <li>Предварительная токенизация (разделение входного текста на слова).</li> <li>Прогон входных данных через модель (использование предварительно токенизированных слов для создания последовательности токенов)</li> <li>Постобработка (добавление специальных токенов токенизатора, генерация маски внимания и идентификаторов типов токенов)</li></ul> <p data-svelte-h="svelte-31tofb">В качестве напоминания вот еще один взгляд на общий процесс:</p> <div class="flex justify-center" data-svelte-h="svelte-oxfng3"><img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface-course/documentation-images/resolve/main/en/chapter6/tokenization_pipeline.svg" alt="The tokenization pipeline."> <img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface-course/documentation-images/resolve/main/en/chapter6/tokenization_pipeline-dark.svg" alt="The tokenization pipeline."></div> <p data-svelte-h="svelte-3ncb60">Библиотека 🤗 Tokenizers была создана для того, чтобы предоставить несколько вариантов каждого из этих шагов, которые вы можете смешивать и сочетать между собой. В этом разделе мы рассмотрим, как можно создать токенизатор с нуля, а не обучать новый токенизатор на основе старого, как мы делали в <a href="../chapter6/2">разделе 2</a>. После этого вы сможете создать любой токенизатор, который только сможете придумать!</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/MR8tZm5ViWU" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-1jjcw0i">Точнее, библиотека построена вокруг центрального класса <code>Tokenizer</code>, а строительные блоки сгруппированы в подмодули:</p> <ul data-svelte-h="svelte-400a5j"><li><code>normalizers</code> содержит все возможные типы нормализаторов текста <code>Normalizer</code>, которые вы можете использовать (полный список <a href="https://huggingface.co/docs/tokenizers/api/normalizers" rel="nofollow">здесь</a>).</li> <li><code>pre_tokenizers</code> содержит все возможные типы предварительных токенизаторов <code>PreTokenizer</code>, которые вы можете использовать (полный список <a href="https://huggingface.co/docs/tokenizers/api/pre-tokenizers" rel="nofollow">здесь</a>).</li> <li><code>models</code> содержит различные типы моделей <code>Model</code>, которые вы можете использовать, такие как <code>BPE</code>, <code>WordPiece</code> и <code>Unigram</code> (полный список <a href="https://huggingface.co/docs/tokenizers/api/models" rel="nofollow">здесь</a>).</li> <li><code>trainers</code> содержит все различные типы <code>Trainer</code>, которые вы можете использовать для обучения модели на корпусе (по одному на каждый тип модели; полный список <a href="https://huggingface.co/docs/tokenizers/api/trainers" rel="nofollow">здесь</a>).</li> <li><code>post_processors</code> содержит различные типы постпроцессоров <code>PostProcessor</code>, которые вы можете использовать (полный список <a href="https://huggingface.co/docs/tokenizers/api/post-processors" rel="nofollow">здесь</a>).</li> <li><code>decoders</code> содержит различные типы декодеров <code>Decoder</code>, которые вы можете использовать для декодирования результатов токенизации (полный список <a href="https://huggingface.co/docs/tokenizers/components#decoders" rel="nofollow">здесь</a>).</li></ul> <p data-svelte-h="svelte-1dqwrlj">Весь список блоков вы можете найти <a href="https://huggingface.co/docs/tokenizers/components" rel="nofollow">здесь</a>.</p> <h2 class="relative group"><a id="acquiring-a-corpus" 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="#acquiring-a-corpus"><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>Получение корпуса текста</span></h2> <p data-svelte-h="svelte-7jvs84">Для обучения нашего нового токенизатора мы будем использовать небольшой корпус текстов (чтобы примеры выполнялись быстро). Шаги по сбору корпуса аналогичны тем, что мы делали в <a href="../chapter6/2">начале этой главы</a>, но на этот раз мы будем использовать набор данных <a href="https://huggingface.co/datasets/wikitext" rel="nofollow">WikiText-2</a>:</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
dataset = load_dataset(<span class="hljs-string">&quot;wikitext&quot;</span>, name=<span class="hljs-string">&quot;wikitext-2-raw-v1&quot;</span>, split=<span class="hljs-string">&quot;train&quot;</span>)
<span class="hljs-keyword">def</span> <span class="hljs-title function_">get_training_corpus</span>():
<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">0</span>, <span class="hljs-built_in">len</span>(dataset), <span class="hljs-number">1000</span>):
<span class="hljs-keyword">yield</span> dataset[i : i + <span class="hljs-number">1000</span>][<span class="hljs-string">&quot;text&quot;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1gsy8ta">Функция <code>get_training_corpus()</code> - это генератор, который выдает батч из 1000 текстов, которые мы будем использовать для обучения токенизатора.</p> <p data-svelte-h="svelte-1wtem0y">🤗 Токенизаторы также можно обучать непосредственно на текстовых файлах. Вот как мы можем сгенерировать текстовый файл, содержащий все тексты/входы из WikiText-2, который мы можем использовать локально:</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">with</span> <span class="hljs-built_in">open</span>(<span class="hljs-string">&quot;wikitext-2.txt&quot;</span>, <span class="hljs-string">&quot;w&quot;</span>, encoding=<span class="hljs-string">&quot;utf-8&quot;</span>) <span class="hljs-keyword">as</span> f:
<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>(dataset)):
