**Transformer原理+代码实现机器翻译示例
（注意：Encoder_input，Decoder_input，Decoder_output：训练标签设定，设定模式不能出错，否则模型训练将极其难达到想要的效果，即使loss已经很低了，甚至模型非常优化也不能达到效果）
Transformer原理：
inputs:Encoder_input
Outputs:Decoder_input
Outputs probility:Decoder_output
## 关键部分代码实现：
masked Loss：
（一）
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy
# 关于Word embedding,序列建模为例
# 考source sentence和target sentence
# 构建序列，序列字符以其在此表中索引形式表示（通常还有start < 字符，本篇中各part省略）
batch_size = 2
# 单词表大小(样本最大单词数)
max_num_src_words = 8
max_num_tgt_words = 8
model_dim = 8
# 序列最大长度
max_src_seg_len = 5
max_tgt_seg_len = 5
max_position_len = 5
# src_len = torch.randint(2,5,(batch_size,))
# tgt_len = torch.randint(2,5,(batch_size,))
src_len = torch.Tensor([2, 4]).to(torch.int32)
tgt_len = torch.Tensor([4, 3]).to(torch.int32)
# 单词索引构成源句子和目标句子，且做了padding，默认padding=0
src_seq = torch.cat(
[torch.unsqueeze(F.pad(torch.randint(1, max_num_src_words, (L,)), (0, max_src_seg_len - L)), 0) for L in src_len])
tgt_seq = torch.cat(
[torch.unsqueeze(F.pad(torch.randint(1, max_num_tgt_words, (L,)), (0, max_tgt_seg_len - L)), 0) for L in tgt_len])
# 构造Word embedding,+1指padding的embedding编码
src_embedding_table = nn.Embedding(max_num_src_words + 1, model_dim)
tgt_embedding_table = nn.Embedding(max_num_tgt_words + 1, model_dim)
src_embedding = src_embedding_table(src_seq)
tgt_embedding = tgt_embedding_table(tgt_seq)
# 构造position embedding(忽略start < 字符)
pos_matric = torch.arange(max_position_len).reshape((-1, 1))
# torch.arange(0,8,2)其中8为对应max_src_seg_len或max_tgt_seg_len，依据encoder或decoder
i_matric = torch.pow(10000, torch.arange(0, 8, 2).reshape((1, -1))) / model_dim
pe_embedding_table = torch.zeros(max_position_len, model_dim)
pe_embedding_table[:, 1::2] = torch.sin(pos_matric / i_matric)
pe_embedding_table[:, 0::2] = torch.cos(pos_matric / i_matric)
pe_embedding = nn.Embedding(max_position_len, model_dim)
pe_embedding.weight = nn.Parameter(pe_embedding_table, requires_grad=False)
src_pos = torch.cat([torch.unsqueeze(torch.arange(max(src_len), 0) for _ in src_len)]).to(torch.int32)
tgt_pos = torch.cat([torch.unsqueeze(torch.arange(max(tgt_len), 0) for _ in tgt_len)]).to(torch.int32)
src_pe_embedding = pe_embedding(src_pos)
tgt_pe_embedding = pe_embedding(tgt_pos)
# 构造encoder的self-attention mask
# mask的shape:[batch_size，max_src_len，max_src_len],只为1或-inf
valid_encoder_pos = torch.unsqueeze(
torch.cat([torch.unsqueeze(F.pad(torch.ones(L), (0, max(src_len) - L)), 0) for L in src_len]), 2)
valid_encoder_pos_matrix = torch.bmm(valid_encoder_pos, valid_encoder_pos.transpose(1, 2))
invalid_encoder_pos_matrix = 1 - valid_encoder_pos_matrix
mask_encoder_self_attention = invalid_encoder_pos_matrix.to(torch.bool)
score = torch.randn(batch_size, max(src_len), max(src_len))
masked_score = score.masked_fill(mask_encoder_self_attention, -1e9)
prob_encoder = F.softmax(masked_score, -1)
# # softax演示，scale（根号dk）的重要性
# alphal1 = 0.1
# alphal2 = 10
# score = torch.randn(5)
# prob1 = F.softmax(score*alphal1,-1)
# prob2 = F.softmax(score*alphal2,-1)
# def softmax_func(score):
# return F.softmax(score)
# jaco_matric1 = torch.autograd.functional.jacobian(softmax_func,score*alphal1)
# jaco_matric2 = torch.autograd.functional.jacobian(softmax_func,score*alphal2)
# 构造 intro-attention(cross attention)的mask
# Q @ K^T shape:[batch_size,tgt_seq_len,src_seq_len]
valid_encoder_pos = torch.unsqueeze(
