这是用户在 2024-6-12 12:38 为 https://nlp.seas.harvard.edu/annotated-transformer/#background 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?

The Annotated Transformer
带注释的变压器

Attention is All You Need
注意力就是你所需要的一切

The Transformer has been on a lot of people’s minds over the last year five years. This post presents an annotated version of the paper in the form of a line-by-line implementation. It reorders and deletes some sections from the original paper and adds comments throughout. This document itself is a working notebook, and should be a completely usable implementation. Code is available here.
在过去的一年五年里,变形金刚一直在很多人的脑海中。这篇文章以逐行实现的形式介绍了该论文的注释版本。它重新排序并删除了原始论文中的某些部分,并在整个过程中添加了注释。本文档本身是一个工作笔记本,应该是一个完全可用的实现。代码可在此处获得。

Table of Contents  目录

Prelims 预赛

Skip 

# !pip install -r requirements.txt
# # Uncomment for colab
# #
# !pip install -q torchdata==0.3.0 torchtext==0.12 spacy==3.2 altair GPUtil
# !python -m spacy download de_core_news_sm
# !python -m spacy download en_core_web_sm
import os
from os.path import exists
import torch
import torch.nn as nn
from torch.nn.functional import log_softmax, pad
import math
import copy
import time
from torch.optim.lr_scheduler import LambdaLR
import pandas as pd
import altair as alt
from torchtext.data.functional import to_map_style_dataset
from torch.utils.data import DataLoader
from torchtext.vocab import build_vocab_from_iterator
import torchtext.datasets as datasets
import spacy
import GPUtil
import warnings
from torch.utils.data.distributed import DistributedSampler
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
# Set to False to skip notebook execution (e.g. for debugging) warnings.filterwarnings("ignore") RUN_EXAMPLES = True
# Some convenience helper functions used throughout the notebook


def is_interactive_notebook():
    return __name__ == "__main__"


def show_example(fn, args=[]):
    if __name__ == "__main__" and RUN_EXAMPLES:
        return fn(*args)


def execute_example(fn, args=[]):
    if __name__ == "__main__" and RUN_EXAMPLES:
        fn(*args)


class DummyOptimizer(torch.optim.Optimizer):
    def __init__(self):
        self.param_groups = [{"lr": 0}]
        None

    def step(self):
        None

    def zero_grad(self, set_to_none=False):
        None


class DummyScheduler:
    def step(self):
        None

My comments are blockquoted. The main text is all from the paper itself.
我的评论被屏蔽了。正文全部来自论文本身。

Background 背景

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.
减少顺序计算的目标也构成了扩展神经 GPU、ByteNet 和 ConvS2S 的基础,所有这些都使用卷积神经网络作为基本构建块,并行计算所有输入和输出位置的隐藏表示。在这些模型中,关联来自两个任意输入或输出位置的信号所需的操作数量随着位置之间的距离而增加,ConvS2S 呈线性增长,ByteNet 呈对数增长。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,这被减少到恒定数量的操作,尽管代价是由于平均注意力加权位置而降低了有效分辨率,我们用多头注意力抵消了这种影响。

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. 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.
自注意力,有时称为内注意力,是一种将单个序列的不同位置联系起来的注意机制,以计算序列的表示。自我注意力已成功用于各种任务,包括阅读理解、抽象总结、文本蕴涵和学习与任务无关的句子表示。端到端记忆网络基于递归注意力机制,而不是序列对齐的递归,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。

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.
然而,据我们所知,Transformer 是第一个完全依靠自注意力来计算其输入和输出表示的转导模型,而无需使用序列对齐的 RNN 或卷积。

Part 1: Model Architecture
第 1 部分:模型架构

Model Architecture 模型架构

Most competitive neural sequence transduction models have an encoder-decoder structure (cite). Here, the encoder maps an input sequence of symbol representations (x1,...,xn)(x_1, ..., x_n) to a sequence of continuous representations z=(z1,...,zn)\mathbf{z} = (z_1, ..., z_n). Given z\mathbf{z}, the decoder then generates an output sequence (y1,...,ym)(y_1,...,y_m) of symbols one element at a time. At each step the model is auto-regressive (cite), 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都具有编码器-解码器结构 (cite)。在这里,编码器将 (x1,...,xn)(x_1, ..., x_n) 符号表示的输入序列映射到连续表示序列 z=(z1,...,zn)\mathbf{z} = (z_1, ..., z_n) 。给定 z\mathbf{z} ,然后解码器一次生成一个符号的输出序列 (y1,...,ym)(y_1,...,y_m) 。在每一步中,模型都是自动回归的(引用),在生成下一个步骤时使用先前生成的符号作为附加输入。

class EncoderDecoder(nn.Module):
    """
    A standard Encoder-Decoder architecture. Base for this and many
    other models.
    """
def __init__(self, encoder, decoder, src_embed, tgt_embed, generator): super(EncoderDecoder, self).__init__() self.encoder = encoder self.decoder = decoder self.src_embed = src_embed self.tgt_embed = tgt_embed self.generator = generator
def forward(self, src, tgt, src_mask, tgt_mask): "Take in and process masked src and target sequences." return self.decode(self.encode(src, src_mask), src_mask, tgt, tgt_mask)
def encode(self, src, src_mask): return self.encoder(self.src_embed(src), src_mask)
def decode(self, memory, src_mask, tgt, tgt_mask): return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask)
class Generator(nn.Module):
    "Define standard linear + softmax generation step."
def __init__(self, d_model, vocab): super(Generator, self).__init__() self.proj = nn.Linear(d_model, vocab)
def forward(self, x): return log_softmax(self.proj(x), dim=-1)

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,对编码器和解码器使用堆叠的自注意力层和逐点全连接层,分别如图 1 的左半部分和右半部分所示。

Encoder and Decoder Stacks
编码器和解码器堆栈

Encoder 编码器

The encoder is composed of a stack of N=6N=6 identical layers.
编码器由 N=6N=6 一堆相同的层组成。

def clones(module, N):
    "Produce N identical layers."
    return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])
class Encoder(nn.Module):
    "Core encoder is a stack of N layers"
def __init__(self, layer, N): super(Encoder, self).__init__() self.layers = clones(layer, N) self.norm = LayerNorm(layer.size)
def forward(self, x, mask): "Pass the input (and mask) through each layer in turn." for layer in self.layers: x = layer(x, mask) return self.norm(x)

We employ a residual connection (cite) around each of the two sub-layers, followed by layer normalization (cite).
我们在两个子层中的每一个周围都采用残差连接(cite),然后进行层归一化(cite)。

class LayerNorm(nn.Module):
    "Construct a layernorm module (See citation for details)."
def __init__(self, features, eps=1e-6): super(LayerNorm, self).__init__() self.a_2 = nn.Parameter(torch.ones(features)) self.b_2 = nn.Parameter(torch.zeros(features)) self.eps = eps
def forward(self, x): mean = x.mean(-1, keepdim=True) std = x.std(-1, keepdim=True) return self.a_2 * (x - mean) / (std + self.eps) + self.b_2

That is, the output of each sub-layer is LayerNorm(x+Sublayer(x))\mathrm{LayerNorm}(x + \mathrm{Sublayer}(x)), where Sublayer(x)\mathrm{Sublayer}(x) is the function implemented by the sub-layer itself. We apply dropout (cite) to the output of each sub-layer, before it is added to the sub-layer input and normalized.
也就是说,每个子层的输出是 LayerNorm(x+Sublayer(x))\mathrm{LayerNorm}(x + \mathrm{Sublayer}(x)) ,其中 Sublayer(x)\mathrm{Sublayer}(x) 是子层本身实现的函数。我们将dropout(引用)应用于每个子层的输出,然后再将其添加到子层输入并归一化。

