Transformer (Google 机器模型)
卷积神经网络(CNNS)
CNNs 易于并行化,却不适合捕捉变长序列内的依赖关系。
循环神经网络(RNNS)
RNNs 适合捕捉长距离变长序列的依赖,但是自身的recurrent特性却难以实现并行化处理序列。
整合CNN和RNN的优势,Vaswani et al., 2017 创新性地使用注意力机制设计了 Transformer 模型。
该模型利用 attention 机制实现了并行化捕捉序列依赖,并且同时处理序列的每个位置的 tokens ,上述优势使得 Transformer 模型在性能优异的同时大大减少了训练时间。
如图展示了 Transformer 模型的架构,与机器翻译及其相关技术介绍中介绍的 seq2seq 相似,Transformer同样基于编码器-解码器架构,其区别主要在于以下三点:
- Transformer blocks:–> Transformer Blocks
Transform Blocks模块包含一个多头注意力层(Multi-head Attention Layers)以及两个 position-wise feed-forward networks(FFN)。对于解码器来说,另一个多头注意力层被用于接受编码器的隐藏状态。 - Add and norm:多头注意力层和前馈网络的输出被送到两个“add and norm”层进行处理
该层包含残差结构以及层归一化。 - Position encoding:由于自注意力层并没有区分元素的顺序,所以一个位置编码层被用于向序列元素里添加位置信息。
鉴于新子块第一次出现,在此前 CNNS 和 RNNS 的基础上,实现 Transform 子模块,并且就机器翻译及其相关技术介绍中的英法翻译数据集实现一个新的机器翻译模型。
Transformer
import os
import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import sys
sys.path.append('path to file storge d2lzh1981')
import d2l
masked softmax
def SequenceMask(X, X_len,value=-1e6):
maxlen = X.size(1)
X_len = X_len.to(X.device)
#print(X.size(),torch.arange((maxlen),dtype=torch.float)[None, :],'\n',X_len[:, None] )
mask = torch.arange((maxlen), dtype=torch.float, device=X.device)
mask = mask[None, :] < X_len[:, None]
#print(mask)
X[~mask]=value
return X
def masked_softmax(X, valid_length):
# X: 3-D tensor, valid_length: 1-D or 2-D tensor
softmax = nn.Softmax(dim=-1)
if valid_length is None:
return softmax(X)
else:
shape = X.shape
if valid_length.dim() == 1:
try:
valid_length = torch.FloatTensor(valid_length.numpy().repeat(shape[1], axis=0))#[2,2,3,3]
except:
valid_length = torch.FloatTensor(valid_length.cpu().numpy().repeat(shape[1], axis=0))#[2,2,3,3]
else:
valid_length = valid_length.reshape((-1,))
# fill masked elements with a large negative, whose exp is 0
X = SequenceMask(X.reshape((-1, shape[-1])), valid_length)
return softmax(X).reshape(shape)
