Hugging Face's logo

Buckets:
hf-doc-build
/
doc-dev

Files
xet
hf-doc-build/doc-dev / course /pr_1069 /ru /chapter6 /7.html
rtrm's picture
rtrm
download
raw
106 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Токенизация Unigram&quot;,&quot;local&quot;:&quot;unigram-tokenization&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Алгоритм обучения&quot;,&quot;local&quot;:&quot;training-algorithm&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Алгоритм токенизации&quot;,&quot;local&quot;:&quot;tokenization-algorithm&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Назад к обучению&quot;,&quot;local&quot;:&quot;back-to-training&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Реализация Unigram&quot;,&quot;local&quot;:&quot;implementing-unigram&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/49.fbc2ad7a.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;Токенизация Unigram&quot;,&quot;local&quot;:&quot;unigram-tokenization&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Алгоритм обучения&quot;,&quot;local&quot;:&quot;training-algorithm&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Алгоритм токенизации&quot;,&quot;local&quot;:&quot;tokenization-algorithm&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Назад к обучению&quot;,&quot;local&quot;:&quot;back-to-training&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Реализация Unigram&quot;,&quot;local&quot;:&quot;implementing-unigram&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="unigram-tokenization" 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="#unigram-tokenization"><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>Токенизация Unigram</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/section7.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/section7.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-ztfl7b">Алгоритм Unigram часто используется в SentencePiece, который является алгоритмом токенизации, применяемым в таких моделях, как AlBERT, T5, mBART, Big Bird и XLNet.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/TGZfZVuF9Yc" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-xcsu8n">💡 В этом разделе подробно рассматривается Unigram, вплоть до демонстрации полной реализации. Вы можете пропустить его, если вам нужен только общий обзор алгоритма токенизации.</p></div> <h2 class="relative group"><a id="training-algorithm" 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="#training-algorithm"><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>Алгоритм обучения</span></h2> <p data-svelte-h="svelte-38wf8v">По сравнению с BPE и WordPiece, Unigram работает в другом направлении: он начинает с большого словарного запаса и удаляет из него токены, пока не достигнет желаемого размера словаря. Существует несколько вариантов создания базового словаря: например, мы можем взять наиболее часто встречающиеся подстроки в предварительно токенизированных словах или применить BPE к исходному корпусу с большим объемом словаря.</p> <p data-svelte-h="svelte-9m1yzd">На каждом шаге обучения алгоритм Unigram рассчитывает потери по корпусу с учетом текущего словарного запаса. Затем для каждого символа в словаре алгоритм вычисляет, насколько увеличится общая потеря, если этот символ будет удален, и ищет символы, которые увеличат ее меньше всего. Эти символы оказывают меньшее влияние на общую потерю по корпусу, поэтому в некотором смысле они “менее нужны” и являются лучшими кандидатами на удаление.</p> <p>Это очень дорогостоящая операция, поэтому мы удаляем не просто один символ, связанный с наименьшим увеличением потерь, а<!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>p</mi></mrow><annotation encoding="application/x-tex">p</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.625em;vertical-align:-0.1944em;"></span><span class="mord mathnormal">p</span></span></span></span><!-- HTML_TAG_END --> (\(p\) - гиперпараметр, которым вы можете управлять, обычно 10 или 20) процентов символов, связанных с наименьшим увеличением потерь. Этот процесс повторяется до тех пор, пока словарь не достигнет желаемого размера.</p> <p data-svelte-h="svelte-6aybad">Обратите внимание, что мы никогда не удаляем базовые символы, чтобы убедиться, что любое слово может быть токенизировано.</p> <p data-svelte-h="svelte-1kd5ss2">Итак, все еще немного туманно: основная часть алгоритма заключается в том, чтобы вычислить потери по корпусу и посмотреть, как они изменяются при удалении некоторых токенов из словаря, но мы еще не объяснили, как это сделать. Этот шаг зависит от алгоритма токенизации модели Unigram, поэтому мы рассмотрим его далее.