To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension dmodel=512d_{\text{model}}=512.
为了促进这些残差连接,模型中的所有子层以及嵌入层都会生成 Dimension dmodel=512d_{\text{model}}=512 的输出。

class SublayerConnection(nn.Module):
    """
    A residual connection followed by a layer norm.
    Note for code simplicity the norm is first as opposed to last.
    """
def __init__(self, size, dropout): super(SublayerConnection, self).__init__() self.norm = LayerNorm(size) self.dropout = nn.Dropout(dropout)
def forward(self, x, sublayer): "Apply residual connection to any sublayer with the same size." return x + self.dropout(sublayer(self.norm(x)))

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.
每个图层有两个子图层。第一种是多头自注意力机制,第二种是简单的、位置全连接的前馈网络。

class EncoderLayer(nn.Module):
    "Encoder is made up of self-attn and feed forward (defined below)"
def __init__(self, size, self_attn, feed_forward, dropout): super(EncoderLayer, self).__init__() self.self_attn = self_attn self.feed_forward = feed_forward self.sublayer = clones(SublayerConnection(size, dropout), 2) self.size = size
def forward(self, x, mask): "Follow Figure 1 (left) for connections." x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask)) return self.sublayer[1](x, self.feed_forward)

Decoder 译码器

The decoder is also composed of a stack of N=6N=6 identical layers.
解码器也由 N=6N=6 一堆相同的层组成。

class Decoder(nn.Module):
    "Generic N layer decoder with masking."
def __init__(self, layer, N): super(Decoder, self).__init__() self.layers = clones(layer, N) self.norm = LayerNorm(layer.size)
def forward(self, x, memory, src_mask, tgt_mask): for layer in self.layers: x = layer(x, memory, src_mask, tgt_mask) return self.norm(x)

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.
除了每个编码器层中的两个子层外,解码器还插入第三个子层,该子层对编码器堆栈的输出执行多头注意力。与编码器类似,我们在每个子层周围使用残差连接,然后进行层归一化。

class DecoderLayer(nn.Module):
    "Decoder is made of self-attn, src-attn, and feed forward (defined below)"
def __init__(self, size, self_attn, src_attn, feed_forward, dropout): super(DecoderLayer, self).__init__() self.size = size self.self_attn = self_attn self.src_attn = src_attn self.feed_forward = feed_forward self.sublayer = clones(SublayerConnection(size, dropout), 3)
def forward(self, x, memory, src_mask, tgt_mask): "Follow Figure 1 (right) for connections." m = memory x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask)) x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask)) return self.sublayer[2](x, self.feed_forward)

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 ii can depend only on the known outputs at positions less than ii.
我们还修改了解码器堆栈中的自注意力子层,以防止位置关注后续位置。这种掩码,再加上输出嵌入偏移一个位置的事实,确保了对位置 ii 的预测只能依赖于小于 ii 的已知输出。

def subsequent_mask(size):
    "Mask out subsequent positions."
    attn_shape = (1, size, size)
    subsequent_mask = torch.triu(torch.ones(attn_shape), diagonal=1).type(
        torch.uint8
    )
    return subsequent_mask == 0

Below the attention mask shows the position each tgt word (row) is allowed to look at (column). Words are blocked for attending to future words during training.
注意力掩码下方显示允许每个 tgt 字(行)查看的位置(列)。在训练期间,单词会被阻止以处理未来的单词。

def example_mask():
    LS_data = pd.concat(
        [
            pd.DataFrame(
                {
                    "Subsequent Mask": subsequent_mask(20)[0][x, y].flatten(),
                    "Window": y,
                    "Masking": x,
                }
            )
            for y in range(20)
            for x in range(20)
        ]
    )
return ( alt.Chart(LS_data) .mark_rect() .properties(height=250, width=250) .encode( alt.X("Window:O"), alt.Y("Masking:O"), alt.Color("Subsequent Mask:Q", scale=alt.Scale(scheme="viridis")), ) .interactive() )
show_example(example_mask)

Attention 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输出,其中查询、键、值和输出都是向量。输出计算为值的加权总和,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。

We call our particular attention “Scaled Dot-Product Attention”. The input consists of queries and keys of dimension dkd_k, and values of dimension dvd_v. We compute the dot products of the query with all keys, divide each by dk\sqrt{d_k}, and apply a softmax function to obtain the weights on the values.
我们称我们的特别关注为“缩放点积关注”。输入由 dimension dkd_k 的查询和键以及 dimension 的值组成 dvd_v 。我们计算包含所有键的查询的点积,将每个键除以 dk\sqrt{d_k} ,并应用 softmax 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 QQ. The keys and values are also packed together into matrices KK and VV. We compute the matrix of outputs as:
在实践中,我们同时计算一组查询的注意力函数,并将它们打包成一个矩阵 QQ 。键和值也被打包到矩阵 KKVV .我们将输出矩阵计算为:

Attention(Q,K,V)=softmax(QKTdk)V \mathrm{Attention}(Q, K, V) = \mathrm{softmax}(\frac{QK^T}{\sqrt{d_k}})V

def attention(query, key, value, mask=None, dropout=None):
    "Compute 'Scaled Dot Product Attention'"
    d_k = query.size(-1)
    scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
    if mask is not None:
        scores = scores.masked_fill(mask == 0, -1e9)
    p_attn = scores.softmax(dim=-1)
    if dropout is not None:
        p_attn = dropout(p_attn)
    return torch.matmul(p_attn, value), p_attn

The two most commonly used attention functions are additive attention (cite), and dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor of 1dk\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.
两个最常用的注意力函数是加法注意力(引用)和点积(乘法)注意力。点积注意力与我们的算法相同,只是比例因子为 1dk\frac{1}{\sqrt{d_k}} 。加法注意力使用具有单个隐藏层的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更快、更节省空间,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 dkd_k the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of dkd_k (cite). We suspect that for large values of dkd_k, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients (To illustrate why the dot products get large, assume that the components of qq and kk are independent random variables with mean 00 and variance 11. Then their dot product, qk=i=1dkqikiq \cdot k = \sum_{i=1}^{d_k} q_ik_i, has mean 00 and variance dkd_k.). To counteract this effect, we scale the dot products by 1dk\frac{1}{\sqrt{d_k}}.
虽然对于两种机制的小值 dkd_k 表现相似,但加法注意力优于点积注意力,而无需缩放较大的值 dkd_k (cite)。我们怀疑,对于较大的值 dkd_k ,点积的大小会变大,从而将softmax函数推入梯度极小的区域(为了说明为什么点积变大,假设 qqkk 的分量是具有均值 00 和方差的 11 独立随机变量。然后他们的点积 qk=i=1dkqikiq \cdot k = \sum_{i=1}^{d_k} q_ik_i , 具有均值 00 dkd_k 和方差。为了抵消这种影响,我们按 1dk\frac{1}{\sqrt{d_k}} 缩放点积。

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共同关注来自不同位置的不同表示子空间的信息。对于单个注意力头,平均会抑制这一点。

MultiHead(Q,K,V)=Concat(head1,...,headh)WOwhere headi=Attention(QWiQ,KWiK,VWiV) \mathrm{MultiHead}(Q, K, V) = \mathrm{Concat}(\mathrm{head_1}, ..., \mathrm{head_h})W^O \\ \text{where}~\mathrm{head_i} = \mathrm{Attention}(QW^Q_i, KW^K_i, VW^V_i)