# Save to the d2l package.
class DotProductAttention(nn.Module):
def __init__(self, dropout, **kwargs):
super(DotProductAttention, self).__init__(**kwargs)
self.dropout = nn.Dropout(dropout)
# query: (batch_size, #queries, d)
# key: (batch_size, #kv_pairs, d)
# value: (batch_size, #kv_pairs, dim_v)
# valid_length: either (batch_size, ) or (batch_size, xx)
def forward(self, query, key, value, valid_length=None):
d = query.shape[-1]
# set transpose_b=True to swap the last two dimensions of key
scores = torch.bmm(query, key.transpose(1,2)) / math.sqrt(d)
attention_weights = self.dropout(masked_softmax(scores, valid_length))
return torch.bmm(attention_weights, value)
多头注意力层
引入:自注意力(self-attention)
自注意力模型是一个正规的注意力模型,序列的每一个元素对应的key,value,query是完全一致的。
与循环神经网络相比,自注意力对每个元素输出的计算是并行的,所以我们可以高效的实现这个模块。
多头注意力层包含个并行的自注意力层,每一个这种层被成为一个head。
对每个头来说,在进行注意力计算之前,我们会将query、key和value用三个现行层进行映射,这个注意力头的输出将会被拼接之后输入最后一个线性层进行整合。
假设query,key和value的维度分别是、和。那么对于每一个头,我们可以训练相应的模型权重、和,以得到每个头的输出:
这里的attention可以是任意的attention function,之后我们将所有head对应的输出拼接起来,送入最后一个线性层进行整合,这个层的权重可以表示为
接下来实现多头注意力,假设有h个头,隐藏层权重 与query,key,value的维度一致。除此之外,因为多头注意力层保持输入与输出张量的维度不变,所以输出feature 的维度也设置为 。
MultiHeadAttention class
class MultiHeadAttention(nn.Module):
def __init__(self, input_size, hidden_size, num_heads, dropout, **kwargs):
super(MultiHeadAttention, self).__init__(**kwargs)
self.num_heads = num_heads
self.attention = DotProductAttention(dropout)
self.W_q = nn.Linear(input_size, hidden_size, bias=False)
self.W_k = nn.Linear(input_size, hidden_size, bias=False)
self.W_v = nn.Linear(input_size, hidden_size, bias=False)
self.W_o = nn.Linear(hidden_size, hidden_size, bias=False)
def forward(self, query, key, value, valid_length):
# query, key, and value shape: (batch_size, seq_len, dim),
# where seq_len is the length of input sequence
# valid_length shape is either (batch_size, )
# or (batch_size, seq_len).
# Project and transpose query, key, and value from
# (batch_size, seq_len, hidden_size * num_heads) to
# (batch_size * num_heads, seq_len, hidden_size).
query = transpose_qkv(self.W_q(query), self.num_heads)
key = transpose_qkv(self.W_k(key), self.num_heads)
value = transpose_qkv(self.W_v(value), self.num_heads)
if valid_length is not None:
# Copy valid_length by num_heads times
device = valid_length.device
valid_length = valid_length.cpu().numpy() if valid_length.is_cuda else valid_length.numpy()
if valid_length.ndim == 1:
valid_length = torch.FloatTensor(np.tile(valid_length, self.num_heads))
else:
valid_length = torch.FloatTensor(np.tile(valid_length, (self.num_heads,1)))
valid_length = valid_length.to(device)
output = self.attention(query, key, value, valid_length)
output_concat = transpose_output(output, self.num_heads)
return self.W_o(output_concat)
转置函数
def transpose_qkv(X, num_heads):
# Original X shape: (batch_size, seq_len, hidden_size * num_heads),
# -1 means inferring its value, after first reshape, X shape:
# (batch_size, seq_len, num_heads, hidden_size)
X = X.view(X.shape[0], X.shape[1], num_heads, -1)
# After transpose, X shape: (batch_size, num_heads, seq_len, hidden_size)
X = X.transpose(2, 1).contiguous()
# Merge the first two dimensions. Use reverse=True to infer shape from
# right to left.
# output shape: (batch_size * num_heads, seq_len, hidden_size)
output = X.view(-1, X.shape[2], X.shape[3])
return output
# Saved in the d2l package for later use
def transpose_output(X, num_heads):
# A reversed version of transpose_qkv
X = X.view(-1, num_heads, X.shape[1], X.shape[2])
X = X.transpose(2, 1).contiguous()
return X.view(X.shape[0], X.shape[1], -1)
测试输出
cell = MultiHeadAttention(5, 9, 3, 0.5)
X = torch.ones((2, 4, 5))
valid_length = torch.FloatTensor([2, 3])
cell(X, X, X, valid_length).shape
Position-Wise Feed-Forward Networks
Transformer 模块另一个非常重要的部分就是基于位置的前馈网络(FFN),它接受一个形状为(batch_size,seq_length, feature_size)的三维张量。
Position-wise FFN 由两个全连接层组成,它们作用在最后一维上。因为序列的每个位置的状态都会被单独地更新,所以我们称为 position-wise,其等效于一个 1x1 的卷积。