</p> <p data-svelte-h="svelte-11p2t2p">Мы используем корпус текста из предыдущих примеров:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">&quot;hug&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">10</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">&quot;pug&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">&quot;pun&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">12</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">&quot;bun&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">&quot;hugs&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1vs2ac9">и для этого примера мы возьмем все подстроки из исходного словаря:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-selector-attr">[<span class="hljs-string">&quot;h&quot;</span>, <span class="hljs-string">&quot;u&quot;</span>, <span class="hljs-string">&quot;g&quot;</span>, <span class="hljs-string">&quot;hu&quot;</span>, <span class="hljs-string">&quot;ug&quot;</span>, <span class="hljs-string">&quot;p&quot;</span>, <span class="hljs-string">&quot;pu&quot;</span>, <span class="hljs-string">&quot;n&quot;</span>, <span class="hljs-string">&quot;un&quot;</span>, <span class="hljs-string">&quot;b&quot;</span>, <span class="hljs-string">&quot;bu&quot;</span>, <span class="hljs-string">&quot;s&quot;</span>, <span class="hljs-string">&quot;hug&quot;</span>, <span class="hljs-string">&quot;gs&quot;</span>, <span class="hljs-string">&quot;ugs&quot;</span>]</span><!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="tokenization-algorithm" 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="#tokenization-algorithm"><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>Алгоритм токенизации</span></h2> <p data-svelte-h="svelte-d2q239">Модель Unigram - это тип языковой модели, в которой каждый токен рассматривается как независимый от предшествующих ему. Это самая простая языковая модель в том смысле, что вероятность появления токена X с учетом предыдущего контекста - это просто вероятность появления токена X. Таким образом, если бы мы использовали модель Unigram для генерации текста, мы бы всегда предсказывали наиболее часто встречающийся токен.</p> <p data-svelte-h="svelte-1bxcwfe">Вероятность данного токена - это его частота (количество раз, когда мы его находим) в исходном корпусе, деленная на сумму частот всех токенов в словаре (чтобы убедиться, что суммы вероятностей равны 1). Например, <code>&quot;ug&quot;</code> присутствует в <code>&quot;hug&quot;</code>, <code>&quot;pug&quot;</code> и <code>&quot;hugs&quot;</code>, поэтому его частота в нашем корпусе равна 20.</p> <p data-svelte-h="svelte-3bq9li">Здесь приведены частоты всех возможных подслов 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 -->(<span class="hljs-string">&quot;h&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">15</span>) (<span class="hljs-string">&quot;u&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">36</span>) (<span class="hljs-string">&quot;g&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">20</span>) (<span class="hljs-string">&quot;hu&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">15</span>) (<span class="hljs-string">&quot;ug&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">20</span>) (<span class="hljs-string">&quot;p&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">17</span>) (<span class="hljs-string">&quot;pu&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">17</span>) (<span class="hljs-string">&quot;n&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">16</span>)
(<span class="hljs-string">&quot;un&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">16</span>) (<span class="hljs-string">&quot;b&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>) (<span class="hljs-string">&quot;bu&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>) (<span class="hljs-string">&quot;s&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>) (<span class="hljs-string">&quot;hug&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">15</span>) (<span class="hljs-string">&quot;gs&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>) (<span class="hljs-string">&quot;ugs&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-im93lg">Итак, сумма всех частот равна 210, а вероятность появления подслова <code>&quot;ug&quot;</code>, таким образом, составляет 20/210.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-fk9ipk">✏️ <strong>Теперь ваша очередь!</strong> Напишите код для вычисления вышеуказанных частот и дважды проверьте правильность приведенных результатов, а также общую сумму.</p></div> <p>Теперь для токенизации данного слова мы рассматриваем все возможные сегментации на токены и вычисляем вероятность каждого из них в соответствии с моделью Unigram. Поскольку все токены считаются независимыми, эта вероятность равна произведению вероятностей появления каждого токена. Например, при токенизации <code data-svelte-h="svelte-1n2m4po">[&quot;p&quot;, &quot;u&quot;, &quot;g&quot;]</code> слова <code data-svelte-h="svelte-1gjdq76">&quot;pug&quot;</code> вероятность составляет:
