Transformer

前面我们详细的介绍了Transformer的原理,但是有的细节还是一头雾水,所以我们接下来介绍一下Transformer的实现,主要参考了文章The Annotated Transformergithub地址

本文的代码部分来自于github,而图来源于The Annotated Transformer

Prelims

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
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

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)
RUN_EXAMPLES = True
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
# 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

Model Architecture

大多数的neural sequence transduction模型都使用了encoder-decoder结构,encoder结构将一个用符号(symbols)表示的输入系列$(x_1, …, x_n)$,表示成为连续表征$\mathbf{z} = (z_1, …, z_n)$。给出$\mathbf{z}$,decoder生成输出序列$(y_1,…,y_m)$,并且一次生成一个元素。在每一步,模型都是自回归的(auto-regressive),在生成下一步时,使用先前生成的符号作为附加输入。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
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)

Transformer总体结构如下,encoder和decoder结构都是堆叠self-attention and point-wise, fully connected layers。

png

Encoder

Encoder由$N=6$个一模一样的层(EncoderLayer)组成。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
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)

我们在each of the two sub-layers使用残差连接,并且后接layer normalization。

1
2
3
4
5
6
7
8
9
10
11
12
13
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

因此,每一个sub-layer的输出是$\mathrm{LayerNorm}(x + \mathrm{Sublayer}(x))$。我们还添加了Dropout层。为了facilitate这些残差连接,模型中所有sub-layer和embedding layers的输出维度均是$d_{\text{model}}=512$。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
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)))

每一个layer含有两个sub-layers,第一个是multi-head self-attention mechanism,第二个是simple, position-wise fully connected feed-forward network。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
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

decoder同样由$N=6$个一模一样的层(encoder layer)组成。

1
2
3
4
5
6
7
8
9
10
11
12
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)

在每个encoder layer除了两个 sub-layers 外,还插入了第三个sub-layer,它在encoder stack的输出上执行multi-head attention。与encoder相同,我们在each of the two sub-layers使用残差连接,并且后接layer normalization。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
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)

我们还修改了decoder中的self-attention sub-layer,以防止它利用到后续位置的信息。This masking, combined with fact that the output embeddings are offset by one position, ensures that the predictions for position $i$ can depend only on the known outputs at positions less than $i$ .

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
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

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")),
)
)


show_example(example_mask)

下图展示了each tgt word(row),被允许看到的信息(column)。单词在训练过程中被遮挡,使模型关注预测下一个words。

png

Attention

attention函数可以被描述为 mapping a query and a set of key-value pairs to an output,其中query, keys, values, and output都是向量。output是values的加权求和,其中每个value的权重是通过query with the corresponding key的compatibility function计算得到。

我们将这种特别的attention称为“Scaled Dot-Product Attention”。它的输入由$d_k$维度的queries、keys,$d_v$维度的values组成。 We compute the dot products of the query with all keys, divide each by $\sqrt{d_k}$, and apply a softmax function to obtain the weights on the values。

png

实际上我们会同时在一系列的的queries上计算attention 函数,对应的会有一系列的keys $K$、values $V$。attention函数的输出为:

1
2
3
4
5
6
7
8
9
10
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

最常用的两个attention函数是additive attention和dot-product (multiplicative) attention。后者除了没有缩放$\frac{1}{\sqrt{d_k}}$,其余与我们的相同。而Additive attention computes the compatibility function using a feed-forward network with a single hidden layer. 虽然两者在理论复杂性上相似,但dot-product attention在实践中要快得多,空间效率更高,因为它可以使用高度优化的矩阵乘法代码来实现。

虽然对于较小的$d_k$值,这两种机制的性能相似,但对于较大的$d_k$值,additive attention优于dot product attention。我们怀疑较大的$d_k$值,dot product的幅度会增大,从而将Softmax函数推入其梯度极小的区域。(To illustrate why the dot products get large, assume that the components of $q$ and $k$ are independent random variables with mean $0$ and variance $1$. Then their dot product, $q \cdot k = \sum_{i=1}^{d_k} q_ik_i$, has mean $0$ and variance$d_k$.). 为了抵消这种影响,我们使用$\frac{1}{\sqrt{d_k}}$对dot products进行缩放。

png

Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. 当仅有一个 attention head,平均化抑制了这一点。

Where the projections are parameter matrices $W^Q_i \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W^K_i \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W^V_i \in \mathbb{R}^{d_{\text{model}} \times d_v}$ and $W^O \in \mathbb{R}^{hd_v \times d_{\text{model}}}$。在本工作中,我们使用了$h=8$个平行attention layers, or heads。对于其中每一个,我们使用了$d_k=d_v=d_{\text{model}}/h=64$。由于each head的维度降低了,所以总的计算量与full dimensionality的single-head attention相同。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
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