torch.cat([torch.unsqueeze(F.pad(torch.ones(L), (0, max(src_len) - L)), 0) for L in src_len]), 2)
valid_decoder_pos = torch.unsqueeze(
torch.cat([torch.unsqueeze(F.pad(torch.ones(L), (0, max(tgt_len) - L)), 0) for L in tgt_len]), 2)
valid_cross_pos_matrix = torch.bmm(valid_encoder_pos, valid_decoder_pos.transpose(1, 2))
invalid_cross_pos_matrix = 1 - valid_cross_pos_matrix
mask_cross_attention = invalid_cross_pos_matrix.to(torch.bool)
# 构造decoder的self-attention mask(下三角阵)
valid_decoder_tri_matrix = torch.cat(
[torch.unsqueeze(F.pad(torch.tril(torch.ones(L, L)), (0, max(tgt_len) - L, 0, max(tgt_len) - L)), 0) for L in
tgt_len])
invalid_decoder_tri_matrix = 1 - valid_decoder_tri_matrix
invalid_decoder_tri_matrix = invalid_decoder_tri_matrix.to(torch.bool)
score = torch.randn(batch_size, max(tgt_len), max(tgt_len))
masked_score = score.masked_fill(invalid_decoder_tri_matrix)
prob_decoder = F.softmax(masked_score, -1)
# 构造scaled self-attention(多头的每个注意力头)
def scaled_dot_product_attention(Q, K, V, attn_mask):
# shape of Q,K,V: [batch_size*num_head,seq_len,model_dim/num_head]
torch.bmm(Q, K.transpose(-2, -1)) / torch.sqrt(model_dim)
masked_score = score.masked_fill(attn_mask, -1e9)
prob = F.softmax(masked_score, -1)
context = torch.bmm(prob, V)
return context
（二）
def get_attn_subsequence_mask(seq): # seq: [batch_size, tgt_len]
attn_shape = [seq.size(0), seq.size(1), seq.size(1)] # 生成上三角矩阵,[batch_size, tgt_len, tgt_len]
subsequence_mask = np.triu(np.ones(attn_shape), k=1)
subsequence_mask = torch.from_numpy(subsequence_mask).byte() # [batch_size, tgt_len, tgt_len]
return subsequence_mask
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) # 如果是停用词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
# MultiHeadAttention #
# 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)
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]
def get_attn_pad_mask(seq_q, seq_k):
batch_size, len_q = seq_q.size() # seq_q 用于升维，为了做attention，mask score矩阵用的
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]
#### （一） ####
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 = np.array([
[pos / np.power(10000, 2 * i / d_model) for i in range(d_model)]
if pos != 0 else np.zeros(d_model) for pos in range(max_len)])
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) # 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), :]
return self.dropout(enc_inputs)
#### （二） ####
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term) # 这里需要注意的是pe[:, 0::2]这个用法，就是从0开始到最后面，步长为2，其实代表的就是偶数位置
pe[:, 1::2] = torch.cos(position * div_term) # 这里需要注意的是pe[:, 1::2]这个用法，就是从1开始到最后面，步长为2，其实代表的就是奇数位置
# 下面这行代码之后，我们得到的pe形状是：[max_len * 1 * d_model]
pe = pe.unsqueeze(0).transpose(0, 1)
## 自回归嵌套（一）
self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])
## 自回归嵌套（二）
import copy
def clones(module, N):
"Produce N identical layers."
