Buckets:

hf-doc-build
/

doc

Files

xet

hf-doc-build/doc / transformers /main /es /serialization.html

rtrm

about 1 month ago

download

raw

103 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Exportar modelos 🤗 Transformers","local":"exportar-modelos--transformers","sections":[{"title":"ONNX","local":"onnx","sections":[{"title":"Exportar un model a ONNX","local":"exportar-un-model-a-onnx","sections":[],"depth":3},{"title":"Seleccionar características para diferentes topologías de un modelo","local":"seleccionar-características-para-diferentes-topologías-de-un-modelo","sections":[],"depth":3},{"title":"Exportar un modelo para una arquitectura no compatible","local":"exportar-un-modelo-para-una-arquitectura-no-compatible","sections":[{"title":"Implementar una configuración personalizada en ONNX","local":"implementar-una-configuración-personalizada-en-onnx","sections":[],"depth":4},{"title":"Exportar el modelo","local":"exportar-el-modelo","sections":[],"depth":4},{"title":"Validar los resultados del modelo","local":"validar-los-resultados-del-modelo","sections":[],"depth":4}],"depth":3},{"title":"Contribuir con una nueva configuración a 🤗 Transformers","local":"contribuir-con-una-nueva-configuración-a--transformers","sections":[],"depth":3}],"depth":2},{"title":"TorchScript","local":"torchscript","sections":[{"title":"Indicador de TorchScript y pesos atados","local":"indicador-de-torchscript-y-pesos-atados","sections":[],"depth":3},{"title":"Entradas ficticias y longitudes estándar","local":"entradas-ficticias-y-longitudes-estándar","sections":[],"depth":3},{"title":"Usar TorchScript en Python","local":"usar-torchscript-en-python","sections":[{"title":"Guardando un modelo","local":"guardando-un-modelo","sections":[],"depth":4},{"title":"Cargar un modelo","local":"cargar-un-modelo","sections":[],"depth":4},{"title":"Usar un modelo rastreado para la inferencia","local":"usar-un-modelo-rastreado-para-la-inferencia","sections":[],"depth":4}],"depth":3},{"title":"Implementar los modelos HuggingFace TorchScript en AWS mediante Neuron SDK","local":"implementar-los-modelos-huggingface-torchscript-en-aws-mediante-neuron-sdk","sections":[{"title":"Implicaciones","local":"implicaciones","sections":[],"depth":4},{"title":"Dependencias","local":"dependencias","sections":[],"depth":4},{"title":"Convertir un modelo a AWS Neuron","local":"convertir-un-modelo-a-aws-neuron","sections":[],"depth":4}],"depth":3}],"depth":2}],"depth":1}">
	<link href="/docs/transformers/main/es/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/entry/start.c5d32e04.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/scheduler.402ed373.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/singletons.da2848c5.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/index.feadd4f6.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/paths.f62bcf27.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/entry/app.9352f273.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/preload-helper.dd42f084.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/index.1afa24b4.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/nodes/0.81c32647.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/each.e59479a4.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/nodes/30.2f59a90e.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/Tip.6d309f99.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/CodeBlock.b62addad.js">
	<link rel="modulepreload" href="/docs/transformers/main/es/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.f34e66e1.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"Exportar modelos 🤗 Transformers","local":"exportar-modelos--transformers","sections":[{"title":"ONNX","local":"onnx","sections":[{"title":"Exportar un model a ONNX","local":"exportar-un-model-a-onnx","sections":[],"depth":3},{"title":"Seleccionar características para diferentes topologías de un modelo","local":"seleccionar-características-para-diferentes-topologías-de-un-modelo","sections":[],"depth":3},{"title":"Exportar un modelo para una arquitectura no compatible","local":"exportar-un-modelo-para-una-arquitectura-no-compatible","sections":[{"title":"Implementar una configuración personalizada en ONNX","local":"implementar-una-configuración-personalizada-en-onnx","sections":[],"depth":4},{"title":"Exportar el modelo","local":"exportar-el-modelo","sections":[],"depth":4},{"title":"Validar los resultados del modelo","local":"validar-los-resultados-del-modelo","sections":[],"depth":4}],"depth":3},{"title":"Contribuir con una nueva configuración a 🤗 Transformers","local":"contribuir-con-una-nueva-configuración-a--transformers","sections":[],"depth":3}],"depth":2},{"title":"TorchScript","local":"torchscript","sections":[{"title":"Indicador de TorchScript y pesos atados","local":"indicador-de-torchscript-y-pesos-atados","sections":[],"depth":3},{"title":"Entradas ficticias y longitudes estándar","local":"entradas-ficticias-y-longitudes-estándar","sections":[],"depth":3},{"title":"Usar TorchScript en Python","local":"usar-torchscript-en-python","sections":[{"title":"Guardando un modelo","local":"guardando-un-modelo","sections":[],"depth":4},{"title":"Cargar un modelo","local":"cargar-un-modelo","sections":[],"depth":4},{"title":"Usar un modelo rastreado para la inferencia","local":"usar-un-modelo-rastreado-para-la-inferencia","sections":[],"depth":4}],"depth":3},{"title":"Implementar los modelos HuggingFace TorchScript en AWS mediante Neuron SDK","local":"implementar-los-modelos-huggingface-torchscript-en-aws-mediante-neuron-sdk","sections":[{"title":"Implicaciones","local":"implicaciones","sections":[],"depth":4},{"title":"Dependencias","local":"dependencias","sections":[],"depth":4},{"title":"Convertir un modelo a AWS Neuron","local":"convertir-un-modelo-a-aws-neuron","sections":[],"depth":4}],"depth":3}],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="exportar-modelos--transformers" 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="#exportar-modelos--transformers"><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>Exportar modelos 🤗 Transformers</span></h1> <p data-svelte-h="svelte-1fx3vq8">Si necesitas implementar modelos 🤗 Transformers en entornos de producción, te
	recomendamos exportarlos a un formato serializado que se pueda cargar y ejecutar
	en tiempos de ejecución y hardware especializados. En esta guía, te mostraremos cómo
	exportar modelos 🤗 Transformers en dos formatos ampliamente utilizados: ONNX y TorchScript.</p> <p data-svelte-h="svelte-v0pjtx">Una vez exportado, un modelo puede optimizarse para la inferencia a través de técnicas
	como la cuantización y <em>pruning</em>. Si estás interesado en optimizar tus modelos para
	que funcionen con la máxima eficiencia, consulta la
	<a href="https://github.com/huggingface/optimum" rel="nofollow">biblioteca de 🤗 Optimum</a>.</p> <h2 class="relative group"><a id="onnx" 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="#onnx"><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>ONNX</span></h2> <p data-svelte-h="svelte-8kqvys">El proyecto <a href="http://onnx.ai" rel="nofollow">ONNX (Open Neural Network eXchange)</a> es un
	estándar abierto que define un conjunto común de operadores y un formato
	de archivo común para representar modelos de aprendizaje profundo en una
	amplia variedad de <em>frameworks</em>, incluidos PyTorch y TensorFlow. Cuando un modelo
	se exporta al formato ONNX, estos operadores se usan para construir un
	grafo computacional (a menudo llamado <em>representación intermedia</em>) que
	representa el flujo de datos a través de la red neuronal.</p> <p data-svelte-h="svelte-7p5plo">Al exponer un grafo con operadores y tipos de datos estandarizados, ONNX facilita
	el cambio entre frameworks. Por ejemplo, un modelo entrenado en PyTorch se puede
	exportar a formato ONNX y luego importar en TensorFlow (y viceversa).</p> <p data-svelte-h="svelte-rxyuqc">🤗 Transformers proporciona un paquete llamado <code>transformers.onnx</code>, el cual permite convertir
	los checkpoints de un modelo en un grafo ONNX aprovechando los objetos de configuración.
	Estos objetos de configuración están hechos a la medida de diferentes arquitecturas de modelos
	y están diseñados para ser fácilmente extensibles a otras arquitecturas.</p> <p data-svelte-h="svelte-b7160f">Las configuraciones a la medida incluyen las siguientes arquitecturas:</p> <ul data-svelte-h="svelte-vz4lhh"><li>ALBERT</li> <li>BART</li> <li>BEiT</li> <li>BERT</li> <li>BigBird</li> <li>BigBird-Pegasus</li> <li>Blenderbot</li> <li>BlenderbotSmall</li> <li>BLOOM</li> <li>CamemBERT</li> <li>CLIP</li> <li>CodeGen</li> <li>ConvBERT</li> <li>ConvNeXT</li> <li>ConvNeXTV2</li> <li>Data2VecText</li> <li>Data2VecVision</li> <li>DeBERTa</li> <li>DeBERTa-v2</li> <li>DeiT</li> <li>DETR</li> <li>DistilBERT</li> <li>ELECTRA</li> <li>FlauBERT</li> <li>GPT Neo</li> <li>GPT-J</li> <li>I-BERT</li> <li>LayoutLM</li> <li>LayoutLMv3</li> <li>LeViT</li> <li>LongT5</li> <li>M2M100</li> <li>Marian</li> <li>mBART</li> <li>MobileBERT</li> <li>MobileViT</li> <li>MT5</li> <li>OpenAI GPT-2</li> <li>Perceiver</li> <li>PLBart</li> <li>ResNet</li> <li>RoBERTa</li> <li>RoFormer</li> <li>SqueezeBERT</li> <li>T5</li> <li>ViT</li> <li>XLM</li> <li>XLM-RoBERTa</li> <li>XLM-RoBERTa-XL</li> <li>YOLOS</li></ul> <p data-svelte-h="svelte-194migv">En las próximas dos secciones, te mostraremos cómo:</p> <ul data-svelte-h="svelte-1223amv"><li>Exportar un modelo compatible utilizando el paquete <code>transformers.onnx</code>.</li> <li>Exportar un modelo personalizado para una arquitectura no compatible.</li></ul> <h3 class="relative group"><a id="exportar-un-model-a-onnx" 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="#exportar-un-model-a-onnx"><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>Exportar un model a ONNX</span></h3> <p data-svelte-h="svelte-1rezqc2">Para exportar un modelo 🤗 Transformers a ONNX, tienes que instalar primero algunas
	dependencias extra:</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 -->pip install transformers[onnx]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-5qhi33">El paquete <code>transformers.onnx</code> puede ser usado luego como un módulo de Python:</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 -->python -m transformers.onnx --<span class="hljs-built_in">help</span>

