Attention Is All You Need

Ashish Vaswani∗ Google Brain avaswani@google.com
Noam Shazeer∗ Google Brain noam@google.com
Niki Parmar∗ Google Research nikip@google.com
Jakob Uszkoreit∗ Google Research usz@google.com
Llion Jones∗ Google Research llion@google.com
Aidan N. Gomez∗ † University of Toronto aidan@cs.toronto.edu
Łukasz Kaiser∗ Google Brain lukaszkaiser@google.com
Illia Polosukhin∗ ‡ illia.polosukhin@gmail.com

Abstract

The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.

主流的序列转换模型基于复杂的递归或卷积神经网络，其中包括一个编码器和一个解码器。性能最好的模型还通过注意力机制连接编码器和解码器。我们提出了一种新的简单网络架构—Transformer，它完全基于注意力机制，无需递归和卷积。在两项机器翻译任务上的实验表明，这些模型的质量更优，同时可并行化程度更高，所需的训练时间也大大减少。在 WMT 2014 英德翻译任务中，我们的模型达到了 28.4 BLEU，比包括集合在内的现有最佳结果提高了 2 BLEU 以上。在 WMT 2014 英法翻译任务中，我们的模型在 8 个 GPU 上进行了 3.5 天的训练后，取得了 41.8 分的单模型最新 BLEU 分数，这只是文献中最佳模型训练成本的一小部分。我们将 Transformer 成功地应用于英语选区解析，并同时使用大量和有限的训练数据，从而证明 Transformer 可以很好地推广到其他任务中。

1 Introduction

Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15].

递归神经网络，尤其是长短期记忆[13]和门控递归[7]神经网络，已被牢固确立为序列建模和转译问题（如语言建模和机器翻译）的最先进方法[35, 2, 5]。自此以后，许多人继续努力推动递归语言模型和编码器-解码器架构的发展[38, 24, 15]。

Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states $h_t$, as a function of the previous hidden state $h_{t-1}$ and the input for position $t$. This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. Recent work has achieved significant improvements in computational efficiency through factorization tricks [21] and conditional computation [32], while also improving model performance in case of the latter. The fundamental constraint of sequential computation, however, remains.

递归模型通常沿着输入和输出序列的符号位置进行计算。将位置与计算时间的步长对齐，它们会生成隐藏状态 $h_t$的序列，作为前一个隐藏状态$h_{t-1}$ 和位置 $t$的输入的函数。这种固有的序列性质排除了训练示例内的并行化，这在序列长度较长时变得至关重要，因为内存约束限制了跨示例的批处理。最近的研究通过因式分解技巧[21]和条件计算[32]显著提高了计算效率，同时也改善了后者的模型性能。然而，顺序计算的基本约束依然存在。

Attention mechanisms have become an integral part of compelling sequence modeling and transduction models in various tasks, allowing modeling of dependencies without regard to their distance in the input or output sequences [2, 19]. In all but a few cases [27], however, such attention mechanisms are used in conjunction with a recurrent network.

注意机制已成为各种任务中令人信服的序列建模和转导模型的一个组成部分，它允许对依赖关系进行建模，而无需考虑它们在输入或输出序列中的距离[2, 19]。然而，除少数情况 [27] 外，这种注意力机制都是与递归网络结合使用的。

In this work we propose the Transformer, a model architecture eschewing recurrence and instead relying entirely on an attention mechanism to draw global dependencies between input and output. The Transformer allows for significantly more parallelization and can reach a new state of the art in translation quality after being trained for as little as twelve hours on eight P100 GPUs.

在这项工作中，我们提出了Transformer这一模型架构，它摒弃了递归机制，而是完全依赖注意力机制来绘制输入和输出之间的全局依赖关系。Transformer 可大大提高并行化程度，在 8 个 P100 GPU 上只需短短 12 个小时的训练，就能在翻译质量方面达到新的技术水平。

2 Background

The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU [16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building block, computing hidden representations in parallel for all input and output positions. In these models, the number of operations required to relate signals from two arbitrary input or output positions grows in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes it more difficult to learn dependencies between distant positions [12]. In the Transformer this is reduced to a constant number of operations, albeit at the cost of reduced effective resolution due to averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as described in section 3.2.

