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.

Recurrent neural networks, long short-term memory and gated recurrent 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. Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures. 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.

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. In all but a few cases, however, such attention mechanisms are used in conjunction with a recurrent network. 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.

The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU, ByteNet and ConvS2S, 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. 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.

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. 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 sequence-aligned RNNs or convolution.

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 around each of the two sub-layers, followed by layer normalization. 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.

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 the 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.

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. We call our particular attention "Scaled Dot-Product Attention". 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 the square root of d_k, and apply a softmax function to obtain the weights on the values: Attention(Q, K, V) = softmax(QK^T / sqrt(d_k))V.

The two most commonly used attention functions are additive attention and dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm except for the scaling factor. 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. 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 1/sqrt(d_k).

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, 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. Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. We employ h = 8 parallel attention layers, or heads. For each of these we use d_k = d_v = 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.

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: FFN(x) = max(0, xW_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. The dimensionality of input and output is d_model = 512, and the inner-layer has dimensionality d_ff = 2048.

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. In the embedding layers, we multiply those weights by sqrt(d_model).

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. We use sine and cosine functions of different frequencies: PE(pos, 2i) = sin(pos/10000^(2i/d_model)) and PE(pos, 2i+1) = cos(pos/10000^(2i/d_model)). 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). We also experimented with using learned positional embeddings, and found that the two versions produced nearly identical results.

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. We consider three desiderata: the total computational complexity per layer, the amount of computation that can be parallelized (measured by the minimum number of sequential operations required), and 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.

A self-attention layer connects all positions with a constant number of sequentially executed operations (O(1)), 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. For very long sequences, self-attention could be restricted to considering only a neighborhood of size r in the input sequence, which would increase the maximum path length to O(n/r). As a side benefit, self-attention could yield more interpretable models: we inspect attention distributions from our models and present and discuss examples in the appendix. Individual attention heads clearly learn to perform different tasks, many of which appear to relate to the syntactic and semantic structure of the sentences.

We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs. For English-French, we used the significantly larger WMT 2014 English-French dataset consisting of 36M sentences. Sentences were encoded using byte-pair encoding, which has a shared source-target vocabulary of about 37000 tokens for English-German and 32000 tokens for English-French. We trained on one machine with 8 NVIDIA P100 GPUs. For our base models, 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, step time was 1.0 seconds. The big models were trained for 300,000 steps (3.5 days).

We used the Adam optimizer with beta_1 = 0.9, beta_2 = 0.98 and epsilon = 10^{-9}. We varied the learning rate over the course of training according to a formula that increases the learning rate linearly for the first warmup_steps = 4000 training steps, and decreases it thereafter proportionally to the inverse square root of the step number. We employed three types of regularization: Residual Dropout with a rate of P_drop = 0.1 applied to the output of each sub-layer and to the sums of the embeddings and positional encodings; attention dropout; and Label Smoothing of value epsilon_ls = 0.1. Label smoothing hurt perplexity but improved accuracy and BLEU score.

On the WMT 2014 English-to-German translation task, the big transformer model 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. On the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.8, outperforming all of the previously published single models, at less than 1/4 the training cost of the previous state-of-the-art model.

In Table 3, we vary various components of the Transformer to evaluate their importance. Row (A) shows that single-head attention is 0.9 BLEU worse than the best setting, while quality also drops off with too many heads. This confirms the value of multi-head attention. Row (B) indicates that reducing the attention key size d_k hurts model quality, suggesting that determining compatibility is not easy and that a more sophisticated compatibility function than dot product may be beneficial. Rows (C) and (D) demonstrate that 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, and observe nearly identical results to the base model.

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. Despite the lack of task-specific tuning, the results show that the Transformer with 4 layers achieves a semi-supervised F1 score of 92.7, outperforming the BerkeleyParser even when training only on the WSJ training set of approximately 40K sentences. This demonstrates that the Transformer can generalize well to English constituency parsing, a task with output noticeably different from machine translation.

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. 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.

We are excited about the future of attention-based models and 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 goal of ours. The code we used to train and evaluate our models is available at github.com/tensorflow/tensor2tensor.