return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])
self.encoder = Encoder(EncoderLayer(config.d_model, deepcopy(attn), deepcopy(ff), dropout), N)
self.src_embed = nn.Sequential(Embeddings(config.d_model, src_vocab),
deepcopy(position)) # Embeddings followed by PE
#### 其他按pytorch API中来即可
**机器翻译源码示例（代码均可跑通）：
注意：Encoder_input，Decoder_input，Decoder_output：训练标签设定，设定模式不能出错，否则模型训练将极其难达到想要的效果，即使loss已经很低了，甚至模型非常优化也不能达到效果
PLUS：训练和测试的Decoder_input不同，训练时是下三角阵mask，测试时一个一个输入
示例一（Encoder_input，Decoder_input，Decoder_output分为正确设定/错误设定，实验结果图片中展现效果差异）：
import math
import torch
import numpy as np
import torch.nn as nn
import torch.optim as optim
import torch.utils.data as Data
# 自制数据集
# #正确设定EG： 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 pad补0
# ['ils regardent . P P','S they are watching .','they are watching . E']]
##----注意：Encoder_input,Decoder_input,Decoder_output的设定模式不能出错，否则模型训练将极其难达到想要的效果，即使loss已经很低了-----##
# 错误设定EG: Encoder_input Decoder_input Decoder_output
sentences = [['我 是 学 生 P', 'I am a student E', 'I am a student E'], # S: 开始符号
['我 喜 欢 学 习', 'I like learning E P', 'I like learning E P'], # E: 结束符号
['我 是 男 生 P', 'I am a boy E', '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 pad补0
['ils regardent . P P','they are watching . E','they are watching . E']]
src_vocab = {'P': 0, '我': 1, '是': 2, '学': 3, '生': 4, '喜': 5, '欢': 6, '习': 7, '男': 8,'ils':9,'regardent':10,'.':11} # 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,'they':10,'are':11,'watching':12,'.':13}
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 5
tgt_len = len(sentences[0][1].split(" ")) # Decoder输入输出最大长度 5
# 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)
# print(dec_inputs)
# print(dec_outputs)
# 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)
d_model = 512 # 字 Embedding 的维度
d_ff = 2048 # 前向传播隐藏层维度
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
###############################构建 Transformer ：type2 #######################
def get_attn_subsequence_mask(seq): # seq: [batch_size, tgt_len]
attn_shape = [seq.size(0), seq.size(1), seq.size(1)] # 生成上三角矩阵,[batch_size, tgt_len, tgt_len]
subsequence_mask = np.triu(np.ones(attn_shape), k=1)
subsequence_mask = torch.from_numpy(subsequence_mask).byte() # [batch_size, tgt_len, tgt_len]
return subsequence_mask
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) # 如果是停用词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)
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)
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)(output + residual), attn
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)(output + residual) # [batch_size, seq_len, d_model]
def get_attn_pad_mask(seq_q, seq_k):
batch_size, len_q = seq_q.size() # seq_q 用于升维，为了做attention，mask score矩阵用的
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]
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 = np.array([
[pos / np.power(10000, 2 * i / d_model) for i in range(d_model)]
if pos != 0 else np.zeros(d_model) for pos in range(max_len)])
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) # 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), :]
return self.dropout(enc_inputs)
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
class Encoder(nn.Module):
def __init__(self):
super(Encoder, self).__init__()
self.src_emb = nn.Embedding(src_vocab_size, d_model)
self.pos_emb = PositionalEncoding(d_model)
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) # [batch_size, src_len, d_model]
enc_outputs = self.pos_emb(enc_outputs.transpose(0, 1)).transpose(0, 1) # [batch_size, src_len, d_model]
enc_self_attn_mask = get_attn_pad_mask(enc_inputs, enc_inputs) # [batch_size, src_len, src_len]
enc_self_attns = []
for layer in self.layers:
# enc_outputs: [batch_size, src_len, d_model], enc_self_attn: [batch_size, n_heads, src_len, src_len]
enc_outputs, enc_self_attn = layer(enc_outputs, enc_self_attn_mask)
enc_self_attns.append(enc_self_attn)
return enc_outputs, enc_self_attns
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
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: [batch_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.transpose(0, 1)).transpose(0, 1) # [batch_size, tgt_len, d_model]
# Decoder输入序列的pad mask矩阵（这个例子中decoder是没有加pad的，实际应用中都是有pad填充的）
dec_self_attn_pad_mask = get_attn_pad_mask(dec_inputs, dec_inputs) # [batch_size, tgt_len, tgt_len]
# Masked Self_Attention：当前时刻是看不到未来的信息的
dec_self_attn_subsequence_mask = get_attn_subsequence_mask(dec_inputs) # [batch_size, tgt_len, tgt_len]
