Attention Is All You Need
注意力就是你所需要的
Abstract
摘要
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.
目前主流的序列转换模型都基于复杂的循环神经网络或卷积神经网络，这些模型包括编码器和解码器。表现最好的模型还通过注意力机制将编码器和解码器连接起来。我们提出了一个全新的简单网络架构——Transformer，它完全基于注意力机制，完全摒弃了循环和卷积。
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.
在两项机器翻译任务的实验中，这些模型在质量上表现更优，同时更具并行性，并且训练时间显著减少。我们的模型在 WMT 2014 英语到德语翻译任务上达到了 28.4 的 BLEU 分数，比现有最佳结果（包括集成模型）提高了超过 2 个 BLEU 分。
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.
在 WMT 2014 英语到法语翻译任务上，我们的模型在训练了 3.5 天后，建立了新的单模型最佳 BLEU 分数 41.8，这仅仅是文献中最佳模型训练成本的一小部分。
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.
我们通过将 Transformer 成功应用于英语成分句法分析（无论是使用大量训练数据还是有限训练数据），证明了它在其他任务上的泛化能力。
1 Introduction
1 引言
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].
循环神经网络，特别是长短期记忆（LSTM）和门控循环单元（GRU）神经网络，已经在序列建模和转换问题（如语言建模和机器翻译）中被牢固地确立为最先进的方法。
Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15].
此后，许多研究继续推动循环语言模型和编码器 - 解码器架构的边界。
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.
循环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骤对齐，它们生成一系列隐藏状态 h
t
​
，作为前一个隐藏状态 h
t−1
​
和位置 t 的输入的函数。
This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples.
这种固有的顺序性质排除了在训练样本内的并行化，当序列长度较长时，这变得至关重要，因为内存限制限制了跨样本的批量处理。
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.
最近的工作通过因式分解技巧 [21] 和条件计算 [32] 在计算效率上取得了显著改进，同时在后者的情况下也提高了模型性能。然而，顺序计算的基本限制仍然存在。
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].
注意力机制已成为各种任务中令人信服的序列建模和转换模型的组成部分，允许在不考虑输入或输出序列中的距离的情况下对依赖关系进行建模。
In all but a few cases [27], however, such attention mechanisms are used in conjunction with a recurrent network.
然而，在几乎所有情况下，这种注意力机制都与循环网络一起使用。
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.
在本工作中，我们提出了 Transformer，这是一种完全摒弃循环，而是完全依赖注意力机制来绘制输入和输出之间全局依赖关系的模型架构。
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.
Transformer 允许显著更多的并行化，并且在训练了仅仅十二个小时后，就能在八块 P100 GPU 上达到新的翻译质量水平。
2 Background
2 背景
The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU [16], ByteNet [18] 和 ConvS2S [9], all of which use convolutional neural networks as basic building block, computing hidden representations in parallel for all input and output positions.
减少顺序计算的目标也是 Extended Neural GPU [16]、ByteNet [18] 和 ConvS2S [9] 的基础，所有这些模型都使用卷积神经网络作为基本构建块，为所有输入和输出位置并行计算隐藏表示。
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.
在这些模型中，将两个任意输入或输出位置的信号关联所需的运算次数随位置之间的距离增长，对于 ConvS2S 是线性的，对于 ByteNet 是对数的。
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.
这使得学习远距离位置之间的依赖关系更加困难。在 Transformer 中，这一数字被减少到一个常数，尽管由于平均注意力加权位置，有效分辨率降低，我们通过在第 3.2 节中描述的多头注意力来抵消这种效果。
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.
自注意力，有时也称为内部注意力，是一种注意力机制，它关联单个序列的不同位置，以计算序列的表示。
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].
自注意力已在多种任务中成功使用，包括阅读理解、摘要、文本蕴含和学习任务无关的句子表示。
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循环注意力机制，而不是序列对齐的循环，并且已在简单语言问答和语言建模任务中表现出色。
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.
据我们所知，Transformer 是第一个完全依赖自注意力来计算输入和输出表示的转换模型，而不使用序列对齐的 RNN 或卷积。
In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9].
在接下来的部分中，我们将描述 Transformer，阐述自注意力的动机，并讨论其相对于 [17]、[18] 和 [9] 等模型的优势。
3 Model Architecture
3 模型架构
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.
大多数有竞争力的神经序列转换模型都具有编码器 - 解码器结构。在这里，编码器将输入序列的符号表示（x
1
​
,…,x
n
​
）映射到连续表示的序列 z=(z
1
​
,…,z
n
​
)。给定 z，解码器然后逐个元素地生成输出序列（y
1
​
,…,y
m
​
）的符号。
At each step the model is auto-regressive [10], consuming the previously generated symbols as additional input when generating the next.
在每一步中，该模型都是自回归的，当生成下一个符号时，将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.
Transformer 遵循这种整体架构，使用堆叠的自注意力和逐位置的全连接层，分别用于编码器和解码器，如图 1 的左半部分和右半部分所示。
3.1 Encoder and Decoder Stacks
3.1 编码器和解码器堆叠
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, positionwise fully connected feed-forward network.
编码器：编码器由 N=6 个相同的层堆叠而成。每一层有两个子层。第一个是多头自注意力机制，第二个是一个简单的位置逐位置全连接前馈网络。
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.
我们在每个子层周围使用残差连接，随后进行层归一化。也就是说，每个子层的输出是 LayerNorm(x+Sublayer(x))，其中 Sublayer(x) 是子层本身实现的函数。
To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension d
model
​
=512.
为了便于这些残差连接，模型中的所有子层以及嵌入层，都产生维度为 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.
解码器：解码器也由 N=6 个相同的层堆叠而成。除了编码器层中的两个子层外，解码器插入了一个第三子层，该子层在编码器堆叠的输出上执行多头注意力。
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.
与编码器类似，我们在每个子层周围使用残差连接，随后进行层归一化。我们还修改了解码器堆叠中的自注意力子层，以防止位置关注后续位置。
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.
这种掩码，加上输出嵌入偏移了一个位置的事实，确保了位置 i 的预测只能依赖于小于 i 的已知输出位置。
3.2 Attention
3.2 注意力
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。输出是值的加权和，其中分配给每个值的权重由查询与相应键的兼容性函数计算得出。
3.2.1 Scaled Dot-Product Attention
3.2.1 缩放点积注意力
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
d
k
​