Transformer以三种不同的方式使用了multi-head attention。

  1. 在“encoder-decoder attention” layers中,queries来自于之前的decoder layer, memory keys and values 来自encoder的输出。This allows every position in the decoder to attend over all positions in the input sequence. 这模仿了sequence-to-sequence模型中典型的encoder-decoder attention机制。
  2. encoder中的self-attention layers. 这里的self-attention layers中,所有的 keys, values and queries均来自于上一层的输出。Each position in the encoder can attend to all positions in the previous layer of the encoder.
  3. decoder中的self-attention layers. 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(up to and including:直到并包括)。我们需要防止信息在decoder中向左流动,以保持自回归(auto-regressive)特性。我们在scaled dot-product attention中实现了这一点,通过屏蔽Softmax输入中对应于非法连接的所有值(设置为$-\infty$)。

Position-wise Feed-Forward Networks

除了attention sub-layers,encoder and decoder中的每一层都包含一个fully connected feed-forward network(完全连接的前馈网络),该网络分别且相同地应用于每个位置。它由两个线性变换组成,中间有一个ReLU激活。

虽然在不同位置上都是线性变换,但它们在不同的层之间使用不同的参数。另一种描述方式是将其描述为核大小为1的两个卷积。input和output的维度为$d_{\text{model}}=512$,内层的维度为$d_{ff}=2048$。

1
2
3
4
5
6
7
8
9
10
11
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

与其它的sequence transduction models相似,we use learned embeddings to convert the input tokens and output tokens to vectors of dimension $d_{\text{model}}$。我们还使用常用的 linear transformation and softmax function 将 decoder output转换为 predicted next-token probabilities. 在我们的模型中,我们在two embedding layers 和pre-softmax linear transformation共享相同的权重矩阵。在embedding layers,我们multiply those weights by $\sqrt{d_{\text{model}}}$。

1
2
3
4
5
6
7
8
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

由于我们的模型不包含recurrence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tokens in the sequence的相对或绝对位置的信息。为此,我们将“positional encodings”添加到 encoder 和 decoder 堆栈底部的input embeddings中。 positional encodings具有与embeddings相同的维度$d_{\text{model}}$,因此这两个模型可以求和。positional encodings有许多选择,学习的和固定的。

在这项工作中,我们使用不同频率的正弦和余弦函数。其中$pos$表示单词在句子中的位置,$2i$ 表示偶数的维度,$2i+1$ 表示奇数维度。也就是说,位置编码的每个维度对应于一个正弦。波长形成从2π到10000⋅2π的几何级数。我们选择这个函数是因为我们假设它将允许模型更容易学习相对位置,因为对于任何固定的偏移量$k$,$PE_{pos+k}$可以表示为$PE_{pos}$的线性函数。

初次之外,我们还将dropout应用于the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. 这里dropout的比例为$P_{drop}=0.1$。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
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.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
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")
)


show_example(example_positional)

png

我们还试验了使用学习的positional embeddings,发现两个版本产生的结果几乎相同。我们选择正弦版本是因为它可能允许模型推广到比训练期间遇到的序列长度更长的序列长度。

Full Model

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
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

img

Inference

在这里,我们执行一个forward,以生成模型的预测。我们尝试使用我们的transformer来记忆输入。正如您将看到的,由于模型尚未经过训练,因此输出是随机生成的。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记住从1到10的数字。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
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)

Training

接下来我们来介绍训练流程,在此之前我们先介绍train a standard encoder decoder model所需的工具。首先我们定义一个batch object保存用于训练的 src and target sentences、masks。

Batches and Masking

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
class Batch:
"""Object for holding a batch of data with mask during training."""

def __init__(self, src, tgt=None, pad=2): # 2 = <blank> for IWST
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

Training Loop

接下来我们创建一个通用的训练和打分函数,来跟踪损失。我们传入一个损失函数,它还会执行参数更新。

1
2
3
4
5
6
7
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
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
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

我们在标准的WMT 2014英语-德语数据集上进行了训练,该数据集由大约450万个句子对组成。Sentences were encoded using byte-pair encoding, which has a shared source-target vocabulary of about 37000 tokens. 对于英语-法语,我们使用了更大的2014年WMT英语-法语数据集,包括3600万个句子和split tokens into a 32000 word-piece vocabulary.