# Decoder中把两种mask矩阵相加（既屏蔽了pad的信息，也屏蔽了未来时刻的信息）
dec_self_attn_mask = torch.gt((dec_self_attn_pad_mask + dec_self_attn_subsequence_mask),
0) # [batch_size, tgt_len, tgt_len]
# 这个mask主要用于encoder-decoder attention层
# get_attn_pad_mask主要是enc_inputs的pad mask矩阵(因为enc是处理K,V的，求Attention时是用v1,v2,..vm去加权的，
# 要把pad对应的v_i的相关系数设为0，这样注意力就不会关注pad向量)
# dec_inputs只是提供expand的size的
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
class Transformer(nn.Module):
def __init__(self):
super(Transformer, self).__init__()
self.Encoder = Encoder()
self.Decoder = Decoder()
self.projection = nn.Linear(d_model, tgt_vocab_size, bias=False)
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 enc_outputs, dec_logits.view(-1, dec_logits.size(-1))
###################################################################################
model = Transformer()
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)
for epoch in range(20):
for enc_inputs, dec_inputs, dec_outputs in loader:
enc_outputs ,outputs = model(enc_inputs, dec_inputs)
loss = criterion(outputs, dec_outputs.view(-1))
print('Epoch:', '%04d' % (epoch + 1), 'loss =', '{:.6f}'.format(loss))
optimizer.zero_grad()
loss.backward()
optimizer.step()
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[1].view(1, -1), start_symbol=tgt_vocab["S"])
_,predict = model(enc_inputs[1].view(1, -1), predict_dec_input)
predict = predict.data.max(1, keepdim=True)[1]
print([src_idx2word[int(i)] for i in enc_inputs[1]], '->',[idx2word[n.item()] for n in predict.squeeze()])
实验结果（正确设定）：
实验结果（错误设定）：
对应上述图片三。
示例二（Encoder_input，Decoder_input，Decoder_output设定是错误的，有热情可以将其改为正确的）：
import collections
import os
import io
import math
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torchtext.vocab as Vocab
import torch.utils.data as Data
import numpy as np
import sys
from torch.utils.data import DataLoader
PAD, BOS, EOS = '<pad>', '<bos>', '<eos>'
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 将一个序列中所有的词记录在all_tokens中以便之后构造词典
# 然后在该序列后面添加PAD直到序列长度为max_seq_len
# 然后将序列保存在all_seqs中
def process_one_seq(seq_tokens, all_tokens, all_seqs, max_seq_len):
all_tokens.extend(seq_tokens)
seq_tokens += [EOS] + [PAD] * (max_seq_len - len(seq_tokens) - 1)
all_seqs.append(seq_tokens)
# 使用所有的词构造词典。并将所有序列中的词变为索引后构造Tensor
def build_data(all_tokens, all_seqs):
vocab = Vocab.Vocab(collections.Counter(all_tokens), specials=[PAD, BOS, EOS])
indices = [[vocab.stoi[w] for w in seq] for seq in all_seqs]
return vocab, torch.tensor(indices)
def read_data(max_seq_len):
# in和out分别是input和output的缩写
in_tokens, out_tokens, in_seqs, out_seqs = [], [], [], []
with io.open('fr-en-small.txt') as f:
lines = f.readlines()
for line in lines:
in_seq, out_seq = line.rstrip().split('\t')
in_seq_tokens = in_seq.split(' ')
out_seq_tokens = out_seq.split(' ')
# print(out_seq_tokens)
if max(len(in_seq_tokens), len(out_seq_tokens)) > max_seq_len - 1:
# 如果加上EOS后长于max_seq_len，则忽略掉此样本
continue
process_one_seq(in_seq_tokens, in_tokens, in_seqs, max_seq_len)
process_one_seq(out_seq_tokens, out_tokens, out_seqs, max_seq_len)
in_vocab, in_data = build_data(in_tokens, in_seqs)
out_vocab, out_data = build_data(out_tokens, out_seqs)
return in_vocab, out_vocab, Data.TensorDataset(in_data, out_data)
############################构建Transformer ：type1#################################
# get_attn_subsequent_mask的实现
# ----------------------------------#
def get_attn_subsequent_mask(seq):
"""
seq: [batch_size, tgt_len]
"""
attn_shape = [seq.size(0), seq.size(1), seq.size(1)]
# attn_shape: [batch_size, tgt_len, tgt_len]
subsequence_mask = np.triu(np.ones(attn_shape), k=1) # 生成一个上三角矩阵
subsequence_mask = torch.from_numpy(subsequence_mask).byte()
return subsequence_mask # [batch_size, tgt_len, tgt_len]
# ----------------------------------#
# ScaledDotProductAttention的实现
# ----------------------------------#
class ScaledDotProductAttention(nn.Module):
def __init__(self):
super(ScaledDotProductAttention, self).__init__()
def forward(self, Q, K, V, attn_mask):
# 输入进来的维度分别是 [batch_size x n_heads x len_q x d_k] K： [batch_size x n_heads x len_k x d_k]
# V: [batch_size x n_heads x len_k x d_v]
# 首先经过matmul函数得到的scores形状是: [batch_size x n_heads x len_q x len_k]
scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(d_k)
# 然后最关键的地方来了，下面这个就是用到了我们之前重点讲的attn_mask，把被mask的地方置为无限小，softmax之后基本就是0，对其他单词就不会起作用
scores.masked_fill_(attn_mask, -1e9) # Fills elements of self tensor with value where mask is one.