f.write(dataset[i][<span class="hljs-string">&quot;text&quot;</span>] + <span class="hljs-string">&quot;\n&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-gnegz7">Далее мы покажем вам, как блок за блоком построить собственные токенизаторы BERT, GPT-2 и XLNet. Это даст нам пример каждого из трех основных алгоритмов токенизации: WordPiece, BPE и Unigram. Начнем с BERT!</p> <h2 class="relative group"><a id="building-a-wordpiece-tokenizer-from-scratch" 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="#building-a-wordpiece-tokenizer-from-scratch"><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>Создание токенизатора WordPiece с нуля</span></h2> <p data-svelte-h="svelte-rdvi8a">Чтобы создать токенизатор с помощью библиотеки 🤗 Tokenizers, мы начнем с инстанцирования объектов <code>Tokenizer</code> и <code>model</code>, затем установим для их атрибутов <code>normalizer</code>, <code>pre_tokenizer</code>, <code>post_processor</code> и <code>decoder</code> нужные нам значения.</p> <p data-svelte-h="svelte-17zu5lc">Для этого примера мы создадим <code>Tokenizer</code> с моделью WordPiece:</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> tokenizers <span class="hljs-keyword">import</span> (
decoders,
models,
normalizers,
pre_tokenizers,
processors,
trainers,
Tokenizer,
)
tokenizer = Tokenizer(models.WordPiece(unk_token=<span class="hljs-string">&quot;[UNK]&quot;</span>))<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1uc1t5u">Мы должны указать <code>unk_token</code>, чтобы модель знала, что возвращать, когда она встречает символы, которых раньше не видела. Другие аргументы, которые мы можем задать здесь, включают <code>vocab</code> нашей модели (мы собираемся обучать модель, поэтому нам не нужно его задавать) и <code>max_input_chars_per_word</code>, который определяет максимальную длину для каждого слова (слова длиннее переданного значения будут разбиты на части).</p> <p data-svelte-h="svelte-14au9ct">Первым шагом токенизации является нормализация, поэтому начнем с нее. Поскольку BERT широко используется, существует <code>BertNormalizer</code> с классическими параметрами, которые мы можем установить для BERT: <code>lowercase</code> и <code>strip_accents</code>, которые не требуют пояснений; <code>clean_text</code> для удаления всех управляющих символов и замены повторяющихся пробелов на один; и <code>handle_chinese_chars</code>, который расставляет пробелы вокруг китайских символов. Чтобы повторить токенизатор <code>bert-base-uncased</code>, мы можем просто установить этот нормализатор:</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 -->tokenizer.normalizer = normalizers.BertNormalizer(lowercase=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-13vk9t1">Однако, как правило, при создании нового токенизатора у вас не будет доступа к такому удобному нормализатору, уже реализованному в библиотеке 🤗 Tokenizers, поэтому давайте посмотрим, как создать нормализатор BERT вручную. Библиотека предоставляет нормализатор <code>Lowercase</code> и нормализатор <code>StripAccents</code>, и вы можете комбинировать несколько нормализаторов с помощью <code>Sequence</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 -->tokenizer.normalizer = normalizers.<span class="hljs-type">Sequence</span>(
[normalizers.NFD(), normalizers.Lowercase(), normalizers.StripAccents()]
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-eabqpa">Мы также используем нормализатор Unicode <code>NFD</code>, поскольку в противном случае нормализатор <code>StripAccents</code> не сможет правильно распознать акцентированные символы и, следовательно, не удалит их.</p> <p data-svelte-h="svelte-1cmtqtp">Как мы уже видели ранее, мы можем использовать метод <code>normalize_str()</code> нормализатора, чтобы проверить, как он влияет на данный текст:</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>(tokenizer.normalizer.normalize_str(<span class="hljs-string">&quot;Héllò hôw are ü?&quot;</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 -->hello how are u?<!-- 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-kju4cm"><strong>Далее</strong> если вы протестируете две версии предыдущих нормализаторов на строке, содержащей символ Unicode <code>u&quot;\u0085&quot;</code>, то наверняка заметите, что эти два нормализатора не совсем эквивалентны.<br>
Чтобы не усложнять версию с <code>normalizers.Sequence</code>, мы не включили в нее Regex-замены, которые требует <code>BertNormalizer</code>, когда аргумент <code>clean_text</code> установлен в <code>True</code>, что является поведением по умолчанию. Но не волнуйтесь: можно получить точно такую же нормализацию без использования удобного <code>BertNormalizer</code>, добавив два <code>normalizers.Replace</code> в последовательность нормализаторов.</p></div> <p data-svelte-h="svelte-1x52eyd">Далее следует этап предварительной токенизации. Опять же, есть готовый <code>BertPreTokenizer</code>, который мы можем использовать:</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 -->tokenizer.pre_tokenizer = pre_tokenizers.BertPreTokenizer()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-gv0viq">Или мы можем создать его с нуля:</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 -->tokenizer.pre_tokenizer = pre_tokenizers.Whitespace()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-iy2ywz">Обратите внимание, что токенизатор <code>Whitespace</code> разделяет пробельные символы и все символы, которые не являются буквами, цифрами или символом подчеркивания, поэтому технически он разделяет пробельные символы и знаки пунктуации:</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 -->tokenizer.pre_tokenizer.pre_tokenize_str(<span class="hljs-string">&quot;Let&#x27;s test my pre-tokenizer.