Transformer
- 关于Transformer的理论学习:Transformer详解
- 本文的源代码地址:Transformer代码地址
- Colab实验地址:Colab
0. 架构图
1. 数据准备
import torch
import numpy as np
import torch.nn as nn
import torch.optim as optim
import torch.utils.data as Data
# Encoder_input Decoder_input Decoder_output
sentences = [['我 是 学 生 P', 'S I am a student', 'I am a student E'], # S: 开始符号
['我 喜 欢 学 习', 'S I like learning P', 'I like learning P E'], # E: 结束符号
['我 是 男 生 P', 'S I am a boy', 'I am a boy E']] # P: 占位符号,如果当前句子不足固定长度用P占位
src_vocab = {'P': 0, '我': 1, '是': 2, '学': 3, '生': 4, '喜': 5, '欢': 6, '习': 7, '男': 8} # 词源字典 字:索引
src_idx2word = {src_vocab[key]: key for key in src_vocab}
src_vocab_size = len(src_vocab) # 字典字的个数
tgt_vocab = {'S': 0, 'E': 1, 'P': 2, 'I': 3, 'am': 4, 'a': 5, 'student': 6, 'like': 7, 'learning': 8, 'boy': 9}
idx2word = {tgt_vocab[key]: key for key in tgt_vocab} # 把目标字典转换成 索引:字的形式
tgt_vocab_size = len(tgt_vocab) # 目标字典尺寸
src_len = len(sentences[0][0].split(" ")) # Encoder输入的最大长度
tgt_len = len(sentences[0][1].split(" ")) # Decoder输入输出最大长度
# 把sentences 转换成字典索引
def make_data(sentences):
enc_inputs, dec_inputs, dec_outputs = [], [], []
for i in range(len(sentences)):
enc_input = [[src_vocab[n] for n in sentences[i][0].split()]]
dec_input = [[tgt_vocab[n] for n in sentences[i][1].split()]]
dec_output = [[tgt_vocab[n] for n in sentences[i][2].split()]]
enc_inputs.extend(enc_input)
dec_inputs.extend(dec_input)
dec_outputs.extend(dec_output)
return torch.LongTensor(enc_inputs), torch.LongTensor(dec_inputs), torch.LongTensor(dec_outputs)
enc_inputs, dec_inputs, dec_outputs = make_data(sentences)
# print(enc_inputs, enc_inputs.shape)
# 自定义数据集函数
class MyDataSet(Data.Dataset):
def __init__(self, enc_inputs, dec_inputs, dec_outputs):
super(MyDataSet, self).__init__()
self.enc_inputs = enc_inputs
self.dec_inputs = dec_inputs
self.dec_outputs = dec_outputs
def __len__(self):
return self.enc_inputs.shape[0]
def __getitem__(self, idx):
return self.enc_inputs[idx], self.dec_inputs[idx], self.dec_outputs[idx]
loader = Data.DataLoader(MyDataSet(enc_inputs, dec_inputs, dec_outputs), 2, True)
# 参数设置
d_model = 512 # 字 Embedding 的维度
d_ff = 2048 # 前向传播隐藏层维度 d_ff = d_model * 4
d_k = d_v = 64 # K(=Q), V的维度
n_layers = 6 # 有多少个encoder和decoder
n_heads = 8 # Multi-Head Attention设置为8 d_k * n_heads = d_model
sentences 里一共有三个训练数据,中文->英文。把Encoder_input、Decoder_input、Decoder_output转换成字典索引,例如"学"->3、"student"->6。再把数据转换成batch大小为2的分组数据,3句话一共可以分成两组,一组2句话、一组1句话。src_len表示中文句子固定最大长度,tgt_len 表示英文句子固定最大长度。
2. 定义位置信息
# Positional Embedding
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout=0.1, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
# pos_table 第一个维度对应第几个词 pos,第二个维度对应每个词的词嵌入维度 i,[max_len, d_model]
pos_table = np.array([
[pos / np.power(10000, 2 * i / d_model) for i in range(d_model)] # i 针对词嵌入的维度
if pos != 0 else np.zeros(d_model) for pos in range(max_len)]) # pos 针对第几个词
pos_table[1:, 0::2] = np.sin(pos_table[1:, 0::2]) # 字嵌入维度为偶数时
pos_table[1:, 1::2] = np.cos(pos_table[1:, 1::2]) # 字嵌入维度为奇数时
self.pos_table = torch.FloatTensor(pos_table).cuda() # enc_inputs: [seq_len, d_model]
def forward(self, enc_inputs): # enc_inputs: [batch_size, seq_len, d_model]
enc_inputs += self.pos_table[:enc_inputs.size(1), :] # 注意,tensor的size(n)表示第n的维度大小,与numpy的array.size不一样,np只是个数值
return self.dropout(enc_inputs.cuda())
生成位置信息矩阵pos_table,直接加上输入的enc_inputs上,得到带有位置信息的字向量,pos_table是一个固定值的矩阵。这里矩阵加法利用到了广播机制.