attn = nn.Softmax(dim=-1)(scores)
context = torch.matmul(attn, V)
return context, attn
# -------------------------#
# MultiHeadAttention的实现
# -------------------------#
class MultiHeadAttention(nn.Module):
def __init__(self):
super(MultiHeadAttention, self).__init__()
# 输入进来的QKV是相等的，我们会使用linear做一个映射得到参数矩阵Wq, Wk,Wv
self.W_Q = nn.Linear(d_model, d_k * n_heads)
self.W_K = nn.Linear(d_model, d_k * n_heads)
self.W_V = nn.Linear(d_model, d_v * n_heads)
self.linear = nn.Linear(n_heads * d_v, d_model)
self.layer_norm = nn.LayerNorm(d_model)
def forward(self, Q, K, V, attn_mask):
# 这个多头注意力机制分为这几个步骤，首先映射分头，然后计算atten_scores，然后计算atten_value;
# 输入进来的数据形状： Q: [batch_size x len_q x d_model], K: [batch_size x len_k x d_model],
# V: [batch_size x len_k x d_model]
residual, batch_size = Q, Q.size(0)
# (B, S, D) -proj-> (B, S, D) -split-> (B, S, H, W) -trans-> (B, H, S, W)
# 下面这个就是先映射，后分头；一定要注意的是q和k分头之后维度是一致额，所以一看这里都是dk
# q_s: [batch_size x n_heads x len_q x d_k]
q_s = self.W_Q(Q).view(batch_size, -1, n_heads, d_k).transpose(1, 2)
# k_s: [batch_size x n_heads x len_k x d_k]
k_s = self.W_K(K).view(batch_size, -1, n_heads, d_k).transpose(1, 2)
# v_s: [batch_size x n_heads x len_k x d_v]
v_s = self.W_V(V).view(batch_size, -1, n_heads, d_v).transpose(1, 2)
# 输入进来的attn_mask形状是batch_size x len_q x len_k，然后经过下面这个代码得到
# 新的attn_mask: [batch_size x n_heads x len_q x len_k]，就是把pad信息重复到了n个头上
attn_mask = attn_mask.unsqueeze(1).repeat(1, n_heads, 1, 1)
# 然后我们运行ScaledDotProductAttention这个函数
# 得到的结果有两个：context: [batch_size x n_heads x len_q x d_v], attn: [batch_size x n_heads x len_q x len_k]
context, attn = ScaledDotProductAttention()(q_s, k_s, v_s, attn_mask)
# context: [batch_size x len_q x n_heads * d_v]
context = context.transpose(1, 2).contiguous().view(batch_size, -1, n_heads * d_v)
output = self.linear(context)
return self.layer_norm(output + residual), attn # output: [batch_size x len_q x d_model]
# ----------------------------#
# PoswiseFeedForwardNet的实现
# ----------------------------#
class PoswiseFeedForwardNet(nn.Module):
def __init__(self):
super(PoswiseFeedForwardNet, self).__init__()
self.conv1 = nn.Conv1d(in_channels=d_model, out_channels=d_ff, kernel_size=1)
self.conv2 = nn.Conv1d(in_channels=d_ff, out_channels=d_model, kernel_size=1)
self.layer_norm = nn.LayerNorm(d_model)
def forward(self, inputs):
residual = inputs # inputs : [batch_size, len_q, d_model]
output = nn.ReLU()(self.conv1(inputs.transpose(1, 2)))
output = self.conv2(output).transpose(1, 2)
return self.layer_norm(output + residual)
# --------------------------#
# get_attn_pad_mask的实现：
# --------------------------#
# 比如说，我现在的句子长度是5，在后面注意力机制的部分，我们在计算出来QK转置除以根号之后，softmax之前，我们得到的形状len_input * len*input
# 代表每个单词对其余包含自己的单词的影响力。所以这里我需要有一个同等大小形状的矩阵，告诉我哪个位置是PAD部分，之后在计算softmax之前会把这里置
# 为无穷大；一定需要注意的是这里得到的矩阵形状是batch_size x len_q x len_k，我们是对k中的pad符号进行标识，并没有对k中的做标识，因为没必要。
# seq_q和seq_k不一定一致，在交互注意力，q来自解码端，k来自编码端，所以告诉模型编码这边的pad符号信息就可以，解码端的pad信息在交互注意力层是
# 没有用到的；
def get_attn_pad_mask(seq_q, seq_k):
batch_size, len_q = seq_q.size()
batch_size, len_k = seq_k.size()
# eq(zero) is PAD token
pad_attn_mask = seq_k.data.eq(0).unsqueeze(1) # batch_size x 1 x len_k, one is masking
return pad_attn_mask.expand(batch_size, len_q, len_k) # batch_size x len_q x len_k
# ------------------------------#
# Positional Encoding的代码实现
# ------------------------------#
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout, max_len=5000):
super(PositionalEncoding, self).__init__()