&quot;</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">&#x27;Let&#x27;</span>, (<span class="hljs-number">0</span>, <span class="hljs-number">3</span>)), (<span class="hljs-string">&quot;&#x27;&quot;</span>, (<span class="hljs-number">3</span>, <span class="hljs-number">4</span>)), (<span class="hljs-string">&#x27;s&#x27;</span>, (<span class="hljs-number">4</span>, <span class="hljs-number">5</span>)), (<span class="hljs-string">&#x27;test&#x27;</span>, (<span class="hljs-number">6</span>, <span class="hljs-number">10</span>)), (<span class="hljs-string">&#x27;my&#x27;</span>, (<span class="hljs-number">11</span>, <span class="hljs-number">13</span>)), (<span class="hljs-string">&#x27;pre&#x27;</span>, (<span class="hljs-number">14</span>, <span class="hljs-number">17</span>)),
(<span class="hljs-string">&#x27;-&#x27;</span>, (<span class="hljs-number">17</span>, <span class="hljs-number">18</span>)), (<span class="hljs-string">&#x27;tokenizer&#x27;</span>, (<span class="hljs-number">18</span>, <span class="hljs-number">27</span>)), (<span class="hljs-string">&#x27;.&#x27;</span>, (<span class="hljs-number">27</span>, <span class="hljs-number">28</span>))]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1hk0h2e">Если вы хотите выполнять разделение только по пробельным символам, то вместо этого следует использовать предварительный токенизатор <code>WhitespaceSplit</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 -->pre_tokenizer = pre_tokenizers.WhitespaceSplit()
pre_tokenizer.pre_tokenize_str(<span class="hljs-string">&quot;Let&#x27;s test my pre-tokenizer.&quot;</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">&quot;Let&#x27;s&quot;</span>, (<span class="hljs-number">0</span>, <span class="hljs-number">5</span>)), (<span class="hljs-string">&#x27;test&#x27;</span>, (<span class="hljs-number">6</span>, <span class="hljs-number">10</span>)), (<span class="hljs-string">&#x27;my&#x27;</span>, (<span class="hljs-number">11</span>, <span class="hljs-number">13</span>)), (<span class="hljs-string">&#x27;pre-tokenizer.&#x27;</span>, (<span class="hljs-number">14</span>, <span class="hljs-number">28</span>))]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1be6pgh">Как и в случае с нормализаторами, вы можете использовать <code>Sequence</code> для комбинирования нескольких предварительных токенизаторов:</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 -->pre_tokenizer = pre_tokenizers.<span class="hljs-type">Sequence</span>(
[pre_tokenizers.WhitespaceSplit(), pre_tokenizers.Punctuation()]
)
pre_tokenizer.pre_tokenize_str(<span class="hljs-string">&quot;Let&#x27;s test my pre-tokenizer.&quot;</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">&#x27;Let&#x27;</span>, (<span class="hljs-number">0</span>, <span class="hljs-number">3</span>)), (<span class="hljs-string">&quot;&#x27;&quot;</span>, (<span class="hljs-number">3</span>, <span class="hljs-number">4</span>)), (<span class="hljs-string">&#x27;s&#x27;</span>, (<span class="hljs-number">4</span>, <span class="hljs-number">5</span>)), (<span class="hljs-string">&#x27;test&#x27;</span>, (<span class="hljs-number">6</span>, <span class="hljs-number">10</span>)), (<span class="hljs-string">&#x27;my&#x27;</span>, (<span class="hljs-number">11</span>, <span class="hljs-number">13</span>)), (<span class="hljs-string">&#x27;pre&#x27;</span>, (<span class="hljs-number">14</span>, <span class="hljs-number">17</span>)),
(<span class="hljs-string">&#x27;-&#x27;</span>, (<span class="hljs-number">17</span>, <span class="hljs-number">18</span>)), (<span class="hljs-string">&#x27;tokenizer&#x27;</span>, (<span class="hljs-number">18</span>, <span class="hljs-number">27</span>)), (<span class="hljs-string">&#x27;.&#x27;</span>, (<span class="hljs-number">27</span>, <span class="hljs-number">28</span>))]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-ornxe5">Следующий шаг в конвейере токенизации - обработка входных данных с помощью модели. Мы уже указали нашу модель в инициализации, но нам все еще нужно обучить ее, для чего потребуется <code>WordPieceTrainer</code>. Главное, что нужно помнить при инстанцировании тренера в 🤗 Tokenizers, это то, что вам нужно передать ему все специальные токены, которые вы собираетесь использовать - иначе он не добавит их в словарь, поскольку их нет в обучающем корпусе:</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 -->special_tokens = [<span class="hljs-string">&quot;[UNK]&quot;</span>, <span class="hljs-string">&quot;[PAD]&quot;</span>, <span class="hljs-string">&quot;[CLS]&quot;</span>, <span class="hljs-string">&quot;[SEP]&quot;</span>, <span class="hljs-string">&quot;[MASK]&quot;</span>]
trainer = trainers.WordPieceTrainer(vocab_size=<span class="hljs-number">25000</span>, special_tokens=special_tokens)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-14vfm6y">Помимо указания <code>vocab_size</code> и <code>special_tokens</code>, мы можем задать <code>min_frequency</code> (количество раз, которое должен встретиться токен, чтобы быть включенным в словарь) или изменить <code>continuing_subword_prefix</code> (если мы хотим использовать что-то отличное от <code>##</code>).</p> <p data-svelte-h="svelte-93abaw">Чтобы обучить нашу модель с помощью итератора, который мы определили ранее, достаточно выполнить эту команду:</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 -->tokenizer.train_from_iterator(get_training_corpus(), trainer=trainer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-y1gxvq">Мы также можем использовать текстовые файлы для обучения нашего токенизатора, что будет выглядеть следующим образом (предварительно мы повторно инициализируем модель с пустым <code>WordPiece</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 -->tokenizer.model = models.WordPiece(unk_token=<span class="hljs-string">&quot;[UNK]&quot;</span>)