Position-wise FFN class
# Save to the d2l package.
class PositionWiseFFN(nn.Module):
def __init__(self, input_size, ffn_hidden_size, hidden_size_out, **kwargs):
super(PositionWiseFFN, self).__init__(**kwargs)
self.ffn_1 = nn.Linear(input_size, ffn_hidden_size)
self.ffn_2 = nn.Linear(ffn_hidden_size, hidden_size_out)
def forward(self, X):
return self.ffn_2(F.relu(self.ffn_1(X)))
与多头注意力层相似,FFN层同样只会对最后一维的大小进行改变;除此之外,对于两个完全相同的输入,FFN层的输出也将相等。
ffn = PositionWiseFFN(4, 4, 8)
out = ffn(torch.ones((2,3,4)))
print(out, out.shape)
Add and Norm
Transformer还有一个重要的相加归一化层
它可以平滑地整合输入和其他层的输出,因此我们在每个多头注意力层和 FFN 层后面都添加一个含残差连接的Layer Norm层。
唯一的区别
- Batch Norm是对于batch size这个维度进行计算均值和方差的,
- Layer Norm则是对最后一维进行计算。
layernorm = nn.LayerNorm(normalized_shape=2, elementwise_affine=True)
batchnorm = nn.BatchNorm1d(num_features=2, affine=True)
X = torch.FloatTensor([[1,2], [3,4]])
print('layer norm:', layernorm(X))
print('batch norm:', batchnorm(X))
层归一化可以防止层内的数值变化过大,从而有利于加快训练速度并且提高泛化性能。
AddNorm class
class AddNorm(nn.Module):
def __init__(self, hidden_size, dropout, **kwargs):
super(AddNorm, self).__init__(**kwargs)
self.dropout = nn.Dropout(dropout)
self.norm = nn.LayerNorm(hidden_size)
def forward(self, X, Y):
return self.norm(self.dropout(Y) + X)
【注意】:由于残差连接,X和Y需要有相同的维度。
模块测试
add_norm = AddNorm(4, 0.5)
add_norm(torch.ones((2,3,4)), torch.ones((2,3,4))).shape
以上是Transformer 模型的三个模块,还记得 Transformer 模型高效并行的特性,得益于:
多头注意力网络还是前馈神经网络都是独立地对每个位置的元素进行更新
但是在这种特性下,却丢失了重要的序列顺序的信息,为了更好的捕捉序列信息,需要一种可以保持序列元素位置的模块,因而引入位置编码。
位置编码
假设输入序列的嵌入表示 , 序列长度为嵌入向量维度为,则其位置编码为 ,输出的向量就是二者相加 。
位置编码是一个二维的矩阵,i对应着序列中的顺序,j对应其embedding vector内部的维度索引。我们可以通过以下等式计算位置编码:
PositionalEncoding class
class PositionalEncoding(nn.Module):
def __init__(self, embedding_size, dropout, max_len=1000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(dropout)
self.P = np.zeros((1, max_len, embedding_size))
X = np.arange(0, max_len).reshape(-1, 1) / np.power(
10000, np.arange(0, embedding_size, 2)/embedding_size)
self.P[:, :, 0::2] = np.sin(X)
self.P[:, :, 1::2] = np.cos(X)
self.P = torch.FloatTensor(self.P)
def forward(self, X):
if X.is_cuda and not self.P.is_cuda:
self.P = self.P.cuda()
X = X + self.P[:, :X.shape[1], :]
return self.dropout(X)
模块测试
可视化其中四个维度,可以看到,第 4 维和第 5 维有相同的频率但偏置不同。第 6 维和第 7 维具有更低的频率;因此 positional encoding 对于不同维度具有可区分性。
import numpy as np
pe = PositionalEncoding(20, 0)
Y = pe(torch.zeros((1, 100, 20))).numpy()
d2l.plot(np.arange(100), Y[0, :, 4:8].T, figsize=(6, 2.5),
legend=["dim %d" % p for p in [4, 5, 6, 7]])
![](https://img-blog.csdnimg.cn/20200219115650513.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L1Jva29CYXNpbGlzaw==,size_16,color_FFFFFF,t_70)
编码器
有了组成Transformer的各个模块,可以搭建一个编码器。
编码器包含一个多头注意力层,一个position-wise FFN,和两个 Add and Norm层。