Where the projections are parameter matrices WiQRdmodel×dkW^Q_i \in \mathbb{R}^{d_{\text{model}} \times d_k}, WiKRdmodel×dkW^K_i \in \mathbb{R}^{d_{\text{model}} \times d_k}, WiVRdmodel×dvW^V_i \in \mathbb{R}^{d_{\text{model}} \times d_v} and WORhdv×dmodelW^O \in \mathbb{R}^{hd_v \times d_{\text{model}}}.
其中投影是参数矩阵 WiQRdmodel×dkW^Q_i \in \mathbb{R}^{d_{\text{model}} \times d_k}WiKRdmodel×dkW^K_i \in \mathbb{R}^{d_{\text{model}} \times d_k}WiVRdmodel×dvW^V_i \in \mathbb{R}^{d_{\text{model}} \times d_v} WORhdv×dmodelW^O \in \mathbb{R}^{hd_v \times d_{\text{model}}}

In this work we employ h=8h=8 parallel attention layers, or heads. For each of these we use dk=dv=dmodel/h=64d_k=d_v=d_{\text{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.
在这项工作中,我们采用了 h=8h=8 平行的注意力层或头部。对于其中的每一个,我们使用 dk=dv=dmodel/h=64d_k=d_v=d_{\text{model}}/h=64 .由于每个头的尺寸都缩小了,总计算成本与全维的单头注意力相似。

class MultiHeadedAttention(nn.Module):
    def __init__(self, h, d_model, dropout=0.1):
        "Take in model size and number of heads."
        super(MultiHeadedAttention, self).__init__()
        assert d_model % h == 0
        # We assume d_v always equals d_k
        self.d_k = d_model // h
        self.h = h
        self.linears = clones(nn.Linear(d_model, d_model), 4)
        self.attn = None
        self.dropout = nn.Dropout(p=dropout)

    def forward(self, query, key, value, mask=None):
        "Implements Figure 2"
        if mask is not None:
            # Same mask applied to all h heads.
            mask = mask.unsqueeze(1)
        nbatches = query.size(0)

        # 1) Do all the linear projections in batch from d_model => h x d_k
        query, key, value = [
            lin(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)
            for lin, x in zip(self.linears, (query, key, value))
        ]

        # 2) Apply attention on all the projected vectors in batch.
        x, self.attn = attention(
            query, key, value, mask=mask, dropout=self.dropout
        )

        # 3) "Concat" using a view and apply a final linear.
        x = (
            x.transpose(1, 2)
            .contiguous()
            .view(nbatches, -1, self.h * self.d_k)
        )
        del query
        del key
        del value
        return self.linears[-1](x)

Applications of Attention in our Model
注意力在我们的模型中的应用

The Transformer uses multi-head attention in three different ways: 1) 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 (cite).
Transformer 以三种不同的方式使用多头注意力:1)在“编码器-解码器注意力”层中,查询来自前一个解码器层,内存键和值来自编码器的输出。这允许解码器中的每个位置都参与输入序列中的所有位置。这模拟了序列到序列模型(如 (cite))中典型的编码器-解码器注意力机制。

  1. 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,所有键、值和查询都来自同一个位置,在本例中,是编码器中前一层的输出。编码器中的每个位置都可以处理编码器上一层中的所有位置。

  2. 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.
    同样,解码器中的自注意力层允许解码器中的每个位置关注解码器中的所有位置,直到并包括该位置。我们需要防止解码器中的信息向左流动,以保留自动回归属性。我们通过屏蔽(设置为 -\infty )softmax输入中与非法连接相对应的所有值,在缩放点积注意力中实现这一点。

Position-wise Feed-Forward Networks
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.
除了注意力子层之外,我们的编码器和解码器中的每一层都包含一个完全连接的前馈网络,该网络分别以相同的方式应用于每个位置。这包括两个线性变换,中间有一个 ReLU 激活。

FFN(x)=max(0,xW1+b1)W2+b2\mathrm{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. Another way of describing this is as two convolutions with kernel size 1. The dimensionality of input and output is dmodel=512d_{\text{model}}=512, and the inner-layer has dimensionality dff=2048d_{ff}=2048.
虽然线性变换在不同位置是相同的,但它们在层与层之间使用不同的参数。另一种描述方式是两个内核大小为 1 的卷积。输入和输出的维数为 dmodel=512d_{\text{model}}=512 ,内层有维数 dff=2048d_{ff}=2048

class PositionwiseFeedForward(nn.Module):
    "Implements FFN equation."
def __init__(self, d_model, d_ff, dropout=0.1): super(PositionwiseFeedForward, self).__init__() self.w_1 = nn.Linear(d_model, d_ff) self.w_2 = nn.Linear(d_ff, d_model) self.dropout = nn.Dropout(dropout)
def forward(self, x): return self.w_2(self.dropout(self.w_1(x).relu()))

Embeddings and Softmax 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 dmodeld_{\text{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 (cite). In the embedding layers, we multiply those weights by dmodel\sqrt{d_{\text{model}}}.
与其他序列转导模型类似,我们使用学习嵌入将输入标记和输出标记转换为维度向量 dmodeld_{\text{model}} 。我们还使用通常学习的线性变换和 softmax 函数将解码器输出转换为预测的下一个令牌概率。在我们的模型中,我们在两个嵌入层和pre-softmax线性变换之间共享相同的权重矩阵,类似于(cite)。在嵌入层中,我们将这些权重乘以 dmodel\sqrt{d_{\text{model}}}

class Embeddings(nn.Module):
    def __init__(self, d_model, vocab):
        super(Embeddings, self).__init__()
        self.lut = nn.Embedding(vocab, d_model)
        self.d_model = d_model
def forward(self, x): return self.lut(x) * math.sqrt(self.d_model)

Positional Encoding 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. 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 dmodeld_{\text{model}} as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed (cite).
由于我们的模型不包含递归和卷积,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 dmodeld_{\text{model}} 与嵌入具有相同的维度,因此可以将两者相加。位置编码有很多选择,学习的和固定的(引用)。

In this work, we use sine and cosine functions of different frequencies:
在这项工作中,我们使用不同频率的正弦和余弦函数:

PE(pos,2i)=sin(pos/100002i/dmodel)PE_{(pos,2i)} = \sin(pos / 10000^{2i/d_{\text{model}}})

PE(pos,2i+1)=cos(pos/100002i/dmodel)PE_{(pos,2i+1)} = \cos(pos / 10000^{2i/d_{\text{model}}})

where pospos is the position and ii is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from 2π2\pi to 100002π10000 \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 kk, PEpos+kPE_{pos+k} can be represented as a linear function of PEposPE_{pos}.
其中 pospos 是位置, ii 是尺寸。也就是说,位置编码的每个维度都对应于一个正弦曲线。波长形成从 2π2\pi 到 的 100002π10000 \cdot 2\pi 几何级数。我们之所以选择这个函数,是因为我们假设它允许模型很容易地学习按相对位置参加,因为对于任何固定偏移 kk 量, PEpos+kPE_{pos+k} 都可以表示为 的 PEposPE_{pos} 线性函数。

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=0.1P_{drop}=0.1.
此外,我们将 dropout 应用于编码器和解码器堆栈中的嵌入和位置编码的总和。对于基本模型,我们使用 的速率为 Pdrop=0.1P_{drop}=0.1

class PositionalEncoding(nn.Module):
    "Implement the PE function."
def __init__(self, d_model, dropout, max_len=5000): super(PositionalEncoding, self).__init__() self.dropout = nn.Dropout(p=dropout)
# Compute the positional encodings once in log space. pe = torch.zeros(max_len, d_model) position = torch.arange(0, max_len).unsqueeze(1) div_term = torch.exp( torch.arange(0, d_model, 2) * -(math.log(10000.0) / d_model) ) pe[:, 0::2] = torch.sin(position * div_term) pe[:, 1::2] = torch.cos(position * div_term) pe = pe.unsqueeze(0) self.register_buffer("pe", pe)
def forward(self, x): x = x + self.pe[:, : x.size(1)].requires_grad_(False) return self.dropout(x)