tokenizer.train([<span class="hljs-string">&quot;wikitext-2.txt&quot;</span>], trainer=trainer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-19sql22">В обоих случаях мы можем проверить работу токенизатора на тексте, вызвав метод <code>encode()</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 -->encoding = tokenizer.encode(<span class="hljs-string">&quot;Let&#x27;s test this tokenizer.&quot;</span>)
<span class="hljs-built_in">print</span>(encoding.tokens)<!-- 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;let&#x27;</span>, <span class="hljs-string">&quot;&#x27;&quot;</span>, <span class="hljs-string">&#x27;s&#x27;</span>, <span class="hljs-string">&#x27;test&#x27;</span>, <span class="hljs-string">&#x27;this&#x27;</span>, <span class="hljs-string">&#x27;tok&#x27;</span>, <span class="hljs-string">&#x27;##eni&#x27;</span>, <span class="hljs-string">&#x27;##zer&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-kyltif">Полученное <code>encoding</code> представляет собой <code>Encoding</code>, которое содержит все необходимые результаты работы токенизатора в разных атрибутах: <code>ids</code>, <code>type_ids</code>, <code>tokens</code>, <code>offsets</code>, <code>attention_mask</code>, <code>special_tokens_mask</code> и <code>overflowing</code>.</p> <p data-svelte-h="svelte-1e0xtn8">Последний шаг в конвейере токенизации - постобработка. Нам нужно добавить токен <code>[CLS]</code> в начале и токен <code>[SEP]</code> в конце (или после каждого предложения, если у нас есть пара предложений). Для этого мы будем использовать <code>TemplateProcessor</code>, но сначала нам нужно узнать идентификаторы токенов <code>[CLS]</code> и <code>[SEP]</code> в словаре:</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 -->cls_token_id = tokenizer.token_to_id(<span class="hljs-string">&quot;[CLS]&quot;</span>)
sep_token_id = tokenizer.token_to_id(<span class="hljs-string">&quot;[SEP]&quot;</span>)
<span class="hljs-built_in">print</span>(cls_token_id, sep_token_id)<!-- 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">2</span>, <span class="hljs-number">3</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1nd2ny2">Чтобы написать шаблон для <code>TemplateProcessor</code>, мы должны указать, как обрабатывать одно предложение и пару предложений. Для обоих случаев мы указываем специальные токены, которые мы хотим использовать; первое (или одиночное) предложение представлено <code>$A</code>, а второе предложение (если кодируется пара) представлено <code>$B</code>. Для каждого из них (специальных токенов и предложений) мы также указываем соответствующий идентификатор типа токена (token type ID) после двоеточия.</p> <p data-svelte-h="svelte-na8mx0">Таким образом, классический шаблон BERT определяется следующим образом:</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 -->tokenizer.post_processor = processors.TemplateProcessing(
single=<span class="hljs-string">f&quot;[CLS]:0 $A:0 [SEP]:0&quot;</span>,
pair=<span class="hljs-string">f&quot;[CLS]:0 $A:0 [SEP]:0 $B:1 [SEP]:1&quot;</span>,
special_tokens=[(<span class="hljs-string">&quot;[CLS]&quot;</span>, cls_token_id), (<span class="hljs-string">&quot;[SEP]&quot;</span>, sep_token_id)],
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-rveu42">Обратите внимание, что нам нужно передать идентификаторы специальных токенов, чтобы токенизатор мог правильно преобразовать их в их идентификаторы.</p> <p data-svelte-h="svelte-8cfq99">Как только это будет добавлено, вернемся к нашему предыдущему примеру:</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 -->encoding = tokenizer.encode(<span class="hljs-string">&quot;Let&#x27;s test this tokenizer.&quot;</span>)
<span class="hljs-built_in">print</span>(encoding.tokens)<!-- 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;[CLS]&#x27;</span>, <span class="hljs-string">&#x27;let&#x27;</span>, <span class="hljs-string">&quot;&#x27;&quot;</span>, <span class="hljs-string">&#x27;s&#x27;</span>, <span class="hljs-string">&#x27;test&#x27;</span>, <span class="hljs-string">&#x27;this&#x27;</span>, <span class="hljs-string">&#x27;tok&#x27;</span>, <span class="hljs-string">&#x27;##eni&#x27;</span>, <span class="hljs-string">&#x27;##zer&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>, <span class="hljs-string">&#x27;[SEP]&#x27;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-utpzay">И на паре предложений мы получаем правильный результат:</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 -->encoding = tokenizer.encode(<span class="hljs-string">&quot;Let&#x27;s test this tokenizer...&quot;</span>, <span class="hljs-string">&quot;on a pair of sentences.&quot;</span>)
<span class="hljs-built_in">print</span>(encoding.tokens)
<span class="hljs-built_in">print</span>(encoding.type_ids)<!-- 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;[CLS]&#x27;</span>, <span class="hljs-string">&#x27;let&#x27;</span>, <span class="hljs-string">&quot;&#x27;&quot;</span>, <span class="hljs-string">&#x27;s&#x27;</span>, <span class="hljs-string">&#x27;test&#x27;</span>, <span class="hljs-string">&#x27;this&#x27;</span>, <span class="hljs-string">&#x27;tok&#x27;</span>, <span class="hljs-string">&#x27;##eni&#x27;</span>, <span class="hljs-string">&#x27;##zer&#x27;</span>, <span class="hljs-string">&#x27;...&#x27;</span>, <span class="hljs-string">&#x27;[SEP]&#x27;</span>, <span class="hljs-string">&#x27;on&#x27;</span>, <span class="hljs-string">&#x27;a&#x27;</span>, <span class="hljs-string">&#x27;pair&#x27;</span>, <span class="hljs-string">&#x27;of&#x27;</span>, <span class="hljs-string">&#x27;sentences&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>, <span class="hljs-string">&#x27;[SEP]&#x27;</span>]