# 位置编码的实现其实很简单，直接对照着公式去敲代码就可以，下面的代码只是其中的一种实现方式；
# 从理解来讲，需要注意的就是偶数和奇数在公式上有一个共同部分，我们使用log函数把次方拿下来，方便计算；
# pos代表的是单词在句子中的索引，这点需要注意；比如max_len是128个，那么索引就是从0，1，2，...,127
# 假设我的d_model是512，2i以步长2从0取到了512，那么i对应取值就是0,1,2...255
self.dropout = nn.Dropout(p=dropout)
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term) # 这里需要注意的是pe[:, 0::2]这个用法，就是从0开始到最后面，步长为2，其实代表的就是偶数位置
pe[:, 1::2] = torch.cos(position * div_term) # 这里需要注意的是pe[:, 1::2]这个用法，就是从1开始到最后面，步长为2，其实代表的就是奇数位置
# 下面这行代码之后，我们得到的pe形状是：[max_len * 1 * d_model]
pe = pe.unsqueeze(0).transpose(0, 1)
self.register_buffer('pe', pe) # 定一个缓冲区，其实简单理解为这个参数不更新就可以
def forward(self, x):
"""
x: [seq_len, batch_size, d_model]
"""
x = x + self.pe[:x.size(0), :]
return self.dropout(x)
# ---------------------------------------------------#
# EncoderLayer：包含两个部分，多头注意力机制和前馈神经网络
# ---------------------------------------------------#
class EncoderLayer(nn.Module):
def __init__(self):
super(EncoderLayer, self).__init__()
self.enc_self_attn = MultiHeadAttention()
self.pos_ffn = PoswiseFeedForwardNet()
def forward(self, enc_inputs, enc_self_attn_mask):
"""
下面这个就是做自注意力层，输入是enc_inputs，形状是[batch_size x seq_len_q x d_model]，需要注意的是最初始的QKV矩阵是等同于这个
输入的，去看一下enc_self_attn函数.
"""
# enc_inputs to same Q,K,V
enc_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs, enc_self_attn_mask)
# enc_outputs: [batch_size x len_q x d_model]
enc_outputs = self.pos_ffn(enc_outputs)
return enc_outputs, attn
# -----------------------------------------------------------------------------#
# Encoder部分包含三个部分：词向量embedding，位置编码部分，自注意力层及后续的前馈神经网络
# -----------------------------------------------------------------------------#
class Encoder(nn.Module):
def __init__(self):
super(Encoder, self).__init__()
# 这行其实就是生成一个矩阵，大小是: src_vocab_size * d_model
self.src_emb = nn.Embedding(src_vocab_size, d_model)
# 位置编码，这里是固定的正余弦函数，也可以使用类似词向量的nn.Embedding获得一个可以更新学习的位置编码
self.pos_emb = PositionalEncoding(d_model, dropout)
# 使用ModuleList对多个encoder进行堆叠，因为后续的encoder并没有使用词向量和位置编码，所以抽离出来；
self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])
def forward(self, enc_inputs):
"""
这里我们的enc_inputs形状是： [batch_size x source_len]
"""
# 下面这行代码通过src_emb进行索引定位，enc_outputs输出形状是[batch_size, src_len, d_model]
enc_outputs = self.src_emb(enc_inputs)
# 这行是位置编码，把两者相加放到了pos_emb函数里面
enc_outputs = self.pos_emb(enc_outputs.transpose(0, 1)).transpose(0, 1)
# get_attn_pad_mask是为了得到句子中pad的位置信息，给到模型后面，在计算自注意力和交互注意力的时候去掉pad符号的影响
enc_self_attn_mask = get_attn_pad_mask(enc_inputs, enc_inputs)
enc_self_attns = []
for layer in self.layers:
# 去看EncoderLayer层函数
enc_outputs, enc_self_attn = layer(enc_outputs, enc_self_attn_mask)
enc_self_attns.append(enc_self_attn)
return enc_outputs, enc_self_attns
# --------------------#
# DecoderLayer的实现
# --------------------#
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_outputs, dec_self_attn = self.dec_self_attn(dec_inputs, dec_inputs, dec_inputs, dec_self_attn_mask)
dec_outputs, dec_enc_attn = self.dec_enc_attn(dec_outputs, enc_outputs, enc_outputs, dec_enc_attn_mask)
dec_outputs = self.pos_ffn(dec_outputs)
return dec_outputs, dec_self_attn, dec_enc_attn
# ----------------#
# Decoder的实现
# ----------------#
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, dropout)
self.layers = nn.ModuleList([DecoderLayer() for _ in range(n_layers)])
def forward(self, dec_inputs, enc_inputs, enc_outputs): # dec_inputs : [batch_size x target_len]
dec_outputs = self.tgt_emb(dec_inputs) # [batch_size, tgt_len, d_model]
dec_outputs = self.pos_emb(dec_outputs.transpose(0, 1)).transpose(0, 1) # [batch_size, tgt_len, d_model]
# get_attn_pad_mask 自注意力层的时候的pad 部分
dec_self_attn_pad_mask = get_attn_pad_mask(dec_inputs, dec_inputs)