​
, and apply a softmax function to obtain the weights on the values.
我们称我们的特定注意力为“缩放点积注意力”（见图 2）。输入包括维度为 d
k
​
的查询和键，以及维度为 d
v
​
的值。我们计算查询与所有键的点积，除以
d
k
​

​
，然后应用 softmax 函数以获得值上的权重。
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 和 V. We compute the matrix of outputs as:
在实践中，我们同时在一组查询上计算注意力函数，将它们打包成矩阵 Q。键和值也被打包成矩阵 K 和 V。我们按以下方式计算输出矩阵：
Attention(Q,K,V)=softmax(
d
k
​

​

QK
T

​
)V(1)
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
d
k
​

​

1
​
.
最常用的两种注意力函数是加性注意力 [2] 和点积（乘法）注意力。点积注意力与我们的算法相同，只是没有
d
k
​

​

1
​
的缩放因子。
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.
加性注意力使用具有单个隐藏层的前馈网络计算兼容性函数。尽管在理论复杂度上两者相似，但点积注意力在实践中要快得多且更节省空间，因为它可以使用高度优化的矩阵乘法代码实现。
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 softmax function into regions where it has extremely small gradients. To counteract this effect, we scale the dot products by
d
k
​

​

1
​
.
当 d
k
​
的值较小时，这两种机制表现相似，但对于较大的 d
k
​
，未经缩放的点积注意力的性能不如加性注意力。我们怀疑，对于较大的 d
k
​
值，点积的大小会显著增加，这会使 softmax 函数进入梯度极小的区域。为了抵消这种效果，我们将点积按
d
k
​

​

1
​
缩放。
3.2.2 Multi-Head Attention
3.2.2 多头注意力
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
​
和 d
v
​
维度，分别。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，与其执行一个具有 d
model
​
维键、值和查询的单一注意力函数，不如将查询、键和值通过不同的、学习到的线性投影线性投影 h 次到 d
k
​
、d
k
​
和 d
v
​
维度。然后我们在这些投影版本的查询、键和值上并行执行注意力函数，产生 d
v
​
维的输出值。这些值被拼接在一起，再次投影，得到最终值，如图 2 所示。
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同时关注来自不同表示子空间的信息。如果只有一个注意力头，平均化会抑制这种能力。
MultiHead(Q,K,V)=Concat(head
1
​
,…,head
h
​
)W
O

