这是用户在 2024-4-25 13:22 为 https://app.immersivetranslate.com/pdf-pro/12f7e126-a6ba-411a-857d-051c28899f95 保存的双语快照页面,由 沉浸式翻译 提供双语支持。了解如何保存?
2024_04_25_128b845e084f1660140dg

Attention Is All You Need
注意力机制就是你所需要的

Ashish Vaswani*Google Brainavaswani@google.com

Noam Shazeer* 诺姆·沙兹尔*Google Brainnoam@google.com

Niki Parmar*Google Research 谷歌研究nikip@google.com

Jakob Uszkoreit*Google Research 谷歌研究usz@google.com

Llion Jones*Google Research 谷歌研究llion@google.com

Aidan N. Gomez* University of Toronto 多伦多大学aidan@cs.toronto. edu

Lukasz Kaiser*Google Brainlukaszkaiser@google.com

Illia Polosukhin* illia.polosukhin@gmail.com

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. 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 Englishto-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.
基于复杂的循环或卷积神经网络的主导序列转换模型包括编码器和解码器。最佳执行模型还通过注意机制将编码器和解码器连接起来。我们提出了一种新的简单的网络架构,Transformer,完全基于注意机制,完全不使用循环和卷积。在两个机器翻译任务中的实验表明,这些模型在质量上优于其他模型,同时更易并行化,训练时间显著缩短。我们的模型在 WMT 2014 英文到德文翻译任务上达到了 28.4 BLEU 的分数,相较于现有的最佳结果(包括集成模型),提升了 2 个 BLEU。在 WMT 2014 英文到法文翻译任务上,我们的模型在训练了 3.5 天的情况下,在八个 GPU 上实现了基于单一模型的 BLEU 分数的最新记录,达到了 41.8。这只是最佳模型的训练成本的一小部分。我们还展示了 Transformer 在使用大量和有限训练数据时成功应用于英文成分句法分析的通用性。

1 Introduction 1 简介

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].
循环神经网络,长短期记忆[13]和门控循环[7]神经网络特别是已经被确定为序列建模和转导问题的最先进方法,例如语言建模和机器翻译[35, 2, 5]。此后,许多努力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 , as a function of the previous hidden state and the input for position . 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.
循环模型通常沿着输入和输出序列的符号位置对计算进行因式分解。将位置与计算时间步骤对齐,它们生成一系列隐藏状态 ,作为前一个隐藏状态 和位置 的输入的函数。这种固有的顺序性质排除了在训练示例中的并行化,这在较长的序列长度下变得至关重要,因为内存限制限制了跨示例的批处理。最近的工作通过因式分解技巧[21]和条件计算[32]实现了计算效率的显着提高,同时在后者的情况下也提高了模型性能。然而,顺序计算的基本约束仍然存在。
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]. In all but a few cases [27], however, such attention mechanisms are used in conjunction with a recurrent network.
注意机制已经成为各种任务中引人注目的序列建模和转导模型的一个组成部分,允许对输入或输出序列中的依赖关系进行建模,而不考虑它们之间的距离[2, 19]。然而,在几乎所有情况下[27],这种注意机制都是与循环网络一起使用的。
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,一个模型架构,摒弃了循环,并完全依赖于注意机制来建立输入和输出之间的全局依赖关系。Transformer 允许更高程度的并行化,在仅经过八个 P100 GPU 进行十二小时训练后,可以达到最新的翻译质量水平。

