Buckets:

hf-doc-build
/

doc-dev

Files

xet

hf-doc-build/doc-dev / transformers /main /pt /serialization.html

rtrm

25 days ago

download

raw

71.2 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Exportando modelos para ONNX","local":"exportando-modelos-para-onnx","sections":[{"title":"Exportando um modelo para ONNX","local":"exportando-um-modelo-para-onnx","sections":[],"depth":2},{"title":"Selecionando features para diferentes tarefas do modelo","local":"selecionando-features-para-diferentes-tarefas-do-modelo","sections":[],"depth":2},{"title":"Exportando um modelo para uma arquitetura sem suporte","local":"exportando-um-modelo-para-uma-arquitetura-sem-suporte","sections":[{"title":"Implementando uma configuração ONNX personalizada","local":"implementando-uma-configuração-onnx-personalizada","sections":[],"depth":3},{"title":"Exportando um modelo","local":"exportando-um-modelo","sections":[],"depth":3},{"title":"Validando a saída dos modelos","local":"validando-a-saída-dos-modelos","sections":[],"depth":3}],"depth":2},{"title":"Contribuindo com uma nova configuração para 🤗 Transformers","local":"contribuindo-com-uma-nova-configuração-para--transformers","sections":[],"depth":2}],"depth":1}">
	<link href="/docs/transformers/main/pt/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/entry/start.3b004bd3.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/scheduler.d586627e.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/singletons.0d7813b8.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/index.b0d7e1f3.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/paths.183a9cf3.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/entry/app.1f3cbd74.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/index.8589a59c.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/nodes/0.30b661ad.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/each.e59479a4.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/nodes/13.e2134bfa.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/Tip.84e2336e.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/CodeBlock.47c46d2c.js">
	<link rel="modulepreload" href="/docs/transformers/main/pt/_app/immutable/chunks/EditOnGithub.073dfa26.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"Exportando modelos para ONNX","local":"exportando-modelos-para-onnx","sections":[{"title":"Exportando um modelo para ONNX","local":"exportando-um-modelo-para-onnx","sections":[],"depth":2},{"title":"Selecionando features para diferentes tarefas do modelo","local":"selecionando-features-para-diferentes-tarefas-do-modelo","sections":[],"depth":2},{"title":"Exportando um modelo para uma arquitetura sem suporte","local":"exportando-um-modelo-para-uma-arquitetura-sem-suporte","sections":[{"title":"Implementando uma configuração ONNX personalizada","local":"implementando-uma-configuração-onnx-personalizada","sections":[],"depth":3},{"title":"Exportando um modelo","local":"exportando-um-modelo","sections":[],"depth":3},{"title":"Validando a saída dos modelos","local":"validando-a-saída-dos-modelos","sections":[],"depth":3}],"depth":2},{"title":"Contribuindo com uma nova configuração para 🤗 Transformers","local":"contribuindo-com-uma-nova-configuração-para--transformers","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="exportando-modelos-para-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="#exportando-modelos-para-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>Exportando modelos para ONNX</span></h1> <p data-svelte-h="svelte-780rag">Se você precisar implantar modelos 🤗 Transformers em ambientes de produção, recomendamos
	exporta-los para um formato serializado que pode ser carregado e executado em
	tempos de execução e hardware. Neste guia, mostraremos como exportar modelos 🤗 Transformers
	para <a href="http://onnx.ai" rel="nofollow">ONNX (Open Neural Network eXchange)</a>.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-17ryloz">Uma vez exportado, um modelo pode ser otimizado para inferência por meio de técnicas como
	quantização e poda. Se você estiver interessado em otimizar seus modelos para serem executados com
	máxima eficiência, confira a biblioteca <a href="https://github.com/huggingface/optimum" rel="nofollow">🤗 Optimum
	</a>.</p></div> <p data-svelte-h="svelte-kzaz5b">ONNX é um padrão aberto que define um conjunto comum de operadores e um formato de arquivo comum
	para representar modelos de aprendizado profundo em uma ampla variedade de estruturas, incluindo PyTorch e
	TensorFlow. Quando um modelo é exportado para o formato ONNX, esses operadores são usados para