	usage: Hugging Face Transformers ONNX exporter [-h] -m MODEL [--feature {causal-lm, ...}] [--opset OPSET] [--atol ATOL] output

	positional arguments:
	output Path indicating <span class="hljs-built_in">where</span> to store generated ONNX model.

	optional arguments:
	-h, --<span class="hljs-built_in">help</span> show this <span class="hljs-built_in">help</span> message and <span class="hljs-built_in">exit</span>
	-m MODEL, --model MODEL
	Model ID on huggingface.co or path on disk to load model from.
	--feature {causal-lm, ...}
	The <span class="hljs-built_in">type</span> of features to <span class="hljs-built_in">export</span> the model with.
	--opset OPSET ONNX opset version to <span class="hljs-built_in">export</span> the model with.
	--atol ATOL Absolute difference tolerence when validating the model.<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1xn293m">Exportar un checkpoint usando una configuración a la medida se puede hacer de la siguiente manera:</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 -->python -m transformers.onnx --model=distilbert/distilbert-base-uncased onnx/<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ekqtp">que debería mostrar los siguientes registros:</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 -->Validating ONNX model...
	-[✓] ONNX model output names match reference model ({<span class="hljs-string">'last_hidden_state'</span>})
	- Validating ONNX Model output <span class="hljs-string">"last_hidden_state"</span>:
	-[✓] (2, 8, 768) matches (2, 8, 768)
	-[✓] all values close (atol: 1e-05)
	All good, model saved at: onnx/model.onnx<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-olsb8t">Esto exporta un grafo ONNX del checkpoint definido por el argumento <code>--model</code>.
	En este ejemplo, es un modelo <code>distilbert/distilbert-base-uncased</code>, pero puede ser cualquier
	checkpoint en Hugging Face Hub o que esté almacenado localmente.</p> <p data-svelte-h="svelte-tcdasa">El archivo <code>model.onnx</code> resultante se puede ejecutar en uno de los
	<a href="https://onnx.ai/supported-tools.html#deployModel" rel="nofollow">muchos aceleradores</a>
	que admiten el estándar ONNX. Por ejemplo, podemos cargar y ejecutar el
	modelo con <a href="https://onnxruntime.ai/" rel="nofollow">ONNX Runtime</a> de la siguiente manera:</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-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer
	<span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> onnxruntime <span class="hljs-keyword">import</span> InferenceSession