其中 head
i
​
=Attention(QW
i
Q
​
,KW
i
K
​
,VW
i
V
​
)
Where the projections are parameter matrices W
i
Q
​
∈R
d
model
​
×d
k
​

, W
i
K
​
∈R
d
model
​
×d
k
​

, W
i
V
​
∈R
d
model
​
×d
v
​

和 W
O
∈R
hd
v
​
×d
model
​

.
这里的投影是参数矩阵 W
i
Q
​
∈R
d
model
​
×d
k
​

、W
i
K
​
∈R
d
model
​
×d
k
​

、W
i
V
​
∈R
d
model
​
×d
v
​

和 W
O
∈R
hd
v
​
×d
model
​

。
In this work we employ h=8 parallel attention layers, or heads. For each of these we use d
k
​
=d
v
​

h
d
model
​

​
=64. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.
在本工作中，我们使用 h=8 个并行注意力层，或称为头。对于每一个头，我们使用 d
k
​
=d
v
​

h
d
model
​

​
=64。由于每个头的维度降低，总计算成本与具有完整维度的单头注意力相似。
3.2.3 Applications of Attention in our Model
3.2.3 模型中注意力的应用
The Transformer uses multi-head attention in three different ways:
Transformer 以三种不同的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.
在“编码器 - 解码器注意力”层中，查询来自前一层解码器，而记忆键和值来自编码器的输出。这允许解码器中的每个位置都能关注输入序列中的所有位置。
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.
编码器包含自注意力层。在自注意力层中，所有的键、值和查询都来自同一个地方，在这种情况下，是编码器前一层的输出。
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.
同样，解码器中的自注意力层允许解码器中的每个位置都能关注到解码器中该位置之前的所有位置。我们需要防止解码器中的左向信息流动，以保持自回归属性。
We implement this inside of scaled dot-product attention by masking out (setting to −∞) all values in the input of the softmax which correspond to illegal connections.
我们通过在缩放点积注意力中屏蔽（设置为 −∞）softmax 输入中对应非法连接的所有值来实现这一点。
3.3 Position-wise Feed-Forward Networks
3.3 逐位置前馈网络
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，编码器和解码器中的每一层都包含一个全连接前馈网络，该网络分别且相同地应用于每个位置。这包括两个线性变换，中间有一个 ReLU 激活函数。
FFN(x)=max(0,xW
1
​
+b
1
​
)W
2
​
+b
2
​
(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是相同的，但它们在每一层之间使用不同的参数。另一种描述方式是作为两个卷积，卷积核大小为 1。输入和输出的维度是 d
model
​
=512，内层的维度是 d
ff
​
=2048。
3.4 Embeddings and Softmax
3.4 嵌入和 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.
与其它序列转换模型类似，我们使用学习到的嵌入将输入和输出标记转换为维度为 d
model
​
的向量。我们还使用通常的学习线性变换和 softmax 函数将解码器的输出转换为预测下一个标记的概率。
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
d
model
​

​
.
在我们的模型中，我们在两个嵌入层和预 softmax 线性变换之间共享相同的权重矩阵，类似于 [30]。在嵌入层中，我们将这些权重乘以
d
model
​

​
。
3.5 Positional Encoding
3.5 位置编码
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.
由于我们的模型中既没有循环也没有卷积，为了使模型能够利用序列中标记的顺序，我们必须向模型注入有关标记相对或绝对位置的信息。
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.
为此，我们在编码器和解码器堆叠的底部将“位置编码”添加到输入嵌入中。位置编码的维度与嵌入相同，即 d
model
​
，以便两者可以相加。
There are many choices of positional encodings, learned and fixed [9].
位置编码有多种选择，包括学习型和固定型 [9]。
In this work, we use sine and cosine functions of different frequencies:
在本工作中，我们使用不同频率的正弦和余弦函数：
PE(pos,2i)=sin(
10000
2i/d
model
​