	construir um grafo computacional (muitas vezes chamado de <em>representação intermediária</em>) que
	representa o fluxo de dados através da rede neural.</p> <p data-svelte-h="svelte-187fdkg">Ao expor um grafo com operadores e tipos de dados padronizados, o ONNX facilita a
	alternar entre os frameworks. Por exemplo, um modelo treinado em PyTorch pode ser exportado para
	formato ONNX e depois importado no TensorFlow (e vice-versa).</p> <p data-svelte-h="svelte-18aznvk">🤗 Transformers fornece um pacote <a href="main_classes/onnx"><code>transformers.onnx</code></a> que permite
	que você converta os checkpoints do modelo em um grafo ONNX aproveitando os objetos de configuração.
	Esses objetos de configuração vêm prontos para várias arquiteturas de modelo e são
	projetado para ser facilmente extensível a outras arquiteturas.</p> <p data-svelte-h="svelte-g9dkld">As configurações prontas incluem as seguintes arquiteturas:</p> <ul data-svelte-h="svelte-hlbcru"><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>Conditional DETR</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>ERNIE</li> <li>FlauBERT</li> <li>GPT Neo</li> <li>GPT-J</li> <li>GroupViT</li> <li>I-BERT</li> <li>LayoutLM</li> <li>LayoutLMv3</li> <li>LeViT</li> <li>Longformer</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>OWL-ViT</li> <li>Perceiver</li> <li>PLBart</li> <li>ResNet</li> <li>RoBERTa</li> <li>RoFormer</li> <li>SegFormer</li> <li>SqueezeBERT</li> <li>Swin Transformer</li> <li>T5</li> <li>Table Transformer</li> <li>Vision Encoder decoder</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-1um2ugk">Nas próximas duas seções, mostraremos como:</p> <ul data-svelte-h="svelte-8ygt46"><li>Exportar um modelo suportado usando o pacote <code>transformers.onnx</code>.</li> <li>Exportar um modelo personalizado para uma arquitetura sem suporte.</li></ul> <h2 class="relative group"><a id="exportando-um-modelo-para-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="#exportando-um-modelo-para-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>Exportando um modelo para ONNX</span></h2> <p data-svelte-h="svelte-ldfupz">Para exportar um modelo 🤗 Transformers para o ONNX, primeiro você precisa instalar algumas
	dependências extras:</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-unqqru">O pacote <code>transformers.onnx</code> pode então ser usado como um módulo 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 tolerance when validating the model.<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-6uusn">A exportação de um checkpoint usando uma configuração pronta pode ser feita da seguinte forma:</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-1urlk3h">Você deve ver os seguintes logs:</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-15ck6bq">Isso exporta um grafo ONNX do ponto de verificação definido pelo argumento <code>--model</code>. Nisso
	Por exemplo, é <code>distilbert/distilbert-base-uncased</code>, mas pode ser qualquer checkpoint no Hugging
	Face Hub ou um armazenado localmente.</p> <p data-svelte-h="svelte-11ias89">O arquivo <code>model.onnx</code> resultante pode ser executado em um dos <a href="https://onnx.ai/supported-tools.html#deployModel" rel="nofollow">muitos
	aceleradores</a> que suportam o ONNX
	padrão. Por exemplo, podemos carregar e executar o modelo com <a href="https://onnxruntime.ai/" rel="nofollow">ONNX
	Tempo de execução</a> da seguinte forma:</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-ehimya">Os nomes de saída necessários (como <code>["last_hidden_state"]</code>) podem ser obtidos pegando uma
	configuração ONNX de cada modelo. Por exemplo, para DistilBERT temos:</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>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1iginlu">O processo é idêntico para os checkpoints do TensorFlow no Hub. Por exemplo, podemos
	exportar um checkpoint TensorFlow puro do <a href="https://huggingface.co/keras-io" rel="nofollow">Keras</a> da seguinte forma:</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-12w4y7b">Para exportar um modelo armazenado localmente, você precisará ter os pesos e
	arquivos tokenizer armazenados em um diretório. Por exemplo, podemos carregar e salvar um checkpoint como:</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-1hwlhij">Uma vez que o checkpoint é salvo, podemos exportá-lo para o ONNX apontando o <code>--model</code>
	argumento do pacote <code>transformers.onnx</code> para o diretório desejado:</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> <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, TFAutoModelForSequenceClassification