Sentence pairs were batched together by approximate sequence length. 每个训练批次包含一组句子对,其中包含大约25000个source tokens和25000个target tokens。

Hardware and Schedule

我们在一台配备8个NVIDIA P100图形处理器的机器上训练了我们的模型。对于使用本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小时的培训。对于我们的大型模型,step time是1.0秒。这些大模型接受了300,000步(3.5天)的训练。

Optimizer

我们使用adam作为优化器,$\beta_1=0.9$, $\beta_2=0.98$ and $\epsilon=10^{-9}$,在训练中学习率也是变化的,变化方式是:

这对应于在前面$warmup_steps$线性增加学习率,此后按与步数的平方根倒数成比例递减。这里$warmup_steps=4000$。

注意:这部分非常重要。 需要使用这种模型设置进行训练。

该模型曲线的示例,用于不同的模型大小和优化超参数。

1
2
3
4
5
6
7
8
9
10
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))
)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
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")
)


example_learning_schedule()

png

Regularization(Label Smoothing)

在训练过程中我们使用了label smoothing,$\epsilon_{ls}=0.1$。虽然模型学会了更多的不确定,但提高了准确性和BLEU的分数。

我们使用KL div loss实现label smoothing. 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.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
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.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
# 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")
),
)
)


show_example(example_label_smoothing)

png

Label smoothing actually starts to penalize the model if it gets very confident about a given choice.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
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",
)
)


show_example(penalization_visualization)

png

A First Example

我们可以从尝试一项简单的抄写任务开始。给定一组来自较小词汇表的随机输入symbols,目标是生成相同的symbols。

Synthetic Data

1
2
3
4
5
6
7
8
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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
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

为简单起见,此代码使用Greedy Decoding来预测翻译。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
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
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
# 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)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
Epoch Step: 1 Loss: 3.023465 Tokens per Sec: 403.074173
Epoch Step: 1 Loss: 1.920030 Tokens per Sec: 641.689380
1.9274832487106324
Epoch Step: 1 Loss: 1.940011 Tokens per Sec: 432.003378
Epoch Step: 1 Loss: 1.699767 Tokens per Sec: 641.979665
1.657595729827881
Epoch Step: 1 Loss: 1.860276 Tokens per Sec: 433.320240
Epoch Step: 1 Loss: 1.546011 Tokens per Sec: 640.537198
1.4888023376464843
Epoch Step: 1 Loss: 1.682198 Tokens per Sec: 432.092305
Epoch Step: 1 Loss: 1.313169 Tokens per Sec: 639.441857
1.3485562801361084
Epoch Step: 1 Loss: 1.278768 Tokens per Sec: 433.568756
Epoch Step: 1 Loss: 1.062384 Tokens per Sec: 642.542067
0.9853351473808288
Epoch Step: 1 Loss: 1.269471 Tokens per Sec: 433.388727
Epoch Step: 1 Loss: 0.590709 Tokens per Sec: 642.862135
0.5686767101287842
Epoch Step: 1 Loss: 0.997076 Tokens per Sec: 433.009746
Epoch Step: 1 Loss: 0.343118 Tokens per Sec: 642.288427
0.34273059368133546
Epoch Step: 1 Loss: 0.459483 Tokens per Sec: 434.594030
Epoch Step: 1 Loss: 0.290385 Tokens per Sec: 642.519464
0.2612409472465515
Epoch Step: 1 Loss: 1.031042 Tokens per Sec: 434.557008
Epoch Step: 1 Loss: 0.437069 Tokens per Sec: 643.630322
0.4323212027549744
Epoch Step: 1 Loss: 0.617165 Tokens per Sec: 436.652626
Epoch Step: 1 Loss: 0.258793 Tokens per Sec: 644.372296
0.27331129014492034

1 2 3 4 5 6 7 8 9 10
[torch.LongTensor of size 1x10]

A Real World Example

现在,我们考虑一个使用IWSLT德语-英语翻译任务的真实世界示例。这项任务比本文考虑的WMT任务小得多,但它能说明整个流程。我们还展示了如何使用多GPU处理来实现真正的速度。

Data Loading

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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
# 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
1
2
3
4
5
6
7
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])
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
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.IWSLT2016(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.IWSLT2016(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])

Iterators

Batching对训练速度很重要。我们希望非常均匀的划分批次(with absolutely minimal padding)。要做到这一点,我们必须修改一下默认的torchtext batching。这段代码修改了它们的默认批处理,以确保我们搜索足够多的句子来找到紧凑的批处理。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
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)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
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.IWSLT2016(
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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
def train_worker(
gpu, ngpus_per_node, vocab_src, vocab_tgt, spacy_de, spacy_en, config
):
print(f"Train worker process using GPU: {gpu} for training", flush=True)
torch.cuda.set_device(gpu)
is_main_process = gpu == 0

pad_idx = vocab_tgt["<blank>"]
d_model = 512
model = make_model(len(vocab_src), len(vocab_tgt), N=6)
model.cuda(gpu)
module = model

dist.init_process_group(
"nccl", init_method="env://", rank=gpu, world_size=ngpus_per_node
)
model = DDP(model, device_ids=[gpu])
module = model.module

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"],
)

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"]):

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)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
def train_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,
),
)


def load_trained_model(create_model):
config = {
"batch_size": 150,
"num_epochs": 8,
"accum_iter": 10,
"base_lr": 1.0,
"max_padding": 72,
"warmup": 3000,
"file_prefix": "iwslt_model_",
}

if create_model:
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("iwslt_model_final.pt"))
return model


if is_interactive_notebook():
model = load_trained_model(create_model=True)

一旦经过训练,我们就可以对模型进行decode,以产生一组翻译。在这里,我们只需翻译验证集中的第一句话。这个数据集非常小,因此使用greedy search的翻译相当准确。

1
2
Translation:	<unk> <unk> . In my language , that means , thank you very much . 
Gold: <unk> <unk> . It means in my language , thank you very much .

