Buckets:

hf-doc-build
/

doc-dev

Files

xet

hf-doc-build/doc-dev / transformers /pr_33956 /zh /attention.html

rtrm

about 2 months ago

download

raw

13.3 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"注意力机制","local":"注意力机制","sections":[{"title":"局部敏感哈希注意力机制（LSH attention）","local":"局部敏感哈希注意力机制lsh-attention","sections":[],"depth":2},{"title":"局部注意力机制（Local attention）","local":"局部注意力机制local-attention","sections":[],"depth":2},{"title":"其他技巧","local":"其他技巧","sections":[{"title":"轴向位置编码","local":"轴向位置编码","sections":[],"depth":3}],"depth":2}],"depth":1}">
	<link href="/docs/transformers/pr_33956/zh/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/entry/start.2075924f.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/chunks/scheduler.9991993c.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/chunks/singletons.de898f32.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/chunks/index.02cfeb18.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/chunks/paths.07a34ce3.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/entry/app.6dbd1b69.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/chunks/index.ed60ef0f.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/nodes/0.90024ce0.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/chunks/each.e59479a4.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/nodes/4.461896b4.js">
	<link rel="modulepreload" href="/docs/transformers/pr_33956/zh/_app/immutable/chunks/EditOnGithub.ba269039.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"注意力机制","local":"注意力机制","sections":[{"title":"局部敏感哈希注意力机制（LSH attention）","local":"局部敏感哈希注意力机制lsh-attention","sections":[],"depth":2},{"title":"局部注意力机制（Local attention）","local":"局部注意力机制local-attention","sections":[],"depth":2},{"title":"其他技巧","local":"其他技巧","sections":[{"title":"轴向位置编码","local":"轴向位置编码","sections":[],"depth":3}],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="注意力机制" 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="#注意力机制"><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>注意力机制</span></h1> <p data-svelte-h="svelte-1txc0zc">大多数 transformer 模型使用完全注意力机制，该机制采用正方形的注意力矩阵。当输入很长的文本时，这将导致巨大的计算瓶颈。Longformer 和 Reformer 是提高注意力机制效率的改进模型，它们使用稀疏化的注意力矩阵来加速训练。</p> <h2 class="relative group"><a id="局部敏感哈希注意力机制lsh-attention" 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="#局部敏感哈希注意力机制lsh-attention"><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>局部敏感哈希注意力机制（LSH attention）</span></h2> <p data-svelte-h="svelte-1lki4fd"><a href="model_doc/reformer">Reformer</a>使用LSH（局部敏感哈希）的注意力机制。在计算softmax(QK^t)时，只有矩阵QK^t中的最大元素（在softmax维度上）会做出有用的贡献。所以对于Q中的每个查询q，我们只需要考虑K中与q接近的键k，这里使用了一个哈希函数来确定q和k是否接近。注意力掩码被修改以掩盖当前的词符（token）（除了第一个位置之外），因为这样会使得查询和键相等（因此非常相似）。由于哈希可能会有些随机性，所以在实践中使用多个哈希函数（由n_rounds参数确定）,然后一起求平均。</p> <h2 class="relative group"><a id="局部注意力机制local-attention" 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="#局部注意力机制local-attention"><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>局部注意力机制（Local attention）</span></h2> <p data-svelte-h="svelte-1k6wuut"><a href="model_doc/longformer">Longformer</a>使用局部注意力机制:通常情况下，局部上下文（例如，左边和右边的两个词符是什么？）对于给定词符的操作已经足够了。此外，通过堆叠具有小窗口的注意力层，最后一层将拥有不仅仅是窗口内词符的感受野，这使得它们能构建整个句子的表示。</p> <p data-svelte-h="svelte-1prs570">一些预先选定的输入词符也被赋予全局注意力:对于这些少数词符，注意力矩阵可以访问所有词符（tokens），并且这个过程是对称的:所有其他词符除了它们局部窗口内的词符之外，也可以访问这些特定的词符。这在论文的图2d中有展示，下面是一个样本注意力掩码：</p> <div class="flex justify-center" data-svelte-h="svelte-d4kpls"><img scale="50 %" align="center" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/local_attention_mask.png"></div> <p data-svelte-h="svelte-fo7hp2">使用参数更少的注意力矩阵，可以让模型处理更长的输入序列。</p> <h2 class="relative group"><a id="其他技巧" 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="#其他技巧"><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>其他技巧</span></h2> <h3 class="relative group"><a id="轴向位置编码" 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="#轴向位置编码"><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>轴向位置编码</span></h3> <p><a href="model_doc/reformer" data-svelte-h="svelte-zf8r8v">Reformer</a>模型使用轴向位置编码：在传统的transformer模型中，位置编码矩阵E的大小是\(l\)乘以\(d\)，其中\(l\)是序列长度，\(d\)是隐藏状态的维度。如果你有非常长的文本，这个矩阵可能会非常大，将会占用大量的GPU显存。为了缓解这个问题，轴向位置编码将这个大矩阵E分解成两个较小的矩阵E1和E2，它们的维度分别是\(l<em data-svelte-h="svelte-a6jgeg">{1} \times d</em>{1}\) 和\(l<em data-svelte-h="svelte-1kso657">{2} \times d</em>{2}\)，满足\(l<em data-svelte-h="svelte-1u2dqkw">{1} \times l</em>{2} = l\)和\(d<em data-svelte-h="svelte-xo83ib">{1} + d</em>{2} = d\)（通过长度的乘积，最终得到的矩阵要小得多）。在E中，对于时间步\(j\) 的嵌入是通过连接E1中时间步<!-- HTML_TAG_START --><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>j</mi><mi mathvariant="normal">%</mi><mi>l</mi><mn>1</mn></mrow><annotation encoding="application/x-tex">j \% l1</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.9444em;vertical-align:-0.1944em;"></span><span class="mord mathnormal" style="margin-right:0.05724em;">j</span><span class="mord">%</span><span class="mord mathnormal" style="margin-right:0.01968em;">l</span><span class="mord">1</span></span></span></span><!-- HTML_TAG_END --> 的嵌入和E2中时间步\(j // l1\)的嵌入来获得的。</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/zh/attention.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_g15out = {
	assets: "/docs/transformers/pr_33956/zh",
	base: "/docs/transformers/pr_33956/zh",
	env: {}
	};

	const element = document.currentScript.parentElement;

	const data = [null,null];

	Promise.all([
	import("/docs/transformers/pr_33956/zh/_app/immutable/entry/start.2075924f.js"),
	import("/docs/transformers/pr_33956/zh/_app/immutable/entry/app.6dbd1b69.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 4],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 13.3 kB
Xet hash:: 1714eb39173ed1e0ec2f07ac6245f0bab6d5f61800f961995a4c46d55db3ec56

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