<!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><mi>P</mi><mo stretchy="false">(</mo><mo stretchy="false">[</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>p</mi><mi mathvariant="normal">&quot;</mi><mo separator="true">,</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>u</mi><mi mathvariant="normal">&quot;</mi><mo separator="true">,</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>g</mi><mi mathvariant="normal">&quot;</mi><mo stretchy="false">]</mo><mo stretchy="false">)</mo><mo>=</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>p</mi><mi mathvariant="normal">&quot;</mi><mo stretchy="false">)</mo><mo>×</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>u</mi><mi mathvariant="normal">&quot;</mi><mo stretchy="false">)</mo><mo>×</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>g</mi><mi mathvariant="normal">&quot;</mi><mo stretchy="false">)</mo><mo>=</mo><mfrac><mn>5</mn><mn>210</mn></mfrac><mo>×</mo><mfrac><mn>36</mn><mn>210</mn></mfrac><mo>×</mo><mfrac><mn>20</mn><mn>210</mn></mfrac><mo>=</mo><mn>0.000389</mn></mrow><annotation encoding="application/x-tex">P([``p&quot;, ``u&quot;, ``g&quot;]) = P(``p&quot;) \times P(``u&quot;) \times P(``g&quot;) = \frac{5}{210} \times \frac{36}{210} \times \frac{20}{210} = 0.000389</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">([</span><span class="mord">‘‘</span><span class="mord mathnormal">p</span><span class="mord">&quot;</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">‘‘</span><span class="mord mathnormal">u</span><span class="mord">&quot;</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">‘‘</span><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="mord">&quot;</span><span class="mclose">])</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal">p</span><span class="mord">&quot;</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal">u</span><span class="mord">&quot;</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="mord">&quot;</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">5</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">36</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">20</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">0.000389</span></span></span></span></span><!-- HTML_TAG_END --></p> <p>Для сравнения, токен <code data-svelte-h="svelte-42m5r0">[&quot;pu&quot;, &quot;g&quot;]</code> имеет вероятность:
<!-- HTML_TAG_START --><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><mi>P</mi><mo stretchy="false">(</mo><mo stretchy="false">[</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>p</mi><mi>u</mi><mi mathvariant="normal">&quot;</mi><mo separator="true">,</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>g</mi><mi mathvariant="normal">&quot;</mi><mo stretchy="false">]</mo><mo stretchy="false">)</mo><mo>=</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>p</mi><mi>u</mi><mi mathvariant="normal">&quot;</mi><mo stretchy="false">)</mo><mo>×</mo><mi>P</mi><mo stretchy="false">(</mo><mi mathvariant="normal">‘</mi><mi mathvariant="normal">‘</mi><mi>g</mi><mi mathvariant="normal">&quot;</mi><mo stretchy="false">)</mo><mo>=</mo><mfrac><mn>5</mn><mn>210</mn></mfrac><mo>×</mo><mfrac><mn>20</mn><mn>210</mn></mfrac><mo>=</mo><mn>0.0022676</mn></mrow><annotation encoding="application/x-tex">P([``pu&quot;, ``g&quot;]) = P(``pu&quot;) \times P(``g&quot;) = \frac{5}{210} \times \frac{20}{210} = 0.0022676</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">([</span><span class="mord">‘‘</span><span class="mord mathnormal">p</span><span class="mord mathnormal">u</span><span class="mord">&quot;</span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.1667em;"></span><span class="mord">‘‘</span><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="mord">&quot;</span><span class="mclose">])</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal">p</span><span class="mord mathnormal">u</span><span class="mord">&quot;</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord mathnormal" style="margin-right:0.13889em;">P</span><span class="mopen">(</span><span class="mord">‘‘</span><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="mord">&quot;</span><span class="mclose">)</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">5</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:2.0074em;vertical-align:-0.686em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3214em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">210</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">20</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">0.0022676</span></span></span></span></span><!-- HTML_TAG_END --></p> <p data-svelte-h="svelte-1t4lryo">так что один из них гораздо более вероятен. В целом, токенизации с наименьшим количеством токенов будут иметь наибольшую вероятность (из-за деления на 210, повторяющегося для каждого токена), что соответствует интуитивному желанию: разбить слово на наименьшее количество токенов.</p> <p data-svelte-h="svelte-x62a9z">Токенизация слова с помощью модели Unigram - это токенизация с наибольшей вероятностью. В примере с <code>&quot;pug&quot;</code> приведены вероятности, которые мы получили бы для каждой возможной сегментации:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[<span class="hljs-string">&quot;p&quot;</span>, <span class="hljs-string">&quot;u&quot;</span>, <span class="hljs-string">&quot;g&quot;</span>] : 0.000389
[<span class="hljs-string">&quot;p&quot;</span>, <span class="hljs-string">&quot;ug&quot;</span>] : 0.0022676