Below the positional encoding will add in a sine wave based on position. The frequency and offset of the wave is different for each dimension.
在位置编码下方将添加基于位置的正弦波。每个维度的波的频率和偏移量都不同。

def example_positional():
    pe = PositionalEncoding(20, 0)
    y = pe.forward(torch.zeros(1, 100, 20))
data = pd.concat( [ pd.DataFrame( { "embedding": y[0, :, dim], "dimension": dim, "position": list(range(100)), } ) for dim in [4, 5, 6, 7] ] )
return ( alt.Chart(data) .mark_line() .properties(width=800) .encode(x="position", y="embedding", color="dimension:N") .interactive() )
show_example(example_positional)

We also experimented with using learned positional embeddings (cite) instead, and found that the two versions produced nearly identical results. 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(cite),发现这两个版本产生了几乎相同的结果。我们之所以选择正弦版本,是因为它可能允许模型推断出比训练期间遇到的序列长度更长的序列长度。

Full Model 完整模型

Here we define a function from hyperparameters to a full model.
在这里,我们定义一个从超参数到完整模型的函数。

def make_model(
    src_vocab, tgt_vocab, N=6, d_model=512, d_ff=2048, h=8, dropout=0.1
):
    "Helper: Construct a model from hyperparameters."
    c = copy.deepcopy
    attn = MultiHeadedAttention(h, d_model)
    ff = PositionwiseFeedForward(d_model, d_ff, dropout)
    position = PositionalEncoding(d_model, dropout)
    model = EncoderDecoder(
        Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),
        Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),
        nn.Sequential(Embeddings(d_model, src_vocab), c(position)),
        nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)),
        Generator(d_model, tgt_vocab),
    )
# This was important from their code. # Initialize parameters with Glorot / fan_avg. for p in model.parameters(): if p.dim() > 1: nn.init.xavier_uniform_(p) return model

Inference: 推理:

Here we make a forward step to generate a prediction of the model. We try to use our transformer to memorize the input. As you will see the output is randomly generated due to the fact that the model is not trained yet. In the next tutorial we will build the training function and try to train our model to memorize the numbers from 1 to 10.
在这里,我们向前迈出一步,生成模型的预测。我们尝试使用变压器来记忆输入。正如你所看到的,由于模型尚未训练,输出是随机生成的。在下一个教程中,我们将构建训练函数,并尝试训练我们的模型来记住从 1 到 10 的数字。

def inference_test():
    test_model = make_model(11, 11, 2)
    test_model.eval()
    src = torch.LongTensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]])
    src_mask = torch.ones(1, 1, 10)
memory = test_model.encode(src, src_mask) ys = torch.zeros(1, 1).type_as(src)
for i in range(9): out = test_model.decode( memory, src_mask, ys, subsequent_mask(ys.size(1)).type_as(src.data) ) prob = test_model.generator(out[:, -1]) _, next_word = torch.max(prob, dim=1) next_word = next_word.data[0] ys = torch.cat( [ys, torch.empty(1, 1).type_as(src.data).fill_(next_word)], dim=1 )
print("Example Untrained Model Prediction:", ys)
def run_tests(): for _ in range(10): inference_test()
show_example(run_tests)
Example Untrained Model Prediction: tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]])
Example Untrained Model Prediction: tensor([[0, 3, 4, 4, 4, 4, 4, 4, 4, 4]])
Example Untrained Model Prediction: tensor([[ 0, 10, 10, 10,  3,  2,  5,  7,  9,  6]])
Example Untrained Model Prediction: tensor([[ 0,  4,  3,  6, 10, 10,  2,  6,  2,  2]])
Example Untrained Model Prediction: tensor([[ 0,  9,  0,  1,  5, 10,  1,  5, 10,  6]])
Example Untrained Model Prediction: tensor([[ 0,  1,  5,  1, 10,  1, 10, 10, 10, 10]])
Example Untrained Model Prediction: tensor([[ 0,  1, 10,  9,  9,  9,  9,  9,  1,  5]])
Example Untrained Model Prediction: tensor([[ 0,  3,  1,  5, 10, 10, 10, 10, 10, 10]])
Example Untrained Model Prediction: tensor([[ 0,  3,  5, 10,  5, 10,  4,  2,  4,  2]])
Example Untrained Model Prediction: tensor([[0, 5, 6, 2, 5, 6, 2, 6, 2, 2]])

Part 2: Model Training
第 2 部分:模型训练

Training 训练

This section describes the training regime for our models.
本节介绍模型的训练制度。

We stop for a quick interlude to introduce some of the tools needed to train a standard encoder decoder model. First we define a batch object that holds the src and target sentences for training, as well as constructing the masks.
我们停下来快速介绍一些训练标准编码器解码器模型所需的工具。首先,我们定义一个批处理对象,该对象保存用于训练的 src 和目标句子,以及构造掩码。

Batches and Masking 批处理和遮蔽

class Batch:
    """Object for holding a batch of data with mask during training."""
def __init__(self, src, tgt=None, pad=2): # 2 = <blank> self.src = src self.src_mask = (src != pad).unsqueeze(-2) if tgt is not None: self.tgt = tgt[:, :-1] self.tgt_y = tgt[:, 1:] self.tgt_mask = self.make_std_mask(self.tgt, pad) self.ntokens = (self.tgt_y != pad).data.sum()
@staticmethod def make_std_mask(tgt, pad): "Create a mask to hide padding and future words." tgt_mask = (tgt != pad).unsqueeze(-2) tgt_mask = tgt_mask & subsequent_mask(tgt.size(-1)).type_as( tgt_mask.data ) return tgt_mask

Next we create a generic training and scoring function to keep track of loss. We pass in a generic loss compute function that also handles parameter updates.
接下来,我们创建一个通用的训练和评分函数来跟踪损失。我们传入了一个通用的损失计算函数,该函数也处理参数更新。

Training Loop 训练循环

class TrainState:
    """Track number of steps, examples, and tokens processed"""
step: int = 0 # Steps in the current epoch accum_step: int = 0 # Number of gradient accumulation steps samples: int = 0 # total # of examples used tokens: int = 0 # total # of tokens processed
def run_epoch(
    data_iter,
    model,
    loss_compute,
    optimizer,
    scheduler,
    mode="train",
    accum_iter=1,
    train_state=TrainState(),
):
    """Train a single epoch"""
    start = time.time()
    total_tokens = 0
    total_loss = 0
    tokens = 0
    n_accum = 0
    for i, batch in enumerate(data_iter):
        out = model.forward(
            batch.src, batch.tgt, batch.src_mask, batch.tgt_mask
        )
        loss, loss_node = loss_compute(out, batch.tgt_y, batch.ntokens)
        # loss_node = loss_node / accum_iter
        if mode == "train" or mode == "train+log":
            loss_node.backward()
            train_state.step += 1
            train_state.samples += batch.src.shape[0]
            train_state.tokens += batch.ntokens
            if i % accum_iter == 0:
                optimizer.step()
                optimizer.zero_grad(set_to_none=True)
                n_accum += 1
                train_state.accum_step += 1
            scheduler.step()

        total_loss += loss
        total_tokens += batch.ntokens
        tokens += batch.ntokens
        if i % 40 == 1 and (mode == "train" or mode == "train+log"):
            lr = optimizer.param_groups[0]["lr"]
            elapsed = time.time() - start
            print(
                (
                    "Epoch Step: %6d | Accumulation Step: %3d | Loss: %6.2f "
                    + "| Tokens / Sec: %7.1f | Learning Rate: %6.1e"
                )
                % (i, n_accum, loss / batch.ntokens, tokens / elapsed, lr)
            )
            start = time.time()
            tokens = 0
        del loss
        del loss_node
    return total_loss / total_tokens, train_state