对于attention模型以及FFN模型,由于残差连接导致输出维度都是与 embedding 维度一致的,因此要将前一层的输出与原始输入相加并归一化。
Encoder Block基础块
class EncoderBlock(nn.Module):
def __init__(self, embedding_size, ffn_hidden_size, num_heads,
dropout, **kwargs):
super(EncoderBlock, self).__init__(**kwargs)
self.attention = MultiHeadAttention(embedding_size, embedding_size, num_heads, dropout)
self.addnorm_1 = AddNorm(embedding_size, dropout)
self.ffn = PositionWiseFFN(embedding_size, ffn_hidden_size, embedding_size)
self.addnorm_2 = AddNorm(embedding_size, dropout)
def forward(self, X, valid_length):
Y = self.addnorm_1(X, self.attention(X, X, X, valid_length))
return self.addnorm_2(Y, self.ffn(Y))
# batch_size = 2, seq_len = 100, embedding_size = 24
# ffn_hidden_size = 48, num_head = 8, dropout = 0.5
X = torch.ones((2, 100, 24))
encoder_blk = EncoderBlock(24, 48, 8, 0.5)
encoder_blk(X, valid_length).shape
整个编码器由 n 个 Encoder Block 堆叠而成,利用 Encoder Block 基础块实现 Transformer 编码器。
两个注意点:
- 残差连接的缘故,中间状态的维度始终与嵌入向量的维度 d 一致;
- 同时注意到我们把嵌入向量乘以 以防止其值过小。
class TransformerEncoder(d2l.Encoder):
def __init__(self, vocab_size, embedding_size, ffn_hidden_size,
num_heads, num_layers, dropout, **kwargs):
super(TransformerEncoder, self).__init__(**kwargs)
self.embedding_size = embedding_size
self.embed = nn.Embedding(vocab_size, embedding_size)
self.pos_encoding = PositionalEncoding(embedding_size, dropout)
self.blks = nn.ModuleList()
for i in range(num_layers):
self.blks.append(
EncoderBlock(embedding_size, ffn_hidden_size,
num_heads, dropout))
def forward(self, X, valid_length, *args):
X = self.pos_encoding(self.embed(X) * math.sqrt(self.embedding_size))
for blk in self.blks:
X = blk(X, valid_length)
return X
# test encoder
encoder = TransformerEncoder(200, 24, 48, 8, 2, 0.5)
encoder(torch.ones((2, 100)).long(), valid_length).shape
解码器
Transformer 模型的解码器与编码器结构类似,然而,除了之前介绍的几个模块之外,解码器部分有另一个子模块。该模块也是多头注意力层,接受编码器的输出作为key和value,decoder的状态作为query。
与编码器部分相类似,解码器同样是使用了add and norm机制,用残差和层归一化将各个子层的输出相连。
在第t个时间步,当前输入是query,那么self attention接受了第t步以及前t-1步的所有输入。在训练时,由于第t位置的输入可以观测到全部的序列,这与预测阶段的情形项矛盾,所以我们要通过将第t个时间步所对应的可观测长度设置为t,以消除不需要看到的未来的信息。
Decoder Block 基础块
class DecoderBlock(nn.Module):
def __init__(self, embedding_size, ffn_hidden_size, num_heads,dropout,i,**kwargs):
super(DecoderBlock, self).__init__(**kwargs)
self.i = i
self.attention_1 = MultiHeadAttention(embedding_size, embedding_size, num_heads, dropout)
self.addnorm_1 = AddNorm(embedding_size, dropout)
self.attention_2 = MultiHeadAttention(embedding_size, embedding_size, num_heads, dropout)
self.addnorm_2 = AddNorm(embedding_size, dropout)
self.ffn = PositionWiseFFN(embedding_size, ffn_hidden_size, embedding_size)
self.addnorm_3 = AddNorm(embedding_size, dropout)
def forward(self, X, state):
enc_outputs, enc_valid_length = state[0], state[1]