3. Mask掉停用词
# Mask掉停用词
def get_attn_pad_mask(seq_q, seq_k): # seq_q: [batch_size, seq_len] ,seq_k: [batch_size, seq_len]
batch_size, len_q = seq_q.size()
batch_size, len_k = seq_k.size()
pad_attn_mask = seq_k.data.eq(0).unsqueeze(1) # 判断 输入那些含有P(=0),用1标记 ,[batch_size, 1, len_k]
return pad_attn_mask.expand(batch_size, len_q, len_k) # 扩展成多维度, [batch_size, len_q, len_k]
# Decoder 输入 Mask
def get_attn_subsequence_mask(seq): # seq: [batch_size, tgt_len]
attn_shape = [seq.size(0), seq.size(1), seq.size(1)]
subsequence_mask = np.triu(np.ones(attn_shape), k=1) # 生成上三角矩阵,[batch_size, tgt_len, tgt_len], 数值为1
subsequence_mask = torch.from_numpy(subsequence_mask).byte() # [batch_size, tgt_len, tgt_len]
return subsequence_mask
-
Mask句子中没有实际意义的占位符,例如'我 是 学 生 P' ,P对应句子没有实际意义,所以需要被Mask,Encoder_input 和Decoder_input占位符都需要被Mask。
-
用来Mask未来输入信息,返回的是一个上三角矩阵。比如我们在中英文翻译时候,会先把"我是学生"整个句子输入到Encoder中,得到最后一层的输出后,才会在Decoder输入"S I am a student"(s表示开始),但是"S I am a student"这个句子我们不会一起输入,而是在T0时刻先输入"S"预测,预测第一个词"I";在下一个T1时刻,同时输入"S"和"I"到Decoder预测下一个单词"am";然后在T2时刻把"S,I,am"同时输入到Decoder预测下一个单词"a",依次把整个句子输入到Decoder,预测出"I am a student E"。
4. 计算注意力信息、残差和归一化、前馈神经网络
# 计算注意力信息、残差和归一化
class ScaledDotProductAttention(nn.Module):
def __init__(self):
super(ScaledDotProductAttention, self).__init__()
def forward(self, Q, K, V, attn_mask):
# Q: [batch_size, n_heads, len_q, d_k]
# K: [batch_size, n_heads, len_k, d_k]
# V: [batch_size, n_heads, len_v(=len_k), d_v]
# attn_mask: [batch_size, n_heads, seq_len, seq_len]
scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(d_k) # scores : [batch_size, n_heads, len_q, len_k]
scores.masked_fill_(attn_mask, -1e9) # 如果时停用词P就等于 0
attn = nn.Softmax(dim=-1)(scores)
context = torch.matmul(attn, V) # [batch_size, n_heads, len_q, d_v]
return context, attn
class MultiHeadAttention(nn.Module):
def __init__(self):
super(MultiHeadAttention, self).__init__()
self.W_Q = nn.Linear(d_model, d_k * n_heads, bias=False)
self.W_K = nn.Linear(d_model, d_k * n_heads, bias=False)
self.W_V = nn.Linear(d_model, d_v * n_heads, bias=False)
self.fc = nn.Linear(n_heads * d_v, d_model, bias=False) # multi-head 最后concat后需要乘上一个W_O得到一个 总输出b
def forward(self, input_Q, input_K, input_V, attn_mask):
# input_Q: [batch_size, len_q, d_model]
# input_K: [batch_size, len_k, d_model]
# input_V: [batch_size, len_v(=len_k), d_model]
# attn_mask: [batch_size, seq_len, seq_len]
residual, batch_size = input_Q, input_Q.size(0)
# view()函数相当于reshape函数
Q = self.W_Q(input_Q).view(batch_size, -1, n_heads, d_k).transpose(1, 2) # Q: [batch_size, n_heads, len_q, d_k]
K = self.W_K(input_K).view(batch_size, -1, n_heads, d_k).transpose(1, 2) # K: [batch_size, n_heads, len_k, d_k]
V = self.W_V(input_V).view(batch_size, -1, n_heads, d_v).transpose(1, 2) # V: [batch_size, n_heads, len_v(=len_k), d_v]
attn_mask = attn_mask.unsqueeze(1).repeat(1, n_heads, 1, 1) # attn_mask : [batch_size, n_heads, seq_len, seq_len]
context, attn = ScaledDotProductAttention()(Q, K, V, attn_mask) # context: [batch_size, n_heads, len_q, d_v]
# attn: [batch_size, n_heads, len_q, len_k]
context = context.transpose(1, 2).reshape(batch_size, -1,
n_heads * d_v) # context: [batch_size, len_q, n_heads * d_v]
output = self.fc(context) # [batch_size, len_q, d_model]