2 Background 2 背景

The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU [16], ByteNet [18] 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构成了 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模块,在所有输入和输出位置并行计算隐藏表示。在这些模型中,从两个任意输入或输出位置关联信号所需的操作数量随着位置之间的距离增加而增加,对于 ConvS2S 是线性增长,对于 ByteNet 是对数增长。这使得学习远距离位置之间的依赖关系变得更加困难。在 Transformer 中,这被减少为一定数量的操作,尽管由于平均注意力加权位置而导致有效分辨率降低,我们通过第 3.2 节中描述的 Multi-Head Attention 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].
自注意力,有时称为内部注意力,是一种关联单个序列的不同位置以计算序列表示的注意机制。 自我注意力已成功用于各种任务,包括阅读理解,摘要制作,文本推理和学习任务独立的句子表示[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].
端到端的记忆网络是基于循环注意力机制而不是序列对齐的循环,并且已被证明在简单语言问答和语言建模任务中表现良好 [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].
就我们目前的了解而言,Transformer 是第一个完全依靠自注意力计算其输入和输出表示的转录模型,而不使用序列对齐的 RNN 或卷积。在下面的章节中,我们将描述 Transformer,解释自注意力并讨论其相对于模型 [17, 18] 和 [9] 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 to a sequence of continuous representations . Given , the decoder then generates an output sequence 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导模型具有编码器-解码器结构[5, 2, 35]。在这里,编码器将一个符号表示的输入序列 映射到一个连续表示的序列 。给定 ,解码器然后逐个元素地生成一个符号输出序列 。在每一步中,模型是自回归的[10],在生成下一个符号时,消耗先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 遵循这一整体架构,使用堆叠的自注意力和点式、全连接层分别用于编码器和解码器,如图 1 左右两半所示。
Figure 1: The Transformer - model architecture.
图 1:变压器-模型架构。

3.1 Encoder and Decoder Stacks
3.1 编码器和解码器堆栈

Encoder: The encoder is composed of a stack of 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. 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 , where 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 .
编码器:编码器由 个相同层堆叠而成。每个层包含两个子层。第一个是多头自注意力机制,第二个是简单的、位置逐层全连接的前馈网络。我们在两个子层周围采用了残差连接[11],然后进行层归一化[1]。也就是说,每个子层的输出是 LayerNorm ,其中 是子层本身实现的函数。为了促进这些残差连接,模型中的所有子层以及嵌入层产生维度为 的输出。
Decoder: The decoder is also composed of a stack of 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 can depend only on the known outputs at positions less than .
解码器:解码器也由 个相同层组成。除了每个编码器层之外的两个子层之外,解码器还插入了第三个子层,它对编码器堆栈的输出进行多头注意力计算。类似于编码器,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注意子层,以防止位置关注之后的位置。这种遮罩操作,结合输出嵌入向量偏移一个位置的事实,确保位置 的预测仅依赖于位置小于 处的已知输出。

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.
Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel.
图 2: (左) 缩放点积注意力。 (右) 多头注意力由多个并行运行的注意力层组成。

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 , and values of dimension . We compute the dot products of the query with all keys, divide each by , and apply a softmax function to obtain the weights on the values.
我们将我们的特别关注称为“缩放点积注意力”(图 2)。输入由维度为 的查询和键以及维度为 的值组成。我们计算查询与所有键的点积,将每个除以 ,并应用 softmax 函数以获得值的权重。
In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix . The keys and values are also packed together into matrices and . We compute the matrix of outputs as:
在实践中,我们同时对一组查询计算注意函数,将其打包成矩阵 。键和值也被打包到矩阵 中。我们计算输出矩阵如下:
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 . 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.
两种最常用的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,除了缩放因子 。加性注意力使用具有单隐藏层的前馈网络计算兼容性函数。虽然在理论复杂性上两者相似,但点积注意力在实践中更快速、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 the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of . We suspect that for large values of , 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 .
的值较小时,两种机制表现相似,但对于较大的 值,加性注意力胜过无缩放的点积注意力。我们怀疑对于较大的 值,点积会变得很大,将 softmax 函数推入具有极小梯度的区域 为了抵消这种影响,我们通过 对点积进行缩放。