	<span class="hljs-meta">>>> </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">"distilbert/distilbert-base-uncased"</span>)
	<span class="hljs-meta">>>> </span>session = InferenceSession(<span class="hljs-string">"onnx/model.onnx"</span>)
	<span class="hljs-meta">>>> </span><span class="hljs-comment"># ONNX Runtime expects NumPy arrays as input</span>
	<span class="hljs-meta">>>> </span>inputs = tokenizer(<span class="hljs-string">"Using DistilBERT with ONNX Runtime!"</span>, return_tensors=<span class="hljs-string">"np"</span>)
	<span class="hljs-meta">>>> </span>outputs = session.run(output_names=[<span class="hljs-string">"last_hidden_state"</span>], input_feed=<span class="hljs-built_in">dict</span>(inputs))<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1hbp2km">Los nombres necesarios de salida (es decir, <code>["last_hidden_state"]</code>) se pueden obtener
	echando un vistazo a la configuración ONNX de cada modelo. Por ejemplo, para DistilBERT tenemos:</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-meta">>>> </span><span class="hljs-keyword">from</span> transformers.models.distilbert <span class="hljs-keyword">import</span> DistilBertConfig, DistilBertOnnxConfig

	<span class="hljs-meta">>>> </span>config = DistilBertConfig()
	<span class="hljs-meta">>>> </span>onnx_config = DistilBertOnnxConfig(config)
	<span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(<span class="hljs-built_in">list</span>(onnx_config.outputs.keys()))
	[<span class="hljs-string">"last_hidden_state"</span>]s<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1h30j8i">El proceso es idéntico para los checkpoints de TensorFlow en Hub.
	Por ejemplo, podemos exportar un checkpoint puro de TensorFlow desde
	<a href="https://huggingface.co/keras-io" rel="nofollow">Keras</a> de la siguiente manera:</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 -->python -m transformers.onnx --model=keras-io/transformers-qa onnx/<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1xwvftw">Para exportar un modelo que está almacenado localmente, deberás tener los pesos
	y tokenizadores del modelo almacenados en un directorio. Por ejemplo, podemos cargar
	y guardar un checkpoint de la siguiente manera:</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-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, AutoModelForSequenceClassification

	<span class="hljs-meta">>>> </span><span class="hljs-comment"># Load tokenizer and PyTorch weights form the Hub</span>
	<span class="hljs-meta">>>> </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">"distilbert/distilbert-base-uncased"</span>)
	<span class="hljs-meta">>>> </span>pt_model = AutoModelForSequenceClassification.from_pretrained(<span class="hljs-string">"distilbert/distilbert-base-uncased"</span>)
	<span class="hljs-meta">>>> </span><span class="hljs-comment"># Save to disk</span>
	<span class="hljs-meta">>>> </span>tokenizer.save_pretrained(<span class="hljs-string">"local-pt-checkpoint"</span>)
	<span class="hljs-meta">>>> </span>pt_model.save_pretrained(<span class="hljs-string">"local-pt-checkpoint"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-c1bn59">Una vez que se guarda el checkpoint, podemos exportarlo a ONNX usando el argumento <code>--model</code>
	del paquete <code>transformers.onnx</code> al directorio deseado:</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 -->python -m transformers.onnx --model=local-pt-checkpoint onnx/<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="seleccionar-características-para-diferentes-topologías-de-un-modelo" 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="#seleccionar-características-para-diferentes-topologías-de-un-modelo"><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>Seleccionar características para diferentes topologías de un modelo</span></h3> <p data-svelte-h="svelte-p07qxg">Cada configuración a la medida viene con un conjunto de <em>características</em> que te permiten exportar
	modelos para diferentes tipos de topologías o tareas. Como se muestra en la siguiente tabla, cada
	función está asociada con una auto-clase de automóvil diferente:</p> <table data-svelte-h="svelte-19g4mid"><thead><tr><th>Feature</th> <th>Auto Class</th></tr></thead> <tbody><tr><td><code>causal-lm</code>, <code>causal-lm-with-past</code></td> <td><code>AutoModelForCausalLM</code></td></tr> <tr><td><code>default</code>, <code>default-with-past</code></td> <td><code>AutoModel</code></td></tr> <tr><td><code>masked-lm</code></td> <td><code>AutoModelForMaskedLM</code></td></tr> <tr><td><code>question-answering</code></td> <td><code>AutoModelForQuestionAnswering</code></td></tr> <tr><td><code>seq2seq-lm</code>, <code>seq2seq-lm-with-past</code></td> <td><code>AutoModelForSeq2SeqLM</code></td></tr> <tr><td><code>sequence-classification</code></td> <td><code>AutoModelForSequenceClassification</code></td></tr> <tr><td><code>token-classification</code></td> <td><code>AutoModelForTokenClassification</code></td></tr></tbody></table> <p data-svelte-h="svelte-z84cjd">Para cada configuración, puedes encontrar la lista de funciones admitidas a través de <code>FeaturesManager</code>.
	Por ejemplo, para DistilBERT tenemos:</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-meta">>>> </span><span class="hljs-keyword">from</span> transformers.onnx.features <span class="hljs-keyword">import</span> FeaturesManager