[<span class="hljs-string">&quot;pu&quot;</span>, <span class="hljs-string">&quot;g&quot;</span>] : 0.0022676<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-825n4p">Так, <code>&quot;pug&quot;</code> будет токенизировано как <code>[&quot;p&quot;, &quot;ug&quot;]</code> или <code>[&quot;pu&quot;, &quot;g&quot;]</code>, в зависимости от того, какая из этих сегментаций встретится первой (отметим, что в большом корпусе подобные случаи равенства будут редки).</p> <p data-svelte-h="svelte-1d4638c">В данном случае было легко найти все возможные сегментации и вычислить их вероятности, но в общем случае это будет немного сложнее. Для этого используется классический алгоритм, который называется <em>алгоритм Витерби (Viterbi algorithm)</em>. По сути, мы можем построить граф для выявления возможных сегментаций данного слова, сказав, что существует ветвь от символа <em>a</em> до символа <em>b</em>, если подслово от <em>a</em> до <em>b</em> есть в словаре, и приписать этой ветви вероятность подслова.</p> <p data-svelte-h="svelte-cnoenp">Чтобы найти путь в этом графе, который будет иметь наилучшую оценку, алгоритм Витерби определяет для каждой позиции в слове сегментацию с наилучшей оценкой, которая заканчивается на этой позиции. Поскольку мы идем от начала к концу, этот лучший результат можно найти, перебирая все подслова, заканчивающиеся на текущей позиции, а затем используя лучший результат токенизации с позиции, на которой начинается это подслово. Затем нужно просто развернуть путь, чтобы прийти к концу.</p> <p data-svelte-h="svelte-14pw0oa">Давайте рассмотрим пример с использованием нашего словаря и слова <code>&quot;unhug&quot;</code>. Для каждой позиции подслова с наилучшими оценками заканчиваются 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 --><span class="hljs-attribute">Character</span> <span class="hljs-number">0</span> (u): <span class="hljs-string">&quot;u&quot;</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">171429</span>)
<span class="hljs-attribute">Character</span> <span class="hljs-number">1</span> (n): <span class="hljs-string">&quot;un&quot;</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">076191</span>)
<span class="hljs-attribute">Character</span> <span class="hljs-number">2</span> (h): <span class="hljs-string">&quot;un&quot;</span> <span class="hljs-string">&quot;h&quot;</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">005442</span>)
<span class="hljs-attribute">Character</span> <span class="hljs-number">3</span> (u): <span class="hljs-string">&quot;un&quot;</span> <span class="hljs-string">&quot;hu&quot;</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">005442</span>)
<span class="hljs-attribute">Character</span> <span class="hljs-number">4</span> (g): <span class="hljs-string">&quot;un&quot;</span> <span class="hljs-string">&quot;hug&quot;</span> (score <span class="hljs-number">0</span>.<span class="hljs-number">005442</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-12uxsr1">Таким образом, <code>&quot;unhug&quot;</code> будет токенизировано как <code>[&quot;un&quot;, &quot;hug&quot;]</code>.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1al7y90">✏️ <strong>Теперь ваша очередь!</strong> Определите токенизацию слова <code>&quot; huggun&quot;</code> и его оценку.</p></div> <h2 class="relative group"><a id="back-to-training" 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="#back-to-training"><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>Назад к обучению</span></h2> <p data-svelte-h="svelte-tcwcgb">Теперь, когда мы увидели, как работает токенизация, мы можем немного глубже изучить потери, используемые во время обучения. На любом этапе эта потеря вычисляется путем токенизации каждого слова в корпусе с использованием текущего словаря и модели Unigram, определяемой частотами каждого токена в корпусе (как было показано ранее).</p> <p data-svelte-h="svelte-1vdkm7">Каждое слово в корпусе имеет оценку, а потеря - это отрицательное логарифмическое правдоподобие этих оценок, то есть сумма для всех слов в корпусе всех <code>-log(P(word))</code>.</p> <p data-svelte-h="svelte-9v0eeq">Давайте вернемся к нашему примеру со следующим корпусом:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">&quot;hug&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">10</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">&quot;pug&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">&quot;pun&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">12</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">&quot;bun&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">&quot;hugs&quot;</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1xzvd9h">Токенизация каждого слова с соответствующими оценками:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-string">&quot;hug&quot;</span>: [<span class="hljs-string">&quot;hug&quot;</span>] <span class="hljs-comment">(score 0.071428)</span>
<span class="hljs-string">&quot;pug&quot;</span>: [<span class="hljs-string">&quot;pu&quot;</span>, <span class="hljs-string">&quot;g&quot;</span>] <span class="hljs-comment">(score 0.007710)</span>
<span class="hljs-string">&quot;pun&quot;</span>: [<span class="hljs-string">&quot;pu&quot;</span>, <span class="hljs-string">&quot;n&quot;</span>] <span class="hljs-comment">(score 0.006168)</span>