[<span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-7ww78z">Мы почти закончили создание этого токенизатора с нуля - остался последний шаг - добавить декодер:</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 -->tokenizer.decoder = decoders.WordPiece(prefix=<span class="hljs-string">&quot;##&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-106lhsb">Давайте проверим его на нашем предыдущем <code>encoding</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 -->tokenizer.decode(encoding.ids)<!-- 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">&quot;let&#x27;s test this tokenizer... on a pair of sentences.&quot;</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-189m9gi">Отлично! Мы можем сохранить наш токенизатор в единственном JSON-файле следующим образом:</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 -->tokenizer.save(<span class="hljs-string">&quot;tokenizer.json&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-pm7hiy">Затем мы можем загрузить этот файл в объект <code>Tokenizer</code> с помощью метода <code>from_file()</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 -->new_tokenizer = Tokenizer.from_file(<span class="hljs-string">&quot;tokenizer.json&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1cpidft">Чтобы использовать этот токенизатор в 🤗 Transformers, мы должны обернуть его в <code>PreTrainedTokenizerFast</code>. Мы можем использовать либо общий класс, либо, если наш токенизатор соответствует существующей модели, использовать этот класс (здесь <code>BertTokenizerFast</code>). Если вы используете этот урок для создания нового токенизатора, вам придется использовать первый вариант.</p> <p data-svelte-h="svelte-1dxzktr">Чтобы обернуть токенизатор в <code>PreTrainedTokenizerFast</code>, мы можем либо передать собранный нами токенизатор как <code>tokenizer_object</code>, либо передать сохраненный файл токенизатора как <code>tokenizer_file</code>. Главное помнить, что нам придется вручную задавать все специальные токены, поскольку класс не может определить из объекта <code>tokenizer</code>, какой токен является токеном маски, токеном <code>[CLS]</code> и т. д.:</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> PreTrainedTokenizerFast
wrapped_tokenizer = PreTrainedTokenizerFast(
tokenizer_object=tokenizer,
<span class="hljs-comment"># tokenizer_file=&quot;tokenizer.json&quot;, # В качестве альтернативы можно загрузить из файла токенизатора.</span>
unk_token=<span class="hljs-string">&quot;[UNK]&quot;</span>,
pad_token=<span class="hljs-string">&quot;[PAD]&quot;</span>,
cls_token=<span class="hljs-string">&quot;[CLS]&quot;</span>,
sep_token=<span class="hljs-string">&quot;[SEP]&quot;</span>,
mask_token=<span class="hljs-string">&quot;[MASK]&quot;</span>,
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-ttz8i2">Если вы используете определенный класс токенизатора (например, <code>BertTokenizerFast</code>), вам нужно будет указать только специальные токены, которые отличаются от токенов по умолчанию (здесь их нет):</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> BertTokenizerFast
wrapped_tokenizer = BertTokenizerFast(tokenizer_object=tokenizer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ij2bkd">Затем вы можете использовать этот токенизатор, как и любой другой токенизатор 🤗 Transformers. Вы можете сохранить его с помощью метода <code>save_pretrained()</code> или загрузить на хаб с помощью метода <code>push_to_hub()</code>.</p> <p data-svelte-h="svelte-1x7u3bt">Теперь, когда мы рассмотрели, как создать токенизатор WordPiece, давайте сделаем то же самое для токенизатора BPE. Мы будем двигаться немного быстрее, поскольку вы знаете все шаги, и подчеркнем только различия.</p> <h2 class="relative group"><a id="building-a-bpe-tokenizer-from-scratch" 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="#building-a-bpe-tokenizer-from-scratch"><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>Создание токенизатора BPE с нуля</span></h2> <p data-svelte-h="svelte-ms4jk9">Теперь давайте создадим токенизатор GPT-2. Как и в случае с токенизатором BERT, мы начнем с инициализации <code>Tokenizer</code> с моделью BPE:</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 -->tokenizer = Tokenizer(models.BPE())<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1efr97x">Также, как и в случае с BERT, мы могли бы инициализировать эту модель словарем, если бы он у нас был (в этом случае нам нужно было бы передать <code>vocab</code> и <code>merges</code>), но поскольку мы будем обучать с нуля, нам не нужно этого делать. Нам также не нужно указывать <code>unk_token</code>, потому что GPT-2 использует byte-level BPE, который не требует этого.</p> <p data-svelte-h="svelte-zmu7y2">GPT-2 не использует нормализатор, поэтому мы пропускаем этот шаг и переходим непосредственно к предварительной токенизации:</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 -->tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=<span class="hljs-literal">False</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1cizskw">Опция, которую мы добавили к <code>ByteLevel</code>, заключается в том, чтобы не добавлять пробел в начале предложения (в противном случае это происходит по умолчанию). Мы можем посмотреть на предварительную токенизацию примера текста, как было показано ранее:</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 -->tokenizer.pre_tokenizer.pre_tokenize_str(<span class="hljs-string">&quot;Let&#x27;s test pre-tokenization!