	<span class="hljs-meta">>>> </span>distilbert_features = <span class="hljs-built_in">list</span>(FeaturesManager.get_supported_features_for_model_type(<span class="hljs-string">"distilbert"</span>).keys())
	<span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(distilbert_features)
	[<span class="hljs-string">"default"</span>, <span class="hljs-string">"masked-lm"</span>, <span class="hljs-string">"causal-lm"</span>, <span class="hljs-string">"sequence-classification"</span>, <span class="hljs-string">"token-classification"</span>, <span class="hljs-string">"question-answering"</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1q0ce7p">Le puedes pasar una de estas características al argumento <code>--feature</code> en el paquete <code>transformers.onnx</code>.
	Por ejemplo, para exportar un modelo de clasificación de texto, podemos elegir un modelo ya ajustado del Hub y ejecutar:</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 -->python -m transformers.onnx --model=distilbert/distilbert-base-uncased-finetuned-sst-2-english \
	--feature=sequence-classification onnx/<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-4a0cb6">que mostrará los siguientes registros:</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 -->Validating ONNX model...
	-[✓] ONNX model output names match reference model ({<span class="hljs-string">'logits'</span>})
	- Validating ONNX Model output <span class="hljs-string">"logits"</span>:
	-[✓] (2, 2) matches (2, 2)
	-[✓] all values close (atol: 1e-05)
	All good, model saved at: onnx/model.onnx<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-4svxy9">Ten en cuenta que, en este caso, los nombres de salida del modelo ajustado son <code>logits</code> en lugar de <code>last_hidden_state</code>
	que vimos anteriormente con el checkpoint <code>distilbert/distilbert-base-uncased</code>. Esto es de esperarse ya que el modelo ajustado
	tiene un cabezal de clasificación secuencial.</p> <blockquote class="tip"><p data-svelte-h="svelte-1d2ttlc">Las características que tienen un sufijo ‘with-past’ (por ejemplo, ‘causal-lm-with-past’) corresponden a topologías
	de modelo con estados ocultos precalculados (clave y valores en los bloques de atención) que se pueden usar para una
	decodificación autorregresiva más rápida.</p></blockquote> <h3 class="relative group"><a id="exportar-un-modelo-para-una-arquitectura-no-compatible" 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="#exportar-un-modelo-para-una-arquitectura-no-compatible"><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>Exportar un modelo para una arquitectura no compatible</span></h3> <p data-svelte-h="svelte-6zjh54">Si deseas exportar un modelo cuya arquitectura no es compatible de forma nativa
	con la biblioteca, debes seguir tres pasos principales:</p> <ol data-svelte-h="svelte-utsrgp"><li>Implementa una configuración personalizada en ONNX.</li> <li>Exporta el modelo a ONNX.</li> <li>Valide los resultados de PyTorch y los modelos exportados.</li></ol> <p data-svelte-h="svelte-uz65rl">En esta sección, veremos cómo se implementó la serialización de DistilBERT
	para mostrar lo que implica cada paso.</p> <h4 class="relative group"><a id="implementar-una-configuración-personalizada-en-onnx" 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="#implementar-una-configuración-personalizada-en-onnx"><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>Implementar una configuración personalizada en ONNX</span></h4> <p data-svelte-h="svelte-1fjrnj0">Comencemos con el objeto de configuración de ONNX. Proporcionamos tres clases abstractas
	de las que debe heredar, según el tipo de arquitectura del modelo que quieras exportar:</p> <ul data-svelte-h="svelte-1oias2v"><li>Modelos basados en el <em>Encoder</em> inherente de <code>OnnxConfig</code></li> <li>Modelos basados en el <em>Decoder</em> inherente de <code>OnnxConfigWithPast</code></li> <li>Modelos <em>Encoder-decoder</em> inherente de <code>OnnxSeq2SeqConfigWithPast</code></li></ul> <blockquote class="tip"><p data-svelte-h="svelte-1nire10">Una buena manera de implementar una configuración personalizada en ONNX es observar la implementación
	existente en el archivo <code>configuration_<model_name>.py</code> de una arquitectura similar.</p></blockquote> <p data-svelte-h="svelte-j69ur5">Dado que DistilBERT es un modelo de tipo <em>encoder</em>, su configuración se hereda de <code>OnnxConfig</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-meta">>>> </span><span class="hljs-keyword">from</span> typing <span class="hljs-keyword">import</span> Mapping, OrderedDict
	<span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers.onnx <span class="hljs-keyword">import</span> OnnxConfig


	<span class="hljs-meta">>>> </span><span class="hljs-keyword">class</span> <span class="hljs-title class_">DistilBertOnnxConfig</span>(<span class="hljs-title class_ inherited__">OnnxConfig</span>):
	<span class="hljs-meta">... </span> @<span class="hljs-built_in">property</span>
	<span class="hljs-meta">... </span> <span class="hljs-keyword">def</span> <span class="hljs-title function_">inputs</span>(<span class="hljs-params">self</span>) -> Mapping[<span class="hljs-built_in">str</span>, Mapping[<span class="hljs-built_in">int</span>, <span class="hljs-built_in">str</span>]]:
	<span class="hljs-meta">... </span> <span class="hljs-keyword">return</span> OrderedDict(
	<span class="hljs-meta">... </span> [
	<span class="hljs-meta">... </span> (<span class="hljs-string">"input_ids"</span>, {<span class="hljs-number">0</span>: <span class="hljs-string">"batch"</span>, <span class="hljs-number">1</span>: <span class="hljs-string">"sequence"</span>}),
	<span class="hljs-meta">... </span> (<span class="hljs-string">"attention_mask"</span>, {<span class="hljs-number">0</span>: <span class="hljs-string">"batch"</span>, <span class="hljs-number">1</span>: <span class="hljs-string">"sequence"</span>}),
	<span class="hljs-meta">... </span> ]
	<span class="hljs-meta">... </span> )<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1b9fj7y">Cada objeto de configuración debe implementar la propiedad <code>inputs</code> y devolver un mapeo,
	donde cada llave corresponde a una entrada esperada y cada valor indica el eje de esa entrada.
	Para DistilBERT, podemos ver que se requieren dos entradas: <code>input_ids</code> y <code>attention_mask</code>.
	Estas entradas tienen la misma forma de <code>(batch_size, sequence_length)</code>, es por lo que vemos
	los mismos ejes utilizados en la configuración.</p> <blockquote class="tip"><p data-svelte-h="svelte-1tas9p7">Observa que la propiedad <code>inputs</code> para <code>DistilBertOnnxConfig</code> devuelve un <code>OrderedDict</code>.
	Esto nos asegura que las entradas coincidan con su posición relativa dentro del método
	<code>PreTrainedModel.forward()</code> al rastrear el grafo. Recomendamos usar un <code>OrderedDict</code>
	para las propiedades <code>inputs</code> y <code>outputs</code> al implementar configuraciones ONNX personalizadas.</p></blockquote> <p data-svelte-h="svelte-19vxzwm">Una vez que hayas implementado una configuración ONNX, puedes crear una
	instancia proporcionando la configuración del modelo base de la siguiente manera:</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-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoConfig

	<span class="hljs-meta">>>> </span>config = AutoConfig.from_pretrained(<span class="hljs-string">"distilbert/distilbert-base-uncased"</span>)
	<span class="hljs-meta">>>> </span>onnx_config = DistilBertOnnxConfig(config)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-2qbtlm">El objeto resultante tiene varias propiedades útiles. Por ejemplo, puedes ver el conjunto de operadores ONNX que se
	utilizará durante la exportación:</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-meta">>>> </span><span class="hljs-built_in">print</span>(onnx_config.default_onnx_opset)
	<span class="hljs-number">11</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ahh71r">También puedes ver los resultados asociados con el modelo de la siguiente manera:</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-meta">>>> </span><span class="hljs-built_in">print</span>(onnx_config.outputs)
	OrderedDict([(<span class="hljs-string">"last_hidden_state"</span>, {<span class="hljs-number">0</span>: <span class="hljs-string">"batch"</span>, <span class="hljs-number">1</span>: <span class="hljs-string">"sequence"</span>})])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-nurpeo">Observa que la propiedad de salidas sigue la misma estructura que las entradas;
	devuelve un objecto <code>OrderedDict</code> de salidas nombradas y sus formas. La estructura
	de salida está vinculada a la elección de la función con la que se inicializa la configuración.
	Por defecto, la configuración de ONNX se inicializa con la función <code>default</code> que
	corresponde a exportar un modelo cargado con la clase <code>AutoModel</code>. Si quieres exportar
	una topología de modelo diferente, simplemente proporciona una característica diferente
	al argumento <code>task</code> cuando inicialices la configuración de ONNX. Por ejemplo, si quisiéramos
	exportar DistilBERT con un cabezal de clasificación de secuencias, podríamos usar:</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-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoConfig

	<span class="hljs-meta">>>> </span>config = AutoConfig.from_pretrained(<span class="hljs-string">"distilbert/distilbert-base-uncased"</span>)
	<span class="hljs-meta">>>> </span>onnx_config_for_seq_clf = DistilBertOnnxConfig(config, task=<span class="hljs-string">"sequence-classification"</span>)
	<span class="hljs-meta">>>> </span><span class="hljs-built_in">print</span>(onnx_config_for_seq_clf.outputs)
	OrderedDict([(<span class="hljs-string">'logits'</span>, {<span class="hljs-number">0</span>: <span class="hljs-string">'batch'</span>})])<!-- HTML_TAG_END --></pre></div> <blockquote class="tip"><p data-svelte-h="svelte-ggia3t">Todas las propiedades base y métodos asociados con <code>OnnxConfig</code> y las
	otras clases de configuración se pueden sobreescribir si es necesario.
	Consulte <code>BartOnnxConfig</code> para ver un ejemplo avanzado.</p></blockquote> <h4 class="relative group"><a id="exportar-el-modelo" 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="#exportar-el-modelo"><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>Exportar el modelo</span></h4> <p data-svelte-h="svelte-1e9iy9f">Una vez que hayas implementado la configuración de ONNX, el siguiente paso es exportar el modelo.
	Aquí podemos usar la función <code>export()</code> proporcionada por el paquete <code>transformers.onnx</code>.
	Esta función espera la configuración de ONNX, junto con el modelo base y el tokenizador,
	y la ruta para guardar el archivo exportado:</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-meta">>>> </span><span class="hljs-keyword">from</span> pathlib <span class="hljs-keyword">import</span> Path
	<span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers.onnx <span class="hljs-keyword">import</span> export
	<span class="hljs-meta">>>> </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, AutoModel

	<span class="hljs-meta">>>> </span>onnx_path = Path(<span class="hljs-string">"model.onnx"</span>)
	<span class="hljs-meta">>>> </span>model_ckpt = <span class="hljs-string">"distilbert/distilbert-base-uncased"</span>
	<span class="hljs-meta">>>> </span>base_model = AutoModel.from_pretrained(model_ckpt)
	<span class="hljs-meta">>>> </span>tokenizer = AutoTokenizer.from_pretrained(model_ckpt)

	<span class="hljs-meta">>>> </span>onnx_inputs, onnx_outputs = export(tokenizer, base_model, onnx_config, onnx_config.default_onnx_opset, onnx_path)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-i9ugge">Los objetos <code>onnx_inputs</code> y <code>onnx_outputs</code> devueltos por la función <code>export()</code>
	son listas de llaves definidas en las propiedades <code>inputs</code> y <code>outputs</code> de la configuración.
	Una vez exportado el modelo, puedes probar que el modelo está bien formado de la siguiente manera:</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-meta">>>> </span><span class="hljs-keyword">import</span> onnx