减少顺序计算的目标也是扩展神经 GPU[16]、ByteNet[18]和 ConvS2S[9]的基础，它们都使用卷积神经网络作为基本构建模块，并行计算所有输入和输出位置的隐藏表示。在这些模型中，将来自两个任意输入或输出位置的信号联系起来所需的运算次数随位置间距离的增加而增加，ConvS2S 是线性增加，ByteNet 是对数增加。这就增加了学习远距离位置之间依赖关系的难度[12]。在 Transformer 中，这种情况会减少到恒定的操作次数，但代价是由于平均注意力加权位置而降低了有效分辨率，我们在 3.2 节中介绍了多头注意力来抵消这种影响。

Self-attention, sometimes called intra-attention is an attention mechanism relating different positions of a single sequence in order to compute a representation of the sequence. Self-attention has been used successfully in a variety of tasks including reading comprehension, abstractive summarization, textual entailment and learning task-independent sentence representations [4, 27, 28, 22].

自我注意（有时也称为内部注意）是一种将单个序列的不同位置联系起来以计算序列表示的注意机制。自我注意已成功应用于多种任务，包括阅读理解、抽象概括、文本引申和学习与任务无关的句子表征[4, 27, 28, 22]。

End-to-end memory networks are based on a recurrent attention mechanism instead of sequencealigned recurrence and have been shown to perform well on simple-language question answering and language modeling tasks [34].

端到端记忆网络基于递归注意机制，而不是序列对齐递归，在简单语言问题解答和语言建模任务中表现出色[34]。

To the best of our knowledge, however, the Transformer is the first transduction model relying entirely on self-attention to compute representations of its input and output without using sequencealigned RNNs or convolution. In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9].

不过，据我们所知，Transformer 是第一个完全依靠自我注意来计算输入和输出表示而不使用序列对齐 RNN 或卷积的Transduction模型。在下面的章节中，我们将描述 Transformer，激发自我注意，并讨论它与 [17, 18] 和 [9] 等模型相比的优势。

3 Model Architecture

Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35]. Here, the encoder maps an input sequence of symbol representations $(x_1,...,x_n)$ to a sequence of continuous representations $z=(z_1,...,z_n)$. Given $z$, the decoder then generates an output sequence $(y_1,...,y_m)$ of symbols one element at a time. At each step the model is auto-regressive [10], consuming the previously generated symbols as additional input when generating the next.

大多数竞争性神经序列转换模型都具有编码器-解码器结构 [5, 2, 35]。在这里，编码器将输入的符号表示序列 $(x_1,...,x_n)$映射到连续表示序列 $z=(z_1,...,z_n)$。给定 $z$，解码器每次生成一个元素的符号输出序列 $(y_1,...,y_m)$。在每一步中，模型都是自动递归的[10]，在生成下一步时，会消耗之前生成的符号作为额外输入。

The Transformer follows this overall architecture using stacked self-attention and point-wise, fully connected layers for both the encoder and decoder, shown in the left and right halves of Figure 1, respectively.

图 1 的左半部分和右半部分分别显示了编码器和解码器的堆叠自注意层和点向全连接层。

3.1 Encoder and Decoder Stacks

Encoder: The encoder is composed of a stack of $N=6$ identical layers. Each layer has two sub-layers. The first is a multi head self-attention mechanism, and the second is a simple, position wise fully connected feed-forward network. We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is $LayerNorm(x+Sublayer(x))$, where $Sublayer(x)$ is the function implemented by the sub-layer itself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension $d_{model}=512$.

编码器 编码器由 $N=6$ 层相同的层堆叠组成。每一层有两个子层。第一层是多头自注意机制，第二层是简单的位置全连接前馈网络。我们在两个子层的每个周围都采用了残差连接[11]，然后进行层归一化[1]。也就是说，每个子层的输出为$LayerNorm(x+Sublayer(x))$，其中$Sublayer(x)$为子层本身实现的函数。为了方便这些残差连接，模型中的所有子层以及嵌入层都会产生维数为 $d_{model}=512$ 的输出。

Decoder: The decoder is also composed of a stack of $N=6$ identical layers. In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization. We also modify the self-attention sub-layer in the decoder stack to prevent positions from attending to subsequent positions. This masking, combined with fact that the output embeddings are offset by one position, ensures that the predictions for position $i$ can depend only on the known outputs at positions less than $i$.