return nn.LayerNorm(d_model).cuda()(output + residual), attn # ADD And Layer Norm
# 前馈神经网络
class PoswiseFeedForwardNet(nn.Module):
def __init__(self):
super(PoswiseFeedForwardNet, self).__init__()
self.fc = nn.Sequential(
nn.Linear(d_model, d_ff, bias=False),
nn.ReLU(),
nn.Linear(d_ff, d_model, bias=False))
def forward(self, inputs): # inputs: [batch_size, seq_len, d_model]
residual = inputs
output = self.fc(inputs)
return nn.LayerNorm(d_model).cuda()(output + residual) # [batch_size, seq_len, d_model]
-
计算注意力信息,W_Q, W_K, W_V矩阵拆分成8个小矩阵。注意传入的input_Q, input_K, input_V,在Encoder和Decoder的第一次调用传入的三个矩阵是相同的,但Decoder的第二次调用传入的三个矩阵input_Q 等于 input_K 不等于 input_V。
-
输入inputs ,经过两个全连接成,得到的结果再加上 inputs ,再做LayerNorm归一化。LayerNorm归一化可以理解层是把Batch中每一句话进行归一化。
5. Encoder
# 单个Encoder
class EncoderLayer(nn.Module):
def __init__(self):
super(EncoderLayer, self).__init__()
self.enc_self_attn = MultiHeadAttention() # 多头注意力机制
self.pos_ffn = PoswiseFeedForwardNet() # 前馈神经网络
def forward(self, enc_inputs, enc_self_attn_mask): # enc_inputs: [batch_size, src_len, d_model]
# 输入3个enc_inputs分别与W_q、W_k、W_v相乘得到Q、K、V # enc_self_attn_mask: [batch_size, src_len, src_len]
enc_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs, # enc_outputs: [batch_size, src_len, d_model],
enc_self_attn_mask) # attn: [batch_size, n_heads, src_len, src_len]
enc_outputs = self.pos_ffn(enc_outputs) # enc_outputs: [batch_size, src_len, d_model]
return enc_outputs, attn
# 整个Encoder
class Encoder(nn.Module):
def __init__(self):
super(Encoder, self).__init__()
self.src_emb = nn.Embedding(src_vocab_size, d_model) # 把字转换字向量
self.pos_emb = PositionalEncoding(d_model) # 加入位置信息
self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])
def forward(self, enc_inputs): # enc_inputs: [batch_size, src_len]
enc_outputs = self.src_emb(enc_inputs) # enc_outputs: [batch_size, src_len, d_model]
''' token embedding 与 positional embedding如何混合? '''
enc_outputs = self.pos_emb(enc_outputs) # enc_outputs: [batch_size, src_len, d_model]
enc_self_attn_mask = get_attn_pad_mask(enc_inputs, enc_inputs) # enc_self_attn_mask: [batch_size, src_len, src_len]
enc_self_attns = []
for layer in self.layers:
enc_outputs, enc_self_attn = layer(enc_outputs, enc_self_attn_mask) # enc_outputs : [batch_size, src_len, d_model],
# enc_self_attn : [batch_size, n_heads, src_len, src_len]
enc_self_attns.append(enc_self_attn)
return enc_outputs, enc_self_attns
- 第一步,中文字索引进行Embedding,转换成512维度的字向量。第二步,在子向量上面加上位置信息。第三步,Mask掉句子中的占位符号。第四步,通过6层的encoder(上一层的输出作为下一层的输入)。
6. Decoder
# 单个decoder
class DecoderLayer(nn.Module):
def __init__(self):
super(DecoderLayer, self).__init__()
self.dec_self_attn = MultiHeadAttention()
self.dec_enc_attn = MultiHeadAttention()
self.pos_ffn = PoswiseFeedForwardNet()
def forward(self, dec_inputs, enc_outputs, dec_self_attn_mask, dec_enc_attn_mask): # dec_inputs: [batch_size, tgt_len, d_model]
# enc_outputs: [batch_size, src_len, d_model]
# dec_self_attn_mask: [batch_size, tgt_len, tgt_len]
# dec_enc_attn_mask: [batch_size, tgt_len, src_len]
dec_outputs, dec_self_attn = self.dec_self_attn(dec_inputs, dec_inputs,
dec_inputs, dec_self_attn_mask) # dec_outputs: [batch_size, tgt_len, d_model]
# dec_self_attn: [batch_size, n_heads, tgt_len, tgt_len]
dec_outputs, dec_enc_attn = self.dec_enc_attn(dec_outputs, enc_outputs,