3.2.2 Multi-Head Attention
3.2.2 多头注意力

Instead of performing a single attention function with -dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values times with different, learned linear projections to and dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding -dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2
不是使用 维的键、值和查询执行单个注意力功能,我们发现用不同的、学习到的线性投影将查询、键和值线性投影 次到分别 维更有益。然后,在这些投影版本的查询、键和值上并行执行注意力功能,产生 维输出值。这些值被连接起来,再次投影,得到最终的值,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同时关注不同位置的不同表示子空间中的信息。使用单个注意力头,平均会受到抑制。
Where the projections are parameter matrices and
投影是参数矩阵
In this work we employ parallel attention layers, or heads. For each of these we use . Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.
在这项工作中,我们使用 个并行注意力层,或头部。对于每个我们使用 。由于每个头部的降维,总的计算成本与具有完整维度的单头部注意力相似。

3.2.3 Applications of Attention in our Model
我们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. This mimics the typical encoder-decoder attention mechanisms in sequence-to-sequence models such as [38, 2, 9].
    在“编码器-解码器注意力”层中,查询来自前一个解码器层,而内存键和值来自编码器的输出。这使得解码器中的每个位置都可以关注输入序列中的所有位置。这模拟了序列到序列模型中典型的编码器-解码器注意力机制,比如 [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.
    编码器包含自注意力层。在自注意力层中,所有的键、值和查询都来源于同一个地方,即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前一层中的所有位置。
  • 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 ) all values in the input of the softmax which correspond to illegal connections. See Figure 2
    同样地,解码器中的自注意力层允许解码器中每个位置都可以关注到解码器中截止到该位置的所有位置。我们需要防止解码器中的左向信息流,以保持自回归性质。我们通过在缩放点积注意力中屏蔽(将其设置为 )与非法连接对应的 softmax 输入中的所有值来实现这一点。请参见图 2。

3.3 Position-wise Feed-Forward Networks
3.3 位置逐层前馈网络

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,我们的编码器(encoder)和解码器(decoder)中的每个层都包含一个全连接前馈网络,该网络独立地作用于每个位置,并且每个位置的操作是相同的。该网络包含两个线性变换,并在中间使用 ReLU 激活函数。
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 , and the inner-layer has dimensionality .
在不同位置之间,线性变换相同,但从一层到另一层使用不同的参数。另一种描述这一点的方式是两个卷积,核大小为 1。输入和输出的维度为 ,内层的维度为

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 . 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 .
与其他序列转导模型类似,我们使用学习到的嵌入来将输入标记和输出标记转换为 维的向量。我们还使用通常的学习线性变换和 softmax 函数来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 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

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
由于我们的模型不包含任何重复和卷积,为了使模型利用序列的顺序,我们必须注入一些关于序列的相对或绝对位置的信息
Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. is the sequence length, is the representation dimension, is the kernel size of convolutions and the size of the neighborhood in restricted self-attention.
表 1:不同层类型的最大路径长度、每层复杂度和最小顺序操作数量。 是序列长度, 是表示维度, 是卷积的核心大小, 是受限自注意力中邻域的大小。
Layer Type Complexity per Layer 每层的复杂性
Sequential
Operations
Maximum Path Length 最大路径长度
Self-Attention
Recurrent
Convolutional
Self-Attention (restricted)
自注意力(受限)
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 as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].
序列中的令牌。为此,我们在编码器和解码器堆栈的底部添加了“位置编码”到输入嵌入中。位置编码与嵌入具有相同的维度 ,因此可以进行相加。关于位置编码,有许多选择,可以是学习的也可以是固定的 [9]。
In this work, we use sine and cosine functions of different frequencies:
在这个工作中,我们使用不同频率的正弦和余弦函数:
where pos is the position and is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from to . 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 can be represented as a linear function of .
其中 pos 是位置, 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 形成一个几何级数。我们选择这个函数是因为我们假设它会让模型轻松学习通过相对位置进行注意力,因为对于任何固定的偏移, 可以表示为 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的位置嵌入[9],发现两个版本产生的结果几乎相同(见表 3 行(E))。我们选择了正弦版本,因为它可能使模型能够推广到比训练过程中遇到的序列长度更长的情况。