解码器： 解码器也是由一叠 $N=6$ 相同的层组成。除了每个编码器层的两个子层外，解码器还插入了第三个子层，对编码器层的输出进行多头关注。与编码器类似，我们在每个子层周围采用残差连接，然后进行层归一化。我们还修改了解码器堆栈中的自我关注子层，以防止位置关注到后续位置。这种屏蔽，加上输出嵌入偏移一个位置的事实，确保了对位置 $i$ 的预测只能依赖于小于 $i$ 位置的已知输出。

3.2 Attention

An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.

注意力函数可以描述为将一个查询和一组键值对映射到一个输出，其中查询、键、值和输出都是向量。输出是以值的加权和来计算的，其中分配给每个值的权重是通过查询与相应键的兼容函数来计算的。

3.2.1 Scaled Dot-Product Attention

We call our particular attention “Scaled Dot-Product Attention” (Figure 2). The input consists of queries and keys of dimension $d_k$, and values of dimension $d_v$. We compute the dot products of the query with all keys, divide each by $\sqrt{d_k}$, and apply a soft-max function to obtain the weights on the values.

我们将这种特殊的注意力称为 “缩放点积注意力”（图 2）。输入包括查询和维度为 $d_k$ 的键，以及维度为 $d_v$ 的值。我们计算查询与所有按键的点积，将每个点积除以 $\sqrt{d_k}$，然后应用软最大函数来获得值的权重。

In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix $Q$. The keys and values are also packed together into matrices $K$ and $V$ . We compute the matrix of outputs as:

在实践中，我们会同时计算一组查询的注意力函数，并将其打包成一个矩阵 $Q$。键和值也被打包成矩阵 $K$ 和 $V$ 。我们计算的输出矩阵为
$$Attention(Q,K,V)=softmax(\frac{Q{K}^T}{\sqrt{d_k}})V$$
The two most commonly used attention functions are additive attention [2], and dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor of $\frac{1}{\sqrt{d_k}}$. Additive attention computes the compatibility function using a feed-forward network with a single hidden layer. While the two are similar in theoretical complexity, dot-product attention is much faster and more space-efficient in practice, since it can be implemented using highly optimized matrix multiplication code.

最常用的两种注意力函数是加法注意力[2]和点积（乘法）注意力。点积注意力与我们的算法相同，只是缩放因子为 $\frac{1}{\sqrt{d_k}}$。加法注意使用单隐层前馈网络计算兼容性函数。虽然两者在理论复杂度上相似，但点积注意力在实践中速度更快，空间效率更高，因为它可以使用高度优化的矩阵乘法代码来实现。

While for small values of $d_k$ the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of $d_k$ [3]. We suspect that for large values of $d_k$, the dot products grow large in magnitude, pushing the soft-max function into regions where it has extremely small gradients 4. To counteract this effect, we scale the dot products by $\frac{1}{\sqrt{d_k}}$.

虽然对于较小的 $d_k$ 值，两种机制的表现类似，但对于较大的 $d_k$ 值，加法注意在不缩放的情况下优于点积注意[3]。我们怀疑，对于较大的 $d_k$ 值，点积的幅度会变大，从而将软-最大函数推向梯度极小的区域 4。为了消除这种影响，我们用 $\frac{1}{sqrt{d}_k}$ 来缩放点积。

3.2.2 Multi-Head Attention

Instead of performing a single attention function with $d_{model}$-dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values $h$ times with different, learned linear projections to $d_k$, $d_k$ and $d_v$ dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding $d_v$-dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.

我们发现，与其使用 $d_{model}$ 维度的键、值和查询执行单个注意力函数，不如将查询、键和值线性投影 $h$ 次，分别投影到 $d_k$、$d_k$ 和 $d_v$ 维度。然后，我们对每个查询、键和值的投影版本并行执行注意力函数，得到 $d_v$ 维的输出值。如图 2 所示，这些值被串联起来并再次投影，从而得到最终值。

Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this.

多头注意力允许模型在不同位置共同关注来自不同表征子空间的信息。而在单注意头的情况下，平均化会抑制这一点。
$$\begin{gathered} MultiHead(Q,K,V)=Concat(hea{d}_1,....,hea{d}_h)W^O, \\ wherehea{d}_i=Attention(Q{W}_i^Q,K{W}_i^K,V{W}_i^V) \end{gathered}$$
Where the projections are parameter matrices$W_i^Q\in R^{d_{model}\times d_k}$,$W_i^K\in R^{d_{model}\times d_k}$,$W_i^V\in R^{d_{model}\times d_v}$ and $W^O\in R^{h{d}_v\times d_{model}}$ .