# Example Demo:
# love dogs ! [EOS]
# | | | |
# Transformer
# Decoder
# | | | |
# I love dogs !
if state[2][self.i] is None:
key_values = X
else:
# shape of key_values = (batch_size, t, hidden_size)
key_values = torch.cat((state[2][self.i], X), dim=1)
state[2][self.i] = key_values
if self.training:
batch_size, seq_len, _ = X.shape
# Shape: (batch_size, seq_len), the values in the j-th column are j+1
valid_length = torch.FloatTensor(np.tile(np.arange(1, seq_len+1), (batch_size, 1)))
valid_length = valid_length.to(X.device)
else:
valid_length = None
X2 = self.attention_1(X, key_values, key_values, valid_length)
Y = self.addnorm_1(X, X2)
Y2 = self.attention_2(Y, enc_outputs, enc_outputs, enc_valid_length)
Z = self.addnorm_2(Y, Y2)
return self.addnorm_3(Z, self.ffn(Z)), state
decoder_blk = DecoderBlock(24, 48, 8, 0.5, 0)
X = torch.ones((2, 100, 24))
state = [encoder_blk(X, valid_length), valid_length, [None]]
decoder_blk(X, state)[0].shape
对于 Transformer 解码器来说,构造方式与编码器一样,除了最后一层添加一个 dense layer 以获得输出的置信度分数。
Transformer Decoder 参数设置:
- 编码器的输出 enc_outputs
- 句子有效长度 enc_valid_length
- 常规的超参数
class TransformerDecoder(d2l.Decoder):
def __init__(self, vocab_size, embedding_size, ffn_hidden_size,
num_heads, num_layers, dropout, **kwargs):
super(TransformerDecoder, self).__init__(**kwargs)
self.embedding_size = embedding_size
self.num_layers = num_layers
self.embed = nn.Embedding(vocab_size, embedding_size)
self.pos_encoding = PositionalEncoding(embedding_size, dropout)
self.blks = nn.ModuleList()
for i in range(num_layers):
self.blks.append(
DecoderBlock(embedding_size, ffn_hidden_size, num_heads,
dropout, i))
self.dense = nn.Linear(embedding_size, vocab_size)
def init_state(self, enc_outputs, enc_valid_length, *args):
return [enc_outputs, enc_valid_length, [None]*self.num_layers]
def forward(self, X, state):
X = self.pos_encoding(self.embed(X) * math.sqrt(self.embedding_size))
for blk in self.blks:
X, state = blk(X, state)
return self.dense(X), state
机器翻译模型 Transformer 训练
import zipfile
import torch
import requests
from io import BytesIO
from torch.utils import data
import sys
import collections
class Vocab(object):
def __init__(self, tokens, min_freq=0, use_special_tokens=False):
# sort by frequency and token
counter = collections.Counter(tokens)
token_freqs = sorted(counter.items(), key=lambda x: x[0])
token_freqs.sort(key=lambda x: x[1], reverse=True)
if use_special_tokens:
# padding, begin of sentence, end of sentence, unknown
self.pad, self.bos, self.eos, self.unk = (0, 1, 2, 3)
tokens = ['', '', '', '']
else:
self.unk = 0
tokens = ['']
tokens += [token for token, freq in token_freqs if freq >= min_freq]
self.idx_to_token = []
self.token_to_idx = dict()
for token in tokens:
self.idx_to_token.append(token)
self.token_to_idx[token] = len(self.idx_to_token) - 1
def __len__(self):
return len(self.idx_to_token)
def __getitem__(self, tokens):
if not isinstance(tokens, (list, tuple)):
return self.token_to_idx.get(tokens, self.unk)
else:
return [self.__getitem__(token) for token in tokens]
def to_tokens(self, indices):
if not isinstance(indices, (list, tuple)):
return self.idx_to_token[indices]
else:
return [self.idx_to_token[index] for index in indices]
def load_data_nmt(batch_size, max_len, num_examples=1000):
"""Download an NMT dataset, return its vocabulary and data iterator."""