Training Data and Batching
训练数据和批处理

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, which has a shared source-target 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.
我们在标准的 WMT 2014 英德数据集上进行了训练,该数据集由大约 450 万个句子对组成。句子使用字节对编码进行编码,该编码具有大约 37000 个标记的共享源目标词汇表。对于英语-法语,我们使用了更大的 WMT 2014 英语-法语数据集,该数据集由 36M 个句子组成,并将标记拆分为 32000 个单词的词汇表。

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.
句子对按近似序列长度批量组合在一起。每个训练批次包含一组句子对,其中包含大约 25000 个源标记和 25000 个目标标记。

Hardware and Schedule 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, 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 NVIDIA P100 GPU 的机器上训练了我们的模型。对于使用整篇论文中描述的超参数的基本模型,每个训练步骤大约需要 0.4 秒。我们对基础模型进行了总共 100,000 步或 12 小时的训练。对于我们的大型模型,步进时间为 1.0 秒。大型模型被训练了 300,000 步(3.5 天)。

Optimizer 优化

We used the Adam optimizer (cite) with β1=0.9\beta_1=0.9, β2=0.98\beta_2=0.98 and ϵ=109\epsilon=10^{-9}. We varied the learning rate over the course of training, according to the formula:
我们将 Adam 优化器 (cite) 与 β1=0.9\beta_1=0.9β2=0.98\beta_2=0.98 ϵ=109\epsilon=10^{-9} 一起使用。我们根据以下公式在培训过程中改变了学习率:

lrate=dmodel0.5min(step_num0.5,step_numwarmup_steps1.5) lrate = d_{\text{model}}^{-0.5} \cdot \min({step\_num}^{-0.5}, {step\_num} \cdot {warmup\_steps}^{-1.5})

This corresponds to increasing the learning rate linearly for the first warmup_stepswarmup\_steps training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used warmup_steps=4000warmup\_steps=4000.
这相当于在第一个 warmup_stepswarmup\_steps 训练步骤中线性增加学习率,然后根据步骤数的平方反比根按比例降低学习率。我们使用了 warmup_steps=4000warmup\_steps=4000 .

Note: This part is very important. Need to train with this setup of the model.
注意:这部分非常重要。需要使用这种模型设置进行训练。

Example of the curves of this model for different model sizes and for optimization hyperparameters.
此模型针对不同模型大小和优化超参数的曲线示例。

def rate(step, model_size, factor, warmup):
    """
    we have to default the step to 1 for LambdaLR function
    to avoid zero raising to negative power.
    """
    if step == 0:
        step = 1
    return factor * (
        model_size ** (-0.5) * min(step ** (-0.5), step * warmup ** (-1.5))
    )
def example_learning_schedule():
    opts = [
        [512, 1, 4000],  # example 1
        [512, 1, 8000],  # example 2
        [256, 1, 4000],  # example 3
    ]

    dummy_model = torch.nn.Linear(1, 1)
    learning_rates = []

    # we have 3 examples in opts list.
    for idx, example in enumerate(opts):
        # run 20000 epoch for each example
        optimizer = torch.optim.Adam(
            dummy_model.parameters(), lr=1, betas=(0.9, 0.98), eps=1e-9
        )
        lr_scheduler = LambdaLR(
            optimizer=optimizer, lr_lambda=lambda step: rate(step, *example)
        )
        tmp = []
        # take 20K dummy training steps, save the learning rate at each step
        for step in range(20000):
            tmp.append(optimizer.param_groups[0]["lr"])
            optimizer.step()
            lr_scheduler.step()
        learning_rates.append(tmp)

    learning_rates = torch.tensor(learning_rates)

    # Enable altair to handle more than 5000 rows
    alt.data_transformers.disable_max_rows()

    opts_data = pd.concat(
        [
            pd.DataFrame(
                {
                    "Learning Rate": learning_rates[warmup_idx, :],
                    "model_size:warmup": ["512:4000", "512:8000", "256:4000"][
                        warmup_idx
                    ],
                    "step": range(20000),
                }
            )
            for warmup_idx in [0, 1, 2]
        ]
    )

    return (
        alt.Chart(opts_data)
        .mark_line()
        .properties(width=600)
        .encode(x="step", y="Learning Rate", color="model_size:warmup:N")
        .interactive()
    )


example_learning_schedule()

Regularization 正规化

Label Smoothing 标签平滑

During training, we employed label smoothing of value ϵls=0.1\epsilon_{ls}=0.1 (cite). This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.
在培训期间,我们采用了值 ϵls=0.1\epsilon_{ls}=0.1 的标签平滑(引用)。这伤害了困惑,因为模型学会了更加不确定,但提高了准确性和 BLEU 分数。

We implement label smoothing using the KL div loss. Instead of using a one-hot target distribution, we create a distribution that has confidence of the correct word and the rest of the smoothing mass distributed throughout the vocabulary.
我们使用 KL div 损失实现标签平滑。我们没有使用单热目标分布,而是创建一个分布,该分布具有 confidence 正确的单词和分布在整个词汇表中的其余 smoothing 质量。

class LabelSmoothing(nn.Module):
    "Implement label smoothing."
def __init__(self, size, padding_idx, smoothing=0.0): super(LabelSmoothing, self).__init__() self.criterion = nn.KLDivLoss(reduction="sum") self.padding_idx = padding_idx self.confidence = 1.0 - smoothing self.smoothing = smoothing self.size = size self.true_dist = None
def forward(self, x, target): assert x.size(1) == self.size true_dist = x.data.clone() true_dist.fill_(self.smoothing / (self.size - 2)) true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence) true_dist[:, self.padding_idx] = 0 mask = torch.nonzero(target.data == self.padding_idx) if mask.dim() > 0: true_dist.index_fill_(0, mask.squeeze(), 0.0) self.true_dist = true_dist return self.criterion(x, true_dist.clone().detach())

Here we can see an example of how the mass is distributed to the words based on confidence.
在这里,我们可以看到一个示例,说明质量如何基于置信度分配给单词。

# Example of label smoothing.


def example_label_smoothing():
    crit = LabelSmoothing(5, 0, 0.4)
    predict = torch.FloatTensor(
        [
            [0, 0.2, 0.7, 0.1, 0],
            [0, 0.2, 0.7, 0.1, 0],
            [0, 0.2, 0.7, 0.1, 0],
            [0, 0.2, 0.7, 0.1, 0],
            [0, 0.2, 0.7, 0.1, 0],
        ]
    )
    crit(x=predict.log(), target=torch.LongTensor([2, 1, 0, 3, 3]))
    LS_data = pd.concat(
        [
            pd.DataFrame(
                {
                    "target distribution": crit.true_dist[x, y].flatten(),
                    "columns": y,
                    "rows": x,
                }
            )
            for y in range(5)
            for x in range(5)
        ]
    )

    return (
        alt.Chart(LS_data)
        .mark_rect(color="Blue", opacity=1)
        .properties(height=200, width=200)
        .encode(
            alt.X("columns:O", title=None),
            alt.Y("rows:O", title=None),
            alt.Color(
                "target distribution:Q", scale=alt.Scale(scheme="viridis")
            ),
        )
        .interactive()
    )


show_example(example_label_smoothing)

Label smoothing actually starts to penalize the model if it gets very confident about a given choice.
如果模型对给定的选择非常有信心,则标签平滑实际上会开始惩罚模型。


def loss(x, crit): d = x + 3 * 1 predict = torch.FloatTensor([[0, x / d, 1 / d, 1 / d, 1 / d]]) return crit(predict.log(), torch.LongTensor([1])).data
def penalization_visualization(): crit = LabelSmoothing(5, 0, 0.1) loss_data = pd.DataFrame( { "Loss": [loss(x, crit) for x in range(1, 100)], "Steps": list(range(99)), } ).astype("float")
return ( alt.Chart(loss_data) .mark_line() .properties(width=350) .encode( x="Steps", y="Loss", ) .interactive() )
show_example(penalization_visualization)