In this work we employ $h=8$ parallel attention layers, or heads. For each of these we use $dk=dv=d_{model}/h=64$. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.

其中的投影是参数矩阵$W_i^Q\in R^{d_{model}\times d_k}$, $W_i^K\in R^{d_{model}\times d_k}$, $W_i^V\in R^{d_{model}\times d_v}$ 和 $W^O\in R^{h{d}_v\times d_{model}}$ 。

在这项工作中，我们采用了 $h=8$ 的并行注意力层（或称 “头”）。我们使用 $dk=dv=d_{model}/h=64$。由于减少了每个头的维度，总计算成本与全维度的单头注意力相似。

3.2.3 Application of Attention in our Model

The Transformer uses multi-head attention in three different ways:

变换器以三种不同方式使用多头注意力：

• In “encoder-decoder attention” layers, the queries come from the previous decoder layer, and the memory keys and values come from the output of the encoder. This allows every position in the decoder to attend over all positions in the input sequence. This mimics the typical encoder-decoder attention mechanisms in sequence-to-sequence models such as [38, 2, 9].

  在 “编码器-解码器注意 ”层中，查询来自前一个解码器层，而记忆键和记忆值则来自编码器的输出。这使得解码器中的每个位置都能关注输入序列中的所有位置。这模仿了序列到序列模型中典型的编码器-解码器注意机制，例如 [38, 2, 9]。

• The encoder contains self-attention layers. In a self-attention layer all of the keys, values and queries come from the same place, in this case, the output of the previous layer in the encoder. Each position in the encoder can attend to all positions in the previous layer of the encoder.

  编码器包含自我注意层。在自我注意层中，所有的键、值和查询都来自同一个地方，在本例中就是编码器中前一层的输出。编码器中的每个位置都可以关注编码器上一层的所有位置。

• Similarly, self-attention layers in the decoder allow each position in the decoder to attend to all positions in the decoder up to and including that position. We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to $-\infty$) all values in the input of the softmax which correspond to illegal connections. See Figure 2.

  同样，解码器中的自关注层允许解码器中的每个位置关注解码器中包括该位置在内的所有位置。我们需要防止解码器中的信息向左流动，以保持自动回归特性。我们通过屏蔽（设置为 $-\infty$）softmax 输入中所有与非法连接相对应的值，在缩放点积关注内部实现了这一点。见图 2。

3.3 Position-wise Feed-Forward Network

In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.

除了注意力子层外，我们的编码器和解码器中的每一层都包含一个全连接的前馈网络，该网络分别对每个位置进行相同的处理。这包括两个线性变换，中间有一个 ReLU 激活。
$$FFN(x)=max(0,x{W}_1+b_1)W_2+b_2$$
While the linear transformations are the same across different positions, they use different parameters from layer to layer. Another way of describing this is as two convolutions with kernel size 1. The dimensionality of input and output is $d_{model}=512$, and the inner-layer has dimensionality $d_{ff}=2048$.

虽然不同位置的线性变换相同，但各层使用的参数不同。另一种描述方法是两个内核大小为 1 的卷积。输入和输出的维度为 $d_{model}=512$，内层的维度为 $d_{ff}=2048$。

3.4 Embedding and softmax

Similarly to other sequence transduction models, we use learned embeddings to convert the input tokens and output tokens to vectors of dimension $d_{model}$. We also use the usual learned linear transformation and softmax function to convert the decoder output to predicted next-token probabilities. In our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation, similar to [30]. In the embedding layers, we multiply those weights by $\sqrt{d_{model}}$.

与其他序列转换模型类似，我们使用学习到的嵌入将输入标记和输出标记转换为维数为 $d_{model}$ 的向量。我们还使用通常的学习线性变换和 softmax 函数将解码器输出转换为预测的下一个标记概率。在我们的模型中，我们在两个嵌入层和预软最大线性变换之间共享相同的权重矩阵，这与 [30] 相似。在嵌入层中，我们将这些权重乘以 $\sqrt{d_{model}}$。

3.5 Position Encoding

Since our model contains no recurrence and no convolution, in order for the model to make use of the order of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence. To this end, we add “positional encodings” to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension $d_{model}$ as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].