	<span class="hljs-meta">>>> </span>onnx_model = onnx.load(<span class="hljs-string">"model.onnx"</span>)
	<span class="hljs-meta">>>> </span>onnx.checker.check_model(onnx_model)<!-- HTML_TAG_END --></pre></div> <blockquote class="tip"><p data-svelte-h="svelte-1ms6fox">Si tu modelo tiene más de 2GB, verás que se crean muchos archivos adicionales durante la exportación.
	Esto es <em>esperado</em> porque ONNX usa <a href="https://developers.google.com/protocol-buffers/" rel="nofollow">Búferes de protocolo</a>
	para almacenar el modelo y éstos tienen un límite de tamaño de 2 GB. Consulta la
	<a href="https://github.com/onnx/onnx/blob/master/docs/ExternalData.md" rel="nofollow">documentación de ONNX</a> para obtener
	instrucciones sobre cómo cargar modelos con datos externos.</p></blockquote> <h4 class="relative group"><a id="validar-los-resultados-del-modelo" 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="#validar-los-resultados-del-modelo"><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>Validar los resultados del modelo</span></h4> <p data-svelte-h="svelte-1cws23w">El paso final es validar que los resultados del modelo base y exportado coincidan dentro
	de cierta tolerancia absoluta. Aquí podemos usar la función <code>validate_model_outputs()</code>
	proporcionada por el paquete <code>transformers.onnx</code> de la siguiente manera:</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-meta">>>> </span><span class="hljs-keyword">from</span> transformers.onnx <span class="hljs-keyword">import</span> validate_model_outputs

	<span class="hljs-meta">>>> </span>validate_model_outputs(
	<span class="hljs-meta">... </span> onnx_config, tokenizer, base_model, onnx_path, onnx_outputs, onnx_config.atol_for_validation
	<span class="hljs-meta">... </span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-e8b61v">Esta función usa el método <code>OnnxConfig.generate_dummy_inputs()</code> para generar entradas para el modelo base
	y exportado, y la tolerancia absoluta se puede definir en la configuración. En general, encontramos una
	concordancia numérica en el rango de 1e-6 a 1e-4, aunque es probable que cualquier valor menor que 1e-3 esté bien.</p> <h3 class="relative group"><a id="contribuir-con-una-nueva-configuración-a--transformers" 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="#contribuir-con-una-nueva-configuración-a--transformers"><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>Contribuir con una nueva configuración a 🤗 Transformers</span></h3> <p data-svelte-h="svelte-gmawhv">¡Estamos buscando expandir el conjunto de configuraciones a la medida para usar y agradecemos las contribuciones de la comunidad!
	Si deseas contribuir con su colaboración a la biblioteca, deberás:</p> <ul data-svelte-h="svelte-1l017x9"><li>Implementa la configuración de ONNX en el archivo <code>configuration_<model_name>.py</code> correspondiente</li> <li>Incluye la arquitectura del modelo y las características correspondientes en <code>~onnx.features.FeatureManager</code></li> <li>Agrega tu arquitectura de modelo a las pruebas en <code>test_onnx_v2.py</code></li></ul> <p data-svelte-h="svelte-5jmim9">Revisa cómo fue la contribución para la <a href="https://github.com/huggingface/transformers/pull/14868/files" rel="nofollow">configuración de IBERT</a>
	y así tener una idea de lo que necesito.</p> <h2 class="relative group"><a id="torchscript" 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="#torchscript"><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>TorchScript</span></h2> <blockquote class="tip"><p data-svelte-h="svelte-1u0h9mk">Este es el comienzo de nuestros experimentos con TorchScript y todavía estamos explorando sus capacidades con modelos de
	tamaño de entrada variable. Es un tema de interés y profundizaremos nuestro análisis en las próximas
	versiones, con más ejemplos de código, una implementación más flexible y puntos de referencia que comparen códigos
	basados en Python con TorchScript compilado.</p></blockquote> <p data-svelte-h="svelte-jnclu5">Según la documentación de PyTorch: “TorchScript es una forma de crear modelos serializables y optimizables a partir del
	código de PyTorch”. Los dos módulos de Pytorch <a href="https://pytorch.org/docs/stable/jit.html" rel="nofollow">JIT y TRACE</a> permiten al
	desarrollador exportar su modelo para reutilizarlo en otros programas, como los programas C++ orientados a la eficiencia.</p> <p data-svelte-h="svelte-1tbohqs">Hemos proporcionado una interfaz que permite exportar modelos de 🤗 Transformers a TorchScript para que puedan reutilizarse
	en un entorno diferente al de un programa Python basado en PyTorch. Aquí explicamos cómo exportar y usar nuestros modelos
	usando TorchScript.</p> <p data-svelte-h="svelte-14y3y0y">Exportar un modelo requiere de dos cosas:</p> <ul data-svelte-h="svelte-1pnqwfl"><li>un pase hacia adelante con entradas ficticias.</li> <li>instanciación del modelo con la indicador <code>torchscript</code>.</li></ul> <p data-svelte-h="svelte-1xnblp0">Estas necesidades implican varias cosas con las que los desarrolladores deben tener cuidado. Éstas se detallan a continuación.</p> <h3 class="relative group"><a id="indicador-de-torchscript-y-pesos-atados" 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="#indicador-de-torchscript-y-pesos-atados"><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>Indicador de TorchScript y pesos atados</span></h3> <p data-svelte-h="svelte-oteso6">Este indicador es necesario porque la mayoría de los modelos de lenguaje en este repositorio tienen pesos vinculados entre su capa
	de <code>Embedding</code> y su capa de <code>Decoding</code>. TorchScript no permite la exportación de modelos que tengan pesos atados, por lo que es
	necesario desvincular y clonar los pesos previamente.</p> <p data-svelte-h="svelte-16q5iky">Esto implica que los modelos instanciados con el indicador <code>torchscript</code> tienen su capa <code>Embedding</code> y <code>Decoding</code> separadas,
	lo que significa que no deben entrenarse más adelante. El entrenamiento desincronizaría las dos capas, lo que generaría
	resultados inesperados.</p> <p data-svelte-h="svelte-1z0i5sf">Este no es el caso de los modelos que no tienen un cabezal de modelo de lenguaje, ya que no tienen pesos atados.
	Estos modelos se pueden exportar de forma segura sin el indicador <code>torchscript</code>.</p> <h3 class="relative group"><a id="entradas-ficticias-y-longitudes-estándar" 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="#entradas-ficticias-y-longitudes-estándar"><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>Entradas ficticias y longitudes estándar</span></h3> <p data-svelte-h="svelte-1g52tin">Las entradas ficticias se utilizan para crear un modelo de pase hacia adelante. Mientras los valores de las entradas se
	propagan a través de las capas, PyTorch realiza un seguimiento de las diferentes operaciones ejecutadas en cada tensor.
	Estas operaciones registradas se utilizan luego para crear el “rastro” del modelo.</p> <p data-svelte-h="svelte-1k2z473">El rastro se crea en relación con las dimensiones de las entradas. Por lo tanto, está limitado por las dimensiones de la
	entrada ficticia y no funcionará para ninguna otra longitud de secuencia o tamaño de lote. Al intentar con un tamaño diferente,
	un error como:</p> <p data-svelte-h="svelte-w9fo8c"><code>The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2</code></p> <p data-svelte-h="svelte-d0488i">aparecerá. Por lo tanto, se recomienda rastrear el modelo con un tamaño de entrada ficticia al menos tan grande como la
	entrada más grande que se alimentará al modelo durante la inferencia. El <em>padding</em> se puede realizar para completar los
	valores que faltan. Sin embargo, como el modelo se habrá rastreado con un tamaño de entrada grande, las dimensiones de
	las diferentes matrices también serán grandes, lo que dará como resultado más cálculos.</p> <p data-svelte-h="svelte-1udnc8x">Se recomienda tener cuidado con el número total de operaciones realizadas en cada entrada y seguir de cerca el rendimiento
	al exportar modelos de longitud de secuencia variable.</p> <h3 class="relative group"><a id="usar-torchscript-en-python" 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="#usar-torchscript-en-python"><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>Usar TorchScript en Python</span></h3> <p data-svelte-h="svelte-z2sk6s">A continuación se muestra un ejemplo que muestra cómo guardar, cargar modelos y cómo usar el rastreo para la inferencia.</p> <h4 class="relative group"><a id="guardando-un-modelo" 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="#guardando-un-modelo"><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>Guardando un modelo</span></h4> <p data-svelte-h="svelte-f2qgin">Este fragmento muestra cómo usar TorchScript para exportar un <code>BertModel</code>. Aquí, el <code>BertModel</code> se instancia de acuerdo
	con la clase <code>BertConfig</code> y luego se guarda en el disco con el nombre de archivo <code>traced_bert.pt</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> BertModel, BertTokenizer, BertConfig
	<span class="hljs-keyword">import</span> torch