A First Example 第一个例子

We can begin by trying out a simple copy-task. Given a random set of input symbols from a small vocabulary, the goal is to generate back those same symbols.
我们可以从尝试一个简单的复制任务开始。给定一组来自小词汇表的随机输入符号,目标是生成回这些相同的符号。

Synthetic Data 合成数据

def data_gen(V, batch_size, nbatches):
    "Generate random data for a src-tgt copy task."
    for i in range(nbatches):
        data = torch.randint(1, V, size=(batch_size, 10))
        data[:, 0] = 1
        src = data.requires_grad_(False).clone().detach()
        tgt = data.requires_grad_(False).clone().detach()
        yield Batch(src, tgt, 0)

Loss Computation 损失计算

class SimpleLossCompute:
    "A simple loss compute and train function."
def __init__(self, generator, criterion): self.generator = generator self.criterion = criterion
def __call__(self, x, y, norm): x = self.generator(x) sloss = ( self.criterion( x.contiguous().view(-1, x.size(-1)), y.contiguous().view(-1) ) / norm ) return sloss.data * norm, sloss

Greedy Decoding 贪婪解码

This code predicts a translation using greedy decoding for simplicity.
为简单起见,此代码使用贪婪解码来预测翻译。

def greedy_decode(model, src, src_mask, max_len, start_symbol):
    memory = model.encode(src, src_mask)
    ys = torch.zeros(1, 1).fill_(start_symbol).type_as(src.data)
    for i in range(max_len - 1):
        out = model.decode(
            memory, src_mask, ys, subsequent_mask(ys.size(1)).type_as(src.data)
        )
        prob = model.generator(out[:, -1])
        _, next_word = torch.max(prob, dim=1)
        next_word = next_word.data[0]
        ys = torch.cat(
            [ys, torch.zeros(1, 1).type_as(src.data).fill_(next_word)], dim=1
        )
    return ys
# Train the simple copy task.


def example_simple_model():
    V = 11
    criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0.0)
    model = make_model(V, V, N=2)

    optimizer = torch.optim.Adam(
        model.parameters(), lr=0.5, betas=(0.9, 0.98), eps=1e-9
    )
    lr_scheduler = LambdaLR(
        optimizer=optimizer,
        lr_lambda=lambda step: rate(
            step, model_size=model.src_embed[0].d_model, factor=1.0, warmup=400
        ),
    )

    batch_size = 80
    for epoch in range(20):
        model.train()
        run_epoch(
            data_gen(V, batch_size, 20),
            model,
            SimpleLossCompute(model.generator, criterion),
            optimizer,
            lr_scheduler,
            mode="train",
        )
        model.eval()
        run_epoch(
            data_gen(V, batch_size, 5),
            model,
            SimpleLossCompute(model.generator, criterion),
            DummyOptimizer(),
            DummyScheduler(),
            mode="eval",
        )[0]

    model.eval()
    src = torch.LongTensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])
    max_len = src.shape[1]
    src_mask = torch.ones(1, 1, max_len)
    print(greedy_decode(model, src, src_mask, max_len=max_len, start_symbol=0))


# execute_example(example_simple_model)

Part 3: A Real World Example
第 3 部分:真实世界的例子

Now we consider a real-world example using the Multi30k German-English Translation task. This task is much smaller than the WMT task considered in the paper, but it illustrates the whole system. We also show how to use multi-gpu processing to make it really fast.
现在,我们考虑一个使用 Multi30k 德语-英语翻译任务的真实示例。这个任务比论文中考虑的WMT任务要小得多,但它说明了整个系统。我们还展示了如何使用多 GPU 处理来使其变得非常快。

Data Loading

We will load the dataset using torchtext and spacy for tokenization.

# Load spacy tokenizer models, download them if they haven't been
# downloaded already
def load_tokenizers():
try: spacy_de = spacy.load("de_core_news_sm") except IOError: os.system("python -m spacy download de_core_news_sm") spacy_de = spacy.load("de_core_news_sm")
try: spacy_en = spacy.load("en_core_web_sm") except IOError: os.system("python -m spacy download en_core_web_sm") spacy_en = spacy.load("en_core_web_sm")
return spacy_de, spacy_en
def tokenize(text, tokenizer):
    return [tok.text for tok in tokenizer.tokenizer(text)]
def yield_tokens(data_iter, tokenizer, index): for from_to_tuple in data_iter: yield tokenizer(from_to_tuple[index])


def build_vocabulary(spacy_de, spacy_en):
    def tokenize_de(text):
        return tokenize(text, spacy_de)

    def tokenize_en(text):
        return tokenize(text, spacy_en)

    print("Building German Vocabulary ...")
    train, val, test = datasets.Multi30k(language_pair=("de", "en"))
    vocab_src = build_vocab_from_iterator(
        yield_tokens(train + val + test, tokenize_de, index=0),
        min_freq=2,
        specials=["<s>", "</s>", "<blank>", "<unk>"],
    )

    print("Building English Vocabulary ...")
    train, val, test = datasets.Multi30k(language_pair=("de", "en"))
    vocab_tgt = build_vocab_from_iterator(
        yield_tokens(train + val + test, tokenize_en, index=1),
        min_freq=2,
        specials=["<s>", "</s>", "<blank>", "<unk>"],
    )

    vocab_src.set_default_index(vocab_src["<unk>"])
    vocab_tgt.set_default_index(vocab_tgt["<unk>"])

    return vocab_src, vocab_tgt


def load_vocab(spacy_de, spacy_en):
    if not exists("vocab.pt"):
        vocab_src, vocab_tgt = build_vocabulary(spacy_de, spacy_en)
        torch.save((vocab_src, vocab_tgt), "vocab.pt")
    else:
        vocab_src, vocab_tgt = torch.load("vocab.pt")
    print("Finished.\nVocabulary sizes:")
    print(len(vocab_src))
    print(len(vocab_tgt))
    return vocab_src, vocab_tgt


if is_interactive_notebook():
    # global variables used later in the script
    spacy_de, spacy_en = show_example(load_tokenizers)
    vocab_src, vocab_tgt = show_example(load_vocab, args=[spacy_de, spacy_en])
Finished.
Vocabulary sizes:
59981
36745

Batching matters a ton for speed. We want to have very evenly divided batches, with absolutely minimal padding. To do this we have to hack a bit around the default torchtext batching. This code patches their default batching to make sure we search over enough sentences to find tight batches.