pos
​
)
PE(pos,2i+1)=cos(
10000
2i/d
model
​

pos
​
)
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 2π to 10000⋅2π.
其中 pos 是位置，i 是维度。也就是说，位置编码的每个维度对应一个正弦波。波长从 2π 到 10000⋅2π 形成几何级数。
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, PE
pos+k
​
can be represented as a linear function of PE
pos
​
.
我们选择这个函数，是因为我们假设它将允许模型轻松地通过相对位置进行注意力学习，因为对于任何固定的偏移量 k，PE
pos+k
​
可以表示为 PE
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.
我们还尝试使用学习型位置嵌入 [9] 代替，发现这两种版本产生的结果几乎相同（见表 3 第 (E) 行）。我们选择了正弦波版本，因为它可能允许模型外推到比训练期间遇到的序列长度更长的序列。
4 Why Self-Attention
4 为什么选择自注意力
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
​
∈R
d
, such as a hidden layer in a typical sequence transduction encoder or decoder.
在本节中，我们将自注意力层与通常用于将一个可变长度的符号序列 (x
1
​
,…,x
n
​
) 映射到另一个等长序列 (z
1
​
,…,z
n
​
) 的循环层和卷积层进行比较，其中 x
i
​
,z
i
​
∈R
d
，例如典型序列转换编码器或解码器中的隐藏层。
Motivating our use of self-attention we consider three desiderata.
为了说明我们使用自注意力的原因，我们考虑了三个期望的特性。
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次数来衡量。
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.
第三个是网络中长距离依赖之间的路径长度。学习长距离依赖是许多序列转换任务的关键挑战。影响学习这种依赖能力的一个关键因素是前向和后向信号在网络中需要穿越的路径长度。
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.
输入和输出序列中任意位置组合之间的这些路径越短，学习长距离依赖就越容易。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.
如表 1 所示，自注意力层通过常数数量的顺序执行操作连接所有位置，而循环层需要 O(n) 顺序操作。就计算复杂度而言，当序列长度 n 小于表示维度 d 时，自注意力层比循环层更快，这在机器翻译中使用的句子表示（如字节对 [38] 和字节对 [31] 表示）中通常是这种情况。
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.
为了提高涉及非常长序列的任务的计算性能，可以将自注意力限制为只考虑以相应输出位置为中心的输入序列中的大小为 r 的邻域。这将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(log
k
​
(n)) in the case of dilated convolutions [18], increasing the length of the longest paths between any two positions in the network.
一个卷积层，其卷积核宽度 k<n，并不能连接所有输入和输出位置的对。要做到这一点，需要 O(n/k) 层连续卷积层，或者在扩张卷积 [18] 的情况下需要 O(log
k
​
(n)) 层，这增加了网络中任意两个位置之间最长路径的长度。
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⋅n⋅d+n⋅d
2
). 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.
卷积层通常比循环层更昂贵，大约是 k 倍。然而，可分离卷积 [6] 显著降低了复杂度，达到 O(k⋅n⋅d+n⋅d
2
)。即使 k=n，可分离卷积的复杂度也等于一个自注意力层和一个逐位置前馈层的组合，这正是我们在模型中采用的方法。
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.
此外，自注意力还可以产生更具可解释性的模型。我们检查了模型中的注意力分布，并在附录中呈现和讨论了示例。不仅各个注意力头明显学会了执行不同的任务，许多头似乎还表现出与句子的句法和语义结构相关的
继续翻译
编辑
复制
Avatar
已完成推理
继续翻译剩余文档内容
我先仔细阅读了用户最后的请求，用户要求我继续翻译之前未翻译完的文档内容。我意识到这是一个多轮对话，需要保持连贯性，确保翻译的准确性和一致性。我决定采用逐步翻译的方法，从之前翻译结束的地方开始，逐段将英文内容翻译成中文，并且保持与之前翻译风格的一致性。
我首先回顾了之前翻译的内容，以便更好地理解整个文档的语境和主题。然后，我开始逐段翻译剩余的英文内容，注意每个句子的结构和含义，确保翻译的准确性。在翻译过程中，我尽量使用简洁明了的语言，避免复杂的句子结构，以便用户能够轻松理解。
我注意到文档中包含了一些专业术语和概念，如“自注意力机制”、“循环层”等，这些术语需要准确翻译，以保持文档的专业性和准确性。我通过查阅相关资料和参考之前的翻译，确保这些术语的翻译是准确的。