<span class="hljs-string">&quot;bun&quot;</span>: [<span class="hljs-string">&quot;bu&quot;</span>, <span class="hljs-string">&quot;n&quot;</span>] <span class="hljs-comment">(score 0.001451)</span>
<span class="hljs-string">&quot;hugs&quot;</span>: [<span class="hljs-string">&quot;hug&quot;</span>, <span class="hljs-string">&quot;s&quot;</span>] <span class="hljs-comment">(score 0.001701)</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1chhfbr">Таким образом, потери будут:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-attribute">10</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">071428</span>)) + <span class="hljs-number">5</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">007710</span>)) + <span class="hljs-number">12</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">006168</span>)) + <span class="hljs-number">4</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">001451</span>)) + <span class="hljs-number">5</span> * (-log(<span class="hljs-number">0</span>.<span class="hljs-number">001701</span>)) = <span class="hljs-number">169</span>.<span class="hljs-number">8</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-51rubi">Теперь нам нужно вычислить, как удаление каждого токена влияет на потери. Это довольно утомительно, поэтому мы просто сделаем это для двух токенов и оставим весь процесс на потом, когда у нас будет код, чтобы помочь нам. В этом (очень) конкретном случае у нас есть две эквивалентные токенизации всех слов: как мы видели ранее, например, <code>&quot;pug&quot;</code> может быть токенизировано <code>[&quot;p&quot;, &quot;ug&quot;]</code> с тем же результатом. Таким образом, удаление токена <code>&quot;pu&quot;</code> из словаря приведет к точно таким же потерям.</p> <p data-svelte-h="svelte-u3c0y6">С другой стороны, удаление <code>&quot; hug&quot;</code> усугубит потери, потому что токенизация <code>&quot;hug&quot;</code> и <code>&quot;hugs&quot;</code> станет:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-string">&quot;hug&quot;</span>: [<span class="hljs-string">&quot;hu&quot;</span>, <span class="hljs-string">&quot;g&quot;</span>] <span class="hljs-comment">(score 0.006802)</span>
<span class="hljs-string">&quot;hugs&quot;</span>: [<span class="hljs-string">&quot;hu&quot;</span>, <span class="hljs-string">&quot;gs&quot;</span>] <span class="hljs-comment">(score 0.001701)</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-122nrhg">Эти изменения приведут к увеличению потерь:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->- <span class="hljs-number">10</span> * (<span class="hljs-name">-log</span>(<span class="hljs-number">0.071428</span>)) + <span class="hljs-number">10</span> * (<span class="hljs-name">-log</span>(<span class="hljs-number">0.006802</span>)) = <span class="hljs-number">23.5</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-quha3z">Поэтому токен <code>&quot;pu&quot;</code>, вероятно, будет удален из словаря, но не <code>&quot;hug&quot;</code>.</p> <h2 class="relative group"><a id="implementing-unigram" 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="#implementing-unigram"><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>Реализация Unigram</span></h2> <p data-svelte-h="svelte-1ypl538">Теперь давайте реализуем все, что мы видели до сих пор, в коде. Как и в случае с BPE и WordPiece, это не эффективная реализация алгоритма Unigram (совсем наоборот), но она должна помочь вам понять его немного лучше.</p> <p data-svelte-h="svelte-11wpc69">В качестве примера мы будем использовать тот же корпус текста, что и раньше:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->corpus = [
 <span class="hljs-string">&quot;This is the Hugging Face Course.&quot;</span>,
 <span class="hljs-string">&quot;This chapter is about tokenization.&quot;</span>,
 <span class="hljs-string">&quot;This section shows several tokenizer algorithms.&quot;</span>,
 <span class="hljs-string">&quot;Hopefully, you will be able to understand how they are trained and generate tokens.&quot;</span>,
]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1wore2o">На этот раз в качестве модели мы будем использовать <code>xlnet-base-cased</code>:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;xlnet-base-cased&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-114dh07">Как и в случае с BPE и WordPiece, мы начинаем с подсчета количества вхождений каждого слова в корпус:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> collections <span class="hljs-keyword">import</span> defaultdict
word_freqs = defaultdict(<span class="hljs-built_in">int</span>)
<span class="hljs-keyword">for</span> text <span class="hljs-keyword">in</span> corpus:
 words_with_offsets = tokenizer.backend_tokenizer.pre_tokenizer.pre_tokenize_str(text)
 new_words = [word <span class="hljs-keyword">for</span> word, offset <span class="hljs-keyword">in</span> words_with_offsets]
 <span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> new_words:
 word_freqs[word] += <span class="hljs-number">1</span>