Iterators

def collate_batch(
    batch,
    src_pipeline,
    tgt_pipeline,
    src_vocab,
    tgt_vocab,
    device,
    max_padding=128,
    pad_id=2,
):
    bs_id = torch.tensor([0], device=device)  # <s> token id
    eos_id = torch.tensor([1], device=device)  # </s> token id
    src_list, tgt_list = [], []
    for (_src, _tgt) in batch:
        processed_src = torch.cat(
            [
                bs_id,
                torch.tensor(
                    src_vocab(src_pipeline(_src)),
                    dtype=torch.int64,
                    device=device,
                ),
                eos_id,
            ],
            0,
        )
        processed_tgt = torch.cat(
            [
                bs_id,
                torch.tensor(
                    tgt_vocab(tgt_pipeline(_tgt)),
                    dtype=torch.int64,
                    device=device,
                ),
                eos_id,
            ],
            0,
        )
        src_list.append(
            # warning - overwrites values for negative values of padding - len
            pad(
                processed_src,
                (
                    0,
                    max_padding - len(processed_src),
                ),
                value=pad_id,
            )
        )
        tgt_list.append(
            pad(
                processed_tgt,
                (0, max_padding - len(processed_tgt)),
                value=pad_id,
            )
        )

    src = torch.stack(src_list)
    tgt = torch.stack(tgt_list)
    return (src, tgt)
def create_dataloaders(
    device,
    vocab_src,
    vocab_tgt,
    spacy_de,
    spacy_en,
    batch_size=12000,
    max_padding=128,
    is_distributed=True,
):
    # def create_dataloaders(batch_size=12000):
    def tokenize_de(text):
        return tokenize(text, spacy_de)

    def tokenize_en(text):
        return tokenize(text, spacy_en)

    def collate_fn(batch):
        return collate_batch(
            batch,
            tokenize_de,
            tokenize_en,
            vocab_src,
            vocab_tgt,
            device,
            max_padding=max_padding,
            pad_id=vocab_src.get_stoi()["<blank>"],
        )

    train_iter, valid_iter, test_iter = datasets.Multi30k(
        language_pair=("de", "en")
    )

    train_iter_map = to_map_style_dataset(
        train_iter
    )  # DistributedSampler needs a dataset len()
    train_sampler = (
        DistributedSampler(train_iter_map) if is_distributed else None
    )
    valid_iter_map = to_map_style_dataset(valid_iter)
    valid_sampler = (
        DistributedSampler(valid_iter_map) if is_distributed else None
    )

    train_dataloader = DataLoader(
        train_iter_map,
        batch_size=batch_size,
        shuffle=(train_sampler is None),
        sampler=train_sampler,
        collate_fn=collate_fn,
    )
    valid_dataloader = DataLoader(
        valid_iter_map,
        batch_size=batch_size,
        shuffle=(valid_sampler is None),
        sampler=valid_sampler,
        collate_fn=collate_fn,
    )
    return train_dataloader, valid_dataloader

Training the System

def train_worker(
    gpu,
    ngpus_per_node,
    vocab_src,
    vocab_tgt,
    spacy_de,
    spacy_en,
    config,
    is_distributed=False,
):
    print(f"Train worker process using GPU: {gpu} for training", flush=True)
    torch.cuda.set_device(gpu)

    pad_idx = vocab_tgt["<blank>"]
    d_model = 512
    model = make_model(len(vocab_src), len(vocab_tgt), N=6)
    model.cuda(gpu)
    module = model
    is_main_process = True
    if is_distributed:
        dist.init_process_group(
            "nccl", init_method="env://", rank=gpu, world_size=ngpus_per_node
        )
        model = DDP(model, device_ids=[gpu])
        module = model.module
        is_main_process = gpu == 0

    criterion = LabelSmoothing(
        size=len(vocab_tgt), padding_idx=pad_idx, smoothing=0.1
    )
    criterion.cuda(gpu)

    train_dataloader, valid_dataloader = create_dataloaders(
        gpu,
        vocab_src,
        vocab_tgt,
        spacy_de,
        spacy_en,
        batch_size=config["batch_size"] // ngpus_per_node,
        max_padding=config["max_padding"],
        is_distributed=is_distributed,
    )

    optimizer = torch.optim.Adam(
        model.parameters(), lr=config["base_lr"], betas=(0.9, 0.98), eps=1e-9
    )
    lr_scheduler = LambdaLR(
        optimizer=optimizer,
        lr_lambda=lambda step: rate(
            step, d_model, factor=1, warmup=config["warmup"]
        ),
    )
    train_state = TrainState()

    for epoch in range(config["num_epochs"]):
        if is_distributed:
            train_dataloader.sampler.set_epoch(epoch)
            valid_dataloader.sampler.set_epoch(epoch)

        model.train()
        print(f"[GPU{gpu}] Epoch {epoch} Training ====", flush=True)
        _, train_state = run_epoch(
            (Batch(b[0], b[1], pad_idx) for b in train_dataloader),
            model,
            SimpleLossCompute(module.generator, criterion),
            optimizer,
            lr_scheduler,
            mode="train+log",
            accum_iter=config["accum_iter"],
            train_state=train_state,
        )

        GPUtil.showUtilization()
        if is_main_process:
            file_path = "%s%.2d.pt" % (config["file_prefix"], epoch)
            torch.save(module.state_dict(), file_path)
        torch.cuda.empty_cache()

        print(f"[GPU{gpu}] Epoch {epoch} Validation ====", flush=True)
        model.eval()
        sloss = run_epoch(
            (Batch(b[0], b[1], pad_idx) for b in valid_dataloader),
            model,
            SimpleLossCompute(module.generator, criterion),
            DummyOptimizer(),
            DummyScheduler(),
            mode="eval",
        )
        print(sloss)
        torch.cuda.empty_cache()

    if is_main_process:
        file_path = "%sfinal.pt" % config["file_prefix"]
        torch.save(module.state_dict(), file_path)
def train_distributed_model(vocab_src, vocab_tgt, spacy_de, spacy_en, config):
    from the_annotated_transformer import train_worker

    ngpus = torch.cuda.device_count()
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "12356"
    print(f"Number of GPUs detected: {ngpus}")
    print("Spawning training processes ...")
    mp.spawn(
        train_worker,
        nprocs=ngpus,
        args=(ngpus, vocab_src, vocab_tgt, spacy_de, spacy_en, config, True),
    )


def train_model(vocab_src, vocab_tgt, spacy_de, spacy_en, config):
    if config["distributed"]:
        train_distributed_model(
            vocab_src, vocab_tgt, spacy_de, spacy_en, config
        )
    else:
        train_worker(
            0, 1, vocab_src, vocab_tgt, spacy_de, spacy_en, config, False
        )


def load_trained_model():
    config = {
        "batch_size": 32,
        "distributed": False,
        "num_epochs": 8,
        "accum_iter": 10,
        "base_lr": 1.0,
        "max_padding": 72,
        "warmup": 3000,
        "file_prefix": "multi30k_model_",
    }
    model_path = "multi30k_model_final.pt"
    if not exists(model_path):
        train_model(vocab_src, vocab_tgt, spacy_de, spacy_en, config)

    model = make_model(len(vocab_src), len(vocab_tgt), N=6)
    model.load_state_dict(torch.load("multi30k_model_final.pt"))
    return model


if is_interactive_notebook():
    model = load_trained_model()

Once trained we can decode the model to produce a set of translations. Here we simply translate the first sentence in the validation set. This dataset is pretty small so the translations with greedy search are reasonably accurate.

Additional Components: BPE, Search, Averaging

So this mostly covers the transformer model itself. There are four aspects that we didn’t cover explicitly. We also have all these additional features implemented in OpenNMT-py.

  1. BPE/ Word-piece: We can use a library to first preprocess the data into subword units. See Rico Sennrich’s subword-nmt implementation. These models will transform the training data to look like this:

▁Die ▁Protokoll datei ▁kann ▁ heimlich ▁per ▁E - Mail ▁oder ▁FTP ▁an ▁einen ▁bestimmte n ▁Empfänger ▁gesendet ▁werden .

  1. Shared Embeddings: When using BPE with shared vocabulary we can share the same weight vectors between the source / target / generator. See the (cite) for details. To add this to the model simply do this:
if False:
    model.src_embed[0].lut.weight = model.tgt_embeddings[0].lut.weight
    model.generator.lut.weight = model.tgt_embed[0].lut.weight
  1. Beam Search: This is a bit too complicated to cover here. See the OpenNMT-py for a pytorch implementation.
  1. Model Averaging: The paper averages the last k checkpoints to create an ensembling effect. We can do this after the fact if we have a bunch of models:
def average(model, models):
    "Average models into model"
    for ps in zip(*[m.params() for m in [model] + models]):
        ps[0].copy_(torch.sum(*ps[1:]) / len(ps[1:]))

Results

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.