Additional Components: BPE, Search, Averaging

上述代码主要介绍了Transformer自身的实现,还有四个函数我们没有实现。

BPE/ Word-piece

我们使用了subword units库对数据进行预处理。它将把训练数据转换成如下所示的形式:▁Die ▁Protokoll datei ▁kann ▁ heimlich ▁per ▁E - Mail ▁oder ▁FTP ▁an ▁einen ▁bestimmte n ▁Empfänger ▁gesendet ▁werden .

Shared Embeddings

当使用共享vocabulary的BPE时,我们可以在source / target / generator之间共享权重向量。详细信息可以阅读cite。要将此想法实现到模型中,只需执行以下操作:

1
2
3
if False:
model.src_embed[0].lut.weight = model.tgt_embeddings[0].lut.weight
model.generator.lut.weight = model.tgt_embed[0].lut.weight

详情可以看OpenNMT-py

Model Averaging

文章中平均了最后k个checkpoints来达到集成效果。如果我们有一堆checkpoint,我们可以在事后做这件事:

1
2
3
4
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

在WMT 2014英德翻译任务中,big transformer模型(表2中的Transformer (big) )的表现超过了已有的最好的模型(包括集成),BLEU的表现高上了2.0%,创造了最好的BLEU 28.4分的新纪录。表3的底部列出了该模型参数配置。在8个P100 GPU上进行了3.5天的训练。甚至我们的base model模型也超过了已有的所有模型和集成模型,而训练成本只是任何已有模型的一小部分。

在2014年WMT英法翻译任务中,我们的大模型达到了BLEU的41.0分,超过了之前已有的所有单一模型,培训成本不到以前最好模型的四分之一。用于英法翻译的Transformer (big) 的dropout 系数等于0.1,而不是0.3。

我们在这里编写的代码是base model的一个版本。完整版本可以看 (Example Models)

使用上以小节的附加扩展,OpenNMT-py 在EN-DE WMT数据集上达到了26.9的BLEU。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
# 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("iwslt_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)
1
Translation:	<s> ▁Die ▁Protokoll datei ▁kann ▁ heimlich ▁per ▁E - Mail ▁oder ▁FTP ▁an ▁einen ▁bestimmte n ▁Empfänger ▁gesendet ▁werden .

Attention Visualization

即使使用greedy decoder,翻译看起来也很好。我们可以进一步把它形象化,看看注意力的每一层都在发生什么。

Encoder Self Attention

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
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=200, width=200)
.interactive()
)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
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)

Decoder Self Attention

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
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)

Decoder Src Attention

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
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)

Question

Transformer mask

在《the annotated transformer》中有多个mask,这里总结一下。

整个模型中使用到的mask主要就是source mask和target mask,其各自的作用如下所示:

  1. source mask:
  • source长短不一而无法形成batch,因此引入了pad。将source mask传入到encoder中,让attention在计算$\mathrm{softmax}(\frac{QK^T}{\sqrt{d_k}})$时,pad位置的值不起作用。

  • 同时这个mask还需要传入每个decoderLayer第二个multi-head attention模块中,就是防止来自encoder的key和来自decoder的query在计算多头注意力的时候算了target中的词和source中pad的权重

  1. target mask:需要分training和testing进行讨论
  • 训练时,用于防止target的ground truth长短不一引入pad造成的误差,以及避免在自回归时看到正在预测的字和以后字的ground truth
  • 测试时,逻辑上decoder不需要target mask,但出于编程方便的考虑引入mask,假装用于防止看到后面的ground truth,target mask的最后两维的shape和目前生成出来的序列长度相同,但实际上每次都会有一些重复运算在里面,比如目前在预测第10个词时,第1-9个词还需要重新算一遍。核心原因是:模型在写的时候主要考虑的是训练,执行一次attention函数翻译完一个batch的所有句子,而测试时必须是单个或多个句子word by word进行计算

参考

annotated-transformer
the annotated transformer中的关于mask的问题 - lumino的文章 - 知乎

------ 本文结束------
坚持原创技术分享,您的支持将鼓励我继续创作!

欢迎关注我的其它发布渠道