由于我们的模型不包含递归和卷积，为了让模型能够利用序列的顺序，我们必须注入一些关于序列中标记的相对或绝对位置的信息。为此，我们在编码器和解码器堆栈底部的输入嵌入中添加了 “位置编码”。位置编码的维度 $d_{model}$ 与嵌入式编码的维度 $d_{model}$相同，因此两者可以相加。位置编码有多种选择，包括学习编码和固定编码 [9]。

In this work, we use $sine$ and $cosine$ functions of different frequencies:
$$\begin{gathered} P{E}_{(pos,2i)}=sin(pos/10000^{2i/d_{model}}) \\ P{E}_{(pos,2i+1)}=cos(pos/10000^{2i/d_{model}}) \end{gathered}$$
where pos is the position and $i$ is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from $2\pi$ to $10000\cdot 2\pi$. We chose this function because we hypothesized it would allow the model to easily learn to attend by relative positions, since for any fixed offset $k$, $P{E}_{pos+k}$ can be represented as a linear function of $P{E}_{pos}$.

其中，pos 是位置，$i$ 是维数。也就是说，位置编码的每个维度对应一个正弦波。波长形成一个从 $2\pi$ 到 $10000-2\pi$ 的几何级数。我们之所以选择这个函数，是因为我们假设它可以让模型轻松地学习到相对位置，因为对于任何固定的偏移$k$来说，$P{E}_{pos+k}$都可以表示为$P{E}_{pos}$的线性函数。

We also experimented with using learned positional embeddings [9] instead, and found that the two versions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training.

我们还尝试了使用学习到的位置嵌入[9]，发现两个版本产生的结果几乎相同（见表 3 第（E）行）。我们之所以选择正弦波版本，是因为它可以让模型推断出比训练时遇到的序列长度更长的序列。

4 Why Self-Attention

In this section we compare various aspects of self-attention layers to the recurrent and convolutional layers commonly used for mapping one variable-length sequence of symbol representations $(x_1,...,x_n)$ to another sequence of equal length $(z_1,...,z_n)$, with $x_i,z_i\in R^d$, such as a hidden layer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we consider three desiderata.

在本节中，我们将比较自注意层与递归层和卷积层的各个方面，后者通常用于将一个可变长度的符号表示序列 $(x_1,...,x_n)$ 映射到另一个等长序列 $(z_1,...,z_n)$，其中 $x_i,z_i\in R^d$，例如典型序列转换编码器或解码器中的隐藏层。我们使用自注意的动机有三个。

One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required.

一个是每层的总计算复杂度。另一个是可以并行化的计算量，以所需的最小顺序操作数来衡量。

The third is the path length between long-range dependencies in the network. Learning long-range dependencies is a key challenge in many sequence transduction tasks. One key factor affecting the ability to learn such dependencies is the length of the paths forward and backward signals have to traverse in the network. The shorter these paths between any combination of positions in the input and output sequences, the easier it is to learn long-range dependencies [12]. Hence we also compare the maximum path length between any two input and output positions in networks composed of the different layer types.

第三个是网络中长距离依赖关系之间的路径长度。学习长程依赖关系是许多序列转换任务的关键挑战。影响学习此类依赖关系能力的一个关键因素是前向和后向信号必须在网络中穿越的路径长度。输入和输出序列中任意位置组合之间的路径越短，学习远距离依赖关系就越容易[12]。因此，我们还比较了由不同层类型组成的网络中任意两个输入和输出位置之间的最大路径长度。

As noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially executed operations, whereas a recurrent layer requires $O(n)$ sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence length n is smaller than the representation dimensionality $d$, which is most often the case with sentence representations used by state-of-the-art models in machine translations, such as word-piece [38] and byte-pair [31] representations. To improve computational performance for tasks involving very long sequences, self-attention could be restricted to considering only a neighborhood of size $r$ in the input sequence centered around the respective output position. This would increase the maximum path length to $O(n/r)$. We plan to investigate this approach further in future work.