# get_attn_subsequent_mask 这个做的是自注意层的mask部分，就是当前单词之后看不到，使用一个上三角为1的矩阵
dec_self_attn_subsequent_mask = get_attn_subsequent_mask(dec_inputs)
# 两个矩阵相加，大于0的为1，不大于0的为0，为1的在之后就会被fill到无限小
dec_self_attn_mask = torch.gt((dec_self_attn_pad_mask + dec_self_attn_subsequent_mask), 0)
# 这个做的是交互注意力机制中的mask矩阵，enc的输入是k，我去看这个k里面哪些是pad符号，给到后面的模型；注意哦，我q肯定也是有pad符号，
# 但是这里我不在意的，之前说了好多次了哈
dec_enc_attn_mask = get_attn_pad_mask(dec_inputs, enc_inputs)
dec_self_attns, dec_enc_attns = [], []
for layer in self.layers:
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
# --------------------------------------------------#
# 从整体网络结构来看，分为三个部分：编码层，解码层，输出层
# --------------------------------------------------#
class Transformer(nn.Module):
def __init__(self):
super(Transformer, self).__init__()
self.encoder = Encoder() # 编码层
self.decoder = Decoder() # 解码层
# 输出层的d_model是我们解码层每个token输出的维度大小，之后会做一个tgt_vocab_size大小的softmax
self.projection = nn.Linear(d_model, tgt_vocab_size, bias=False)
def forward(self, enc_inputs, dec_inputs):
"""
这里有两个数据进行输入，一个是enc_inputs，形状为[batch_size, src_len]，主要是作为编码端的输入，一个是dec_inputs，
形状为[batch_size, tgt_len]，主要是作为解码端的输入.
enc_inputs作为输入，形状为[batch_size, src_len]，输出由自己的函数内部指定，想要什么指定输出什么，可以是全部tokens的输出，
可以是特定每一层的输出，也可以是中间某些参数的输出；
"""
# enc_outputs就是编码端的输出，enc_self_attns这里没记错的话是QK转置相乘经softmax之后的矩阵值，代表
# 的是每个单词和其他单词的相关性，即相关性矩阵；
enc_outputs, enc_self_attns = self.encoder(enc_inputs)
# dec_outputs是decoder的主要输出，用于后续的linear映射； dec_self_attns类比于enc_self_attns，
# 是查看每个单词对decoder中输入的其余单词的相关性；dec_enc_attns是decoder中每个单词对encoder中每
# 个单词的相关性；
dec_outputs, dec_self_attns, dec_enc_attns = self.decoder(dec_inputs, enc_inputs, enc_outputs)
# print(dec_outputs,dec_outputs.shape)
# dec_outputs做映射到词表大小
# dec_logits : [batch_size x src_vocab_size x tgt_vocab_size]
dec_logits = self.projection(dec_outputs)
# return dec_logits.view(-1, dec_logits.size(-1)), enc_self_attns, dec_self_attns, dec_enc_attns
# dec_output = self.softmax(dec_outputs).argmax(dim=-1).to(torch.float32)
return enc_outputs, dec_logits.view(-1, dec_logits.size(-1))
##########################################################################################
def translate(model, input_seq, seq_len):
in_tokens = input_seq.split(' ')
in_tokens += [EOS] + [PAD] * (seq_len - len(in_tokens) - 1)
# batch=1
enc_input = torch.tensor([[in_vocab.stoi[tk] for tk in in_tokens]])
print("input sentence:",in_tokens)
batch_sizes = 1
out_tokens = [BOS]
# out_tokens += [BOS] + [PAD] * (max_seq_len - len(out_tokens) - 1)
# print("dec_inputs:", out_tokens)
dec_inputs = torch.tensor([[out_vocab.stoi[tk] for tk in out_tokens]])
output_tokens = []
for _ in range(seq_len):
enc_outputs, dec_output = model(enc_input, dec_inputs)
# print(dec_output.shape)
outputs = dec_output.view(batch_sizes, -1, tgt_vocab_size)
dec_outputs = F.softmax(outputs, dim=-1).argmax(dim=-1)
# print(dec_outputs.shape)
pred = torch.squeeze(dec_outputs, 0)
# print(pred,pred.shape)
pred_words = [out_vocab.itos[int(tk)] for tk in pred]
# print("dec_outputs:", pred_words)
pred_word = out_vocab.itos[int(pred)]
# print("pre_add_word:",pred_word)
# 当任一时间步搜索出现EOS时，输出序列即完成
if EOS == pred_word:
break
if PAD == pred_word:
continue
else:
output_tokens.append(pred_word)
out_token = [BOS] + output_tokens
dec_inputs = torch.tensor([[out_vocab.stoi[tk] for tk in out_tokens]])
# print("dec_inputs:", out_token)
# dec_inputs = torch.unsqueeze(pred,0)
# print("output_words:", output_tokens)
# out_tokens[i] =
# out_tokens[i + 1:] = [PAD]
return output_tokens
def train(model, dataset, lr, batch_size, num_epochs):