4 Why Self-Attention 4 自注意力的原因

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 to another sequence of equal length , with , 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,我们将自注意力层的各个方面与常用于将一个变长符号表示序列 映射到另一个等长序列 的循环和卷积层进行比较,这些层通常用于典型序列传导编码器或解码器中的隐藏层等,通过自注意力的使用动机,我们考虑三个期望条件。
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.
其一是每层的总计算复杂度。另一方面是可以并行计算的量,由最少需要的顺序操作数量来衡量。
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.
第三个是网络中长程依赖之间的路径长度。学习长程依赖是许多序列转换任务中的一个关键挑战。影响学习这类依赖能力的一个关键因素是信号在网络中前向和后向传递所需经过的路径长度。在输入和输出序列的任意位置之间的这些路径越短,学习长程依赖就越容易[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 sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence length is smaller than the representation dimensionality , 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 in
如表 1 所示,自注意力层将所有位置连接起来,执行固定数量的顺序操作,而循环层需要 个顺序操作。就计算复杂度而言,当序列长度 小于表示维度 时,自注意力层比循环层更快,这在最先进的机器翻译模型中经常出现,比如单词片段[38]和字节对[31]表示法。为了提高处理非常长序列任务的计算性能,自注意力可以限制考虑输入序列周围大小为 的邻域。

the input sequence centered around the respective output position. This would increase the maximum path length to . We plan to investigate this approach further in future work.
围绕各自的输出位置。这将增加最大路径长度到 。我们计划在未来的工作中进一步研究这种方法。
A single convolutional layer with kernel width does not connect all pairs of input and output positions. Doing so requires a stack of convolutional layers in the case of contiguous kernels, or 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 . Separable convolutions [6], however, decrease the complexity considerably, to . Even with , 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.
一个宽度为 的单个卷积层不能连接所有输入和输出位置的所有对。为了实现这一点,在连续卷积核的情况下需要一堆 个卷积层,或者在扩张卷积的情况下需要 个[18],这会增加网络中任意两个位置之间最长路径的长度。卷积层通常比循环层更昂贵,比例为 。然而,可分离卷积[6]可以大大降低复杂性,降至 。然而,即使使用 ,可分离卷积的复杂性仍等于自注意力层和逐点前馈层的组合,这是我们模型采用的方法。
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.
作为一个附带的好处,自注意力可以产生更具可解释性的模型。我们检查了我们模型的注意力分布,并在附录中呈现和讨论示例。不仅个别注意力头明显学会执行不同的任务,许多似乎表现出与句子的句法和语义结构相关的行为。

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 训练数据和分批处理

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,其具有约 37000 个令牌的共享源目标词汇。 对于英法语,我们使用了大得多的 WMT 2014 英法数据集,其中包含 36M 句子,并将令牌拆分为一个 32000 个词汇量的词块。 句子对按近似序列长度分组在一起。 每个训练批次包含一组包含大约 25000 个源令牌和 25000 个目标令牌的句子对。

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).
我们在一个拥有 8 个 NVIDIA P100 GPU 的计算机上训练了我们的模型。对于我们的基础模型,使用了整篇论文中描述的超参数,每个训练步骤需要大约 0.4 秒。我们将基础模型训练了总共 100,000 步,或者 12 小时。对于我们的大型模型(在表 3 的最后一行有描述),步骤时间为 1.0 秒。大型模型进行了 300,000 步训练(3.5 天)。

5.3 Optimizer 5.3 优化器

We used the Adam optimizer [20] with and . We varied the learning rate over the course of training, according to the formula:
我们使用了 Adam 优化器[20],设置了学习率为 。在训练过程中,我们根据公式调整学习率:
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 .
这相当于在前 warmup_steps 训练步骤中线性增加学习率,然后在之后的步骤中按步数的倒数平方根成比例减少。我们使用了 warmup_steps