如表 1 所示，自注意层连接所有位置的连续操作数不变，而递归层则需要 $O(n)$ 的连续操作数。就计算复杂度而言，当序列长度 n 小于表示维度 $d$ 时，自注意层比递归层更快，而机器翻译中最先进的模型所使用的句子表示，如单词片[38]和字节对[31]表示，通常就是这种情况。为了提高涉及超长序列的任务的计算性能，可以限制自我注意只考虑输入序列中以相应输出位置为中心的大小为 $r$ 的邻域。这将把最大路径长度增加到 $O(n/r)$。我们计划在今后的工作中进一步研究这种方法。

A single convolutional layer with kernel width $k<n$ does not connect all pairs of input and output positions. Doing so requires a stack of $O(n/k)$ convolutional layers in the case of contiguous kernels, or $O(lo{g}_k(n))$ in the case of dilated convolutions [18], increasing the length of the longest paths between any two positions in the network. Convolutional layers are generally more expensive than recurrent layers, by a factor of $k$. Separable convolutions [6], however, decrease the complexity considerably, to $O(k\cdot n\cdot d+n\cdot d2)$. Even with $k=n$, however, the complexity of a separable convolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer, the approach we take in our model.

内核宽度为 $k<n$ 的单卷积层无法连接所有输入和输出位置对。在内核连续的情况下，这样做需要堆叠 $O(n/k)$ 的卷积层，在扩张卷积的情况下，则需要 $O(lo{g}_k(n))$ [18]，从而增加了网络中任意两个位置之间最长路径的长度。卷积层的成本通常比递归层高 $k$。然而，可分离卷积[6]可将复杂度大幅降低至 $O(k-n-d+n-d2)$。不过，即使在 $k=n$ 的情况下，可分离卷积的复杂度也相当于自注意层和点式前馈层的组合，而我们的模型正是采用了这种方法。

As side benefit, self-attention could yield more interpretable models. We inspect attention distributions from our models and present and discuss examples in the appendix. Not only do individual attention heads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic and semantic structure of the sentences.

作为附带的好处，自我注意可以产生更多可解释的模型。我们从模型中检测注意力分布，并在附录中举例说明和讨论。单个注意力头不仅明显学会了执行不同的任务，而且许多注意力头似乎还表现出了与句子的句法和语义结构相关的行为。

5 Training

This section describes the training regime for our models.

本节将介绍我们模型的训练机制。

5.1 Training Data and batching

We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs. Sentences were encoded using byte-pair encoding [3], which has a shared sourcetarget vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT 2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece vocabulary [38]. Sentence pairs were batched together by approximate sequence length. Each training batch contained a set of sentence pairs containing approximately 25000 source tokens and 25000 target tokens.

我们在标准的 WMT 2014 英德数据集上进行了训练，该数据集包含约 450 万个句子对。句子使用字节对编码[3]进行编码，其中有大约 37000 个共享的源目标词汇。对于English-French，我们使用了规模更大的 WMT 2014 English-French数据集，该数据集包含 3600 万个句子，并将标记拆分为 32000 个词块词汇[38]。句子对按近似序列长度分组。每个训练批包含一组句对，其中包含约 25000 个源词块和 25000 个目标词块。

5.2 hardware and Schedule

We trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using the hyperparameters described throughout the paper, each training step took about 0.4 seconds. We trained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the bottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps (3.5 days).

我们在一台配备 8 个英伟达 P100 GPU 的机器上训练模型。对于我们的基础模型，使用本文所述的超参数，每个训练步骤大约需要 0.4 秒。我们总共训练了 100,000 步或 12 个小时的基础模型。对于我们的大型模型（如表 3 底行所述），每步训练时间为 1.0 秒。大型模型的训练时间为 300,000 步（3.5 天）。

5.3 Optimizer

We used the Adam optimizer [20] with ${\beta }_1=0.9$, ${\beta }_2=0.98$ and $ϵ=10^{-9}$. We varied the learning rate over the course of training, according to the formula:
$$lrate=d_{model}^{-0.5}\cdot \min (step\_nu{m}^{-0.5},step\_num\cdot warmup\_step{s}^{-1.5})$$
This corresponds to increasing the learning rate linearly for the first warmup_steps training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used warmup_steps = 4000.

这相当于在第一个热身_steps 训练步数中线性增加学习率，此后学习率按步数的平方反比例递减。我们使用的是 warmup_steps = 4000。

5.4 Regularization

We employ three types of regularization during training:

在训练过程中，我们采用了三种正则化方法：

Residual Dropout We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. For the base model, we use a rate of $P_{drop}=0.1$.

残差剔除 我们对每个子层的输出进行剔除[33]，然后再将其添加到子层输入中并进行归一化处理。此外，我们还对编码器和解码器堆栈中的嵌入和位置编码之和进行了滤除。对于基础模型，我们使用的比率为 $P_{drop}=0.1$。

Label Smoothing During training, we employed label smoothing of value ${ϵ}_{ls}=0.1$ [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.