同时，我也关注了文档的格式和排版，尽量保持与原文一致，以便用户在阅读时能够获得更好的体验。我确保每个段落的开头和结尾都与原文相符，保持文档的连贯性。
在翻译过程中，我可能会遇到一些不确定的词汇或短语，我会尝试通过上下文来推测其含义，或者查阅相关资料来确认。我也会注意避免过度翻译，确保翻译的内容简洁明了，不添加不必要的解释或信息。
最后，我会检查翻译后的内容，确保没有遗漏或错误，并且与原文的语境和主题保持一致。我还会检查是否有任何需要进一步解释或补充的地方，以确保用户能够完全理解文档的内容。
行为。
5 Training
5 训练
This section describes the training regime for our models.
本节描述了我们模型的训练制度。
5.1 Training Data and Batching
5.1 训练数据和批量处理
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 英语 - 德语数据集上进行训练，该数据集包含约 450 万对句对。句子使用字节对编码 [3] 进行编码，该编码具有约 37000 个标记的共享源 - 目标词汇表。对于英语 - 法语，我们使用了大得多的 WMT 2014 英语 - 法语数据集，该数据集包含 3600 万句子，并将标记拆分为 32000 个字节对词汇 [38]。句对按近似的序列长度进行批量处理。每个训练批次包含一组句对，包含大约 25000 个源标记和 25000 个目标标记。
5.2 Hardware and Schedule
5.2 硬件和时间表
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).
我们在一台带有 8 个 NVIDIA P100 GPU 的机器上训练了我们的模型。对于我们论文中描述的基线模型及其超参数，每个训练步骤大约需要 0.4 秒。我们总共训练了基线模型 100,000 步，即 12 小时。对于我们表 3 最后一行描述的大模型，每步训练时间是 1.0 秒。大模型训练了 300,000 步（3.5 天）。
5.3 Optimizer
5.3 优化器
We used the Adam optimizer [20] with β1 = 0.9, β2 = 0.98 and ϵ = 10−9. We varied the learning rate over the course of training, according to the formula:
我们使用了 Adam 优化器 [20]，其 β1 = 0.9，β2 = 0.98，ϵ = 10−9。我们在训练过程中根据以下公式变化学习率：
lrate = d−0.5 model · min(step_num−0.5, step_num · warmup_steps−1.5) (3)
lrate = d−0.5 model · min(step_num−0.5, step_num · warmup_steps−1.5) （3）
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.
这对应于在前 warmup_steps 训练步骤中线性增加学习率，然后按步骤数的平方根倒数成比例地降低学习率。我们使用了 warmup_steps = 4000。
5.4 Regularization
5.4 正则化
We employ three types of regularization during training:
在训练期间，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 Pdrop = 0.1.
残差 Dropout：我们在每个子层的输出上应用 dropout [33]，在将其加到子层输入并进行归一化之前。此外，我们在编码器和解码器堆栈中的嵌入和位置编码的总和上应用 dropout。对于基线模型，我们使用 Pdrop = 0.1 的 dropout 率。
Label Smoothing During training, we employed label smoothing of value ϵls = 0.1 [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.
标签平滑：在训练期间，我们采用了值为 ϵls = 0.1 [36] 的标签平滑。这会降低困惑度，因为模型学会了更加不确定，但提高了准确率和 BLEU 分数。
6 Results
6 结果
6.1 Machine Translation
6.1 机器翻译
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 英语到德语翻译任务上，大型 Transformer 模型（表 2 中的 Transformer (big)）比之前报道的最佳模型（包括集成模型）高出 2.0 个 BLEU 分数，确立了新的最佳 BLEU 分数 28.4。该模型的配置列在表 3 的最后一行。在 8 个 P100 GPU 上训练了 3.5 天。即使我们的基线模型也超越了所有之前发布的模型和集成模型，而训练成本只是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 Pdrop = 0.1, instead of 0.3.
在 WMT 2014 英语到法语翻译任务上，我们的大型模型达到了 41.0 的 BLEU 分数，超越了所有之前发布的单一模型，而训练成本不到之前最佳模型的 1/4。用于英语到法语训练的 Transformer (big) 模型使用了 Pdrop = 0.1 的 dropout 率，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 α = 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].