	<span class="hljs-meta">>>> </span><span class="hljs-comment"># Load tokenizer and TensorFlow weights from 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>tf_model = TFAutoModelForSequenceClassification.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-tf-checkpoint"</span>)
	<span class="hljs-meta">>>> </span>tf_model.save_pretrained(<span class="hljs-string">"local-tf-checkpoint"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1hwlhij">Uma vez que o checkpoint é salvo, podemos exportá-lo para o ONNX apontando o <code>--model</code>
	argumento do pacote <code>transformers.onnx</code> para o diretório desejado:</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-tf-checkpoint onnx/<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="selecionando-features-para-diferentes-tarefas-do-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="#selecionando-features-para-diferentes-tarefas-do-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>Selecionando features para diferentes tarefas do modelo</span></h2> <p data-svelte-h="svelte-au9t2">Cada configuração pronta vem com um conjunto de <em>features</em> que permitem exportar
	modelos para diferentes tipos de tarefas. Conforme mostrado na tabela abaixo, cada recurso é
	associado a uma <code>AutoClass</code> 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-67rlav">Para cada configuração, você pode encontrar a lista de recursos suportados por meio do
	<code>FeaturesManager</code>. Por exemplo, para DistilBERT temos:</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-zfw7vl">Você pode então passar um desses recursos para o argumento <code>--feature</code> no
	pacote <code>transformers.onnx</code>. Por exemplo, para exportar um modelo de classificação de texto, podemos
	escolher um modelo ajustado no Hub e executar:</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-1qh1xkx">Isso exibe os seguintes logs:</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-1sxlp9n">Observe que, neste caso, os nomes de saída do modelo ajustado são <code>logits</code>
	em vez do <code>last_hidden_state</code> que vimos com o checkpoint <code>distilbert/distilbert-base-uncased</code>
	mais cedo. Isso é esperado, pois o modelo ajustado (fine-tuned) possui uma cabeça de classificação de sequência.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-frt6wh">Os recursos que têm um sufixo <code>with-pass</code> (como <code>causal-lm-with-pass</code>) correspondem a
	classes de modelo com estados ocultos pré-computados (chave e valores nos blocos de atenção)
	que pode ser usado para decodificação autorregressiva rápida.</p></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-7aqol8">Para modelos do tipo <code>VisionEncoderDecoder</code>, as partes do codificador e do decodificador são
	exportados separadamente como dois arquivos ONNX chamados <code>encoder_model.onnx</code> e <code>decoder_model.onnx</code> respectivamente.</p></div> <h2 class="relative group"><a id="exportando-um-modelo-para-uma-arquitetura-sem-suporte" 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="#exportando-um-modelo-para-uma-arquitetura-sem-suporte"><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>Exportando um modelo para uma arquitetura sem suporte</span></h2> <p data-svelte-h="svelte-srdmc5">Se você deseja exportar um modelo cuja arquitetura não é suportada nativamente pela
	biblioteca, há três etapas principais a seguir:</p> <ol data-svelte-h="svelte-119emik"><li>Implemente uma configuração ONNX personalizada.</li> <li>Exporte o modelo para o ONNX.</li> <li>Valide as saídas do PyTorch e dos modelos exportados.</li></ol> <p data-svelte-h="svelte-1tntr09">Nesta seção, veremos como o DistilBERT foi implementado para mostrar o que está envolvido
	em cada passo.</p> <h3 class="relative group"><a id="implementando-uma-configuração-onnx-personalizada" 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="#implementando-uma-configuração-onnx-personalizada"><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>Implementando uma configuração ONNX personalizada</span></h3> <p data-svelte-h="svelte-1yke088">Vamos começar com o objeto de configuração ONNX. Fornecemos três classes abstratas que