标签平滑 在训练过程中，我们使用了标签平滑值 ${ϵ}_{ls}=0.1$ [36]。这会降低复杂度，因为模型会变得更加不确定，但会提高准确度和 BLEU 得分。

6 Results

6.1 Machine Translation

On the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big) in Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0 BLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is listed in the bottom line of Table 3. Training took 3.5 days on 8 P100 GPUs. Even our base model surpasses all previously published models and ensembles, at a fraction of the training cost of any of the competitive models.

在 WMT 2014 英译德翻译任务中，大转换器模型（表 2 中的转换器（big））的 BLEU 值比之前报道的最佳模型（包括集合）高出 2.0 以上，达到了 28.4 的新的最先进 BLEU 值。该模型的配置见表 3 底行。在 8 个 P100 GPU 上的训练耗时 3.5 天。即使是我们的基本模型，也超越了之前发布的所有模型和合集，而训练成本仅为任何竞争模型的一小部分。

On the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0, outperforming all of the previously published single models, at less than 1/4 the training cost of the previous state-of-the-art model. The Transformer (big) model trained for English-to-French used dropout rate $P_{drop}=0.1$, instead of 0.3.

在 WMT 2014 英法互译任务中，我们的大模型获得了 41.0 的 BLEU 分数，超过了之前发布的所有单一模型，而训练成本还不到之前最先进模型的 1/4。针对英译法训练的 Transformer（大）模型使用了 $P_{drop}=0.1$，而不是 0.3。

For the base models, we used a single model obtained by averaging the last 5 checkpoints, which were written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We used beam search with a beam size of 4 and length penalty $\alpha =0.6$ [38]. These hyperparameters were chosen after experimentation on the development set. We set the maximum output length during inference to input length + 50, but terminate early when possible [38].

对于基础模型，我们使用的是通过平均最近 5 个检查点得到的单一模型，这些检查点是以 10 分钟的间隔写入的。对于大型模型，我们取最后 20 个检查点的平均值。我们使用波束搜索，波束大小为 4，长度惩罚为 $\alpha =0.6$ [38]。这些超参数是在开发集上实验后选择的。我们将推理过程中的最大输出长度设定为输入长度 +50，但尽可能提前终止 [38]。

Table 2 summarizes our results and compares our translation quality and training costs to other model architectures from the literature. We estimate the number of floating point operations used to train a model by multiplying the training time, the number of GPUs used, and an estimate of the sustained single-precision floating-point capacity of each GPU 5.

表 2 总结了我们的结果，并将我们的翻译质量和训练成本与文献中的其他模型架构进行了比较。我们通过将训练时间、使用的 GPU 数量和每个 GPU 的持续单精度浮点运算能力的估计值相乘，估算出训练一个模型所使用的浮点运算次数。

6.2 Model Variations

To evaluate the importance of different components of the Transformer, we varied our base model in different ways, measuring the change in performance on English-to-German translation on the development set, newstest2013. We used beam search as described in the previous section, but no checkpoint averaging. We present these results in Table 3.

为了评估 Transformer 不同组件的重要性，我们以不同方式改变了基础模型，测量了开发集 newstest2013 上英译德性能的变化。我们使用了上一节所述的波束搜索，但没有使用检查点平均法。表 3 列出了这些结果。

In Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions, keeping the amount of computation constant, as described in Section 3.2.2. While single-head attention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.

在表 3 行（A）中，我们按照第 3.2.2 节所述，在保持计算量不变的情况下，改变了注意头的数量以及注意键和值的维度。虽然单头注意力比最佳设置差 0.9 BLEU，但如果头数过多，质量也会下降。

In Table 3 rows (B), we observe that reducing the attention key size dk hurts model quality. This suggests that determining compatibility is not easy and that a more sophisticated compatibility function than dot product may be beneficial. We further observe in rows © and (D) that, as expected, bigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our sinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical results to the base model.

在表 3 行（B）中，我们发现减小注意力密钥大小 dk 会降低模型质量。这表明，确定兼容性并不容易，比点积更复杂的兼容性函数可能更有益处。在第（C）行和第（D）行中，我们进一步观察到，正如我们所预期的那样，模型越大越好，而 dropout 对避免过度拟合很有帮助。在第（E）行中，我们用学习到的位置嵌入[9]替换了正弦位置编码，观察到的结果与基础模型几乎完全相同。

6.3 English Constituency Parsing

To evaluate if the Transformer can generalize to other tasks we performed experiments on English constituency parsing. This task presents specific challenges: the output is subject to strong structural constraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence models have not been able to attain state-of-the-art results in small-data regimes [37].