word_freqs<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-pnwlm1">Затем нам нужно инициализировать наш словарь чем-то большим, чем размер словаря, который мы захотим получить в конце. Мы должны включить все основные символы (иначе мы не сможем токенизировать каждое слово), но для больших подстрок мы сохраним только самые распространенные, поэтому мы отсортируем их по частоте:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->char_freqs = defaultdict(<span class="hljs-built_in">int</span>)
subwords_freqs = defaultdict(<span class="hljs-built_in">int</span>)
<span class="hljs-keyword">for</span> word, freq <span class="hljs-keyword">in</span> word_freqs.items():
 <span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">len</span>(word)):
 char_freqs[word[i]] += freq
 <span class="hljs-comment"># Перебираем подслова длиной не менее 2</span>
 <span class="hljs-keyword">for</span> j <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(i + <span class="hljs-number">2</span>, <span class="hljs-built_in">len</span>(word) + <span class="hljs-number">1</span>):
 subwords_freqs[word[i:j]] += freq
<span class="hljs-comment"># Сортировка подслов по частоте</span>
sorted_subwords = <span class="hljs-built_in">sorted</span>(subwords_freqs.items(), key=<span class="hljs-keyword">lambda</span> x: x[<span class="hljs-number">1</span>], reverse=<span class="hljs-literal">True</span>)
sorted_subwords[:<span class="hljs-number">10</span>]<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[(<span class="hljs-string">&#x27;▁t&#x27;</span>, <span class="hljs-number">7</span>), (<span class="hljs-string">&#x27;is&#x27;</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">&#x27;er&#x27;</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">&#x27;▁a&#x27;</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">&#x27;▁to&#x27;</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">&#x27;to&#x27;</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">&#x27;en&#x27;</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">&#x27;▁T&#x27;</span>, <span class="hljs-number">3</span>), (<span class="hljs-string">&#x27;▁Th&#x27;</span>, <span class="hljs-number">3</span>), (<span class="hljs-string">&#x27;▁Thi&#x27;</span>, <span class="hljs-number">3</span>)]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1hgj0lr">Мы группируем символы с лучшими подсловами, чтобы получить начальный словарь размером 300:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->token_freqs = <span class="hljs-built_in">list</span>(char_freqs.items()) + sorted_subwords[: <span class="hljs-number">300</span> - <span class="hljs-built_in">len</span>(char_freqs)]
token_freqs = {token: freq <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs}<!-- HTML_TAG_END --></pre></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1fbawyj">💡 SentencePiece использует более эффективный алгоритм под названием Enhanced Suffix Array (ESA) для создания начального словаря.</p></div> <p data-svelte-h="svelte-1qzj9gd">Далее мы вычисляем сумму всех частот, чтобы преобразовать частоты в вероятности. Для нашей модели мы будем хранить логарифмы вероятностей, потому что численно стабильнее складывать логарифмы, чем перемножать маленькие числа, и это упростит вычисление потерь модели:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> math <span class="hljs-keyword">import</span> log
total_sum = <span class="hljs-built_in">sum</span>([freq <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs.items()])
model = {token: -log(freq / total_sum) <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs.items()}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-73tw6p">Теперь основная функция - это функция токенизации слов с помощью алгоритма Витерби. Как мы уже видели, этот алгоритм вычисляет наилучшую сегментацию каждой подстроки слова, которую мы будем хранить в переменной с именем <code>best_segmentations</code>. Мы будем хранить по одному словарю на каждую позицию в слове (от 0 до его полной длины), с двумя ключами: индекс начала последнего токена в лучшей сегментации и оценка лучшей сегментации. По индексу начала последнего токена мы сможем получить полную сегментацию, когда список будет полностью заполнен.</p> <p data-svelte-h="svelte-1f89qpo">Пополнение списка осуществляется с помощью двух циклов: основной цикл просматривает каждую начальную позицию, а второй цикл перебирает все подстроки, начинающиеся с этой начальной позиции. Если подстрока есть в словаре, мы получаем новую сегментацию слова до этой конечной позиции, которую сравниваем с той, что хранится в <code>best_segmentations</code>.</p> <p data-svelte-h="svelte-1gujlqm">После завершения основного цикла мы просто начинаем с конца и переходим от одной начальной позиции к другой, записывая токены по мере продвижения, пока не достигнем начала слова:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">encode_word</span>(<span class="hljs-params">word, model</span>):
 best_segmentations = [{<span class="hljs-string">&quot;start&quot;</span>: <span class="hljs-number">0</span>, <span class="hljs-string">&quot;score&quot;</span>: <span class="hljs-number">1</span>}] + [
 {<span class="hljs-string">&quot;start&quot;</span>: <span class="hljs-literal">None</span>, <span class="hljs-string">&quot;score&quot;</span>: <span class="hljs-literal">None</span>} <span class="hljs-keyword">for</span> _ <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">len</span>(word))