	você deve herdar, dependendo do tipo de arquitetura de modelo que deseja exportar:</p> <ul data-svelte-h="svelte-16srizl"><li>Modelos baseados em codificador herdam de <code>OnnxConfig</code></li> <li>Modelos baseados em decodificador herdam de <code>OnnxConfigWithPast</code></li> <li>Os modelos codificador-decodificador herdam de <code>OnnxSeq2SeqConfigWithPast</code></li></ul> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-10qekhj">Uma boa maneira de implementar uma configuração ONNX personalizada é observar as
	implementação no arquivo <code>configuration_<model_name>.py</code> de uma arquitetura semelhante.</p></div> <p data-svelte-h="svelte-1w3ieqz">Como o DistilBERT é um modelo baseado em codificador, sua configuração é herdada 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-1dcv64l">Todo objeto de configuração deve implementar a propriedade <code>inputs</code> e retornar um mapeamento,
	onde cada chave corresponde a uma entrada esperada e cada valor indica o eixo
	dessa entrada. Para o DistilBERT, podemos ver que duas entradas são necessárias: <code>input_ids</code> e
	<code>attention_mask</code>. Essas entradas têm a mesma forma de <code>(batch_size, sequence_length)</code>
	é por isso que vemos os mesmos eixos usados na configuração.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-14shdz8">Notice that <code>inputs</code> property for <code>DistilBertOnnxConfig</code> returns an <code>OrderedDict</code>. This
	ensures that the inputs are matched with their relative position within the
	<code>PreTrainedModel.forward()</code> method when tracing the graph. We recommend using an
	<code>OrderedDict</code> for the <code>inputs</code> and <code>outputs</code> properties when implementing custom ONNX
	configurations.</p> <p data-svelte-h="svelte-ts6haw">Observe que a propriedade <code>inputs</code> para <code>DistilBertOnnxConfig</code> retorna um <code>OrderedDict</code>. Este
	garante que as entradas sejam combinadas com sua posição relativa dentro do
	método <code>PreTrainedModel.forward()</code> ao traçar o grafo. Recomendamos o uso de um
	<code>OrderedDict</code> para as propriedades <code>inputs</code> e <code>outputs</code> ao implementar configurações personalizadas ONNX.</p></div> <p data-svelte-h="svelte-1xs5c3l">Depois de implementar uma configuração ONNX, você pode instanciá-la fornecendo a
	configuração do modelo base da seguinte forma:</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-conxji">O objeto resultante tem várias propriedades úteis. Por exemplo, você pode visualizar o conjunto de operadores ONNX
	que será usado durante a exportação:</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-1gla57i">Você também pode visualizar as saídas associadas ao modelo da seguinte forma:</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-eqvgnl">Observe que a propriedade outputs segue a mesma estrutura das entradas; ele retorna um
	<code>OrderedDict</code> de saídas nomeadas e suas formas. A estrutura de saída está ligada a
	escolha do recurso com o qual a configuração é inicializada. Por padrão, a configuração do ONNX
	é inicializada com o recurso <code>default</code> que corresponde à exportação de um
	modelo carregado com a classe <code>AutoModel</code>. Se você deseja exportar um modelo para outra tarefa,
	apenas forneça um recurso diferente para o argumento <code>task</code> quando você inicializar a configuração ONNX
	. Por exemplo, se quisermos exportar o DistilBERT com uma sequência
	de classificação, poderíamos 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> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-js6ttx">Todas as propriedades e métodos básicos associados a <code>OnnxConfig</code> e
	as outras classes de configuração podem ser substituídas se necessário. Confira <code>BartOnnxConfig</code>
	para um exemplo avançado.</p></div> <h3 class="relative group"><a id="exportando-um-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="#exportando-um-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>Exportando um modelo</span></h3> <p data-svelte-h="svelte-scf0rn">Depois de ter implementado a configuração do ONNX, o próximo passo é exportar o modelo.