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.

With the addtional extensions in the last section, the OpenNMT-py replication gets to 26.9 on EN-DE WMT. Here I have loaded in those parameters to our reimplemenation.

# Load data and model for output checks
def check_outputs(
    valid_dataloader,
    model,
    vocab_src,
    vocab_tgt,
    n_examples=15,
    pad_idx=2,
    eos_string="</s>",
):
    results = [()] * n_examples
    for idx in range(n_examples):
        print("\nExample %d ========\n" % idx)
        b = next(iter(valid_dataloader))
        rb = Batch(b[0], b[1], pad_idx)
        greedy_decode(model, rb.src, rb.src_mask, 64, 0)[0]

        src_tokens = [
            vocab_src.get_itos()[x] for x in rb.src[0] if x != pad_idx
        ]
        tgt_tokens = [
            vocab_tgt.get_itos()[x] for x in rb.tgt[0] if x != pad_idx
        ]

        print(
            "Source Text (Input)        : "
            + " ".join(src_tokens).replace("\n", "")
        )
        print(
            "Target Text (Ground Truth) : "
            + " ".join(tgt_tokens).replace("\n", "")
        )
        model_out = greedy_decode(model, rb.src, rb.src_mask, 72, 0)[0]
        model_txt = (
            " ".join(
                [vocab_tgt.get_itos()[x] for x in model_out if x != pad_idx]
            ).split(eos_string, 1)[0]
            + eos_string
        )
        print("Model Output               : " + model_txt.replace("\n", ""))
        results[idx] = (rb, src_tokens, tgt_tokens, model_out, model_txt)
    return results


def run_model_example(n_examples=5):
    global vocab_src, vocab_tgt, spacy_de, spacy_en

    print("Preparing Data ...")
    _, valid_dataloader = create_dataloaders(
        torch.device("cpu"),
        vocab_src,
        vocab_tgt,
        spacy_de,
        spacy_en,
        batch_size=1,
        is_distributed=False,
    )

    print("Loading Trained Model ...")

    model = make_model(len(vocab_src), len(vocab_tgt), N=6)
    model.load_state_dict(
        torch.load("multi30k_model_final.pt", map_location=torch.device("cpu"))
    )

    print("Checking Model Outputs:")
    example_data = check_outputs(
        valid_dataloader, model, vocab_src, vocab_tgt, n_examples=n_examples
    )
    return model, example_data


# execute_example(run_model_example)

Attention Visualization

Even with a greedy decoder the translation looks pretty good. We can further visualize it to see what is happening at each layer of the attention

def mtx2df(m, max_row, max_col, row_tokens, col_tokens):
    "convert a dense matrix to a data frame with row and column indices"
    return pd.DataFrame(
        [
            (
                r,
                c,
                float(m[r, c]),
                "%.3d %s"
                % (r, row_tokens[r] if len(row_tokens) > r else "<blank>"),
                "%.3d %s"
                % (c, col_tokens[c] if len(col_tokens) > c else "<blank>"),
            )
            for r in range(m.shape[0])
            for c in range(m.shape[1])
            if r < max_row and c < max_col
        ],
        # if float(m[r,c]) != 0 and r < max_row and c < max_col],
        columns=["row", "column", "value", "row_token", "col_token"],
    )


def attn_map(attn, layer, head, row_tokens, col_tokens, max_dim=30):
    df = mtx2df(
        attn[0, head].data,
        max_dim,
        max_dim,
        row_tokens,
        col_tokens,
    )
    return (
        alt.Chart(data=df)
        .mark_rect()
        .encode(
            x=alt.X("col_token", axis=alt.Axis(title="")),
            y=alt.Y("row_token", axis=alt.Axis(title="")),
            color="value",
            tooltip=["row", "column", "value", "row_token", "col_token"],
        )
        .properties(height=400, width=400)
        .interactive()
    )
def get_encoder(model, layer):
    return model.encoder.layers[layer].self_attn.attn


def get_decoder_self(model, layer):
    return model.decoder.layers[layer].self_attn.attn


def get_decoder_src(model, layer):
    return model.decoder.layers[layer].src_attn.attn


def visualize_layer(model, layer, getter_fn, ntokens, row_tokens, col_tokens):
    # ntokens = last_example[0].ntokens
    attn = getter_fn(model, layer)
    n_heads = attn.shape[1]
    charts = [
        attn_map(
            attn,
            0,
            h,
            row_tokens=row_tokens,
            col_tokens=col_tokens,
            max_dim=ntokens,
        )
        for h in range(n_heads)
    ]
    assert n_heads == 8
    return alt.vconcat(
        charts[0]
        # | charts[1]
        | charts[2]
        # | charts[3]
        | charts[4]
        # | charts[5]
        | charts[6]
        # | charts[7]
        # layer + 1 due to 0-indexing
    ).properties(title="Layer %d" % (layer + 1))

Encoder Self Attention

def viz_encoder_self():
    model, example_data = run_model_example(n_examples=1)
    example = example_data[
        len(example_data) - 1
    ]  # batch object for the final example
layer_viz = [ visualize_layer( model, layer, get_encoder, len(example[1]), example[1], example[1] ) for layer in range(6) ] return alt.hconcat( layer_viz[0] # & layer_viz[1] & layer_viz[2] # & layer_viz[3] & layer_viz[4] # & layer_viz[5] )
show_example(viz_encoder_self)
Preparing Data ...
Loading Trained Model ...
Checking Model Outputs:
Example 0 ========
Source Text (Input) : <s> Zwei Frauen in pinkfarbenen T-Shirts und <unk> unterhalten sich vor einem <unk> . </s> Target Text (Ground Truth) : <s> Two women wearing pink T - shirts and blue jeans converse outside clothing store . </s> Model Output : <s> Two women in pink shirts and face are talking in front of a <unk> . </s>

Decoder Self Attention

def viz_decoder_self():
    model, example_data = run_model_example(n_examples=1)
    example = example_data[len(example_data) - 1]
layer_viz = [ visualize_layer( model, layer, get_decoder_self, len(example[1]), example[1], example[1], ) for layer in range(6) ] return alt.hconcat( layer_viz[0] & layer_viz[1] & layer_viz[2] & layer_viz[3] & layer_viz[4] & layer_viz[5] )
show_example(viz_decoder_self)
Preparing Data ...
Loading Trained Model ...
Checking Model Outputs:
Example 0 ========
Source Text (Input) : <s> Eine Gruppe von Männern in Kostümen spielt Musik . </s> Target Text (Ground Truth) : <s> A group of men in costume play music . </s> Model Output : <s> A group of men in costumes playing music . </s>

Decoder Src Attention

def viz_decoder_src():
    model, example_data = run_model_example(n_examples=1)
    example = example_data[len(example_data) - 1]
layer_viz = [ visualize_layer( model, layer, get_decoder_src, max(len(example[1]), len(example[2])), example[1], example[2], ) for layer in range(6) ] return alt.hconcat( layer_viz[0] & layer_viz[1] & layer_viz[2] & layer_viz[3] & layer_viz[4] & layer_viz[5] )
show_example(viz_decoder_src)
Preparing Data ...
Loading Trained Model ...
Checking Model Outputs:
Example 0 ========
Source Text (Input) : <s> Ein kleiner Junge verwendet einen Bohrer , um ein Loch in ein Holzstück zu machen . </s> Target Text (Ground Truth) : <s> A little boy using a drill to make a hole in a piece of wood . </s> Model Output : <s> A little boy uses a machine to be working in a hole in a log . </s>

Conclusion

Hopefully this code is useful for future research. Please reach out if you have any issues.

Cheers, Sasha Rush, Austin Huang, Suraj Subramanian, Jonathan Sum, Khalid Almubarak, Stella Biderman