enc_outputs, dec_enc_attn_mask) # dec_outputs: [batch_size, tgt_len, d_model]
# dec_enc_attn: [batch_size, h_heads, tgt_len, src_len]
dec_outputs = self.pos_ffn(dec_outputs) # dec_outputs: [batch_size, tgt_len, d_model]
return dec_outputs, dec_self_attn, dec_enc_attn
# 整个Decoder
class Decoder(nn.Module):
def __init__(self):
super(Decoder, self).__init__()
self.tgt_emb = nn.Embedding(tgt_vocab_size, d_model)
self.pos_emb = PositionalEncoding(d_model)
self.layers = nn.ModuleList([DecoderLayer() for _ in range(n_layers)])
def forward(self, dec_inputs, enc_inputs, enc_outputs): # dec_inputs: [batch_size, tgt_len]
# enc_intpus: [batch_size, src_len]
# enc_outputs: [batsh_size, src_len, d_model]
dec_outputs = self.tgt_emb(dec_inputs) # [batch_size, tgt_len, d_model]
dec_outputs = self.pos_emb(dec_outputs).cuda() # [batch_size, tgt_len, d_model]
dec_self_attn_pad_mask = get_attn_pad_mask(dec_inputs, dec_inputs).cuda() # [batch_size, tgt_len, tgt_len]
dec_self_attn_subsequence_mask = get_attn_subsequence_mask(dec_inputs).cuda() # [batch_size, tgt_len, tgt_len]
# gt(a, b) a>b , return true
# >>> torch.gt(torch.tensor([[1, 2], [3, 4]]), torch.tensor([[1, 1], [4, 4]]))
# tensor([[False, True], [False, False]])
dec_self_attn_mask = torch.gt((dec_self_attn_pad_mask +
dec_self_attn_subsequence_mask), 0).cuda() # [batch_size, tgt_len, tgt_len]
dec_enc_attn_mask = get_attn_pad_mask(dec_inputs, enc_inputs) # [batc_size, tgt_len, src_len]
dec_self_attns, dec_enc_attns = [], []
for layer in self.layers: # dec_outputs: [batch_size, tgt_len, d_model]
# dec_self_attn: [batch_size, n_heads, tgt_len, tgt_len]
# dec_enc_attn: [batch_size, h_heads, tgt_len, src_len]
dec_outputs, dec_self_attn, dec_enc_attn = layer(dec_outputs, enc_outputs, dec_self_attn_mask, dec_enc_attn_mask)
dec_self_attns.append(dec_self_attn)
dec_enc_attns.append(dec_enc_attn)
return dec_outputs, dec_self_attns, dec_enc_attns
- 第一步,英文字索引进行Embedding,转换成512维度的字向量。第二步,在子向量上面加上位置信息。第三步,Mask掉句子中的占位符号和输出顺序细节见步骤3。第四步,通过6层的decoder(上一层的输出作为下一层的输入)。
7. Transformer
# Trasformer
class Transformer(nn.Module):
def __init__(self):
super(Transformer, self).__init__()
self.Encoder = Encoder().cuda()
self.Decoder = Decoder().cuda()
self.projection = nn.Linear(d_model, tgt_vocab_size, bias=False).cuda()
def forward(self, enc_inputs, dec_inputs): # enc_inputs: [batch_size, src_len]
# dec_inputs: [batch_size, tgt_len]
enc_outputs, enc_self_attns = self.Encoder(enc_inputs) # enc_outputs: [batch_size, src_len, d_model],
# enc_self_attns: [n_layers, batch_size, n_heads, src_len, src_len]
dec_outputs, dec_self_attns, dec_enc_attns = self.Decoder(
dec_inputs, enc_inputs, enc_outputs) # dec_outpus : [batch_size, tgt_len, d_model],
# dec_self_attns: [n_layers, batch_size, n_heads, tgt_len, tgt_len],
# dec_enc_attn : [n_layers, batch_size, tgt_len, src_len]
dec_logits = self.projection(dec_outputs) # dec_logits: [batch_size, tgt_len, tgt_vocab_size]
return dec_logits.view(-1, dec_logits.size(-1)), enc_self_attns, dec_self_attns, dec_enc_attns
- Trasformer的整体结构,输入数据先通过Encoder,再同个Decoder,最后把输出进行多分类,分类数为英文字典长度,也就是判断每一个字的概率。
8. 定义网络与训练测试
# 定义网络
model = Transformer().cuda()
criterion = nn.CrossEntropyLoss(ignore_index=0) #忽略 占位符 索引为0.