 ]
 <span class="hljs-keyword">for</span> start_idx <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">len</span>(word)):
 <span class="hljs-comment"># Это должно быть правильно заполнено предыдущими шагами цикла</span>
 best_score_at_start = best_segmentations[start_idx][<span class="hljs-string">&quot;score&quot;</span>]
 <span class="hljs-keyword">for</span> end_idx <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(start_idx + <span class="hljs-number">1</span>, <span class="hljs-built_in">len</span>(word) + <span class="hljs-number">1</span>):
 token = word[start_idx:end_idx]
 <span class="hljs-keyword">if</span> token <span class="hljs-keyword">in</span> model <span class="hljs-keyword">and</span> best_score_at_start <span class="hljs-keyword">is</span> <span class="hljs-keyword">not</span> <span class="hljs-literal">None</span>:
 score = model[token] + best_score_at_start
 <span class="hljs-comment"># Если мы нашли лучшую сегментацию, заканчивающуюся на end_idx, мы обновляем</span>
 <span class="hljs-keyword">if</span> (
 best_segmentations[end_idx][<span class="hljs-string">&quot;score&quot;</span>] <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span>
 <span class="hljs-keyword">or</span> best_segmentations[end_idx][<span class="hljs-string">&quot;score&quot;</span>] &gt; score
 ):
 best_segmentations[end_idx] = {<span class="hljs-string">&quot;start&quot;</span>: start_idx, <span class="hljs-string">&quot;score&quot;</span>: score}
segmentation = best_segmentations[-<span class="hljs-number">1</span>]
 <span class="hljs-keyword">if</span> segmentation[<span class="hljs-string">&quot;score&quot;</span>] <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span>:
 <span class="hljs-comment"># Мы не нашли токенизацию слова -&gt; возвращаем unknown</span>
 <span class="hljs-keyword">return</span> [<span class="hljs-string">&quot;&lt;unk&gt;&quot;</span>], <span class="hljs-literal">None</span>
score = segmentation[<span class="hljs-string">&quot;score&quot;</span>]
 start = segmentation[<span class="hljs-string">&quot;start&quot;</span>]
 end = <span class="hljs-built_in">len</span>(word)
 tokens = []
 <span class="hljs-keyword">while</span> start != <span class="hljs-number">0</span>:
 tokens.insert(<span class="hljs-number">0</span>, word[start:end])
 next_start = best_segmentations[start][<span class="hljs-string">&quot;start&quot;</span>]
 end = start
 start = next_start
 tokens.insert(<span class="hljs-number">0</span>, word[start:end])
 <span class="hljs-keyword">return</span> tokens, score<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1mshvqh">Мы уже можем опробовать нашу первоначальную модель на некоторых словах:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-built_in">print</span>(encode_word(<span class="hljs-string">&quot;Hopefully&quot;</span>, model))
<span class="hljs-built_in">print</span>(encode_word(<span class="hljs-string">&quot;This&quot;</span>, model))<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->([<span class="hljs-string">&#x27;H&#x27;</span>, <span class="hljs-string">&#x27;o&#x27;</span>, <span class="hljs-string">&#x27;p&#x27;</span>, <span class="hljs-string">&#x27;e&#x27;</span>, <span class="hljs-string">&#x27;f&#x27;</span>, <span class="hljs-string">&#x27;u&#x27;</span>, <span class="hljs-string">&#x27;ll&#x27;</span>, <span class="hljs-string">&#x27;y&#x27;</span>], <span class="hljs-number">41.5157494601402</span>)
([<span class="hljs-string">&#x27;This&#x27;</span>], <span class="hljs-number">6.288267030694535</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-obvrxs">Теперь легко вычислить потери модели на 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 --><span class="hljs-keyword">def</span> <span class="hljs-title function_">compute_loss</span>(<span class="hljs-params">model</span>):
 loss = <span class="hljs-number">0</span>
 <span class="hljs-keyword">for</span> word, freq <span class="hljs-keyword">in</span> word_freqs.items():
 _, word_loss = encode_word(word, model)
 loss += freq * word_loss
 <span class="hljs-keyword">return</span> loss<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1pk5f76">Мы можем проверить его работу на имеющейся у нас модели:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->compute_loss(model)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-number">413.10377642940875</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1qh2n66">Вычисление оценок для каждого токена также не представляет особой сложности; нам просто нужно вычислить потери для модели, полученные при удалении каждого токена:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> copy
<span class="hljs-keyword">def</span> <span class="hljs-title function_">compute_scores</span>(<span class="hljs-params">model</span>):
 scores = {}
 model_loss = compute_loss(model)
 <span class="hljs-keyword">for</span> token, score <span class="hljs-keyword">in</span> model.items():
 <span class="hljs-comment"># Мы всегда храним токены длиной 1</span>
 <span class="hljs-keyword">if</span> <span class="hljs-built_in">len</span>(token) == <span class="hljs-number">1</span>:
 <span class="hljs-keyword">continue</span>
 model_without_token = copy.deepcopy(model)
 _ = model_without_token.pop(token)
 scores[token] = compute_loss(model_without_token) - model_loss
 <span class="hljs-keyword">return</span> scores<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-12dq90h">Мы можем попробовать это на заданном токене:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->scores = compute_scores(model)
<span class="hljs-built_in">print</span>(scores[<span class="hljs-string">&quot;ll&quot;</span>])
<span class="hljs-built_in">print</span>(scores[<span class="hljs-string">&quot;his&quot;</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-pcm8dz">Поскольку <code>&quot;ll&quot;</code> используется в токенизации слова <code>&quot;Hopefully&quot;</code>, и его удаление, вероятно, заставит нас дважды использовать токен <code>&quot;l&quot;</code> вместо этого, мы ожидаем, что он будет иметь положительную потерю. <code>&quot;his&quot;</code> используется только внутри слова <code>&quot;This&quot;</code>, которое токенизируется само по себе, поэтому мы ожидаем, что потери будут нулевыми. Вот результаты:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-number">6.376412403623874</span>
<span class="hljs-number">0.0</span><!-- HTML_TAG_END --></pre></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-ozraz9">💡 Такой подход очень неэффективен, поэтому SentencePiece использует приближенную оценку потерь модели без токена X: вместо того чтобы начинать с нуля, он просто заменяет токен X его сегментацией в оставшемся словаре. Таким образом, все оценки могут быть вычислены одновременно с потерями модели.</p></div> <p data-svelte-h="svelte-1cc5cfv">Когда этот процесс завершиться, останется только добавить в словарь специальные токены, используемые моделью, а затем итерироваться, пока мы не вычеркнем из словаря достаточно токенов, чтобы достичь желаемого размера:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->percent_to_remove = <span class="hljs-number">0.1</span>
<span class="hljs-keyword">while</span> <span class="hljs-built_in">len</span>(model) &gt; <span class="hljs-number">100</span>:
 scores = compute_scores(model)
 sorted_scores = <span class="hljs-built_in">sorted</span>(scores.items(), key=<span class="hljs-keyword">lambda</span> x: x[<span class="hljs-number">1</span>])
 <span class="hljs-comment"># Удалите токены percent_to_remove с наименьшими оценками</span>
 <span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">int</span>(<span class="hljs-built_in">len</span>(model) * percent_to_remove)):
 _ = token_freqs.pop(sorted_scores[i][<span class="hljs-number">0</span>])
total_sum = <span class="hljs-built_in">sum</span>([freq <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs.items()])
 model = {token: -log(freq / total_sum) <span class="hljs-keyword">for</span> token, freq <span class="hljs-keyword">in</span> token_freqs.items()}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-147fp2n">Затем, чтобы токенизировать некоторый текст, нам просто нужно применить предварительную токенизацию, а затем использовать нашу функцию <code>encode_word()</code>:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">tokenize</span>(<span class="hljs-params">text, model</span>):
 words_with_offsets = tokenizer.backend_tokenizer.pre_tokenizer.pre_tokenize_str(text)
 pre_tokenized_text = [word <span class="hljs-keyword">for</span> word, offset <span class="hljs-keyword">in</span> words_with_offsets]
 encoded_words = [encode_word(word, model)[<span class="hljs-number">0</span>] <span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> pre_tokenized_text]
 <span class="hljs-keyword">return</span> <span class="hljs-built_in">sum</span>(encoded_words, [])
tokenize(<span class="hljs-string">&quot;This is the Hugging Face course.&quot;</span>, model)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[<span class="hljs-string">&#x27;▁This&#x27;</span>, <span class="hljs-string">&#x27;▁is&#x27;</span>, <span class="hljs-string">&#x27;▁the&#x27;</span>, <span class="hljs-string">&#x27;▁Hugging&#x27;</span>, <span class="hljs-string">&#x27;▁Face&#x27;</span>, <span class="hljs-string">&#x27;▁&#x27;</span>, <span class="hljs-string">&#x27;c&#x27;</span>, <span class="hljs-string">&#x27;ou&#x27;</span>, <span class="hljs-string">&#x27;r&#x27;</span>, <span class="hljs-string">&#x27;s&#x27;</span>, <span class="hljs-string">&#x27;e&#x27;</span>, <span class="hljs-string">&#x27;.&#x27;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-uhnvvo">Вот и все об Unigram! Надеемся, теперь вы чувствуете себя экспертом во всем, что касается токенизаторов. В следующем разделе мы рассмотрим блоки библиотеки 🤗 Tokenizers и покажем, как их можно использовать для создания собственного токенизатора.</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/chapter6/7.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, 49],
 data,
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
106 kB
·
Xet hash:
82692f789b13330bcb7e01b1197b9a831f6ade78dcabf12e521b8c85979fbeef

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