# Download and preprocess
def preprocess_raw(text):
text = text.replace('\u202f', ' ').replace('\xa0', ' ')
out = ''
for i, char in enumerate(text.lower()):
if char in (',', '!', '.') and text[i-1] != ' ':
out += ' '
out += char
return out
with open('path to data dile', 'r') as f:
raw_text = f.read()
text = preprocess_raw(raw_text)
# Tokenize
source, target = [], []
for i, line in enumerate(text.split('\n')):
if i >= num_examples:
break
parts = line.split('\t')
if len(parts) >= 2:
source.append(parts[0].split(' '))
target.append(parts[1].split(' '))
# Build vocab
def build_vocab(tokens):
tokens = [token for line in tokens for token in line]
return Vocab(tokens, min_freq=3, use_special_tokens=True)
src_vocab, tgt_vocab = build_vocab(source), build_vocab(target)
# Convert to index arrays
def pad(line, max_len, padding_token):
if len(line) > max_len:
return line[:max_len]
return line + [padding_token] * (max_len - len(line))
def build_array(lines, vocab, max_len, is_source):
lines = [vocab[line] for line in lines]
if not is_source:
lines = [[vocab.bos] + line + [vocab.eos] for line in lines]
array = torch.tensor([pad(line, max_len, vocab.pad) for line in lines])
valid_len = (array != vocab.pad).sum(1)
return array, valid_len
src_vocab, tgt_vocab = build_vocab(source), build_vocab(target)
src_array, src_valid_len = build_array(source, src_vocab, max_len, True)
tgt_array, tgt_valid_len = build_array(target, tgt_vocab, max_len, False)
train_data = data.TensorDataset(src_array, src_valid_len, tgt_array, tgt_valid_len)
train_iter = data.DataLoader(train_data, batch_size, shuffle=True)
return src_vocab, tgt_vocab, train_iter
import os
import d2l
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
embed_size, embedding_size, num_layers, dropout = 32, 32, 2, 0.05
batch_size, num_steps = 64, 10
lr, num_epochs, ctx = 0.005, 250, d2l.try_gpu()
print(ctx)
num_hiddens, num_heads = 64, 4
src_vocab, tgt_vocab, train_iter = load_data_nmt(batch_size, num_steps)
encoder = TransformerEncoder(
len(src_vocab), embedding_size, num_hiddens, num_heads, num_layers,
dropout)
decoder = TransformerDecoder(
len(src_vocab), embedding_size, num_hiddens, num_heads, num_layers,
dropout)
model = d2l.EncoderDecoder(encoder, decoder)
d2l.train_s2s_ch9(model, train_iter, lr, num_epochs, ctx)
model.eval()
for sentence in ['Go .', 'Wow !', "I'm OK .", 'I won !']:
print(sentence + ' => ' + d2l.predict_s2s_ch9(
model, sentence, src_vocab, tgt_vocab, num_steps, ctx))
以上是利用 Transformer 机器翻译模型实现翻译 Demo 的全部分块,针对其中的层归一化进行总结:
层归一化
- 层归一化有利于加快收敛,减少训练时间成本
- 层归一化对一个中间层的所有神经元进行归一化
- 层归一化的效果不会受到batch大小的影响
补充:
批归一化
每个神经元的输入数据以mini-batch为单位进行汇总