	Aqui podemos usar a função <code>export()</code> fornecida pelo pacote <code>transformers.onnx</code>.
	Esta função espera a configuração do ONNX, juntamente com o modelo base e o tokenizer,
	e o caminho para salvar o arquivo 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-1clqbx4">Os <code>onnx_inputs</code> e <code>onnx_outputs</code> retornados pela função <code>export()</code> são listas de
	chaves definidas nas propriedades <code>inputs</code> e <code>outputs</code> da configuração. Uma vez que o
	modelo é exportado, você pode testar se o modelo está bem formado da seguinte forma:</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> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1iua1gz">Se o seu modelo for maior que 2GB, você verá que muitos arquivos adicionais são criados
	durante a exportação. Isso é <em>esperado</em> porque o ONNX usa <a href="https://developers.google.com/protocol-buffers/" rel="nofollow">Protocol
	Buffers</a> para armazenar o modelo e estes
	têm um limite de tamanho de 2GB. Veja a <a href="https://github.com/onnx/onnx/blob/master/docs/ExternalData.md" rel="nofollow">ONNX
	documentação</a> para
	instruções sobre como carregar modelos com dados externos.</p></div> <h3 class="relative group"><a id="validando-a-saída-dos-modelos" 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="#validando-a-saída-dos-modelos"><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>Validando a saída dos modelos</span></h3> <p data-svelte-h="svelte-1fom1t1">A etapa final é validar se as saídas do modelo base e exportado concordam
	dentro de alguma tolerância absoluta. Aqui podemos usar a função <code>validate_model_outputs()</code>
	fornecida pelo pacote <code>transformers.onnx</code> da seguinte forma:</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-1w48q5s">Esta função usa o método <code>generate_dummy_inputs()</code> para
	gerar entradas para o modelo base e o exportado, e a tolerância absoluta pode ser
	definida na configuração. Geralmente encontramos concordância numérica em 1e-6 a 1e-4
	de alcance, embora qualquer coisa menor que 1e-3 provavelmente esteja OK.</p> <h2 class="relative group"><a id="contribuindo-com-uma-nova-configuração-para--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="#contribuindo-com-uma-nova-configuração-para--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>Contribuindo com uma nova configuração para 🤗 Transformers</span></h2> <p data-svelte-h="svelte-nik27b">Estamos procurando expandir o conjunto de configurações prontas e receber contribuições
	da comunidade! Se você gostaria de contribuir para a biblioteca, você
	precisará:</p> <ul data-svelte-h="svelte-aqybbn"><li>Implemente a configuração do ONNX no arquivo <code>configuration_<model_name>.py</code> correspondente
	Arquivo</li> <li>Incluir a arquitetura do modelo e recursos correspondentes em
	<code>~onnx.features.FeatureManager</code></li> <li>Adicione sua arquitetura de modelo aos testes em <code>test_onnx_v2.py</code></li></ul> <p data-svelte-h="svelte-1uvyz3f">Confira como ficou a configuração do <a href="https://github.com/huggingface/transformers/pull/14868/files" rel="nofollow">IBERT</a> para obter uma
	idéia do que está envolvido.</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/pt/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_1jqvrjl = {
	assets: "/docs/transformers/main/pt",
	base: "/docs/transformers/main/pt",
	env: {}
	};

	const element = document.currentScript.parentElement;

	const data = [null,null];

	Promise.all([
	import("/docs/transformers/main/pt/_app/immutable/entry/start.3b004bd3.js"),
	import("/docs/transformers/main/pt/_app/immutable/entry/app.1f3cbd74.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 13],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 71.2 kB
Xet hash:: 4f05e5e9e2dba8958770c6a6a8b92b1bf410dd6179827e419bfb0a99db35232f

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