&quot;</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">&#x27;Let&#x27;</span>, (<span class="hljs-number">0</span>, <span class="hljs-number">3</span>)), (<span class="hljs-string">&quot;&#x27;s&quot;</span>, (<span class="hljs-number">3</span>, <span class="hljs-number">5</span>)), (<span class="hljs-string">&#x27;Ġtest&#x27;</span>, (<span class="hljs-number">5</span>, <span class="hljs-number">10</span>)), (<span class="hljs-string">&#x27;Ġpre&#x27;</span>, (<span class="hljs-number">10</span>, <span class="hljs-number">14</span>)), (<span class="hljs-string">&#x27;-&#x27;</span>, (<span class="hljs-number">14</span>, <span class="hljs-number">15</span>)),
(<span class="hljs-string">&#x27;tokenization&#x27;</span>, (<span class="hljs-number">15</span>, <span class="hljs-number">27</span>)), (<span class="hljs-string">&#x27;!&#x27;</span>, (<span class="hljs-number">27</span>, <span class="hljs-number">28</span>))]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1a0orcp">Далее следует модель, которую нужно обучить. Для GPT-2 единственным специальным токеном является токен конца текста:</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 -->trainer = trainers.BpeTrainer(vocab_size=<span class="hljs-number">25000</span>, special_tokens=[<span class="hljs-string">&quot;&lt;|endoftext|&gt;&quot;</span>])
tokenizer.train_from_iterator(get_training_corpus(), trainer=trainer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1p9vr1i">Как и в случае с <code>WordPieceTrainer</code>, а также <code>vocab_size</code> и <code>special_tokens</code>, мы можем указать <code>min_frequency</code>, если хотим, или если у нас есть суффикс конца слова (например, <code>&lt;/w&gt;</code>), мы можем задать его с помощью <code>end_of_word_suffix</code>.</p> <p data-svelte-h="svelte-u4s8kx">Этот токенизатор также может быть обучен на текстовых файлах:</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 -->tokenizer.model = models.BPE()
tokenizer.train([<span class="hljs-string">&quot;wikitext-2.txt&quot;</span>], trainer=trainer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-cmzjwq">Давайте посмотрим на пример токенизации текста:</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 -->encoding = tokenizer.encode(<span class="hljs-string">&quot;Let&#x27;s test this tokenizer.&quot;</span>)
<span class="hljs-built_in">print</span>(encoding.tokens)<!-- 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;L&#x27;</span>, <span class="hljs-string">&#x27;et&#x27;</span>, <span class="hljs-string">&quot;&#x27;&quot;</span>, <span class="hljs-string">&#x27;s&#x27;</span>, <span class="hljs-string">&#x27;Ġtest&#x27;</span>, <span class="hljs-string">&#x27;Ġthis&#x27;</span>, <span class="hljs-string">&#x27;Ġto&#x27;</span>, <span class="hljs-string">&#x27;ken&#x27;</span>, <span class="hljs-string">&#x27;izer&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1gd6jqz">Мы применяем постобработку на уровне байтов для токенизатора GPT-2 следующим образом:</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 -->tokenizer.post_processor = processors.ByteLevel(trim_offsets=<span class="hljs-literal">False</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1g95bkh">Опция <code>trim_offsets = False</code> указывает постпроцессору, что мы должны оставить смещения токенов, начинающихся с ‘Ġ’, как есть: таким образом, начало смещения будет указывать на пробел перед словом, а не на первый символ слова (поскольку пробел технически является частью токена). Давайте посмотрим на результат с текстом, который мы только что закодировали, где <code>&#39;Ġtest&#39;</code> - это токен с индексом 4:</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 -->sentence = <span class="hljs-string">&quot;Let&#x27;s test this tokenizer.&quot;</span>
encoding = tokenizer.encode(sentence)
start, end = encoding.offsets[<span class="hljs-number">4</span>]
sentence[start:end]<!-- 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; test&#x27;</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-eerktk">Наконец, мы добавляем декодер на уровне байтов:</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 -->tokenizer.decoder = decoders.ByteLevel()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-yra0gn">и мы сможем перепроверить, правильно ли он работает:</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 -->tokenizer.decode(encoding.ids)<!-- 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">&quot;Let&#x27;s test this tokenizer.&quot;</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-ka6eer">Отлично! Теперь, когда мы закончили, мы можем сохранить токенизатор, как раньше, и обернуть его в <code>PreTrainedTokenizerFast</code> или <code>GPT2TokenizerFast</code>, если мы хотим использовать его в 🤗 Transformers:</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> PreTrainedTokenizerFast
wrapped_tokenizer = PreTrainedTokenizerFast(
tokenizer_object=tokenizer,
bos_token=<span class="hljs-string">&quot;&lt;|endoftext|&gt;&quot;</span>,
eos_token=<span class="hljs-string">&quot;&lt;|endoftext|&gt;&quot;</span>,
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1cmlt11">или:</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> GPT2TokenizerFast