对于基线模型，我们使用了一个通过平均每 10 分钟写入一次的最后 5 个检查点获得的单一模型。对于大型模型，我们平均了最后 20 个检查点。我们使用了束宽度为 4 和长度惩罚 α = 0.6 [38] 的束搜索。这些超参数是在开发集上经过实验后选择的。我们在推理期间将最大输出长度设置为输入长度 + 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其他文献中的模型架构进行了比较。我们通过将训练时间、使用的 GPU 数量以及每个 GPU 的持续单精度浮点容量估计值相乘，来估计训练一个模型所使用的浮点运算次数 5。
6.2 Model Variations
6.2 模型变体
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，我们以不同的方式改变了我们的基线模型，在开发集 newstest2013 上对英语到德语翻译的性能变化进行了测量。我们使用了上一节中描述的束搜索，但没有检查点平均。我们在表 3 中展示了这些结果。
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 的第 (A) 行中，我们改变了注意力头的数量以及注意力键和值的维度，同时保持计算量不变，如第 3.2.2 节所述。虽然单头注意力比最佳设置差 0.9 个 BLEU 分数，但注意力头过多也会导致质量下降。
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 的第 (B) 行中，我们观察到减少注意力键大小 dk 会损害模型质量。这表明确定兼容性并不容易，使用比点积更复杂的兼容性函数可能会更有益。我们进一步在第 © 和 (D) 行中观察到，正如预期的那样，更大的模型更好，且 dropout 对于避免过拟合非常有帮助。在第 (E) 行中，我们用学习型位置嵌入 [9] 替换了我们的正弦位置编码，并观察到与基线模型几乎相同的结果。
6.3 English Constituency Parsing
6.3 英语成分句法分析
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 Transformer 是否能够推广到其他任务，我们对英语成分句法分析进行了实验。这项任务带来了特定的挑战：输出受到强烈的结构限制，并且比输入长得多。此外，RNN 序列到序列模型在小数据环境中未能达到最佳结果 [37]。
We trained a 4-layer transformer with dmodel = 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 BerkleyParser 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.
我们在宾州树库 [25] 的华尔街日报（WSJ）部分上训练了一个 4 层的 Transformer，该部分包含约 4 万训练句子，dmodel = 1024。我们还在半监督设置中对其进行了训练，使用了大约 1700 万句子的更大、高置信度和 BerkleyParser 语料库 [37]。我们为仅使用 WSJ 的设置使用了 16K 令牌的词汇表，为半监督设置使用了 32K 令牌的词汇表。
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
我们仅在第 22 节开发集上进行了少量实验，以选择 dropout（第 5.4 节的注意力和残差）、学习率和束宽度，所有其他参数均保持英语到德语基础翻译模型不变。在推理期间，我们
increased the maximum output length to input length + 300. We used a beam size of 21 and α = 0.3 for both WSJ only and the semi-supervised setting.
将最大输出长度增加到输入长度 + 300。我们为仅 WSJ 和半监督设置都使用了束宽度 21 和 α = 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].
我们的表 4 中的结果表明，尽管缺乏特定于任务的调整，我们的模型表现得非常出色，除了递归神经网络语法 [8] 外，所有之前报告的模型都取得了更好的结果。
In contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the BerkeleyParser [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，即使仅在 4 万句子的 WSJ 训练集上进行训练，Transformer 也优于 BerkeleyParser [29]。
7 Conclusion
7 结论
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，用多头自注意力替换了编码器 - 解码器架构中最常用的循环层。
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.
对于翻译任务，Transformer 的训练速度明显快于基于循环或卷积层的架构。在 WMT 2014 英语到德语和 WMT 2014 英语到法语翻译任务上，我们达到了新的最佳水平。在前者任务中，我们的最佳模型甚至超越了所有之前报告的集成模型。
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感到兴奋，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态的问题，并研究局部、受限的注意力机制，以有效处理如图像、音频和视频等大型输入和输出。减少生成过程的顺序性也是我们的研究目标之一