为了评估转换器是否能推广到其他任务，我们进行了英语选区解析实验。这项任务具有特殊的挑战性：输出结果受到强大的结构约束，而且比输入结果要长得多。此外，RNN 序列到序列模型在小数据环境下也无法达到最先进的结果[37]。

We trained a 4-layer transformer with $d_{model}=1024$ on the Wall Street Journal (WSJ) portion of the Penn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting, using the larger high-confidence and Berkley Parser corpora from with approximately 17M sentences [37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens for the semi-supervised setting.

我们在宾夕法尼亚大学树库（Penn Treebank）[25]的《华尔街日报》（WSJ）部分（约 4 万个训练句子）上训练了一个 $d_{model}=1024$ 的 4 层变换器。我们还在半监督环境下使用更大的高置信度语料库和伯克利解析器语料库（约 177 个句子）[37]对其进行了训练。在仅使用 WSJ 的设置中，我们使用了 16K 个词组的词汇量，而在半监督设置中，我们使用了 32K 个词组的词汇量。

We performed only a small number of experiments to select the dropout, both attention and residual (section 5.4), learning rates and beam size on the Section 22 development set, all other parameters remained unchanged from the English-to-German base translation model. During inference, we increased the maximum output length to input length + 300. We used a beam size of 21 and $\alpha =0.3$ for both WSJ only and the semi-supervised setting.

我们仅在第 22 节开发集上进行了少量实验，以选择注意力和残差（5.4 节）、学习率和波束大小，所有其他参数均与英德基础翻译模型保持一致。在推理过程中，我们将最大输出长度增加到输入长度 + 300。在仅 WSJ 和半监督设置中，我们使用的波束大小为 21，$\alpha =0.3$。

Our results in Table 4 show that despite the lack of task-specific tuning our model performs surprisingly well, yielding better results than all previously reported models with the exception of the Recurrent Neural Network Grammar [8].

表 4 中的结果表明，尽管我们没有针对特定任务进行调整，但我们的模型表现出人意料的好，除了循环神经网络文法 [8] 之外，其结果优于之前报道的所有模型。

In contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the Berkeley-Parser [29] even when training only on the WSJ training set of 40K sentences.

与 RNN 序列到序列模型[37]相比，即使只在由 40K 个句子组成的 WSJ 训练集上进行训练，Transformer 的表现也优于 Berkeley-Parser [29]。

7 Conclusion

In this work, we presented the Transformer, the first sequence transduction model based entirely on attention, replacing the recurrent layers most commonly used in encoder-decoder architectures with multi-headed self-attention.

在这项工作中，我们提出了 Transformer，这是第一个完全基于注意力的序列转换模型，用多头自我注意力取代了编码器-解码器架构中最常用的递归层。

For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014 English-to-French translation tasks, we achieve a new state of the art. In the former task our best model outperforms even all previously reported ensembles.

在翻译任务中，Transformer 的训练速度明显快于基于递归层或卷积层的架构。在 WMT 2014 英译德和 WMT 2014 英译法翻译任务中，我们都达到了新的技术水平。在前一项任务中，我们的最佳模型甚至优于之前报告的所有集合。

We are excited about the future of attention-based models and plan to apply them to other tasks. We plan to extend the Transformer to problems involving input and output modalities other than text and to investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs such as images, audio and video. Making generation less sequential is another research goals of ours.

我们对基于注意力的模型的未来充满期待，并计划将其应用于其他任务。我们计划将 Transformer 扩展到涉及文本以外的输入和输出模式的问题上，并研究局部的、受限的注意力机制，以有效处理大型输入和输出，如图像、音频和视频。我们的另一个研究目标是减少生成的顺序。

The code we used to train and evaluate our models is available at https://github.com/tensorflow/tensor2tensor.

我们用于训练和评估模型的代码可在 https://github.com/tensorflow/tensor2tensor 上获取。

Acknowledgements We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful comments, corrections and inspiration.

致谢 我们感谢 Nal Kalchbrenner 和 Stephan Gouws 富有成效的评论、纠正和启发。