wrapped_tokenizer = GPT2TokenizerFast(tokenizer_object=tokenizer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-upbbw9">В качестве последнего примера мы покажем вам, как создать токенизатор Unigram с нуля.</p> <h2 class="relative group"><a id="building-a-unigram-tokenizer-from-scratch" 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="#building-a-unigram-tokenizer-from-scratch"><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>Создание токенизатора Unigram с нуля</span></h2> <p data-svelte-h="svelte-gprlmq">Теперь давайте построим токенизатор XLNet. Как и в предыдущих токенизаторах, мы начнем с инициализации <code>Tokenizer</code> с моделью Unigram:</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 -->tokenizer = Tokenizer(models.Unigram())<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1yy918s">Опять же, мы могли бы инициализировать эту модель словарем, если бы он у нас был.</p> <p data-svelte-h="svelte-1n71160">Для нормализации XLNet использует несколько замен (которые пришли из SentencePiece):</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> tokenizers <span class="hljs-keyword">import</span> Regex
tokenizer.normalizer = normalizers.<span class="hljs-type">Sequence</span>(
[
normalizers.Replace(<span class="hljs-string">&quot;``&quot;</span>, <span class="hljs-string">&#x27;&quot;&#x27;</span>),
normalizers.Replace(<span class="hljs-string">&quot;&#x27;&#x27;&quot;</span>, <span class="hljs-string">&#x27;&quot;&#x27;</span>),
normalizers.NFKD(),
normalizers.StripAccents(),
normalizers.Replace(Regex(<span class="hljs-string">&quot; {2,}&quot;</span>), <span class="hljs-string">&quot; &quot;</span>),
]
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-197ybm4">Он заменяет <code></code> и <code></code> with <code></code> и любую последовательность из двух или более пробелов на один пробел, а также удаляет ударения в токенезируемых текстах.</p> <p data-svelte-h="svelte-f4i3wg">Предварительный токенизатор, который должен использоваться для любого токенизатора SentencePiece, - это <code>Metaspace</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 -->tokenizer.pre_tokenizer = pre_tokenizers.Metaspace()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-xj6oo">Мы можем посмотреть на предварительную токенизацию примера текста, как было показано ранее:</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 -->tokenizer.pre_tokenizer.pre_tokenize_str(<span class="hljs-string">&quot;Let&#x27;s test the pre-tokenizer!&quot;</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">&quot;▁Let&#x27;s&quot;</span>, (<span class="hljs-number">0</span>, <span class="hljs-number">5</span>)), (<span class="hljs-string">&#x27;▁test&#x27;</span>, (<span class="hljs-number">5</span>, <span class="hljs-number">10</span>)), (<span class="hljs-string">&#x27;▁the&#x27;</span>, (<span class="hljs-number">10</span>, <span class="hljs-number">14</span>)), (<span class="hljs-string">&#x27;▁pre-tokenizer!&#x27;</span>, (<span class="hljs-number">14</span>, <span class="hljs-number">29</span>))]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1tjrbj7">Далее следует модель, которую нужно обучить. В XLNet довольно много специальных токенов:</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 -->special_tokens = [<span class="hljs-string">&quot;&lt;cls&gt;&quot;</span>, <span class="hljs-string">&quot;&lt;sep&gt;&quot;</span>, <span class="hljs-string">&quot;&lt;unk&gt;&quot;</span>, <span class="hljs-string">&quot;&lt;pad&gt;&quot;</span>, <span class="hljs-string">&quot;&lt;mask&gt;&quot;</span>, <span class="hljs-string">&quot;&lt;s&gt;&quot;</span>, <span class="hljs-string">&quot;&lt;/s&gt;&quot;</span>]
trainer = trainers.UnigramTrainer(
vocab_size=<span class="hljs-number">25000</span>, special_tokens=special_tokens, unk_token=<span class="hljs-string">&quot;&lt;unk&gt;&quot;</span>
)
tokenizer.train_from_iterator(get_training_corpus(), trainer=trainer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-z62fgj">Очень важный аргумент, который не стоит забывать для <code>UnigramTrainer</code> - это <code>unk_token</code>. Мы также можем передавать другие аргументы, специфичные для алгоритма Unigram, такие как <code>shrinking_factor</code> для каждого шага удаления токенов (по умолчанию 0.75) или <code>max_piece_length</code> для указания максимальной длины данного токена (по умолчанию 16).</p> <p data-svelte-h="svelte-u4s8kx">Этот токенизатор также может быть обучен на текстовых файлах:</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 -->tokenizer.model = models.Unigram()
tokenizer.train([<span class="hljs-string">&quot;wikitext-2.txt&quot;</span>], trainer=trainer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1gxe1cq">Давайте посмотрим на токенизацию примера текста:</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 -->encoding = tokenizer.encode(<span class="hljs-string">&quot;Let&#x27;s test this tokenizer.&quot;</span>)
<span class="hljs-built_in">print</span>(encoding.tokens)<!-- 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;▁Let&#x27;</span>, <span class="hljs-string">&quot;&#x27;&quot;</span>, <span class="hljs-string">&#x27;s&#x27;</span>, <span class="hljs-string">&#x27;▁test&#x27;</span>, <span class="hljs-string">&#x27;▁this&#x27;</span>, <span class="hljs-string">&#x27;▁to&#x27;</span>, <span class="hljs-string">&#x27;ken&#x27;</span>, <span class="hljs-string">&#x27;izer&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-hd4plk">Особенностью XLNet является то, что он помещает токен <code>&lt;cls&gt;</code> в конец предложения с идентификатором типа 2 (чтобы отличить его от других токенов). В результате он помещается слева. Мы можем разобраться со всеми специальными токенами и идентификаторами типов токенов с помощью шаблона, как в BERT, но сначала нам нужно получить идентификаторы токенов <code>&lt;cls&gt;</code> и <code>&lt;sep&gt;</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 -->cls_token_id = tokenizer.token_to_id(<span class="hljs-string">&quot;&lt;cls&gt;&quot;</span>)