data_loader = DataLoader(dataset=dataset, batch_size=batch_size, shuffle=True, drop_last=False)
criterion = nn.CrossEntropyLoss(ignore_index=0)
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.99)
for epoch in range(num_epochs):
losses = 0
for enc_inputs, dec_inputs in data_loader:
target_batch = dec_inputs
enc_outputs, outputs = model(enc_inputs, dec_inputs)
# print(outputs.shape)
loss = criterion(outputs, target_batch.view(-1)) # 训练集、测试集和标签的设定对模型效果影响很大
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses = losses + loss.item()
if (epoch + 1) % 5 == 0:
print('Epoch:', '%04d' % (epoch + 1), 'loss =', '{:.6f}'.format(losses / (data_loader.__len__())))
if __name__ == '__main__':
# 模型参数
max_seq_len = 7
batch_size = 2
d_model = 512 # 每个字符转换为Embedding的时候的大小
d_ff = 2048 # 前馈神经网络中Linear层映射到多少维度
d_k = d_v = 64 # dimension of K(=Q), V
n_layers = 3 # 2个encoder/decoder
n_heads = 8 # 多头注意力机制的时候把我的头分为几个
lr = 0.001
num_epochs = 20
dropout = 0.5
# src_len = 5 # 输入长度
# tgt_len = 5 # 解码端输入长度
# ------------------------------------------------------------------------------#
in_vocab, out_vocab, dataset = read_data(max_seq_len)
# print("in_vocab:", in_vocab.stoi["ils"], in_vocab.stoi)
# print("# ----------------------------------#")
# print("out_vocab:",out_vocab.stoi)
# print(dataset)
src_vocab_size = in_vocab.__len__()
tgt_vocab_size = out_vocab.__len__()
model = Transformer()
train(model, dataset, lr, batch_size, num_epochs)
# input_seq = 'ils regardent .'
# 评价
def bleu(pred_tokens, label_tokens, k):
len_pred, len_label = len(pred_tokens), len(label_tokens)
if len_label != 0:
score = math.exp(min(0, 1 - len_label / len_pred))
for n in range(1, k + 1):
num_matches, label_subs = 0, collections.defaultdict(int)
for i in range(len_label - n + 1):
label_subs[''.join(label_tokens[i: i + n])] += 1
for i in range(len_pred - n + 1):
if label_subs[''.join(pred_tokens[i: i + n])] > 0:
num_matches += 1
label_subs[''.join(pred_tokens[i: i + n])] -= 1
if (len_pred - n + 1) != 0:
score *= math.pow(num_matches / (len_pred - n + 1), math.pow(0.5, n))
else:
return "output error!"
else:
return "output error!"
return score
def score(input_seq, label_seq, k):
pred_tokens = translate(model, input_seq, result_seq_len)
print("pred_sen:",pred_tokens)
# pred_tokens = [out_vocab.stoi[tk] for tk in pred_tokens]
label_tokens = label_seq.split(' ')
print('bleu %.3f, predict: %s' % (bleu(pred_tokens, label_tokens, k), ' '.join(pred_tokens)))
result_seq_len = 4
score('ils regardent .', 'they are watching .', k=2)
score('ils se japonaise .', 'they are japanese .', k=2)
实验数据集：fr-en-small.txt
elle est vieille . she is old .
elle est tranquille . she is quiet .
elle a tort . she is wrong .
elle est canadienne . she is canadian .
elle est japonaise . she is japanese .
ils sont russes . they are russian .
ils se disputent . they are arguing .
ils regardent . they are watching .
ils sont acteurs . they are actors .
elles sont crevees . they are exhausted .
il est mon genre ! he is my type !
il a des ennuis . he is in trouble .
c est mon frere . he is my brother .
c est mon oncle . he is my uncle .
il a environ mon age . he is about my age .
elles sont toutes deux bonnes . they are both good .
elle est bonne nageuse . she is a good swimmer .
c est une personne adorable . he is a lovable person .
il fait du velo . he is riding a bicycle .
ils sont de grands amis . they are great friends .
实验结果：
测试时一个一个输入示例结果（输入和标签设置未修改）：
其他有关于Transformer：
https://www.zhihu.com/question/337886108/answer/2364160309
https://zhuanlan.zhihu.com/p/82312421
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