5.4 Regularization 5.4 正则化

We employ three types of regularization during training:
我们在训练过程中使用三种类型的正则化:
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 .
残差丢弃:我们在每个子层的输出之前应用丢弃[33],然后将其添加到子层的输入并进行归一化。此外,我们还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丢弃。对于基础模型,我们使用速率为
Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the English-to-German and English-to-French newstest2014 tests at a fraction of the training cost.
表 2:Transformer 在英语到德语和英语到法语 newstest2014 测试中取得比以往最先进模型更好的 BLEU 分数,且训练成本仅为一小部分。
Model BLEU Training Cost (FLOPs) 训练成本(FLOPs)
EN-DE EN-FR EN-DE EN-FR
ByteNet [18] 23.75
Deep-Att + PosUnk [39]
深度注意力 + PosUnk [39]
39.2
GNMT + RL [38] 24.6 39.92
ConvS2S [9] 25.16 40.46
MoE [32] 26.03 40.56
Deep-Att + PosUnk Ensemble [39]
深度注意力 + 位置未知合集 [39]
40.4
GNMT + RL Ensemble [38]
GNMT + RL 合集 [38]
26.30 41.16
ConvS2S Ensemble [9] ConvS2S 集合[9] 26.36
Transformer (base model)
Transformer(基础模型)
27.3 38.1
Transformer (big) Transformer(大型)
Label Smoothing During training, we employed label smoothing of value [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.
在训练过程中,我们采用了值为 的标签平滑[36]。这会降低困惑度,因为模型学会更加不确定,但会提高准确性和 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中,大型变压器模型(表 2 中的 Transformer(big))的表现优于先前报告的最佳模型(包括集成模型)超过 2.0 BLEU,建立了新的最先进的 BLEU 得分 28.4。该模型的配置列在表 3 的底部行中。在 8 个 P100 GPU 上训练耗时 3.5 天。即使是我们的基础模型也超过了以前发布的所有模型和集成模型,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 the training cost of the previous state-of-the-art model. The Transformer (big) model trained for English-to-French used dropout rate , instead of 0.3 .
在 WMT 2014 年的英法翻译任务中,我们的大模型取得了 41.0 的 BLEU 分数,超越了之前所有已发布的单一模型,在比之前最先进模型的训练成本少于 的情况下。英法转换器(大)模型使用了 的辍学率,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 [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,我们使用了通过对最近的 5 个检查点取平均得到的单个模型,这些检查点每 10 分钟写入一次。对于大模型,我们取最近的 20 个检查点的平均值。我们使用了带有大小为 4 的波束搜索和长度惩罚 [38]。这些超参数是在开发集上进行实验后选择的。在推理过程中,我们将最大输出长度设置为输入长度加 50,但在可能的情况下提前终止 [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
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 的持续单精度浮点容量的估计相乘来估计训练模型所使用的浮点运算次数