sep_token_id = tokenizer.token_to_id(<span class="hljs-string">&quot;&lt;sep&gt;&quot;</span>)
<span class="hljs-built_in">print</span>(cls_token_id, sep_token_id)<!-- 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">0</span> <span class="hljs-number">1</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1b1rase">Шаблон выглядит следующим образом:</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 -->tokenizer.post_processor = processors.TemplateProcessing(
single=<span class="hljs-string">&quot;$A:0 &lt;sep&gt;:0 &lt;cls&gt;:2&quot;</span>,
pair=<span class="hljs-string">&quot;$A:0 &lt;sep&gt;:0 $B:1 &lt;sep&gt;:1 &lt;cls&gt;:2&quot;</span>,
special_tokens=[(<span class="hljs-string">&quot;&lt;sep&gt;&quot;</span>, sep_token_id), (<span class="hljs-string">&quot;&lt;cls&gt;&quot;</span>, cls_token_id)],
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-45g554">И мы можем проверить, как это работает, закодировав пару предложений:</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 -->encoding = tokenizer.encode(<span class="hljs-string">&quot;Let&#x27;s test this tokenizer...&quot;</span>, <span class="hljs-string">&quot;on a pair of sentences!&quot;</span>)
<span class="hljs-built_in">print</span>(encoding.tokens)
<span class="hljs-built_in">print</span>(encoding.type_ids)<!-- 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;▁Let&#x27;</span>, <span class="hljs-string">&quot;&#x27;&quot;</span>, <span class="hljs-string">&#x27;s&#x27;</span>, <span class="hljs-string">&#x27;▁test&#x27;</span>, <span class="hljs-string">&#x27;▁this&#x27;</span>, <span class="hljs-string">&#x27;▁to&#x27;</span>, <span class="hljs-string">&#x27;ken&#x27;</span>, <span class="hljs-string">&#x27;izer&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>, <span class="hljs-string">&#x27;&lt;sep&gt;&#x27;</span>, <span class="hljs-string">&#x27;&#x27;</span>, <span class="hljs-string">&#x27;on&#x27;</span>, <span class="hljs-string">&#x27;&#x27;</span>, <span class="hljs-string">&#x27;a&#x27;</span>, <span class="hljs-string">&#x27;▁pair&#x27;</span>,
<span class="hljs-string">&#x27;▁of&#x27;</span>, <span class="hljs-string">&#x27;▁sentence&#x27;</span>, <span class="hljs-string">&#x27;s&#x27;</span>, <span class="hljs-string">&#x27;!&#x27;</span>, <span class="hljs-string">&#x27;&lt;sep&gt;&#x27;</span>, <span class="hljs-string">&#x27;&lt;cls&gt;&#x27;</span>]
[<span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">2</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1b0rt28">Наконец, мы добавляем декодер <code>Metaspace</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 -->tokenizer.decoder = decoders.Metaspace()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-tmcl78">и мы закончили работу с этим токенизатором! Мы можем сохранить токенизатор, как и раньше, и обернуть его в <code>PreTrainedTokenizerFast</code> или <code>XLNetTokenizerFast</code>, если мы хотим использовать его в 🤗 Transformers. При использовании <code>PreTrainedTokenizerFast</code> следует обратить внимание на то, что помимо специальных токенов, нам нужно указать библиотеке 🤗 Transformers на то, чтобы они располагались слева:</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> PreTrainedTokenizerFast
wrapped_tokenizer = PreTrainedTokenizerFast(
tokenizer_object=tokenizer,
bos_token=<span class="hljs-string">&quot;&lt;s&gt;&quot;</span>,
eos_token=<span class="hljs-string">&quot;&lt;/s&gt;&quot;</span>,
unk_token=<span class="hljs-string">&quot;&lt;unk&gt;&quot;</span>,
pad_token=<span class="hljs-string">&quot;&lt;pad&gt;&quot;</span>,
cls_token=<span class="hljs-string">&quot;&lt;cls&gt;&quot;</span>,
sep_token=<span class="hljs-string">&quot;&lt;sep&gt;&quot;</span>,
mask_token=<span class="hljs-string">&quot;&lt;mask&gt;&quot;</span>,
padding_side=<span class="hljs-string">&quot;left&quot;</span>,
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-b94jpj">Или альтернативно:</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> XLNetTokenizerFast
wrapped_tokenizer = XLNetTokenizerFast(tokenizer_object=tokenizer)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-15vwsr5">Теперь, когда вы увидели, как различные блоки используются для создания существующих токенизаторов, вы должны быть в состоянии написать любой токенизатор, который вы хотите, с помощью библиотеки 🤗 Tokenizers и использовать его в 🤗 Transformers.</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/ru/chapter6/8.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_j8s7wf = {
assets: "/docs/course/pr_1069/ru",
base: "/docs/course/pr_1069/ru",
env: {}
};
const element = document.currentScript.parentElement;
const data = [null,null];
Promise.all([
import("/docs/course/pr_1069/ru/_app/immutable/entry/start.48687cc8.js"),
import("/docs/course/pr_1069/ru/_app/immutable/entry/app.b79a803d.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 50],
data,
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
149 kB
·
Xet hash:
51a561440215b464445725f3e63fa182708166391f7e519dc20825e9521d8402

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