	enc = BertTokenizer.from_pretrained(<span class="hljs-string">"google-bert/bert-base-uncased"</span>)

	<span class="hljs-comment"># Tokenizing input text</span>
	text = <span class="hljs-string">"[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"</span>
	tokenized_text = enc.tokenize(text)

	<span class="hljs-comment"># Masking one of the input tokens</span>
	masked_index = <span class="hljs-number">8</span>
	tokenized_text[masked_index] = <span class="hljs-string">"[MASK]"</span>
	indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
	segments_ids = [<span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">0</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>, <span class="hljs-number">1</span>]

	<span class="hljs-comment"># Creating a dummy input</span>
	tokens_tensor = torch.tensor([indexed_tokens])
	segments_tensors = torch.tensor([segments_ids])
	dummy_input = [tokens_tensor, segments_tensors]

	<span class="hljs-comment"># Initializing the model with the torchscript flag</span>
	<span class="hljs-comment"># Flag set to True even though it is not necessary as this model does not have an LM Head.</span>
	config = BertConfig(
	vocab_size_or_config_json_file=<span class="hljs-number">32000</span>,
	hidden_size=<span class="hljs-number">768</span>,
	num_hidden_layers=<span class="hljs-number">12</span>,
	num_attention_heads=<span class="hljs-number">12</span>,
	intermediate_size=<span class="hljs-number">3072</span>,
	torchscript=<span class="hljs-literal">True</span>,
	)

	<span class="hljs-comment"># Instantiating the model</span>
	model = BertModel(config)

	<span class="hljs-comment"># The model needs to be in evaluation mode</span>
	model.<span class="hljs-built_in">eval</span>()

	<span class="hljs-comment"># If you are instantiating the model with from_pretrained you can also easily set the TorchScript flag</span>
	model = BertModel.from_pretrained(<span class="hljs-string">"google-bert/bert-base-uncased"</span>, torchscript=<span class="hljs-literal">True</span>)

	<span class="hljs-comment"># Creating the trace</span>
	traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
	torch.jit.save(traced_model, <span class="hljs-string">"traced_bert.pt"</span>)<!-- HTML_TAG_END --></pre></div> <h4 class="relative group"><a id="cargar-un-modelo" 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="#cargar-un-modelo"><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>Cargar un modelo</span></h4> <p data-svelte-h="svelte-cu3t2j">Este fragmento muestra cómo cargar el <code>BertModel</code> que se guardó previamente en el disco con el nombre <code>traced_bert.pt</code>.
	Estamos reutilizando el <code>dummy_input</code> previamente inicializado.</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 -->loaded_model = torch.jit.load(<span class="hljs-string">"traced_bert.pt"</span>)
	loaded_model.<span class="hljs-built_in">eval</span>()