optimizer = optim.SGD(model.parameters(), lr=1e-3, momentum=0.99)
# 训练Transformer
for epoch in range(50):
for enc_inputs, dec_inputs, dec_outputs in loader: # enc_inputs : [batch_size, src_len]
# dec_inputs : [batch_size, tgt_len]
# dec_outputs: [batch_size, tgt_len]
enc_inputs, dec_inputs, dec_outputs = enc_inputs.cuda(), dec_inputs.cuda(), dec_outputs.cuda()
outputs, enc_self_attns, dec_self_attns, dec_enc_attns = model(enc_inputs, dec_inputs)
# outputs: [batch_size * tgt_len, tgt_vocab_size]
loss = criterion(outputs, dec_outputs.view(-1))
print('Epoch:', '%04d' % (epoch + 1), 'loss =', '{:.6f}'.format(loss))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# 测试模型
def test(model, enc_input, start_symbol):
enc_outputs, enc_self_attns = model.Encoder(enc_input)
dec_input = torch.zeros(1, tgt_len).type_as(enc_input.data)
next_symbol = start_symbol
for i in range(0, tgt_len):
dec_input[0][i] = next_symbol
dec_outputs, _, _ = model.Decoder(dec_input, enc_input, enc_outputs)
projected = model.projection(dec_outputs)
prob = projected.squeeze(0).max(dim=-1, keepdim=False)[1]
next_word = prob.data[i]
next_symbol = next_word.item()
return dec_input
enc_inputs, _, _ = next(iter(loader))
predict_dec_input = test(model, enc_inputs[0].view(1, -1).cuda(), start_symbol=tgt_vocab["S"])
predict, _, _, _ = model(enc_inputs[0].view(1, -1).cuda(), predict_dec_input)
predict = predict.data.max(1, keepdim=True)[1]
print([src_idx2word[int(i)] for i in enc_inputs[0]], '->',
[idx2word[n.item()] for n in predict.squeeze()])
9. 一些细节
-
positional embedding在Transformer和bert中方式不一样,positional embbeding的设计是一个课题
-
注意Multi-Head Self-Attention相比正常的Self-Attention只在最后concat的时候多出参数,同时在最后concat后需要经过一个Linear层
-
Wq,Wk,Wv是要训练的参数,其尺寸取决于输入的embedding向量长度和句子长度
-
区分Layer Normalization与Batch Normalization
a = [[[3, 4, 6], [4, 5, 6]],[[4, 4, 4], [4, 5, 6]], [[3, 4, 6], [4, 5, 6]], [[3, 4, 6], [4, 5, 6]]] e = torch.FloatTensor(a) print(e.size(0), e.size(1), e.size(2)) f = torch.nn.LayerNorm(3)(e) m = torch.nn.BatchNorm1d(2, 3) g = m(e)
-
Feed Forward是一个两层的线性层,使用Relu函数
-
Encoder与Decoder连接的attention,Encoder提供k和v,Decoder提供q
-
Decoder的Multi-Head Self-Attention带Mask,Mask在矩阵计算中相当于实现一个上三角矩阵
-
Positional Embedding与Token Embedding相加的操作细节
-
Multi-Head在切分的时候,其数量Head_Num是在Embedding向量维度上切分的
-
参数设置中需要满足的一些条件:d_k(一个head对应的计算矩阵维度) * n_heads(head数目) = d_model(embedding维度);Relu层中的隐藏层神经元数目是embedding维度的4倍:d_ff = d_model * 4
网友评论