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,我们以不同方式改变基础模型,测量在开发集 newstest2013 上英语到德语翻译性能的变化。我们使用前一节中描述的 beam search,但没有使用检查点平均。我们将这些结果呈现在表 3 中。
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 worse than the best setting, quality also drops off with too many heads.
在表 3 的(A)行中,我们改变注意力头的数量以及注意力键和值的维度,保持计算量恒定,如第 3.2.2 节所述。虽然单头注意力比最佳设置差 ,但过多的头也会导致质量下降。
Table 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base model. All metrics are on the English-to-German translation development set, newstest2013. Listed perplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to per-word perplexities.
表 3: Transformer 架构的变种。未列出的数值与基础模型相同。所有指标都是在英语到德语翻译的开发集 newstest2013 上计算的,列出的困惑度是每个词片段的困惑度,根据我们的字节对编码(Byte-pair encoding)计算,不应与每个词的困惑度进行比较。
train
steps
PPL
(dev)
BLEU
params
base 6 512 2048 8 64 64 0.1 0.1 4.92 25.8 65
(A) 1 512 512 5.29 24.9
4 128 128 5.00 25.5
16 32 32 4.91 25.8
32 16 16 5.01 25.4
(B) 16 5.16 25.1 58
32 5.01 25.4 60
(C) 2 6.11 23.7 36
4 5.19 25.3 50
8 4.88 25.5 80
256 32 32 5.75 24.5 28
1024 128 128 4.66 26.0 168
1024 5.12 25.4 53
4096 4.75 26.2 90
(D) 0.0 5.77 24.6
0.2 4.95 25.5
0.0 4.67 25.3
0.2 5.47 25.7
(E) positional embedding instead of sinusoids
使用位置嵌入代替正弦波
4.92 25.7
big 6 1024 4096 16 0.3 4.33 26.4 213
Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)
表 4:Transformer 在英语成分解析中表现良好(结果在 WSJ 第 23 节中)
Parser Training WSJ 23 F1
Vinyals & Kaiser el al. (2014) [37]
Vinyals&Kaiser 等人(2014 年)[37]
WSJ only, discriminative
仅限 WSJ,具有区分性
88.3
Petrov et al. (2006) [29]
Petrov 等(2006)[29]
WSJ only, discriminative
仅限 WSJ,具有区分性
90.4
Zhu et al. (2013)[40]
朱等人(2013)[40]
WSJ only, discriminative
仅限 WSJ,具有区分性
90.4
Dyer et al. (2016)[8]
2016 年的 Dyer 等人[8]
WSJ only, discriminative
仅限 WSJ,具有区分性
91.7
Transformer (4 layers) 变压器(4 层) WSJ only, discriminative
仅限 WSJ,具有区分性
91.3
Zhu et al. (2013)[40]
朱等人(2013)[40]
semi-supervised 91.3
Huang & Harper (2009)[14]
黄和哈珀(2009)[14]
semi-supervised 91.3
McClosky et al. (2006) [26]
McClosky 等人(2006 年)[26]
semi-supervised 92.1
Vinyals & Kaiser el al. (2014) [37]
Vinyals&Kaiser 等人(2014 年)[37]
semi-supervised 92.1
Transformer (4 layers) 变压器(4 层) semi-supervised 92.7
Luong et al. (2015)[23]
2015 年的 Luong 等人[23]
multi-task 93.0
Dyer et al. (2016)[8]
2016 年的 Dyer 等人[8]
generative 93.3
In Table 3 rows (B), we observe that reducing the attention key size 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 (C) 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,我们观察到减小注意力键大小 会损害模型质量。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会更有益。我们进一步观察到在行(C)和(D)中,更大的模型表现更好,丢弃对于避免过拟合非常有帮助。在行(E)中,我们用学习的位置嵌入[9]替换了我们的正弦位置编码,并观察到与基础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
为了评估 Transformer 是否能够推广到其他任务,我们在英语成分解析上进行了实验。这项任务提出了特定的挑战:输出受到强烈的结构约束,并且明显比输入更长。此外,在小数据情况下,RNN 序列到序列模型尚未能够达到最先进的结果[37]。

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].
注意力函数可以被描述为将查询和一组键值对映射到输出,其中查询、键、值和输出都是向量。输出被计算为值的加权和,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。
We trained a 4-layer transformer with 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 sentences [37]. We used a vocabulary of tokens for the WSJ only setting and a vocabulary of tokens for the semi-supervised setting.
我们使用 Penn Treebank 的《华尔街日报》(WSJ) 部分进行了一个包含 的 4 层 transformer 训练 [25],大约有 40K 训练句子。我们还在半监督设置下进行训练,使用了更大的高置信度和 BerkleyParser 语料库,其中包含大约 句子 [37]。我们在 WSJ 单独设置下使用了一个包含