	all_encoder_layers, pooled_output = loaded_model(*dummy_input)<!-- HTML_TAG_END --></pre></div> <h4 class="relative group"><a id="usar-un-modelo-rastreado-para-la-inferencia" 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="#usar-un-modelo-rastreado-para-la-inferencia"><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>Usar un modelo rastreado para la inferencia</span></h4> <p data-svelte-h="svelte-j4u541">Usar el modelo rastreado para la inferencia es tan simple como usar su método <code>__call__</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 -->traced_model(tokens_tensor, segments_tensors)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="implementar-los-modelos-huggingface-torchscript-en-aws-mediante-neuron-sdk" 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="#implementar-los-modelos-huggingface-torchscript-en-aws-mediante-neuron-sdk"><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>Implementar los modelos HuggingFace TorchScript en AWS mediante Neuron SDK</span></h3> <p data-svelte-h="svelte-y13wnz">AWS presentó la familia de instancias <a href="https://aws.amazon.com/ec2/instance-types/inf1/" rel="nofollow">Amazon EC2 Inf1</a> para la inferencia
	de aprendizaje automático de bajo costo y alto rendimiento en la nube. Las instancias Inf1 funcionan con el chip AWS
	Inferentia, un acelerador de hardware personalizado, que se especializa en cargas de trabajo de inferencia de aprendizaje
	profundo. <a href="https://awsdocs-neuron.readthedocs-hosted.com/en/latest/#" rel="nofollow">AWS Neuron</a> es el kit de desarrollo para Inferentia
	que admite el rastreo y la optimización de modelos de transformers para su implementación en Inf1. El SDK de Neuron proporciona:</p> <ol data-svelte-h="svelte-j8267o"><li>API fácil de usar con una línea de cambio de código para rastrear y optimizar un modelo de TorchScript para la inferencia en la nube.</li> <li>Optimizaciones de rendimiento listas para usar con un <a href="https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/benchmark/%3E" rel="nofollow">costo-rendimiento mejorado</a></li> <li>Soporte para modelos HuggingFace Transformers construidos con <a href="https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/bert_tutorial/tutorial_pretrained_bert.html" rel="nofollow">PyTorch</a>
	o <a href="https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/tensorflow/huggingface_bert/huggingface_bert.html" rel="nofollow">TensorFlow</a>.</li></ol> <h4 class="relative group"><a id="implicaciones" 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="#implicaciones"><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>Implicaciones</span></h4> <p data-svelte-h="svelte-17gv1st">Los modelos Transformers basados en la arquitectura
	<a href="https://huggingface.co/docs/transformers/main/model_doc/bert" rel="nofollow">BERT (Representaciones de <em>Enconder</em> bidireccional de Transformers)</a>,
	o sus variantes, como <a href="https://huggingface.co/docs/transformers/main/model_doc/distilbert" rel="nofollow">distilBERT</a> y
	<a href="https://huggingface.co/docs/transformers/main/model_doc/roberta" rel="nofollow">roBERTa</a>, se ejecutarán mejor en Inf1 para tareas no
	generativas, como la respuesta extractiva de preguntas, la clasificación de secuencias y la clasificación de tokens.
	Como alternativa, las tareas de generación de texto se pueden adaptar para ejecutarse en Inf1, según este
	<a href="https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/transformers-marianmt.html" rel="nofollow">tutorial de AWS Neuron MarianMT</a>.
	Puedes encontrar más información sobre los modelos que están listos para usarse en Inferentia en la
	<a href="https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/models/models-inferentia.html#models-inferentia" rel="nofollow">sección <em>Model Architecture Fit</em> de la documentación de Neuron</a>.</p> <h4 class="relative group"><a id="dependencias" 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="#dependencias"><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>Dependencias</span></h4> <p data-svelte-h="svelte-xgquje">Usar AWS Neuron para convertir modelos requiere las siguientes dependencias y entornos:</p> <ul data-svelte-h="svelte-ctgjcu"><li>Un <a href="https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/neuron-frameworks/pytorch-neuron/index.html#installation-guide" rel="nofollow">entorno Neuron SDK</a>,
	que viene preconfigurado en <a href="https://docs.aws.amazon.com/dlami/latest/devguide/tutorial-inferentia-launching.html" rel="nofollow">AWS Deep Learning AMI</a>.</li></ul> <h4 class="relative group"><a id="convertir-un-modelo-a-aws-neuron" 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="#convertir-un-modelo-a-aws-neuron"><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>Convertir un modelo a AWS Neuron</span></h4> <p data-svelte-h="svelte-rrgk4r">Con el mismo script usado en <a href="https://huggingface.co/docs/transformers/main/es/serialization#using-torchscript-in-python" rel="nofollow">Uso de TorchScript en Python</a>
	para rastrear un “BertModel”, puedes importar la extensión del <em>framework</em> <code>torch.neuron</code> para acceder a los componentes
	del SDK de Neuron a través de una API de Python.</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> BertModel, BertTokenizer, BertConfig
	<span class="hljs-keyword">import</span> torch
	<span class="hljs-keyword">import</span> torch.neuron<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-17bbjt2">Y modificando la línea de código de rastreo de:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->torch.jit.trace(model, [tokens_tensor, segments_tensors])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-157z9d8">con lo siguiente:</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 -->torch.neuron.trace(model, [token_tensor, segments_tensors])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-nhaqqn">Este cambio permite a Neuron SDK rastrear el modelo y optimizarlo para ejecutarse en instancias Inf1.</p> <p data-svelte-h="svelte-rtq77g">Para obtener más información sobre las funciones, las herramientas, los tutoriales de ejemplo y las últimas actualizaciones
	de AWS Neuron SDK, consulte la <a href="https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html" rel="nofollow">documentación de AWS NeuronSDK</a>.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/transformers/blob/main/docs/source/es/serialization.md" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span data-svelte-h="svelte-x0xyl0">></span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>

	<script>
	{
	__sveltekit_1n12h89 = {
	assets: "/docs/transformers/main/es",
	base: "/docs/transformers/main/es",
	env: {}
	};

	const element = document.currentScript.parentElement;

	const data = [null,null];

	Promise.all([
	import("/docs/transformers/main/es/_app/immutable/entry/start.c5d32e04.js"),
	import("/docs/transformers/main/es/_app/immutable/entry/app.9352f273.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 30],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 103 kB
Xet hash:: 7351970195326ea9893bcd51b2c5cbfdc7c4c2479e4b162b20ee3b395cb6d007

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